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.
270 ``linkonce_odr_auto_hide``
271 Similar to "``linkonce_odr``", but nothing in the translation unit
272 takes the address of this definition. For instance, functions that
273 had an inline definition, but the compiler decided not to inline it.
274 ``linkonce_odr_auto_hide`` may have only ``default`` visibility. The
275 symbols are removed by the linker from the final linked image
276 (executable or dynamic library).
278 If none of the above identifiers are used, the global is externally
279 visible, meaning that it participates in linkage and can be used to
280 resolve external symbol references.
282 The next two types of linkage are targeted for Microsoft Windows
283 platform only. They are designed to support importing (exporting)
284 symbols from (to) DLLs (Dynamic Link Libraries).
287 "``dllimport``" linkage causes the compiler to reference a function
288 or variable via a global pointer to a pointer that is set up by the
289 DLL exporting the symbol. On Microsoft Windows targets, the pointer
290 name is formed by combining ``__imp_`` and the function or variable
293 "``dllexport``" linkage causes the compiler to provide a global
294 pointer to a pointer in a DLL, so that it can be referenced with the
295 ``dllimport`` attribute. On Microsoft Windows targets, the pointer
296 name is formed by combining ``__imp_`` and the function or variable
299 For example, since the "``.LC0``" variable is defined to be internal, if
300 another module defined a "``.LC0``" variable and was linked with this
301 one, one of the two would be renamed, preventing a collision. Since
302 "``main``" and "``puts``" are external (i.e., lacking any linkage
303 declarations), they are accessible outside of the current module.
305 It is illegal for a function *declaration* to have any linkage type
306 other than ``external``, ``dllimport`` or ``extern_weak``.
313 LLVM :ref:`functions <functionstructure>`, :ref:`calls <i_call>` and
314 :ref:`invokes <i_invoke>` can all have an optional calling convention
315 specified for the call. The calling convention of any pair of dynamic
316 caller/callee must match, or the behavior of the program is undefined.
317 The following calling conventions are supported by LLVM, and more may be
320 "``ccc``" - The C calling convention
321 This calling convention (the default if no other calling convention
322 is specified) matches the target C calling conventions. This calling
323 convention supports varargs function calls and tolerates some
324 mismatch in the declared prototype and implemented declaration of
325 the function (as does normal C).
326 "``fastcc``" - The fast calling convention
327 This calling convention attempts to make calls as fast as possible
328 (e.g. by passing things in registers). This calling convention
329 allows the target to use whatever tricks it wants to produce fast
330 code for the target, without having to conform to an externally
331 specified ABI (Application Binary Interface). `Tail calls can only
332 be optimized when this, the GHC or the HiPE convention is
333 used. <CodeGenerator.html#id80>`_ This calling convention does not
334 support varargs and requires the prototype of all callees to exactly
335 match the prototype of the function definition.
336 "``coldcc``" - The cold calling convention
337 This calling convention attempts to make code in the caller as
338 efficient as possible under the assumption that the call is not
339 commonly executed. As such, these calls often preserve all registers
340 so that the call does not break any live ranges in the caller side.
341 This calling convention does not support varargs and requires the
342 prototype of all callees to exactly match the prototype of the
344 "``cc 10``" - GHC convention
345 This calling convention has been implemented specifically for use by
346 the `Glasgow Haskell Compiler (GHC) <http://www.haskell.org/ghc>`_.
347 It passes everything in registers, going to extremes to achieve this
348 by disabling callee save registers. This calling convention should
349 not be used lightly but only for specific situations such as an
350 alternative to the *register pinning* performance technique often
351 used when implementing functional programming languages. At the
352 moment only X86 supports this convention and it has the following
355 - On *X86-32* only supports up to 4 bit type parameters. No
356 floating point types are supported.
357 - On *X86-64* only supports up to 10 bit type parameters and 6
358 floating point parameters.
360 This calling convention supports `tail call
361 optimization <CodeGenerator.html#id80>`_ but requires both the
362 caller and callee are using it.
363 "``cc 11``" - The HiPE calling convention
364 This calling convention has been implemented specifically for use by
365 the `High-Performance Erlang
366 (HiPE) <http://www.it.uu.se/research/group/hipe/>`_ compiler, *the*
367 native code compiler of the `Ericsson's Open Source Erlang/OTP
368 system <http://www.erlang.org/download.shtml>`_. It uses more
369 registers for argument passing than the ordinary C calling
370 convention and defines no callee-saved registers. The calling
371 convention properly supports `tail call
372 optimization <CodeGenerator.html#id80>`_ but requires that both the
373 caller and the callee use it. It uses a *register pinning*
374 mechanism, similar to GHC's convention, for keeping frequently
375 accessed runtime components pinned to specific hardware registers.
376 At the moment only X86 supports this convention (both 32 and 64
378 "``cc <n>``" - Numbered convention
379 Any calling convention may be specified by number, allowing
380 target-specific calling conventions to be used. Target specific
381 calling conventions start at 64.
383 More calling conventions can be added/defined on an as-needed basis, to
384 support Pascal conventions or any other well-known target-independent
387 .. _visibilitystyles:
392 All Global Variables and Functions have one of the following visibility
395 "``default``" - Default style
396 On targets that use the ELF object file format, default visibility
397 means that the declaration is visible to other modules and, in
398 shared libraries, means that the declared entity may be overridden.
399 On Darwin, default visibility means that the declaration is visible
400 to other modules. Default visibility corresponds to "external
401 linkage" in the language.
402 "``hidden``" - Hidden style
403 Two declarations of an object with hidden visibility refer to the
404 same object if they are in the same shared object. Usually, hidden
405 visibility indicates that the symbol will not be placed into the
406 dynamic symbol table, so no other module (executable or shared
407 library) can reference it directly.
408 "``protected``" - Protected style
409 On ELF, protected visibility indicates that the symbol will be
410 placed in the dynamic symbol table, but that references within the
411 defining module will bind to the local symbol. That is, the symbol
412 cannot be overridden by another module.
419 LLVM IR allows you to specify name aliases for certain types. This can
420 make it easier to read the IR and make the IR more condensed
421 (particularly when recursive types are involved). An example of a name
426 %mytype = type { %mytype*, i32 }
428 You may give a name to any :ref:`type <typesystem>` except
429 ":ref:`void <t_void>`". Type name aliases may be used anywhere a type is
430 expected with the syntax "%mytype".
432 Note that type names are aliases for the structural type that they
433 indicate, and that you can therefore specify multiple names for the same
434 type. This often leads to confusing behavior when dumping out a .ll
435 file. Since LLVM IR uses structural typing, the name is not part of the
436 type. When printing out LLVM IR, the printer will pick *one name* to
437 render all types of a particular shape. This means that if you have code
438 where two different source types end up having the same LLVM type, that
439 the dumper will sometimes print the "wrong" or unexpected type. This is
440 an important design point and isn't going to change.
447 Global variables define regions of memory allocated at compilation time
450 Global variables definitions must be initialized, may have an explicit section
451 to be placed in, and may have an optional explicit alignment specified.
453 Global variables in other translation units can also be declared, in which
454 case they don't have an initializer.
456 A variable may be defined as ``thread_local``, which means that it will
457 not be shared by threads (each thread will have a separated copy of the
458 variable). Not all targets support thread-local variables. Optionally, a
459 TLS model may be specified:
462 For variables that are only used within the current shared library.
464 For variables in modules that will not be loaded dynamically.
466 For variables defined in the executable and only used within it.
468 The models correspond to the ELF TLS models; see `ELF Handling For
469 Thread-Local Storage <http://people.redhat.com/drepper/tls.pdf>`_ for
470 more information on under which circumstances the different models may
471 be used. The target may choose a different TLS model if the specified
472 model is not supported, or if a better choice of model can be made.
474 A variable may be defined as a global ``constant``, which indicates that
475 the contents of the variable will **never** be modified (enabling better
476 optimization, allowing the global data to be placed in the read-only
477 section of an executable, etc). Note that variables that need runtime
478 initialization cannot be marked ``constant`` as there is a store to the
481 LLVM explicitly allows *declarations* of global variables to be marked
482 constant, even if the final definition of the global is not. This
483 capability can be used to enable slightly better optimization of the
484 program, but requires the language definition to guarantee that
485 optimizations based on the 'constantness' are valid for the translation
486 units that do not include the definition.
488 As SSA values, global variables define pointer values that are in scope
489 (i.e. they dominate) all basic blocks in the program. Global variables
490 always define a pointer to their "content" type because they describe a
491 region of memory, and all memory objects in LLVM are accessed through
494 Global variables can be marked with ``unnamed_addr`` which indicates
495 that the address is not significant, only the content. Constants marked
496 like this can be merged with other constants if they have the same
497 initializer. Note that a constant with significant address *can* be
498 merged with a ``unnamed_addr`` constant, the result being a constant
499 whose address is significant.
501 A global variable may be declared to reside in a target-specific
502 numbered address space. For targets that support them, address spaces
503 may affect how optimizations are performed and/or what target
504 instructions are used to access the variable. The default address space
505 is zero. The address space qualifier must precede any other attributes.
507 LLVM allows an explicit section to be specified for globals. If the
508 target supports it, it will emit globals to the section specified.
510 By default, global initializers are optimized by assuming that global
511 variables defined within the module are not modified from their
512 initial values before the start of the global initializer. This is
513 true even for variables potentially accessible from outside the
514 module, including those with external linkage or appearing in
515 ``@llvm.used``. This assumption may be suppressed by marking the
516 variable with ``externally_initialized``.
518 An explicit alignment may be specified for a global, which must be a
519 power of 2. If not present, or if the alignment is set to zero, the
520 alignment of the global is set by the target to whatever it feels
521 convenient. If an explicit alignment is specified, the global is forced
522 to have exactly that alignment. Targets and optimizers are not allowed
523 to over-align the global if the global has an assigned section. In this
524 case, the extra alignment could be observable: for example, code could
525 assume that the globals are densely packed in their section and try to
526 iterate over them as an array, alignment padding would break this
529 For example, the following defines a global in a numbered address space
530 with an initializer, section, and alignment:
534 @G = addrspace(5) constant float 1.0, section "foo", align 4
536 The following example just declares a global variable
540 @G = external global i32
542 The following example defines a thread-local global with the
543 ``initialexec`` TLS model:
547 @G = thread_local(initialexec) global i32 0, align 4
549 .. _functionstructure:
554 LLVM function definitions consist of the "``define``" keyword, an
555 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
556 style <visibility>`, an optional :ref:`calling convention <callingconv>`,
557 an optional ``unnamed_addr`` attribute, a return type, an optional
558 :ref:`parameter attribute <paramattrs>` for the return type, a function
559 name, a (possibly empty) argument list (each with optional :ref:`parameter
560 attributes <paramattrs>`), optional :ref:`function attributes <fnattrs>`,
561 an optional section, an optional alignment, an optional :ref:`garbage
562 collector name <gc>`, an optional :ref:`prefix <prefixdata>`, an opening
563 curly brace, a list of basic blocks, and a closing curly brace.
565 LLVM function declarations consist of the "``declare``" keyword, an
566 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
567 style <visibility>`, an optional :ref:`calling convention <callingconv>`,
568 an optional ``unnamed_addr`` attribute, a return type, an optional
569 :ref:`parameter attribute <paramattrs>` for the return type, a function
570 name, a possibly empty list of arguments, an optional alignment, an optional
571 :ref:`garbage collector name <gc>` and an optional :ref:`prefix <prefixdata>`.
573 A function definition contains a list of basic blocks, forming the CFG (Control
574 Flow Graph) for the function. Each basic block may optionally start with a label
575 (giving the basic block a symbol table entry), contains a list of instructions,
576 and ends with a :ref:`terminator <terminators>` instruction (such as a branch or
577 function return). If an explicit label is not provided, a block is assigned an
578 implicit numbered label, using the next value from the same counter as used for
579 unnamed temporaries (:ref:`see above<identifiers>`). For example, if a function
580 entry block does not have an explicit label, it will be assigned label "%0",
581 then the first unnamed temporary in that block will be "%1", etc.
583 The first basic block in a function is special in two ways: it is
584 immediately executed on entrance to the function, and it is not allowed
585 to have predecessor basic blocks (i.e. there can not be any branches to
586 the entry block of a function). Because the block can have no
587 predecessors, it also cannot have any :ref:`PHI nodes <i_phi>`.
589 LLVM allows an explicit section to be specified for functions. If the
590 target supports it, it will emit functions to the section specified.
592 An explicit alignment may be specified for a function. If not present,
593 or if the alignment is set to zero, the alignment of the function is set
594 by the target to whatever it feels convenient. If an explicit alignment
595 is specified, the function is forced to have at least that much
596 alignment. All alignments must be a power of 2.
598 If the ``unnamed_addr`` attribute is given, the address is know to not
599 be significant and two identical functions can be merged.
603 define [linkage] [visibility]
605 <ResultType> @<FunctionName> ([argument list])
606 [fn Attrs] [section "name"] [align N]
607 [gc] [prefix Constant] { ... }
614 Aliases act as "second name" for the aliasee value (which can be either
615 function, global variable, another alias or bitcast of global value).
616 Aliases may have an optional :ref:`linkage type <linkage>`, and an optional
617 :ref:`visibility style <visibility>`.
621 @<Name> = alias [Linkage] [Visibility] <AliaseeTy> @<Aliasee>
623 The linkage must be one of ``private``, ``linker_private``,
624 ``linker_private_weak``, ``internal``, ``linkonce``, ``weak``,
625 ``linkonce_odr``, ``weak_odr``, ``linkonce_odr_auto_hide``, ``external``. Note
626 that some system linkers might not correctly handle dropping a weak symbol that
627 is aliased by a non weak alias.
629 .. _namedmetadatastructure:
634 Named metadata is a collection of metadata. :ref:`Metadata
635 nodes <metadata>` (but not metadata strings) are the only valid
636 operands for a named metadata.
640 ; Some unnamed metadata nodes, which are referenced by the named metadata.
641 !0 = metadata !{metadata !"zero"}
642 !1 = metadata !{metadata !"one"}
643 !2 = metadata !{metadata !"two"}
645 !name = !{!0, !1, !2}
652 The return type and each parameter of a function type may have a set of
653 *parameter attributes* associated with them. Parameter attributes are
654 used to communicate additional information about the result or
655 parameters of a function. Parameter attributes are considered to be part
656 of the function, not of the function type, so functions with different
657 parameter attributes can have the same function type.
659 Parameter attributes are simple keywords that follow the type specified.
660 If multiple parameter attributes are needed, they are space separated.
665 declare i32 @printf(i8* noalias nocapture, ...)
666 declare i32 @atoi(i8 zeroext)
667 declare signext i8 @returns_signed_char()
669 Note that any attributes for the function result (``nounwind``,
670 ``readonly``) come immediately after the argument list.
672 Currently, only the following parameter attributes are defined:
675 This indicates to the code generator that the parameter or return
676 value should be zero-extended to the extent required by the target's
677 ABI (which is usually 32-bits, but is 8-bits for a i1 on x86-64) by
678 the caller (for a parameter) or the callee (for a return value).
680 This indicates to the code generator that the parameter or return
681 value should be sign-extended to the extent required by the target's
682 ABI (which is usually 32-bits) by the caller (for a parameter) or
683 the callee (for a return value).
685 This indicates that this parameter or return value should be treated
686 in a special target-dependent fashion during while emitting code for
687 a function call or return (usually, by putting it in a register as
688 opposed to memory, though some targets use it to distinguish between
689 two different kinds of registers). Use of this attribute is
692 This indicates that the pointer parameter should really be passed by
693 value to the function. The attribute implies that a hidden copy of
694 the pointee is made between the caller and the callee, so the callee
695 is unable to modify the value in the caller. This attribute is only
696 valid on LLVM pointer arguments. It is generally used to pass
697 structs and arrays by value, but is also valid on pointers to
698 scalars. The copy is considered to belong to the caller not the
699 callee (for example, ``readonly`` functions should not write to
700 ``byval`` parameters). This is not a valid attribute for return
703 The byval attribute also supports specifying an alignment with the
704 align attribute. It indicates the alignment of the stack slot to
705 form and the known alignment of the pointer specified to the call
706 site. If the alignment is not specified, then the code generator
707 makes a target-specific assumption.
710 This indicates that the pointer parameter specifies the address of a
711 structure that is the return value of the function in the source
712 program. This pointer must be guaranteed by the caller to be valid:
713 loads and stores to the structure may be assumed by the callee
714 not to trap and to be properly aligned. This may only be applied to
715 the first parameter. This is not a valid attribute for return
718 This indicates that pointer values :ref:`based <pointeraliasing>` on
719 the argument or return value do not alias pointer values which are
720 not *based* on it, ignoring certain "irrelevant" dependencies. For a
721 call to the parent function, dependencies between memory references
722 from before or after the call and from those during the call are
723 "irrelevant" to the ``noalias`` keyword for the arguments and return
724 value used in that call. The caller shares the responsibility with
725 the callee for ensuring that these requirements are met. For further
726 details, please see the discussion of the NoAlias response in `alias
727 analysis <AliasAnalysis.html#MustMayNo>`_.
729 Note that this definition of ``noalias`` is intentionally similar
730 to the definition of ``restrict`` in C99 for function arguments,
731 though it is slightly weaker.
733 For function return values, C99's ``restrict`` is not meaningful,
734 while LLVM's ``noalias`` is.
736 This indicates that the callee does not make any copies of the
737 pointer that outlive the callee itself. This is not a valid
738 attribute for return values.
743 This indicates that the pointer parameter can be excised using the
744 :ref:`trampoline intrinsics <int_trampoline>`. This is not a valid
745 attribute for return values and can only be applied to one parameter.
748 This indicates that the function always returns the argument as its return
749 value. This is an optimization hint to the code generator when generating
750 the caller, allowing tail call optimization and omission of register saves
751 and restores in some cases; it is not checked or enforced when generating
752 the callee. The parameter and the function return type must be valid
753 operands for the :ref:`bitcast instruction <i_bitcast>`. This is not a
754 valid attribute for return values and can only be applied to one parameter.
758 Garbage Collector Names
759 -----------------------
761 Each function may specify a garbage collector name, which is simply a
766 define void @f() gc "name" { ... }
768 The compiler declares the supported values of *name*. Specifying a
769 collector which will cause the compiler to alter its output in order to
770 support the named garbage collection algorithm.
777 Prefix data is data associated with a function which the code generator
778 will emit immediately before the function body. The purpose of this feature
779 is to allow frontends to associate language-specific runtime metadata with
780 specific functions and make it available through the function pointer while
781 still allowing the function pointer to be called. To access the data for a
782 given function, a program may bitcast the function pointer to a pointer to
783 the constant's type. This implies that the IR symbol points to the start
786 To maintain the semantics of ordinary function calls, the prefix data must
787 have a particular format. Specifically, it must begin with a sequence of
788 bytes which decode to a sequence of machine instructions, valid for the
789 module's target, which transfer control to the point immediately succeeding
790 the prefix data, without performing any other visible action. This allows
791 the inliner and other passes to reason about the semantics of the function
792 definition without needing to reason about the prefix data. Obviously this
793 makes the format of the prefix data highly target dependent.
795 Prefix data is laid out as if it were an initializer for a global variable
796 of the prefix data's type. No padding is automatically placed between the
797 prefix data and the function body. If padding is required, it must be part
800 A trivial example of valid prefix data for the x86 architecture is ``i8 144``,
801 which encodes the ``nop`` instruction:
805 define void @f() prefix i8 144 { ... }
807 Generally prefix data can be formed by encoding a relative branch instruction
808 which skips the metadata, as in this example of valid prefix data for the
809 x86_64 architecture, where the first two bytes encode ``jmp .+10``:
813 %0 = type <{ i8, i8, i8* }>
815 define void @f() prefix %0 <{ i8 235, i8 8, i8* @md}> { ... }
817 A function may have prefix data but no body. This has similar semantics
818 to the ``available_externally`` linkage in that the data may be used by the
819 optimizers but will not be emitted in the object file.
826 Attribute groups are groups of attributes that are referenced by objects within
827 the IR. They are important for keeping ``.ll`` files readable, because a lot of
828 functions will use the same set of attributes. In the degenerative case of a
829 ``.ll`` file that corresponds to a single ``.c`` file, the single attribute
830 group will capture the important command line flags used to build that file.
832 An attribute group is a module-level object. To use an attribute group, an
833 object references the attribute group's ID (e.g. ``#37``). An object may refer
834 to more than one attribute group. In that situation, the attributes from the
835 different groups are merged.
837 Here is an example of attribute groups for a function that should always be
838 inlined, has a stack alignment of 4, and which shouldn't use SSE instructions:
842 ; Target-independent attributes:
843 attributes #0 = { alwaysinline alignstack=4 }
845 ; Target-dependent attributes:
846 attributes #1 = { "no-sse" }
848 ; Function @f has attributes: alwaysinline, alignstack=4, and "no-sse".
849 define void @f() #0 #1 { ... }
856 Function attributes are set to communicate additional information about
857 a function. Function attributes are considered to be part of the
858 function, not of the function type, so functions with different function
859 attributes can have the same function type.
861 Function attributes are simple keywords that follow the type specified.
862 If multiple attributes are needed, they are space separated. For
867 define void @f() noinline { ... }
868 define void @f() alwaysinline { ... }
869 define void @f() alwaysinline optsize { ... }
870 define void @f() optsize { ... }
873 This attribute indicates that, when emitting the prologue and
874 epilogue, the backend should forcibly align the stack pointer.
875 Specify the desired alignment, which must be a power of two, in
878 This attribute indicates that the inliner should attempt to inline
879 this function into callers whenever possible, ignoring any active
880 inlining size threshold for this caller.
882 This indicates that the callee function at a call site should be
883 recognized as a built-in function, even though the function's declaration
884 uses the ``nobuiltin`` attribute. This is only valid at call sites for
885 direct calls to functions which are declared with the ``nobuiltin``
888 This attribute indicates that this function is rarely called. When
889 computing edge weights, basic blocks post-dominated by a cold
890 function call are also considered to be cold; and, thus, given low
893 This attribute indicates that the source code contained a hint that
894 inlining this function is desirable (such as the "inline" keyword in
895 C/C++). It is just a hint; it imposes no requirements on the
898 This attribute suggests that optimization passes and code generator
899 passes make choices that keep the code size of this function as small
900 as possible and perform optimizations that may sacrifice runtime
901 performance in order to minimize the size of the generated code.
903 This attribute disables prologue / epilogue emission for the
904 function. This can have very system-specific consequences.
906 This indicates that the callee function at a call site is not recognized as
907 a built-in function. LLVM will retain the original call and not replace it
908 with equivalent code based on the semantics of the built-in function, unless
909 the call site uses the ``builtin`` attribute. This is valid at call sites
910 and on function declarations and definitions.
912 This attribute indicates that calls to the function cannot be
913 duplicated. A call to a ``noduplicate`` function may be moved
914 within its parent function, but may not be duplicated within
917 A function containing a ``noduplicate`` call may still
918 be an inlining candidate, provided that the call is not
919 duplicated by inlining. That implies that the function has
920 internal linkage and only has one call site, so the original
921 call is dead after inlining.
923 This attributes disables implicit floating point instructions.
925 This attribute indicates that the inliner should never inline this
926 function in any situation. This attribute may not be used together
927 with the ``alwaysinline`` attribute.
929 This attribute suppresses lazy symbol binding for the function. This
930 may make calls to the function faster, at the cost of extra program
931 startup time if the function is not called during program startup.
933 This attribute indicates that the code generator should not use a
934 red zone, even if the target-specific ABI normally permits it.
936 This function attribute indicates that the function never returns
937 normally. This produces undefined behavior at runtime if the
938 function ever does dynamically return.
940 This function attribute indicates that the function never returns
941 with an unwind or exceptional control flow. If the function does
942 unwind, its runtime behavior is undefined.
944 This function attribute indicates that the function is not optimized
945 by any optimization or code generator passes with the
946 exception of interprocedural optimization passes.
947 This attribute cannot be used together with the ``alwaysinline``
948 attribute; this attribute is also incompatible
949 with the ``minsize`` attribute and the ``optsize`` attribute.
951 The inliner should never inline this function in any situation.
952 Only functions with the ``alwaysinline`` attribute are valid
953 candidates for inlining inside the body of this function.
955 This attribute suggests that optimization passes and code generator
956 passes make choices that keep the code size of this function low,
957 and otherwise do optimizations specifically to reduce code size as
958 long as they do not significantly impact runtime performance.
960 On a function, this attribute indicates that the function computes its
961 result (or decides to unwind an exception) based strictly on its arguments,
962 without dereferencing any pointer arguments or otherwise accessing
963 any mutable state (e.g. memory, control registers, etc) visible to
964 caller functions. It does not write through any pointer arguments
965 (including ``byval`` arguments) and never changes any state visible
966 to callers. This means that it cannot unwind exceptions by calling
967 the ``C++`` exception throwing methods.
969 On an argument, this attribute indicates that the function does not
970 dereference that pointer argument, even though it may read or write the
971 memory that the pointer points to if accessed through other pointers.
973 On a function, this attribute indicates that the function does not write
974 through any pointer arguments (including ``byval`` arguments) or otherwise
975 modify any state (e.g. memory, control registers, etc) visible to
976 caller functions. It may dereference pointer arguments and read
977 state that may be set in the caller. A readonly function always
978 returns the same value (or unwinds an exception identically) when
979 called with the same set of arguments and global state. It cannot
980 unwind an exception by calling the ``C++`` exception throwing
983 On an argument, this attribute indicates that the function does not write
984 through this pointer argument, even though it may write to the memory that
985 the pointer points to.
987 This attribute indicates that this function can return twice. The C
988 ``setjmp`` is an example of such a function. The compiler disables
989 some optimizations (like tail calls) in the caller of these
992 This attribute indicates that AddressSanitizer checks
993 (dynamic address safety analysis) are enabled for this function.
995 This attribute indicates that MemorySanitizer checks (dynamic detection
996 of accesses to uninitialized memory) are enabled for this function.
998 This attribute indicates that ThreadSanitizer checks
999 (dynamic thread safety analysis) are enabled for this function.
1001 This attribute indicates that the function should emit a stack
1002 smashing protector. It is in the form of a "canary" --- a random value
1003 placed on the stack before the local variables that's checked upon
1004 return from the function to see if it has been overwritten. A
1005 heuristic is used to determine if a function needs stack protectors
1006 or not. The heuristic used will enable protectors for functions with:
1008 - Character arrays larger than ``ssp-buffer-size`` (default 8).
1009 - Aggregates containing character arrays larger than ``ssp-buffer-size``.
1010 - Calls to alloca() with variable sizes or constant sizes greater than
1011 ``ssp-buffer-size``.
1013 If a function that has an ``ssp`` attribute is inlined into a
1014 function that doesn't have an ``ssp`` attribute, then the resulting
1015 function will have an ``ssp`` attribute.
1017 This attribute indicates that the function should *always* emit a
1018 stack smashing protector. This overrides the ``ssp`` function
1021 If a function that has an ``sspreq`` attribute is inlined into a
1022 function that doesn't have an ``sspreq`` attribute or which has an
1023 ``ssp`` or ``sspstrong`` attribute, then the resulting function will have
1024 an ``sspreq`` attribute.
1026 This attribute indicates that the function should emit a stack smashing
1027 protector. This attribute causes a strong heuristic to be used when
1028 determining if a function needs stack protectors. The strong heuristic
1029 will enable protectors for functions with:
1031 - Arrays of any size and type
1032 - Aggregates containing an array of any size and type.
1033 - Calls to alloca().
1034 - Local variables that have had their address taken.
1036 This overrides the ``ssp`` function attribute.
1038 If a function that has an ``sspstrong`` attribute is inlined into a
1039 function that doesn't have an ``sspstrong`` attribute, then the
1040 resulting function will have an ``sspstrong`` attribute.
1042 This attribute indicates that the ABI being targeted requires that
1043 an unwind table entry be produce for this function even if we can
1044 show that no exceptions passes by it. This is normally the case for
1045 the ELF x86-64 abi, but it can be disabled for some compilation
1050 Module-Level Inline Assembly
1051 ----------------------------
1053 Modules may contain "module-level inline asm" blocks, which corresponds
1054 to the GCC "file scope inline asm" blocks. These blocks are internally
1055 concatenated by LLVM and treated as a single unit, but may be separated
1056 in the ``.ll`` file if desired. The syntax is very simple:
1058 .. code-block:: llvm
1060 module asm "inline asm code goes here"
1061 module asm "more can go here"
1063 The strings can contain any character by escaping non-printable
1064 characters. The escape sequence used is simply "\\xx" where "xx" is the
1065 two digit hex code for the number.
1067 The inline asm code is simply printed to the machine code .s file when
1068 assembly code is generated.
1070 .. _langref_datalayout:
1075 A module may specify a target specific data layout string that specifies
1076 how data is to be laid out in memory. The syntax for the data layout is
1079 .. code-block:: llvm
1081 target datalayout = "layout specification"
1083 The *layout specification* consists of a list of specifications
1084 separated by the minus sign character ('-'). Each specification starts
1085 with a letter and may include other information after the letter to
1086 define some aspect of the data layout. The specifications accepted are
1090 Specifies that the target lays out data in big-endian form. That is,
1091 the bits with the most significance have the lowest address
1094 Specifies that the target lays out data in little-endian form. That
1095 is, the bits with the least significance have the lowest address
1098 Specifies the natural alignment of the stack in bits. Alignment
1099 promotion of stack variables is limited to the natural stack
1100 alignment to avoid dynamic stack realignment. The stack alignment
1101 must be a multiple of 8-bits. If omitted, the natural stack
1102 alignment defaults to "unspecified", which does not prevent any
1103 alignment promotions.
1104 ``p[n]:<size>:<abi>:<pref>``
1105 This specifies the *size* of a pointer and its ``<abi>`` and
1106 ``<pref>``\erred alignments for address space ``n``. All sizes are in
1107 bits. Specifying the ``<pref>`` alignment is optional. If omitted, the
1108 preceding ``:`` should be omitted too. The address space, ``n`` is
1109 optional, and if not specified, denotes the default address space 0.
1110 The value of ``n`` must be in the range [1,2^23).
1111 ``i<size>:<abi>:<pref>``
1112 This specifies the alignment for an integer type of a given bit
1113 ``<size>``. The value of ``<size>`` must be in the range [1,2^23).
1114 ``v<size>:<abi>:<pref>``
1115 This specifies the alignment for a vector type of a given bit
1117 ``f<size>:<abi>:<pref>``
1118 This specifies the alignment for a floating point type of a given bit
1119 ``<size>``. Only values of ``<size>`` that are supported by the target
1120 will work. 32 (float) and 64 (double) are supported on all targets; 80
1121 or 128 (different flavors of long double) are also supported on some
1123 ``a<size>:<abi>:<pref>``
1124 This specifies the alignment for an aggregate type of a given bit
1126 ``s<size>:<abi>:<pref>``
1127 This specifies the alignment for a stack object of a given bit
1129 ``n<size1>:<size2>:<size3>...``
1130 This specifies a set of native integer widths for the target CPU in
1131 bits. For example, it might contain ``n32`` for 32-bit PowerPC,
1132 ``n32:64`` for PowerPC 64, or ``n8:16:32:64`` for X86-64. Elements of
1133 this set are considered to support most general arithmetic operations
1136 When constructing the data layout for a given target, LLVM starts with a
1137 default set of specifications which are then (possibly) overridden by
1138 the specifications in the ``datalayout`` keyword. The default
1139 specifications are given in this list:
1141 - ``E`` - big endian
1142 - ``p:64:64:64`` - 64-bit pointers with 64-bit alignment.
1143 - ``p[n]:64:64:64`` - Other address spaces are assumed to be the
1144 same as the default address space.
1145 - ``S0`` - natural stack alignment is unspecified
1146 - ``i1:8:8`` - i1 is 8-bit (byte) aligned
1147 - ``i8:8:8`` - i8 is 8-bit (byte) aligned
1148 - ``i16:16:16`` - i16 is 16-bit aligned
1149 - ``i32:32:32`` - i32 is 32-bit aligned
1150 - ``i64:32:64`` - i64 has ABI alignment of 32-bits but preferred
1151 alignment of 64-bits
1152 - ``f16:16:16`` - half is 16-bit aligned
1153 - ``f32:32:32`` - float is 32-bit aligned
1154 - ``f64:64:64`` - double is 64-bit aligned
1155 - ``f128:128:128`` - quad is 128-bit aligned
1156 - ``v64:64:64`` - 64-bit vector is 64-bit aligned
1157 - ``v128:128:128`` - 128-bit vector is 128-bit aligned
1158 - ``a0:0:64`` - aggregates are 64-bit aligned
1160 When LLVM is determining the alignment for a given type, it uses the
1163 #. If the type sought is an exact match for one of the specifications,
1164 that specification is used.
1165 #. If no match is found, and the type sought is an integer type, then
1166 the smallest integer type that is larger than the bitwidth of the
1167 sought type is used. If none of the specifications are larger than
1168 the bitwidth then the largest integer type is used. For example,
1169 given the default specifications above, the i7 type will use the
1170 alignment of i8 (next largest) while both i65 and i256 will use the
1171 alignment of i64 (largest specified).
1172 #. If no match is found, and the type sought is a vector type, then the
1173 largest vector type that is smaller than the sought vector type will
1174 be used as a fall back. This happens because <128 x double> can be
1175 implemented in terms of 64 <2 x double>, for example.
1177 The function of the data layout string may not be what you expect.
1178 Notably, this is not a specification from the frontend of what alignment
1179 the code generator should use.
1181 Instead, if specified, the target data layout is required to match what
1182 the ultimate *code generator* expects. This string is used by the
1183 mid-level optimizers to improve code, and this only works if it matches
1184 what the ultimate code generator uses. If you would like to generate IR
1185 that does not embed this target-specific detail into the IR, then you
1186 don't have to specify the string. This will disable some optimizations
1187 that require precise layout information, but this also prevents those
1188 optimizations from introducing target specificity into the IR.
1195 A module may specify a target triple string that describes the target
1196 host. The syntax for the target triple is simply:
1198 .. code-block:: llvm
1200 target triple = "x86_64-apple-macosx10.7.0"
1202 The *target triple* string consists of a series of identifiers delimited
1203 by the minus sign character ('-'). The canonical forms are:
1207 ARCHITECTURE-VENDOR-OPERATING_SYSTEM
1208 ARCHITECTURE-VENDOR-OPERATING_SYSTEM-ENVIRONMENT
1210 This information is passed along to the backend so that it generates
1211 code for the proper architecture. It's possible to override this on the
1212 command line with the ``-mtriple`` command line option.
1214 .. _pointeraliasing:
1216 Pointer Aliasing Rules
1217 ----------------------
1219 Any memory access must be done through a pointer value associated with
1220 an address range of the memory access, otherwise the behavior is
1221 undefined. Pointer values are associated with address ranges according
1222 to the following rules:
1224 - A pointer value is associated with the addresses associated with any
1225 value it is *based* on.
1226 - An address of a global variable is associated with the address range
1227 of the variable's storage.
1228 - The result value of an allocation instruction is associated with the
1229 address range of the allocated storage.
1230 - A null pointer in the default address-space is associated with no
1232 - An integer constant other than zero or a pointer value returned from
1233 a function not defined within LLVM may be associated with address
1234 ranges allocated through mechanisms other than those provided by
1235 LLVM. Such ranges shall not overlap with any ranges of addresses
1236 allocated by mechanisms provided by LLVM.
1238 A pointer value is *based* on another pointer value according to the
1241 - A pointer value formed from a ``getelementptr`` operation is *based*
1242 on the first operand of the ``getelementptr``.
1243 - The result value of a ``bitcast`` is *based* on the operand of the
1245 - A pointer value formed by an ``inttoptr`` is *based* on all pointer
1246 values that contribute (directly or indirectly) to the computation of
1247 the pointer's value.
1248 - The "*based* on" relationship is transitive.
1250 Note that this definition of *"based"* is intentionally similar to the
1251 definition of *"based"* in C99, though it is slightly weaker.
1253 LLVM IR does not associate types with memory. The result type of a
1254 ``load`` merely indicates the size and alignment of the memory from
1255 which to load, as well as the interpretation of the value. The first
1256 operand type of a ``store`` similarly only indicates the size and
1257 alignment of the store.
1259 Consequently, type-based alias analysis, aka TBAA, aka
1260 ``-fstrict-aliasing``, is not applicable to general unadorned LLVM IR.
1261 :ref:`Metadata <metadata>` may be used to encode additional information
1262 which specialized optimization passes may use to implement type-based
1267 Volatile Memory Accesses
1268 ------------------------
1270 Certain memory accesses, such as :ref:`load <i_load>`'s,
1271 :ref:`store <i_store>`'s, and :ref:`llvm.memcpy <int_memcpy>`'s may be
1272 marked ``volatile``. The optimizers must not change the number of
1273 volatile operations or change their order of execution relative to other
1274 volatile operations. The optimizers *may* change the order of volatile
1275 operations relative to non-volatile operations. This is not Java's
1276 "volatile" and has no cross-thread synchronization behavior.
1278 IR-level volatile loads and stores cannot safely be optimized into
1279 llvm.memcpy or llvm.memmove intrinsics even when those intrinsics are
1280 flagged volatile. Likewise, the backend should never split or merge
1281 target-legal volatile load/store instructions.
1283 .. admonition:: Rationale
1285 Platforms may rely on volatile loads and stores of natively supported
1286 data width to be executed as single instruction. For example, in C
1287 this holds for an l-value of volatile primitive type with native
1288 hardware support, but not necessarily for aggregate types. The
1289 frontend upholds these expectations, which are intentionally
1290 unspecified in the IR. The rules above ensure that IR transformation
1291 do not violate the frontend's contract with the language.
1295 Memory Model for Concurrent Operations
1296 --------------------------------------
1298 The LLVM IR does not define any way to start parallel threads of
1299 execution or to register signal handlers. Nonetheless, there are
1300 platform-specific ways to create them, and we define LLVM IR's behavior
1301 in their presence. This model is inspired by the C++0x memory model.
1303 For a more informal introduction to this model, see the :doc:`Atomics`.
1305 We define a *happens-before* partial order as the least partial order
1308 - Is a superset of single-thread program order, and
1309 - When a *synchronizes-with* ``b``, includes an edge from ``a`` to
1310 ``b``. *Synchronizes-with* pairs are introduced by platform-specific
1311 techniques, like pthread locks, thread creation, thread joining,
1312 etc., and by atomic instructions. (See also :ref:`Atomic Memory Ordering
1313 Constraints <ordering>`).
1315 Note that program order does not introduce *happens-before* edges
1316 between a thread and signals executing inside that thread.
1318 Every (defined) read operation (load instructions, memcpy, atomic
1319 loads/read-modify-writes, etc.) R reads a series of bytes written by
1320 (defined) write operations (store instructions, atomic
1321 stores/read-modify-writes, memcpy, etc.). For the purposes of this
1322 section, initialized globals are considered to have a write of the
1323 initializer which is atomic and happens before any other read or write
1324 of the memory in question. For each byte of a read R, R\ :sub:`byte`
1325 may see any write to the same byte, except:
1327 - If write\ :sub:`1` happens before write\ :sub:`2`, and
1328 write\ :sub:`2` happens before R\ :sub:`byte`, then
1329 R\ :sub:`byte` does not see write\ :sub:`1`.
1330 - If R\ :sub:`byte` happens before write\ :sub:`3`, then
1331 R\ :sub:`byte` does not see write\ :sub:`3`.
1333 Given that definition, R\ :sub:`byte` is defined as follows:
1335 - If R is volatile, the result is target-dependent. (Volatile is
1336 supposed to give guarantees which can support ``sig_atomic_t`` in
1337 C/C++, and may be used for accesses to addresses which do not behave
1338 like normal memory. It does not generally provide cross-thread
1340 - Otherwise, if there is no write to the same byte that happens before
1341 R\ :sub:`byte`, R\ :sub:`byte` returns ``undef`` for that byte.
1342 - Otherwise, if R\ :sub:`byte` may see exactly one write,
1343 R\ :sub:`byte` returns the value written by that write.
1344 - Otherwise, if R is atomic, and all the writes R\ :sub:`byte` may
1345 see are atomic, it chooses one of the values written. See the :ref:`Atomic
1346 Memory Ordering Constraints <ordering>` section for additional
1347 constraints on how the choice is made.
1348 - Otherwise R\ :sub:`byte` returns ``undef``.
1350 R returns the value composed of the series of bytes it read. This
1351 implies that some bytes within the value may be ``undef`` **without**
1352 the entire value being ``undef``. Note that this only defines the
1353 semantics of the operation; it doesn't mean that targets will emit more
1354 than one instruction to read the series of bytes.
1356 Note that in cases where none of the atomic intrinsics are used, this
1357 model places only one restriction on IR transformations on top of what
1358 is required for single-threaded execution: introducing a store to a byte
1359 which might not otherwise be stored is not allowed in general.
1360 (Specifically, in the case where another thread might write to and read
1361 from an address, introducing a store can change a load that may see
1362 exactly one write into a load that may see multiple writes.)
1366 Atomic Memory Ordering Constraints
1367 ----------------------------------
1369 Atomic instructions (:ref:`cmpxchg <i_cmpxchg>`,
1370 :ref:`atomicrmw <i_atomicrmw>`, :ref:`fence <i_fence>`,
1371 :ref:`atomic load <i_load>`, and :ref:`atomic store <i_store>`) take
1372 an ordering parameter that determines which other atomic instructions on
1373 the same address they *synchronize with*. These semantics are borrowed
1374 from Java and C++0x, but are somewhat more colloquial. If these
1375 descriptions aren't precise enough, check those specs (see spec
1376 references in the :doc:`atomics guide <Atomics>`).
1377 :ref:`fence <i_fence>` instructions treat these orderings somewhat
1378 differently since they don't take an address. See that instruction's
1379 documentation for details.
1381 For a simpler introduction to the ordering constraints, see the
1385 The set of values that can be read is governed by the happens-before
1386 partial order. A value cannot be read unless some operation wrote
1387 it. This is intended to provide a guarantee strong enough to model
1388 Java's non-volatile shared variables. This ordering cannot be
1389 specified for read-modify-write operations; it is not strong enough
1390 to make them atomic in any interesting way.
1392 In addition to the guarantees of ``unordered``, there is a single
1393 total order for modifications by ``monotonic`` operations on each
1394 address. All modification orders must be compatible with the
1395 happens-before order. There is no guarantee that the modification
1396 orders can be combined to a global total order for the whole program
1397 (and this often will not be possible). The read in an atomic
1398 read-modify-write operation (:ref:`cmpxchg <i_cmpxchg>` and
1399 :ref:`atomicrmw <i_atomicrmw>`) reads the value in the modification
1400 order immediately before the value it writes. If one atomic read
1401 happens before another atomic read of the same address, the later
1402 read must see the same value or a later value in the address's
1403 modification order. This disallows reordering of ``monotonic`` (or
1404 stronger) operations on the same address. If an address is written
1405 ``monotonic``-ally by one thread, and other threads ``monotonic``-ally
1406 read that address repeatedly, the other threads must eventually see
1407 the write. This corresponds to the C++0x/C1x
1408 ``memory_order_relaxed``.
1410 In addition to the guarantees of ``monotonic``, a
1411 *synchronizes-with* edge may be formed with a ``release`` operation.
1412 This is intended to model C++'s ``memory_order_acquire``.
1414 In addition to the guarantees of ``monotonic``, if this operation
1415 writes a value which is subsequently read by an ``acquire``
1416 operation, it *synchronizes-with* that operation. (This isn't a
1417 complete description; see the C++0x definition of a release
1418 sequence.) This corresponds to the C++0x/C1x
1419 ``memory_order_release``.
1420 ``acq_rel`` (acquire+release)
1421 Acts as both an ``acquire`` and ``release`` operation on its
1422 address. This corresponds to the C++0x/C1x ``memory_order_acq_rel``.
1423 ``seq_cst`` (sequentially consistent)
1424 In addition to the guarantees of ``acq_rel`` (``acquire`` for an
1425 operation which only reads, ``release`` for an operation which only
1426 writes), there is a global total order on all
1427 sequentially-consistent operations on all addresses, which is
1428 consistent with the *happens-before* partial order and with the
1429 modification orders of all the affected addresses. Each
1430 sequentially-consistent read sees the last preceding write to the
1431 same address in this global order. This corresponds to the C++0x/C1x
1432 ``memory_order_seq_cst`` and Java volatile.
1436 If an atomic operation is marked ``singlethread``, it only *synchronizes
1437 with* or participates in modification and seq\_cst total orderings with
1438 other operations running in the same thread (for example, in signal
1446 LLVM IR floating-point binary ops (:ref:`fadd <i_fadd>`,
1447 :ref:`fsub <i_fsub>`, :ref:`fmul <i_fmul>`, :ref:`fdiv <i_fdiv>`,
1448 :ref:`frem <i_frem>`) have the following flags that can set to enable
1449 otherwise unsafe floating point operations
1452 No NaNs - Allow optimizations to assume the arguments and result are not
1453 NaN. Such optimizations are required to retain defined behavior over
1454 NaNs, but the value of the result is undefined.
1457 No Infs - Allow optimizations to assume the arguments and result are not
1458 +/-Inf. Such optimizations are required to retain defined behavior over
1459 +/-Inf, but the value of the result is undefined.
1462 No Signed Zeros - Allow optimizations to treat the sign of a zero
1463 argument or result as insignificant.
1466 Allow Reciprocal - Allow optimizations to use the reciprocal of an
1467 argument rather than perform division.
1470 Fast - Allow algebraically equivalent transformations that may
1471 dramatically change results in floating point (e.g. reassociate). This
1472 flag implies all the others.
1479 The LLVM type system is one of the most important features of the
1480 intermediate representation. Being typed enables a number of
1481 optimizations to be performed on the intermediate representation
1482 directly, without having to do extra analyses on the side before the
1483 transformation. A strong type system makes it easier to read the
1484 generated code and enables novel analyses and transformations that are
1485 not feasible to perform on normal three address code representations.
1487 .. _typeclassifications:
1489 Type Classifications
1490 --------------------
1492 The types fall into a few useful classifications:
1501 * - :ref:`integer <t_integer>`
1502 - ``i1``, ``i2``, ``i3``, ... ``i8``, ... ``i16``, ... ``i32``, ...
1505 * - :ref:`floating point <t_floating>`
1506 - ``half``, ``float``, ``double``, ``x86_fp80``, ``fp128``,
1514 - :ref:`integer <t_integer>`, :ref:`floating point <t_floating>`,
1515 :ref:`pointer <t_pointer>`, :ref:`vector <t_vector>`,
1516 :ref:`structure <t_struct>`, :ref:`array <t_array>`,
1517 :ref:`label <t_label>`, :ref:`metadata <t_metadata>`.
1519 * - :ref:`primitive <t_primitive>`
1520 - :ref:`label <t_label>`,
1521 :ref:`void <t_void>`,
1522 :ref:`integer <t_integer>`,
1523 :ref:`floating point <t_floating>`,
1524 :ref:`x86mmx <t_x86mmx>`,
1525 :ref:`metadata <t_metadata>`.
1527 * - :ref:`derived <t_derived>`
1528 - :ref:`array <t_array>`,
1529 :ref:`function <t_function>`,
1530 :ref:`pointer <t_pointer>`,
1531 :ref:`structure <t_struct>`,
1532 :ref:`vector <t_vector>`,
1533 :ref:`opaque <t_opaque>`.
1535 The :ref:`first class <t_firstclass>` types are perhaps the most important.
1536 Values of these types are the only ones which can be produced by
1544 The primitive types are the fundamental building blocks of the LLVM
1555 The integer type is a very simple type that simply specifies an
1556 arbitrary bit width for the integer type desired. Any bit width from 1
1557 bit to 2\ :sup:`23`\ -1 (about 8 million) can be specified.
1566 The number of bits the integer will occupy is specified by the ``N``
1572 +----------------+------------------------------------------------+
1573 | ``i1`` | a single-bit integer. |
1574 +----------------+------------------------------------------------+
1575 | ``i32`` | a 32-bit integer. |
1576 +----------------+------------------------------------------------+
1577 | ``i1942652`` | a really big integer of over 1 million bits. |
1578 +----------------+------------------------------------------------+
1582 Floating Point Types
1583 ^^^^^^^^^^^^^^^^^^^^
1592 - 16-bit floating point value
1595 - 32-bit floating point value
1598 - 64-bit floating point value
1601 - 128-bit floating point value (112-bit mantissa)
1604 - 80-bit floating point value (X87)
1607 - 128-bit floating point value (two 64-bits)
1617 The x86mmx type represents a value held in an MMX register on an x86
1618 machine. The operations allowed on it are quite limited: parameters and
1619 return values, load and store, and bitcast. User-specified MMX
1620 instructions are represented as intrinsic or asm calls with arguments
1621 and/or results of this type. There are no arrays, vectors or constants
1639 The void type does not represent any value and has no size.
1656 The label type represents code labels.
1673 The metadata type represents embedded metadata. No derived types may be
1674 created from metadata except for :ref:`function <t_function>` arguments.
1688 The real power in LLVM comes from the derived types in the system. This
1689 is what allows a programmer to represent arrays, functions, pointers,
1690 and other useful types. Each of these types contain one or more element
1691 types which may be a primitive type, or another derived type. For
1692 example, it is possible to have a two dimensional array, using an array
1693 as the element type of another array.
1700 Aggregate Types are a subset of derived types that can contain multiple
1701 member types. :ref:`Arrays <t_array>` and :ref:`structs <t_struct>` are
1702 aggregate types. :ref:`Vectors <t_vector>` are not considered to be
1713 The array type is a very simple derived type that arranges elements
1714 sequentially in memory. The array type requires a size (number of
1715 elements) and an underlying data type.
1722 [<# elements> x <elementtype>]
1724 The number of elements is a constant integer value; ``elementtype`` may
1725 be any type with a size.
1730 +------------------+--------------------------------------+
1731 | ``[40 x i32]`` | Array of 40 32-bit integer values. |
1732 +------------------+--------------------------------------+
1733 | ``[41 x i32]`` | Array of 41 32-bit integer values. |
1734 +------------------+--------------------------------------+
1735 | ``[4 x i8]`` | Array of 4 8-bit integer values. |
1736 +------------------+--------------------------------------+
1738 Here are some examples of multidimensional arrays:
1740 +-----------------------------+----------------------------------------------------------+
1741 | ``[3 x [4 x i32]]`` | 3x4 array of 32-bit integer values. |
1742 +-----------------------------+----------------------------------------------------------+
1743 | ``[12 x [10 x float]]`` | 12x10 array of single precision floating point values. |
1744 +-----------------------------+----------------------------------------------------------+
1745 | ``[2 x [3 x [4 x i16]]]`` | 2x3x4 array of 16-bit integer values. |
1746 +-----------------------------+----------------------------------------------------------+
1748 There is no restriction on indexing beyond the end of the array implied
1749 by a static type (though there are restrictions on indexing beyond the
1750 bounds of an allocated object in some cases). This means that
1751 single-dimension 'variable sized array' addressing can be implemented in
1752 LLVM with a zero length array type. An implementation of 'pascal style
1753 arrays' in LLVM could use the type "``{ i32, [0 x float]}``", for
1764 The function type can be thought of as a function signature. It consists of a
1765 return type and a list of formal parameter types. The return type of a function
1766 type is a void type or first class type --- except for :ref:`label <t_label>`
1767 and :ref:`metadata <t_metadata>` types.
1774 <returntype> (<parameter list>)
1776 ...where '``<parameter list>``' is a comma-separated list of type
1777 specifiers. Optionally, the parameter list may include a type ``...``, which
1778 indicates that the function takes a variable number of arguments. Variable
1779 argument functions can access their arguments with the :ref:`variable argument
1780 handling intrinsic <int_varargs>` functions. '``<returntype>``' is any type
1781 except :ref:`label <t_label>` and :ref:`metadata <t_metadata>`.
1786 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1787 | ``i32 (i32)`` | function taking an ``i32``, returning an ``i32`` |
1788 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1789 | ``float (i16, i32 *) *`` | :ref:`Pointer <t_pointer>` to a function that takes an ``i16`` and a :ref:`pointer <t_pointer>` to ``i32``, returning ``float``. |
1790 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1791 | ``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. |
1792 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1793 | ``{i32, i32} (i32)`` | A function taking an ``i32``, returning a :ref:`structure <t_struct>` containing two ``i32`` values |
1794 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1804 The structure type is used to represent a collection of data members
1805 together in memory. The elements of a structure may be any type that has
1808 Structures in memory are accessed using '``load``' and '``store``' by
1809 getting a pointer to a field with the '``getelementptr``' instruction.
1810 Structures in registers are accessed using the '``extractvalue``' and
1811 '``insertvalue``' instructions.
1813 Structures may optionally be "packed" structures, which indicate that
1814 the alignment of the struct is one byte, and that there is no padding
1815 between the elements. In non-packed structs, padding between field types
1816 is inserted as defined by the DataLayout string in the module, which is
1817 required to match what the underlying code generator expects.
1819 Structures can either be "literal" or "identified". A literal structure
1820 is defined inline with other types (e.g. ``{i32, i32}*``) whereas
1821 identified types are always defined at the top level with a name.
1822 Literal types are uniqued by their contents and can never be recursive
1823 or opaque since there is no way to write one. Identified types can be
1824 recursive, can be opaqued, and are never uniqued.
1831 %T1 = type { <type list> } ; Identified normal struct type
1832 %T2 = type <{ <type list> }> ; Identified packed struct type
1837 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1838 | ``{ i32, i32, i32 }`` | A triple of three ``i32`` values |
1839 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1840 | ``{ 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``. |
1841 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1842 | ``<{ i8, i32 }>`` | A packed struct known to be 5 bytes in size. |
1843 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1847 Opaque Structure Types
1848 ^^^^^^^^^^^^^^^^^^^^^^
1853 Opaque structure types are used to represent named structure types that
1854 do not have a body specified. This corresponds (for example) to the C
1855 notion of a forward declared structure.
1868 +--------------+-------------------+
1869 | ``opaque`` | An opaque type. |
1870 +--------------+-------------------+
1880 The pointer type is used to specify memory locations. Pointers are
1881 commonly used to reference objects in memory.
1883 Pointer types may have an optional address space attribute defining the
1884 numbered address space where the pointed-to object resides. The default
1885 address space is number zero. The semantics of non-zero address spaces
1886 are target-specific.
1888 Note that LLVM does not permit pointers to void (``void*``) nor does it
1889 permit pointers to labels (``label*``). Use ``i8*`` instead.
1901 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1902 | ``[4 x i32]*`` | A :ref:`pointer <t_pointer>` to :ref:`array <t_array>` of four ``i32`` values. |
1903 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1904 | ``i32 (i32*) *`` | A :ref:`pointer <t_pointer>` to a :ref:`function <t_function>` that takes an ``i32*``, returning an ``i32``. |
1905 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1906 | ``i32 addrspace(5)*`` | A :ref:`pointer <t_pointer>` to an ``i32`` value that resides in address space #5. |
1907 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1917 A vector type is a simple derived type that represents a vector of
1918 elements. Vector types are used when multiple primitive data are
1919 operated in parallel using a single instruction (SIMD). A vector type
1920 requires a size (number of elements) and an underlying primitive data
1921 type. Vector types are considered :ref:`first class <t_firstclass>`.
1928 < <# elements> x <elementtype> >
1930 The number of elements is a constant integer value larger than 0;
1931 elementtype may be any integer or floating point type, or a pointer to
1932 these types. Vectors of size zero are not allowed.
1937 +-------------------+--------------------------------------------------+
1938 | ``<4 x i32>`` | Vector of 4 32-bit integer values. |
1939 +-------------------+--------------------------------------------------+
1940 | ``<8 x float>`` | Vector of 8 32-bit floating-point values. |
1941 +-------------------+--------------------------------------------------+
1942 | ``<2 x i64>`` | Vector of 2 64-bit integer values. |
1943 +-------------------+--------------------------------------------------+
1944 | ``<4 x i64*>`` | Vector of 4 pointers to 64-bit integer values. |
1945 +-------------------+--------------------------------------------------+
1950 LLVM has several different basic types of constants. This section
1951 describes them all and their syntax.
1956 **Boolean constants**
1957 The two strings '``true``' and '``false``' are both valid constants
1959 **Integer constants**
1960 Standard integers (such as '4') are constants of the
1961 :ref:`integer <t_integer>` type. Negative numbers may be used with
1963 **Floating point constants**
1964 Floating point constants use standard decimal notation (e.g.
1965 123.421), exponential notation (e.g. 1.23421e+2), or a more precise
1966 hexadecimal notation (see below). The assembler requires the exact
1967 decimal value of a floating-point constant. For example, the
1968 assembler accepts 1.25 but rejects 1.3 because 1.3 is a repeating
1969 decimal in binary. Floating point constants must have a :ref:`floating
1970 point <t_floating>` type.
1971 **Null pointer constants**
1972 The identifier '``null``' is recognized as a null pointer constant
1973 and must be of :ref:`pointer type <t_pointer>`.
1975 The one non-intuitive notation for constants is the hexadecimal form of
1976 floating point constants. For example, the form
1977 '``double 0x432ff973cafa8000``' is equivalent to (but harder to read
1978 than) '``double 4.5e+15``'. The only time hexadecimal floating point
1979 constants are required (and the only time that they are generated by the
1980 disassembler) is when a floating point constant must be emitted but it
1981 cannot be represented as a decimal floating point number in a reasonable
1982 number of digits. For example, NaN's, infinities, and other special
1983 values are represented in their IEEE hexadecimal format so that assembly
1984 and disassembly do not cause any bits to change in the constants.
1986 When using the hexadecimal form, constants of types half, float, and
1987 double are represented using the 16-digit form shown above (which
1988 matches the IEEE754 representation for double); half and float values
1989 must, however, be exactly representable as IEEE 754 half and single
1990 precision, respectively. Hexadecimal format is always used for long
1991 double, and there are three forms of long double. The 80-bit format used
1992 by x86 is represented as ``0xK`` followed by 20 hexadecimal digits. The
1993 128-bit format used by PowerPC (two adjacent doubles) is represented by
1994 ``0xM`` followed by 32 hexadecimal digits. The IEEE 128-bit format is
1995 represented by ``0xL`` followed by 32 hexadecimal digits. Long doubles
1996 will only work if they match the long double format on your target.
1997 The IEEE 16-bit format (half precision) is represented by ``0xH``
1998 followed by 4 hexadecimal digits. All hexadecimal formats are big-endian
1999 (sign bit at the left).
2001 There are no constants of type x86mmx.
2003 .. _complexconstants:
2008 Complex constants are a (potentially recursive) combination of simple
2009 constants and smaller complex constants.
2011 **Structure constants**
2012 Structure constants are represented with notation similar to
2013 structure type definitions (a comma separated list of elements,
2014 surrounded by braces (``{}``)). For example:
2015 "``{ i32 4, float 17.0, i32* @G }``", where "``@G``" is declared as
2016 "``@G = external global i32``". Structure constants must have
2017 :ref:`structure type <t_struct>`, and the number and types of elements
2018 must match those specified by the type.
2020 Array constants are represented with notation similar to array type
2021 definitions (a comma separated list of elements, surrounded by
2022 square brackets (``[]``)). For example:
2023 "``[ i32 42, i32 11, i32 74 ]``". Array constants must have
2024 :ref:`array type <t_array>`, and the number and types of elements must
2025 match those specified by the type.
2026 **Vector constants**
2027 Vector constants are represented with notation similar to vector
2028 type definitions (a comma separated list of elements, surrounded by
2029 less-than/greater-than's (``<>``)). For example:
2030 "``< i32 42, i32 11, i32 74, i32 100 >``". Vector constants
2031 must have :ref:`vector type <t_vector>`, and the number and types of
2032 elements must match those specified by the type.
2033 **Zero initialization**
2034 The string '``zeroinitializer``' can be used to zero initialize a
2035 value to zero of *any* type, including scalar and
2036 :ref:`aggregate <t_aggregate>` types. This is often used to avoid
2037 having to print large zero initializers (e.g. for large arrays) and
2038 is always exactly equivalent to using explicit zero initializers.
2040 A metadata node is a structure-like constant with :ref:`metadata
2041 type <t_metadata>`. For example:
2042 "``metadata !{ i32 0, metadata !"test" }``". Unlike other
2043 constants that are meant to be interpreted as part of the
2044 instruction stream, metadata is a place to attach additional
2045 information such as debug info.
2047 Global Variable and Function Addresses
2048 --------------------------------------
2050 The addresses of :ref:`global variables <globalvars>` and
2051 :ref:`functions <functionstructure>` are always implicitly valid
2052 (link-time) constants. These constants are explicitly referenced when
2053 the :ref:`identifier for the global <identifiers>` is used and always have
2054 :ref:`pointer <t_pointer>` type. For example, the following is a legal LLVM
2057 .. code-block:: llvm
2061 @Z = global [2 x i32*] [ i32* @X, i32* @Y ]
2068 The string '``undef``' can be used anywhere a constant is expected, and
2069 indicates that the user of the value may receive an unspecified
2070 bit-pattern. Undefined values may be of any type (other than '``label``'
2071 or '``void``') and be used anywhere a constant is permitted.
2073 Undefined values are useful because they indicate to the compiler that
2074 the program is well defined no matter what value is used. This gives the
2075 compiler more freedom to optimize. Here are some examples of
2076 (potentially surprising) transformations that are valid (in pseudo IR):
2078 .. code-block:: llvm
2088 This is safe because all of the output bits are affected by the undef
2089 bits. Any output bit can have a zero or one depending on the input bits.
2091 .. code-block:: llvm
2102 These logical operations have bits that are not always affected by the
2103 input. For example, if ``%X`` has a zero bit, then the output of the
2104 '``and``' operation will always be a zero for that bit, no matter what
2105 the corresponding bit from the '``undef``' is. As such, it is unsafe to
2106 optimize or assume that the result of the '``and``' is '``undef``'.
2107 However, it is safe to assume that all bits of the '``undef``' could be
2108 0, and optimize the '``and``' to 0. Likewise, it is safe to assume that
2109 all the bits of the '``undef``' operand to the '``or``' could be set,
2110 allowing the '``or``' to be folded to -1.
2112 .. code-block:: llvm
2114 %A = select undef, %X, %Y
2115 %B = select undef, 42, %Y
2116 %C = select %X, %Y, undef
2126 This set of examples shows that undefined '``select``' (and conditional
2127 branch) conditions can go *either way*, but they have to come from one
2128 of the two operands. In the ``%A`` example, if ``%X`` and ``%Y`` were
2129 both known to have a clear low bit, then ``%A`` would have to have a
2130 cleared low bit. However, in the ``%C`` example, the optimizer is
2131 allowed to assume that the '``undef``' operand could be the same as
2132 ``%Y``, allowing the whole '``select``' to be eliminated.
2134 .. code-block:: llvm
2136 %A = xor undef, undef
2153 This example points out that two '``undef``' operands are not
2154 necessarily the same. This can be surprising to people (and also matches
2155 C semantics) where they assume that "``X^X``" is always zero, even if
2156 ``X`` is undefined. This isn't true for a number of reasons, but the
2157 short answer is that an '``undef``' "variable" can arbitrarily change
2158 its value over its "live range". This is true because the variable
2159 doesn't actually *have a live range*. Instead, the value is logically
2160 read from arbitrary registers that happen to be around when needed, so
2161 the value is not necessarily consistent over time. In fact, ``%A`` and
2162 ``%C`` need to have the same semantics or the core LLVM "replace all
2163 uses with" concept would not hold.
2165 .. code-block:: llvm
2173 These examples show the crucial difference between an *undefined value*
2174 and *undefined behavior*. An undefined value (like '``undef``') is
2175 allowed to have an arbitrary bit-pattern. This means that the ``%A``
2176 operation can be constant folded to '``undef``', because the '``undef``'
2177 could be an SNaN, and ``fdiv`` is not (currently) defined on SNaN's.
2178 However, in the second example, we can make a more aggressive
2179 assumption: because the ``undef`` is allowed to be an arbitrary value,
2180 we are allowed to assume that it could be zero. Since a divide by zero
2181 has *undefined behavior*, we are allowed to assume that the operation
2182 does not execute at all. This allows us to delete the divide and all
2183 code after it. Because the undefined operation "can't happen", the
2184 optimizer can assume that it occurs in dead code.
2186 .. code-block:: llvm
2188 a: store undef -> %X
2189 b: store %X -> undef
2194 These examples reiterate the ``fdiv`` example: a store *of* an undefined
2195 value can be assumed to not have any effect; we can assume that the
2196 value is overwritten with bits that happen to match what was already
2197 there. However, a store *to* an undefined location could clobber
2198 arbitrary memory, therefore, it has undefined behavior.
2205 Poison values are similar to :ref:`undef values <undefvalues>`, however
2206 they also represent the fact that an instruction or constant expression
2207 which cannot evoke side effects has nevertheless detected a condition
2208 which results in undefined behavior.
2210 There is currently no way of representing a poison value in the IR; they
2211 only exist when produced by operations such as :ref:`add <i_add>` with
2214 Poison value behavior is defined in terms of value *dependence*:
2216 - Values other than :ref:`phi <i_phi>` nodes depend on their operands.
2217 - :ref:`Phi <i_phi>` nodes depend on the operand corresponding to
2218 their dynamic predecessor basic block.
2219 - Function arguments depend on the corresponding actual argument values
2220 in the dynamic callers of their functions.
2221 - :ref:`Call <i_call>` instructions depend on the :ref:`ret <i_ret>`
2222 instructions that dynamically transfer control back to them.
2223 - :ref:`Invoke <i_invoke>` instructions depend on the
2224 :ref:`ret <i_ret>`, :ref:`resume <i_resume>`, or exception-throwing
2225 call instructions that dynamically transfer control back to them.
2226 - Non-volatile loads and stores depend on the most recent stores to all
2227 of the referenced memory addresses, following the order in the IR
2228 (including loads and stores implied by intrinsics such as
2229 :ref:`@llvm.memcpy <int_memcpy>`.)
2230 - An instruction with externally visible side effects depends on the
2231 most recent preceding instruction with externally visible side
2232 effects, following the order in the IR. (This includes :ref:`volatile
2233 operations <volatile>`.)
2234 - An instruction *control-depends* on a :ref:`terminator
2235 instruction <terminators>` if the terminator instruction has
2236 multiple successors and the instruction is always executed when
2237 control transfers to one of the successors, and may not be executed
2238 when control is transferred to another.
2239 - Additionally, an instruction also *control-depends* on a terminator
2240 instruction if the set of instructions it otherwise depends on would
2241 be different if the terminator had transferred control to a different
2243 - Dependence is transitive.
2245 Poison Values have the same behavior as :ref:`undef values <undefvalues>`,
2246 with the additional affect that any instruction which has a *dependence*
2247 on a poison value has undefined behavior.
2249 Here are some examples:
2251 .. code-block:: llvm
2254 %poison = sub nuw i32 0, 1 ; Results in a poison value.
2255 %still_poison = and i32 %poison, 0 ; 0, but also poison.
2256 %poison_yet_again = getelementptr i32* @h, i32 %still_poison
2257 store i32 0, i32* %poison_yet_again ; memory at @h[0] is poisoned
2259 store i32 %poison, i32* @g ; Poison value stored to memory.
2260 %poison2 = load i32* @g ; Poison value loaded back from memory.
2262 store volatile i32 %poison, i32* @g ; External observation; undefined behavior.
2264 %narrowaddr = bitcast i32* @g to i16*
2265 %wideaddr = bitcast i32* @g to i64*
2266 %poison3 = load i16* %narrowaddr ; Returns a poison value.
2267 %poison4 = load i64* %wideaddr ; Returns a poison value.
2269 %cmp = icmp slt i32 %poison, 0 ; Returns a poison value.
2270 br i1 %cmp, label %true, label %end ; Branch to either destination.
2273 store volatile i32 0, i32* @g ; This is control-dependent on %cmp, so
2274 ; it has undefined behavior.
2278 %p = phi i32 [ 0, %entry ], [ 1, %true ]
2279 ; Both edges into this PHI are
2280 ; control-dependent on %cmp, so this
2281 ; always results in a poison value.
2283 store volatile i32 0, i32* @g ; This would depend on the store in %true
2284 ; if %cmp is true, or the store in %entry
2285 ; otherwise, so this is undefined behavior.
2287 br i1 %cmp, label %second_true, label %second_end
2288 ; The same branch again, but this time the
2289 ; true block doesn't have side effects.
2296 store volatile i32 0, i32* @g ; This time, the instruction always depends
2297 ; on the store in %end. Also, it is
2298 ; control-equivalent to %end, so this is
2299 ; well-defined (ignoring earlier undefined
2300 ; behavior in this example).
2304 Addresses of Basic Blocks
2305 -------------------------
2307 ``blockaddress(@function, %block)``
2309 The '``blockaddress``' constant computes the address of the specified
2310 basic block in the specified function, and always has an ``i8*`` type.
2311 Taking the address of the entry block is illegal.
2313 This value only has defined behavior when used as an operand to the
2314 ':ref:`indirectbr <i_indirectbr>`' instruction, or for comparisons
2315 against null. Pointer equality tests between labels addresses results in
2316 undefined behavior --- though, again, comparison against null is ok, and
2317 no label is equal to the null pointer. This may be passed around as an
2318 opaque pointer sized value as long as the bits are not inspected. This
2319 allows ``ptrtoint`` and arithmetic to be performed on these values so
2320 long as the original value is reconstituted before the ``indirectbr``
2323 Finally, some targets may provide defined semantics when using the value
2324 as the operand to an inline assembly, but that is target specific.
2328 Constant Expressions
2329 --------------------
2331 Constant expressions are used to allow expressions involving other
2332 constants to be used as constants. Constant expressions may be of any
2333 :ref:`first class <t_firstclass>` type and may involve any LLVM operation
2334 that does not have side effects (e.g. load and call are not supported).
2335 The following is the syntax for constant expressions:
2337 ``trunc (CST to TYPE)``
2338 Truncate a constant to another type. The bit size of CST must be
2339 larger than the bit size of TYPE. Both types must be integers.
2340 ``zext (CST to TYPE)``
2341 Zero extend a constant to another type. The bit size of CST must be
2342 smaller than the bit size of TYPE. Both types must be integers.
2343 ``sext (CST to TYPE)``
2344 Sign extend a constant to another type. The bit size of CST must be
2345 smaller than the bit size of TYPE. Both types must be integers.
2346 ``fptrunc (CST to TYPE)``
2347 Truncate a floating point constant to another floating point type.
2348 The size of CST must be larger than the size of TYPE. Both types
2349 must be floating point.
2350 ``fpext (CST to TYPE)``
2351 Floating point extend a constant to another type. The size of CST
2352 must be smaller or equal to the size of TYPE. Both types must be
2354 ``fptoui (CST to TYPE)``
2355 Convert a floating point constant to the corresponding unsigned
2356 integer constant. TYPE must be a scalar or vector integer type. CST
2357 must be of scalar or vector floating point type. Both CST and TYPE
2358 must be scalars, or vectors of the same number of elements. If the
2359 value won't fit in the integer type, the results are undefined.
2360 ``fptosi (CST to TYPE)``
2361 Convert a floating point constant to the corresponding signed
2362 integer constant. TYPE must be a scalar or vector integer type. CST
2363 must be of scalar or vector floating point type. Both CST and TYPE
2364 must be scalars, or vectors of the same number of elements. If the
2365 value won't fit in the integer type, the results are undefined.
2366 ``uitofp (CST to TYPE)``
2367 Convert an unsigned integer constant to the corresponding floating
2368 point constant. TYPE must be a scalar or vector floating point type.
2369 CST must be of scalar or vector integer type. Both CST and TYPE must
2370 be scalars, or vectors of the same number of elements. If the value
2371 won't fit in the floating point type, the results are undefined.
2372 ``sitofp (CST to TYPE)``
2373 Convert a signed integer constant to the corresponding floating
2374 point constant. TYPE must be a scalar or vector floating point type.
2375 CST must be of scalar or vector integer type. Both CST and TYPE must
2376 be scalars, or vectors of the same number of elements. If the value
2377 won't fit in the floating point type, the results are undefined.
2378 ``ptrtoint (CST to TYPE)``
2379 Convert a pointer typed constant to the corresponding integer
2380 constant. ``TYPE`` must be an integer type. ``CST`` must be of
2381 pointer type. The ``CST`` value is zero extended, truncated, or
2382 unchanged to make it fit in ``TYPE``.
2383 ``inttoptr (CST to TYPE)``
2384 Convert an integer constant to a pointer constant. TYPE must be a
2385 pointer type. CST must be of integer type. The CST value is zero
2386 extended, truncated, or unchanged to make it fit in a pointer size.
2387 This one is *really* dangerous!
2388 ``bitcast (CST to TYPE)``
2389 Convert a constant, CST, to another TYPE. The constraints of the
2390 operands are the same as those for the :ref:`bitcast
2391 instruction <i_bitcast>`.
2392 ``getelementptr (CSTPTR, IDX0, IDX1, ...)``, ``getelementptr inbounds (CSTPTR, IDX0, IDX1, ...)``
2393 Perform the :ref:`getelementptr operation <i_getelementptr>` on
2394 constants. As with the :ref:`getelementptr <i_getelementptr>`
2395 instruction, the index list may have zero or more indexes, which are
2396 required to make sense for the type of "CSTPTR".
2397 ``select (COND, VAL1, VAL2)``
2398 Perform the :ref:`select operation <i_select>` on constants.
2399 ``icmp COND (VAL1, VAL2)``
2400 Performs the :ref:`icmp operation <i_icmp>` on constants.
2401 ``fcmp COND (VAL1, VAL2)``
2402 Performs the :ref:`fcmp operation <i_fcmp>` on constants.
2403 ``extractelement (VAL, IDX)``
2404 Perform the :ref:`extractelement operation <i_extractelement>` on
2406 ``insertelement (VAL, ELT, IDX)``
2407 Perform the :ref:`insertelement operation <i_insertelement>` on
2409 ``shufflevector (VEC1, VEC2, IDXMASK)``
2410 Perform the :ref:`shufflevector operation <i_shufflevector>` on
2412 ``extractvalue (VAL, IDX0, IDX1, ...)``
2413 Perform the :ref:`extractvalue operation <i_extractvalue>` on
2414 constants. The index list is interpreted in a similar manner as
2415 indices in a ':ref:`getelementptr <i_getelementptr>`' operation. At
2416 least one index value must be specified.
2417 ``insertvalue (VAL, ELT, IDX0, IDX1, ...)``
2418 Perform the :ref:`insertvalue operation <i_insertvalue>` on constants.
2419 The index list is interpreted in a similar manner as indices in a
2420 ':ref:`getelementptr <i_getelementptr>`' operation. At least one index
2421 value must be specified.
2422 ``OPCODE (LHS, RHS)``
2423 Perform the specified operation of the LHS and RHS constants. OPCODE
2424 may be any of the :ref:`binary <binaryops>` or :ref:`bitwise
2425 binary <bitwiseops>` operations. The constraints on operands are
2426 the same as those for the corresponding instruction (e.g. no bitwise
2427 operations on floating point values are allowed).
2434 Inline Assembler Expressions
2435 ----------------------------
2437 LLVM supports inline assembler expressions (as opposed to :ref:`Module-Level
2438 Inline Assembly <moduleasm>`) through the use of a special value. This
2439 value represents the inline assembler as a string (containing the
2440 instructions to emit), a list of operand constraints (stored as a
2441 string), a flag that indicates whether or not the inline asm expression
2442 has side effects, and a flag indicating whether the function containing
2443 the asm needs to align its stack conservatively. An example inline
2444 assembler expression is:
2446 .. code-block:: llvm
2448 i32 (i32) asm "bswap $0", "=r,r"
2450 Inline assembler expressions may **only** be used as the callee operand
2451 of a :ref:`call <i_call>` or an :ref:`invoke <i_invoke>` instruction.
2452 Thus, typically we have:
2454 .. code-block:: llvm
2456 %X = call i32 asm "bswap $0", "=r,r"(i32 %Y)
2458 Inline asms with side effects not visible in the constraint list must be
2459 marked as having side effects. This is done through the use of the
2460 '``sideeffect``' keyword, like so:
2462 .. code-block:: llvm
2464 call void asm sideeffect "eieio", ""()
2466 In some cases inline asms will contain code that will not work unless
2467 the stack is aligned in some way, such as calls or SSE instructions on
2468 x86, yet will not contain code that does that alignment within the asm.
2469 The compiler should make conservative assumptions about what the asm
2470 might contain and should generate its usual stack alignment code in the
2471 prologue if the '``alignstack``' keyword is present:
2473 .. code-block:: llvm
2475 call void asm alignstack "eieio", ""()
2477 Inline asms also support using non-standard assembly dialects. The
2478 assumed dialect is ATT. When the '``inteldialect``' keyword is present,
2479 the inline asm is using the Intel dialect. Currently, ATT and Intel are
2480 the only supported dialects. An example is:
2482 .. code-block:: llvm
2484 call void asm inteldialect "eieio", ""()
2486 If multiple keywords appear the '``sideeffect``' keyword must come
2487 first, the '``alignstack``' keyword second and the '``inteldialect``'
2493 The call instructions that wrap inline asm nodes may have a
2494 "``!srcloc``" MDNode attached to it that contains a list of constant
2495 integers. If present, the code generator will use the integer as the
2496 location cookie value when report errors through the ``LLVMContext``
2497 error reporting mechanisms. This allows a front-end to correlate backend
2498 errors that occur with inline asm back to the source code that produced
2501 .. code-block:: llvm
2503 call void asm sideeffect "something bad", ""(), !srcloc !42
2505 !42 = !{ i32 1234567 }
2507 It is up to the front-end to make sense of the magic numbers it places
2508 in the IR. If the MDNode contains multiple constants, the code generator
2509 will use the one that corresponds to the line of the asm that the error
2514 Metadata Nodes and Metadata Strings
2515 -----------------------------------
2517 LLVM IR allows metadata to be attached to instructions in the program
2518 that can convey extra information about the code to the optimizers and
2519 code generator. One example application of metadata is source-level
2520 debug information. There are two metadata primitives: strings and nodes.
2521 All metadata has the ``metadata`` type and is identified in syntax by a
2522 preceding exclamation point ('``!``').
2524 A metadata string is a string surrounded by double quotes. It can
2525 contain any character by escaping non-printable characters with
2526 "``\xx``" where "``xx``" is the two digit hex code. For example:
2529 Metadata nodes are represented with notation similar to structure
2530 constants (a comma separated list of elements, surrounded by braces and
2531 preceded by an exclamation point). Metadata nodes can have any values as
2532 their operand. For example:
2534 .. code-block:: llvm
2536 !{ metadata !"test\00", i32 10}
2538 A :ref:`named metadata <namedmetadatastructure>` is a collection of
2539 metadata nodes, which can be looked up in the module symbol table. For
2542 .. code-block:: llvm
2544 !foo = metadata !{!4, !3}
2546 Metadata can be used as function arguments. Here ``llvm.dbg.value``
2547 function is using two metadata arguments:
2549 .. code-block:: llvm
2551 call void @llvm.dbg.value(metadata !24, i64 0, metadata !25)
2553 Metadata can be attached with an instruction. Here metadata ``!21`` is
2554 attached to the ``add`` instruction using the ``!dbg`` identifier:
2556 .. code-block:: llvm
2558 %indvar.next = add i64 %indvar, 1, !dbg !21
2560 More information about specific metadata nodes recognized by the
2561 optimizers and code generator is found below.
2566 In LLVM IR, memory does not have types, so LLVM's own type system is not
2567 suitable for doing TBAA. Instead, metadata is added to the IR to
2568 describe a type system of a higher level language. This can be used to
2569 implement typical C/C++ TBAA, but it can also be used to implement
2570 custom alias analysis behavior for other languages.
2572 The current metadata format is very simple. TBAA metadata nodes have up
2573 to three fields, e.g.:
2575 .. code-block:: llvm
2577 !0 = metadata !{ metadata !"an example type tree" }
2578 !1 = metadata !{ metadata !"int", metadata !0 }
2579 !2 = metadata !{ metadata !"float", metadata !0 }
2580 !3 = metadata !{ metadata !"const float", metadata !2, i64 1 }
2582 The first field is an identity field. It can be any value, usually a
2583 metadata string, which uniquely identifies the type. The most important
2584 name in the tree is the name of the root node. Two trees with different
2585 root node names are entirely disjoint, even if they have leaves with
2588 The second field identifies the type's parent node in the tree, or is
2589 null or omitted for a root node. A type is considered to alias all of
2590 its descendants and all of its ancestors in the tree. Also, a type is
2591 considered to alias all types in other trees, so that bitcode produced
2592 from multiple front-ends is handled conservatively.
2594 If the third field is present, it's an integer which if equal to 1
2595 indicates that the type is "constant" (meaning
2596 ``pointsToConstantMemory`` should return true; see `other useful
2597 AliasAnalysis methods <AliasAnalysis.html#OtherItfs>`_).
2599 '``tbaa.struct``' Metadata
2600 ^^^^^^^^^^^^^^^^^^^^^^^^^^
2602 The :ref:`llvm.memcpy <int_memcpy>` is often used to implement
2603 aggregate assignment operations in C and similar languages, however it
2604 is defined to copy a contiguous region of memory, which is more than
2605 strictly necessary for aggregate types which contain holes due to
2606 padding. Also, it doesn't contain any TBAA information about the fields
2609 ``!tbaa.struct`` metadata can describe which memory subregions in a
2610 memcpy are padding and what the TBAA tags of the struct are.
2612 The current metadata format is very simple. ``!tbaa.struct`` metadata
2613 nodes are a list of operands which are in conceptual groups of three.
2614 For each group of three, the first operand gives the byte offset of a
2615 field in bytes, the second gives its size in bytes, and the third gives
2618 .. code-block:: llvm
2620 !4 = metadata !{ i64 0, i64 4, metadata !1, i64 8, i64 4, metadata !2 }
2622 This describes a struct with two fields. The first is at offset 0 bytes
2623 with size 4 bytes, and has tbaa tag !1. The second is at offset 8 bytes
2624 and has size 4 bytes and has tbaa tag !2.
2626 Note that the fields need not be contiguous. In this example, there is a
2627 4 byte gap between the two fields. This gap represents padding which
2628 does not carry useful data and need not be preserved.
2630 '``fpmath``' Metadata
2631 ^^^^^^^^^^^^^^^^^^^^^
2633 ``fpmath`` metadata may be attached to any instruction of floating point
2634 type. It can be used to express the maximum acceptable error in the
2635 result of that instruction, in ULPs, thus potentially allowing the
2636 compiler to use a more efficient but less accurate method of computing
2637 it. ULP is defined as follows:
2639 If ``x`` is a real number that lies between two finite consecutive
2640 floating-point numbers ``a`` and ``b``, without being equal to one
2641 of them, then ``ulp(x) = |b - a|``, otherwise ``ulp(x)`` is the
2642 distance between the two non-equal finite floating-point numbers
2643 nearest ``x``. Moreover, ``ulp(NaN)`` is ``NaN``.
2645 The metadata node shall consist of a single positive floating point
2646 number representing the maximum relative error, for example:
2648 .. code-block:: llvm
2650 !0 = metadata !{ float 2.5 } ; maximum acceptable inaccuracy is 2.5 ULPs
2652 '``range``' Metadata
2653 ^^^^^^^^^^^^^^^^^^^^
2655 ``range`` metadata may be attached only to loads of integer types. It
2656 expresses the possible ranges the loaded value is in. The ranges are
2657 represented with a flattened list of integers. The loaded value is known
2658 to be in the union of the ranges defined by each consecutive pair. Each
2659 pair has the following properties:
2661 - The type must match the type loaded by the instruction.
2662 - The pair ``a,b`` represents the range ``[a,b)``.
2663 - Both ``a`` and ``b`` are constants.
2664 - The range is allowed to wrap.
2665 - The range should not represent the full or empty set. That is,
2668 In addition, the pairs must be in signed order of the lower bound and
2669 they must be non-contiguous.
2673 .. code-block:: llvm
2675 %a = load i8* %x, align 1, !range !0 ; Can only be 0 or 1
2676 %b = load i8* %y, align 1, !range !1 ; Can only be 255 (-1), 0 or 1
2677 %c = load i8* %z, align 1, !range !2 ; Can only be 0, 1, 3, 4 or 5
2678 %d = load i8* %z, align 1, !range !3 ; Can only be -2, -1, 3, 4 or 5
2680 !0 = metadata !{ i8 0, i8 2 }
2681 !1 = metadata !{ i8 255, i8 2 }
2682 !2 = metadata !{ i8 0, i8 2, i8 3, i8 6 }
2683 !3 = metadata !{ i8 -2, i8 0, i8 3, i8 6 }
2688 It is sometimes useful to attach information to loop constructs. Currently,
2689 loop metadata is implemented as metadata attached to the branch instruction
2690 in the loop latch block. This type of metadata refer to a metadata node that is
2691 guaranteed to be separate for each loop. The loop identifier metadata is
2692 specified with the name ``llvm.loop``.
2694 The loop identifier metadata is implemented using a metadata that refers to
2695 itself to avoid merging it with any other identifier metadata, e.g.,
2696 during module linkage or function inlining. That is, each loop should refer
2697 to their own identification metadata even if they reside in separate functions.
2698 The following example contains loop identifier metadata for two separate loop
2701 .. code-block:: llvm
2703 !0 = metadata !{ metadata !0 }
2704 !1 = metadata !{ metadata !1 }
2706 The loop identifier metadata can be used to specify additional per-loop
2707 metadata. Any operands after the first operand can be treated as user-defined
2708 metadata. For example the ``llvm.vectorizer.unroll`` metadata is understood
2709 by the loop vectorizer to indicate how many times to unroll the loop:
2711 .. code-block:: llvm
2713 br i1 %exitcond, label %._crit_edge, label %.lr.ph, !llvm.loop !0
2715 !0 = metadata !{ metadata !0, metadata !1 }
2716 !1 = metadata !{ metadata !"llvm.vectorizer.unroll", i32 2 }
2721 Metadata types used to annotate memory accesses with information helpful
2722 for optimizations are prefixed with ``llvm.mem``.
2724 '``llvm.mem.parallel_loop_access``' Metadata
2725 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2727 For a loop to be parallel, in addition to using
2728 the ``llvm.loop`` metadata to mark the loop latch branch instruction,
2729 also all of the memory accessing instructions in the loop body need to be
2730 marked with the ``llvm.mem.parallel_loop_access`` metadata. If there
2731 is at least one memory accessing instruction not marked with the metadata,
2732 the loop must be considered a sequential loop. This causes parallel loops to be
2733 converted to sequential loops due to optimization passes that are unaware of
2734 the parallel semantics and that insert new memory instructions to the loop
2737 Example of a loop that is considered parallel due to its correct use of
2738 both ``llvm.loop`` and ``llvm.mem.parallel_loop_access``
2739 metadata types that refer to the same loop identifier metadata.
2741 .. code-block:: llvm
2745 %0 = load i32* %arrayidx, align 4, !llvm.mem.parallel_loop_access !0
2747 store i32 %0, i32* %arrayidx4, align 4, !llvm.mem.parallel_loop_access !0
2749 br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !0
2753 !0 = metadata !{ metadata !0 }
2755 It is also possible to have nested parallel loops. In that case the
2756 memory accesses refer to a list of loop identifier metadata nodes instead of
2757 the loop identifier metadata node directly:
2759 .. code-block:: llvm
2766 %0 = load i32* %arrayidx, align 4, !llvm.mem.parallel_loop_access !0
2768 store i32 %0, i32* %arrayidx4, align 4, !llvm.mem.parallel_loop_access !0
2770 br i1 %exitcond, label %inner.for.end, label %inner.for.body, !llvm.loop !1
2774 %0 = load i32* %arrayidx, align 4, !llvm.mem.parallel_loop_access !0
2776 store i32 %0, i32* %arrayidx4, align 4, !llvm.mem.parallel_loop_access !0
2778 br i1 %exitcond, label %outer.for.end, label %outer.for.body, !llvm.loop !2
2780 outer.for.end: ; preds = %for.body
2782 !0 = metadata !{ metadata !1, metadata !2 } ; a list of loop identifiers
2783 !1 = metadata !{ metadata !1 } ; an identifier for the inner loop
2784 !2 = metadata !{ metadata !2 } ; an identifier for the outer loop
2786 '``llvm.vectorizer``'
2787 ^^^^^^^^^^^^^^^^^^^^^
2789 Metadata prefixed with ``llvm.vectorizer`` is used to control per-loop
2790 vectorization parameters such as vectorization factor and unroll factor.
2792 ``llvm.vectorizer`` metadata should be used in conjunction with ``llvm.loop``
2793 loop identification metadata.
2795 '``llvm.vectorizer.unroll``' Metadata
2796 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2798 This metadata instructs the loop vectorizer to unroll the specified
2799 loop exactly ``N`` times.
2801 The first operand is the string ``llvm.vectorizer.unroll`` and the second
2802 operand is an integer specifying the unroll factor. For example:
2804 .. code-block:: llvm
2806 !0 = metadata !{ metadata !"llvm.vectorizer.unroll", i32 4 }
2808 Note that setting ``llvm.vectorizer.unroll`` to 1 disables unrolling of the
2811 If ``llvm.vectorizer.unroll`` is set to 0 then the amount of unrolling will be
2812 determined automatically.
2814 '``llvm.vectorizer.width``' Metadata
2815 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2817 This metadata sets the target width of the vectorizer to ``N``. Without
2818 this metadata, the vectorizer will choose a width automatically.
2819 Regardless of this metadata, the vectorizer will only vectorize loops if
2820 it believes it is valid to do so.
2822 The first operand is the string ``llvm.vectorizer.width`` and the second
2823 operand is an integer specifying the width. For example:
2825 .. code-block:: llvm
2827 !0 = metadata !{ metadata !"llvm.vectorizer.width", i32 4 }
2829 Note that setting ``llvm.vectorizer.width`` to 1 disables vectorization of the
2832 If ``llvm.vectorizer.width`` is set to 0 then the width will be determined
2835 Module Flags Metadata
2836 =====================
2838 Information about the module as a whole is difficult to convey to LLVM's
2839 subsystems. The LLVM IR isn't sufficient to transmit this information.
2840 The ``llvm.module.flags`` named metadata exists in order to facilitate
2841 this. These flags are in the form of key / value pairs --- much like a
2842 dictionary --- making it easy for any subsystem who cares about a flag to
2845 The ``llvm.module.flags`` metadata contains a list of metadata triplets.
2846 Each triplet has the following form:
2848 - The first element is a *behavior* flag, which specifies the behavior
2849 when two (or more) modules are merged together, and it encounters two
2850 (or more) metadata with the same ID. The supported behaviors are
2852 - The second element is a metadata string that is a unique ID for the
2853 metadata. Each module may only have one flag entry for each unique ID (not
2854 including entries with the **Require** behavior).
2855 - The third element is the value of the flag.
2857 When two (or more) modules are merged together, the resulting
2858 ``llvm.module.flags`` metadata is the union of the modules' flags. That is, for
2859 each unique metadata ID string, there will be exactly one entry in the merged
2860 modules ``llvm.module.flags`` metadata table, and the value for that entry will
2861 be determined by the merge behavior flag, as described below. The only exception
2862 is that entries with the *Require* behavior are always preserved.
2864 The following behaviors are supported:
2875 Emits an error if two values disagree, otherwise the resulting value
2876 is that of the operands.
2880 Emits a warning if two values disagree. The result value will be the
2881 operand for the flag from the first module being linked.
2885 Adds a requirement that another module flag be present and have a
2886 specified value after linking is performed. The value must be a
2887 metadata pair, where the first element of the pair is the ID of the
2888 module flag to be restricted, and the second element of the pair is
2889 the value the module flag should be restricted to. This behavior can
2890 be used to restrict the allowable results (via triggering of an
2891 error) of linking IDs with the **Override** behavior.
2895 Uses the specified value, regardless of the behavior or value of the
2896 other module. If both modules specify **Override**, but the values
2897 differ, an error will be emitted.
2901 Appends the two values, which are required to be metadata nodes.
2905 Appends the two values, which are required to be metadata
2906 nodes. However, duplicate entries in the second list are dropped
2907 during the append operation.
2909 It is an error for a particular unique flag ID to have multiple behaviors,
2910 except in the case of **Require** (which adds restrictions on another metadata
2911 value) or **Override**.
2913 An example of module flags:
2915 .. code-block:: llvm
2917 !0 = metadata !{ i32 1, metadata !"foo", i32 1 }
2918 !1 = metadata !{ i32 4, metadata !"bar", i32 37 }
2919 !2 = metadata !{ i32 2, metadata !"qux", i32 42 }
2920 !3 = metadata !{ i32 3, metadata !"qux",
2922 metadata !"foo", i32 1
2925 !llvm.module.flags = !{ !0, !1, !2, !3 }
2927 - Metadata ``!0`` has the ID ``!"foo"`` and the value '1'. The behavior
2928 if two or more ``!"foo"`` flags are seen is to emit an error if their
2929 values are not equal.
2931 - Metadata ``!1`` has the ID ``!"bar"`` and the value '37'. The
2932 behavior if two or more ``!"bar"`` flags are seen is to use the value
2935 - Metadata ``!2`` has the ID ``!"qux"`` and the value '42'. The
2936 behavior if two or more ``!"qux"`` flags are seen is to emit a
2937 warning if their values are not equal.
2939 - Metadata ``!3`` has the ID ``!"qux"`` and the value:
2943 metadata !{ metadata !"foo", i32 1 }
2945 The behavior is to emit an error if the ``llvm.module.flags`` does not
2946 contain a flag with the ID ``!"foo"`` that has the value '1' after linking is
2949 Objective-C Garbage Collection Module Flags Metadata
2950 ----------------------------------------------------
2952 On the Mach-O platform, Objective-C stores metadata about garbage
2953 collection in a special section called "image info". The metadata
2954 consists of a version number and a bitmask specifying what types of
2955 garbage collection are supported (if any) by the file. If two or more
2956 modules are linked together their garbage collection metadata needs to
2957 be merged rather than appended together.
2959 The Objective-C garbage collection module flags metadata consists of the
2960 following key-value pairs:
2969 * - ``Objective-C Version``
2970 - **[Required]** --- The Objective-C ABI version. Valid values are 1 and 2.
2972 * - ``Objective-C Image Info Version``
2973 - **[Required]** --- The version of the image info section. Currently
2976 * - ``Objective-C Image Info Section``
2977 - **[Required]** --- The section to place the metadata. Valid values are
2978 ``"__OBJC, __image_info, regular"`` for Objective-C ABI version 1, and
2979 ``"__DATA,__objc_imageinfo, regular, no_dead_strip"`` for
2980 Objective-C ABI version 2.
2982 * - ``Objective-C Garbage Collection``
2983 - **[Required]** --- Specifies whether garbage collection is supported or
2984 not. Valid values are 0, for no garbage collection, and 2, for garbage
2985 collection supported.
2987 * - ``Objective-C GC Only``
2988 - **[Optional]** --- Specifies that only garbage collection is supported.
2989 If present, its value must be 6. This flag requires that the
2990 ``Objective-C Garbage Collection`` flag have the value 2.
2992 Some important flag interactions:
2994 - If a module with ``Objective-C Garbage Collection`` set to 0 is
2995 merged with a module with ``Objective-C Garbage Collection`` set to
2996 2, then the resulting module has the
2997 ``Objective-C Garbage Collection`` flag set to 0.
2998 - A module with ``Objective-C Garbage Collection`` set to 0 cannot be
2999 merged with a module with ``Objective-C GC Only`` set to 6.
3001 Automatic Linker Flags Module Flags Metadata
3002 --------------------------------------------
3004 Some targets support embedding flags to the linker inside individual object
3005 files. Typically this is used in conjunction with language extensions which
3006 allow source files to explicitly declare the libraries they depend on, and have
3007 these automatically be transmitted to the linker via object files.
3009 These flags are encoded in the IR using metadata in the module flags section,
3010 using the ``Linker Options`` key. The merge behavior for this flag is required
3011 to be ``AppendUnique``, and the value for the key is expected to be a metadata
3012 node which should be a list of other metadata nodes, each of which should be a
3013 list of metadata strings defining linker options.
3015 For example, the following metadata section specifies two separate sets of
3016 linker options, presumably to link against ``libz`` and the ``Cocoa``
3019 !0 = metadata !{ i32 6, metadata !"Linker Options",
3021 metadata !{ metadata !"-lz" },
3022 metadata !{ metadata !"-framework", metadata !"Cocoa" } } }
3023 !llvm.module.flags = !{ !0 }
3025 The metadata encoding as lists of lists of options, as opposed to a collapsed
3026 list of options, is chosen so that the IR encoding can use multiple option
3027 strings to specify e.g., a single library, while still having that specifier be
3028 preserved as an atomic element that can be recognized by a target specific
3029 assembly writer or object file emitter.
3031 Each individual option is required to be either a valid option for the target's
3032 linker, or an option that is reserved by the target specific assembly writer or
3033 object file emitter. No other aspect of these options is defined by the IR.
3035 .. _intrinsicglobalvariables:
3037 Intrinsic Global Variables
3038 ==========================
3040 LLVM has a number of "magic" global variables that contain data that
3041 affect code generation or other IR semantics. These are documented here.
3042 All globals of this sort should have a section specified as
3043 "``llvm.metadata``". This section and all globals that start with
3044 "``llvm.``" are reserved for use by LLVM.
3048 The '``llvm.used``' Global Variable
3049 -----------------------------------
3051 The ``@llvm.used`` global is an array which has
3052 :ref:`appending linkage <linkage_appending>`. This array contains a list of
3053 pointers to named global variables, functions and aliases which may optionally
3054 have a pointer cast formed of bitcast or getelementptr. For example, a legal
3057 .. code-block:: llvm
3062 @llvm.used = appending global [2 x i8*] [
3064 i8* bitcast (i32* @Y to i8*)
3065 ], section "llvm.metadata"
3067 If a symbol appears in the ``@llvm.used`` list, then the compiler, assembler,
3068 and linker are required to treat the symbol as if there is a reference to the
3069 symbol that it cannot see (which is why they have to be named). For example, if
3070 a variable has internal linkage and no references other than that from the
3071 ``@llvm.used`` list, it cannot be deleted. This is commonly used to represent
3072 references from inline asms and other things the compiler cannot "see", and
3073 corresponds to "``attribute((used))``" in GNU C.
3075 On some targets, the code generator must emit a directive to the
3076 assembler or object file to prevent the assembler and linker from
3077 molesting the symbol.
3079 .. _gv_llvmcompilerused:
3081 The '``llvm.compiler.used``' Global Variable
3082 --------------------------------------------
3084 The ``@llvm.compiler.used`` directive is the same as the ``@llvm.used``
3085 directive, except that it only prevents the compiler from touching the
3086 symbol. On targets that support it, this allows an intelligent linker to
3087 optimize references to the symbol without being impeded as it would be
3090 This is a rare construct that should only be used in rare circumstances,
3091 and should not be exposed to source languages.
3093 .. _gv_llvmglobalctors:
3095 The '``llvm.global_ctors``' Global Variable
3096 -------------------------------------------
3098 .. code-block:: llvm
3100 %0 = type { i32, void ()* }
3101 @llvm.global_ctors = appending global [1 x %0] [%0 { i32 65535, void ()* @ctor }]
3103 The ``@llvm.global_ctors`` array contains a list of constructor
3104 functions and associated priorities. The functions referenced by this
3105 array will be called in ascending order of priority (i.e. lowest first)
3106 when the module is loaded. The order of functions with the same priority
3109 .. _llvmglobaldtors:
3111 The '``llvm.global_dtors``' Global Variable
3112 -------------------------------------------
3114 .. code-block:: llvm
3116 %0 = type { i32, void ()* }
3117 @llvm.global_dtors = appending global [1 x %0] [%0 { i32 65535, void ()* @dtor }]
3119 The ``@llvm.global_dtors`` array contains a list of destructor functions
3120 and associated priorities. The functions referenced by this array will
3121 be called in descending order of priority (i.e. highest first) when the
3122 module is loaded. The order of functions with the same priority is not
3125 Instruction Reference
3126 =====================
3128 The LLVM instruction set consists of several different classifications
3129 of instructions: :ref:`terminator instructions <terminators>`, :ref:`binary
3130 instructions <binaryops>`, :ref:`bitwise binary
3131 instructions <bitwiseops>`, :ref:`memory instructions <memoryops>`, and
3132 :ref:`other instructions <otherops>`.
3136 Terminator Instructions
3137 -----------------------
3139 As mentioned :ref:`previously <functionstructure>`, every basic block in a
3140 program ends with a "Terminator" instruction, which indicates which
3141 block should be executed after the current block is finished. These
3142 terminator instructions typically yield a '``void``' value: they produce
3143 control flow, not values (the one exception being the
3144 ':ref:`invoke <i_invoke>`' instruction).
3146 The terminator instructions are: ':ref:`ret <i_ret>`',
3147 ':ref:`br <i_br>`', ':ref:`switch <i_switch>`',
3148 ':ref:`indirectbr <i_indirectbr>`', ':ref:`invoke <i_invoke>`',
3149 ':ref:`resume <i_resume>`', and ':ref:`unreachable <i_unreachable>`'.
3153 '``ret``' Instruction
3154 ^^^^^^^^^^^^^^^^^^^^^
3161 ret <type> <value> ; Return a value from a non-void function
3162 ret void ; Return from void function
3167 The '``ret``' instruction is used to return control flow (and optionally
3168 a value) from a function back to the caller.
3170 There are two forms of the '``ret``' instruction: one that returns a
3171 value and then causes control flow, and one that just causes control
3177 The '``ret``' instruction optionally accepts a single argument, the
3178 return value. The type of the return value must be a ':ref:`first
3179 class <t_firstclass>`' type.
3181 A function is not :ref:`well formed <wellformed>` if it it has a non-void
3182 return type and contains a '``ret``' instruction with no return value or
3183 a return value with a type that does not match its type, or if it has a
3184 void return type and contains a '``ret``' instruction with a return
3190 When the '``ret``' instruction is executed, control flow returns back to
3191 the calling function's context. If the caller is a
3192 ":ref:`call <i_call>`" instruction, execution continues at the
3193 instruction after the call. If the caller was an
3194 ":ref:`invoke <i_invoke>`" instruction, execution continues at the
3195 beginning of the "normal" destination block. If the instruction returns
3196 a value, that value shall set the call or invoke instruction's return
3202 .. code-block:: llvm
3204 ret i32 5 ; Return an integer value of 5
3205 ret void ; Return from a void function
3206 ret { i32, i8 } { i32 4, i8 2 } ; Return a struct of values 4 and 2
3210 '``br``' Instruction
3211 ^^^^^^^^^^^^^^^^^^^^
3218 br i1 <cond>, label <iftrue>, label <iffalse>
3219 br label <dest> ; Unconditional branch
3224 The '``br``' instruction is used to cause control flow to transfer to a
3225 different basic block in the current function. There are two forms of
3226 this instruction, corresponding to a conditional branch and an
3227 unconditional branch.
3232 The conditional branch form of the '``br``' instruction takes a single
3233 '``i1``' value and two '``label``' values. The unconditional form of the
3234 '``br``' instruction takes a single '``label``' value as a target.
3239 Upon execution of a conditional '``br``' instruction, the '``i1``'
3240 argument is evaluated. If the value is ``true``, control flows to the
3241 '``iftrue``' ``label`` argument. If "cond" is ``false``, control flows
3242 to the '``iffalse``' ``label`` argument.
3247 .. code-block:: llvm
3250 %cond = icmp eq i32 %a, %b
3251 br i1 %cond, label %IfEqual, label %IfUnequal
3259 '``switch``' Instruction
3260 ^^^^^^^^^^^^^^^^^^^^^^^^
3267 switch <intty> <value>, label <defaultdest> [ <intty> <val>, label <dest> ... ]
3272 The '``switch``' instruction is used to transfer control flow to one of
3273 several different places. It is a generalization of the '``br``'
3274 instruction, allowing a branch to occur to one of many possible
3280 The '``switch``' instruction uses three parameters: an integer
3281 comparison value '``value``', a default '``label``' destination, and an
3282 array of pairs of comparison value constants and '``label``'s. The table
3283 is not allowed to contain duplicate constant entries.
3288 The ``switch`` instruction specifies a table of values and destinations.
3289 When the '``switch``' instruction is executed, this table is searched
3290 for the given value. If the value is found, control flow is transferred
3291 to the corresponding destination; otherwise, control flow is transferred
3292 to the default destination.
3297 Depending on properties of the target machine and the particular
3298 ``switch`` instruction, this instruction may be code generated in
3299 different ways. For example, it could be generated as a series of
3300 chained conditional branches or with a lookup table.
3305 .. code-block:: llvm
3307 ; Emulate a conditional br instruction
3308 %Val = zext i1 %value to i32
3309 switch i32 %Val, label %truedest [ i32 0, label %falsedest ]
3311 ; Emulate an unconditional br instruction
3312 switch i32 0, label %dest [ ]
3314 ; Implement a jump table:
3315 switch i32 %val, label %otherwise [ i32 0, label %onzero
3317 i32 2, label %ontwo ]
3321 '``indirectbr``' Instruction
3322 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3329 indirectbr <somety>* <address>, [ label <dest1>, label <dest2>, ... ]
3334 The '``indirectbr``' instruction implements an indirect branch to a
3335 label within the current function, whose address is specified by
3336 "``address``". Address must be derived from a
3337 :ref:`blockaddress <blockaddress>` constant.
3342 The '``address``' argument is the address of the label to jump to. The
3343 rest of the arguments indicate the full set of possible destinations
3344 that the address may point to. Blocks are allowed to occur multiple
3345 times in the destination list, though this isn't particularly useful.
3347 This destination list is required so that dataflow analysis has an
3348 accurate understanding of the CFG.
3353 Control transfers to the block specified in the address argument. All
3354 possible destination blocks must be listed in the label list, otherwise
3355 this instruction has undefined behavior. This implies that jumps to
3356 labels defined in other functions have undefined behavior as well.
3361 This is typically implemented with a jump through a register.
3366 .. code-block:: llvm
3368 indirectbr i8* %Addr, [ label %bb1, label %bb2, label %bb3 ]
3372 '``invoke``' Instruction
3373 ^^^^^^^^^^^^^^^^^^^^^^^^
3380 <result> = invoke [cconv] [ret attrs] <ptr to function ty> <function ptr val>(<function args>) [fn attrs]
3381 to label <normal label> unwind label <exception label>
3386 The '``invoke``' instruction causes control to transfer to a specified
3387 function, with the possibility of control flow transfer to either the
3388 '``normal``' label or the '``exception``' label. If the callee function
3389 returns with the "``ret``" instruction, control flow will return to the
3390 "normal" label. If the callee (or any indirect callees) returns via the
3391 ":ref:`resume <i_resume>`" instruction or other exception handling
3392 mechanism, control is interrupted and continued at the dynamically
3393 nearest "exception" label.
3395 The '``exception``' label is a `landing
3396 pad <ExceptionHandling.html#overview>`_ for the exception. As such,
3397 '``exception``' label is required to have the
3398 ":ref:`landingpad <i_landingpad>`" instruction, which contains the
3399 information about the behavior of the program after unwinding happens,
3400 as its first non-PHI instruction. The restrictions on the
3401 "``landingpad``" instruction's tightly couples it to the "``invoke``"
3402 instruction, so that the important information contained within the
3403 "``landingpad``" instruction can't be lost through normal code motion.
3408 This instruction requires several arguments:
3410 #. The optional "cconv" marker indicates which :ref:`calling
3411 convention <callingconv>` the call should use. If none is
3412 specified, the call defaults to using C calling conventions.
3413 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
3414 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
3416 #. '``ptr to function ty``': shall be the signature of the pointer to
3417 function value being invoked. In most cases, this is a direct
3418 function invocation, but indirect ``invoke``'s are just as possible,
3419 branching off an arbitrary pointer to function value.
3420 #. '``function ptr val``': An LLVM value containing a pointer to a
3421 function to be invoked.
3422 #. '``function args``': argument list whose types match the function
3423 signature argument types and parameter attributes. All arguments must
3424 be of :ref:`first class <t_firstclass>` type. If the function signature
3425 indicates the function accepts a variable number of arguments, the
3426 extra arguments can be specified.
3427 #. '``normal label``': the label reached when the called function
3428 executes a '``ret``' instruction.
3429 #. '``exception label``': the label reached when a callee returns via
3430 the :ref:`resume <i_resume>` instruction or other exception handling
3432 #. The optional :ref:`function attributes <fnattrs>` list. Only
3433 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
3434 attributes are valid here.
3439 This instruction is designed to operate as a standard '``call``'
3440 instruction in most regards. The primary difference is that it
3441 establishes an association with a label, which is used by the runtime
3442 library to unwind the stack.
3444 This instruction is used in languages with destructors to ensure that
3445 proper cleanup is performed in the case of either a ``longjmp`` or a
3446 thrown exception. Additionally, this is important for implementation of
3447 '``catch``' clauses in high-level languages that support them.
3449 For the purposes of the SSA form, the definition of the value returned
3450 by the '``invoke``' instruction is deemed to occur on the edge from the
3451 current block to the "normal" label. If the callee unwinds then no
3452 return value is available.
3457 .. code-block:: llvm
3459 %retval = invoke i32 @Test(i32 15) to label %Continue
3460 unwind label %TestCleanup ; {i32}:retval set
3461 %retval = invoke coldcc i32 %Testfnptr(i32 15) to label %Continue
3462 unwind label %TestCleanup ; {i32}:retval set
3466 '``resume``' Instruction
3467 ^^^^^^^^^^^^^^^^^^^^^^^^
3474 resume <type> <value>
3479 The '``resume``' instruction is a terminator instruction that has no
3485 The '``resume``' instruction requires one argument, which must have the
3486 same type as the result of any '``landingpad``' instruction in the same
3492 The '``resume``' instruction resumes propagation of an existing
3493 (in-flight) exception whose unwinding was interrupted with a
3494 :ref:`landingpad <i_landingpad>` instruction.
3499 .. code-block:: llvm
3501 resume { i8*, i32 } %exn
3505 '``unreachable``' Instruction
3506 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3518 The '``unreachable``' instruction has no defined semantics. This
3519 instruction is used to inform the optimizer that a particular portion of
3520 the code is not reachable. This can be used to indicate that the code
3521 after a no-return function cannot be reached, and other facts.
3526 The '``unreachable``' instruction has no defined semantics.
3533 Binary operators are used to do most of the computation in a program.
3534 They require two operands of the same type, execute an operation on
3535 them, and produce a single value. The operands might represent multiple
3536 data, as is the case with the :ref:`vector <t_vector>` data type. The
3537 result value has the same type as its operands.
3539 There are several different binary operators:
3543 '``add``' Instruction
3544 ^^^^^^^^^^^^^^^^^^^^^
3551 <result> = add <ty> <op1>, <op2> ; yields {ty}:result
3552 <result> = add nuw <ty> <op1>, <op2> ; yields {ty}:result
3553 <result> = add nsw <ty> <op1>, <op2> ; yields {ty}:result
3554 <result> = add nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3559 The '``add``' instruction returns the sum of its two operands.
3564 The two arguments to the '``add``' instruction must be
3565 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3566 arguments must have identical types.
3571 The value produced is the integer sum of the two operands.
3573 If the sum has unsigned overflow, the result returned is the
3574 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
3577 Because LLVM integers use a two's complement representation, this
3578 instruction is appropriate for both signed and unsigned integers.
3580 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3581 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3582 result value of the ``add`` is a :ref:`poison value <poisonvalues>` if
3583 unsigned and/or signed overflow, respectively, occurs.
3588 .. code-block:: llvm
3590 <result> = add i32 4, %var ; yields {i32}:result = 4 + %var
3594 '``fadd``' Instruction
3595 ^^^^^^^^^^^^^^^^^^^^^^
3602 <result> = fadd [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3607 The '``fadd``' instruction returns the sum of its two operands.
3612 The two arguments to the '``fadd``' instruction must be :ref:`floating
3613 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3614 Both arguments must have identical types.
3619 The value produced is the floating point sum of the two operands. This
3620 instruction can also take any number of :ref:`fast-math flags <fastmath>`,
3621 which are optimization hints to enable otherwise unsafe floating point
3627 .. code-block:: llvm
3629 <result> = fadd float 4.0, %var ; yields {float}:result = 4.0 + %var
3631 '``sub``' Instruction
3632 ^^^^^^^^^^^^^^^^^^^^^
3639 <result> = sub <ty> <op1>, <op2> ; yields {ty}:result
3640 <result> = sub nuw <ty> <op1>, <op2> ; yields {ty}:result
3641 <result> = sub nsw <ty> <op1>, <op2> ; yields {ty}:result
3642 <result> = sub nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3647 The '``sub``' instruction returns the difference of its two operands.
3649 Note that the '``sub``' instruction is used to represent the '``neg``'
3650 instruction present in most other intermediate representations.
3655 The two arguments to the '``sub``' instruction must be
3656 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3657 arguments must have identical types.
3662 The value produced is the integer difference of the two operands.
3664 If the difference has unsigned overflow, the result returned is the
3665 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
3668 Because LLVM integers use a two's complement representation, this
3669 instruction is appropriate for both signed and unsigned integers.
3671 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3672 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3673 result value of the ``sub`` is a :ref:`poison value <poisonvalues>` if
3674 unsigned and/or signed overflow, respectively, occurs.
3679 .. code-block:: llvm
3681 <result> = sub i32 4, %var ; yields {i32}:result = 4 - %var
3682 <result> = sub i32 0, %val ; yields {i32}:result = -%var
3686 '``fsub``' Instruction
3687 ^^^^^^^^^^^^^^^^^^^^^^
3694 <result> = fsub [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3699 The '``fsub``' instruction returns the difference of its two operands.
3701 Note that the '``fsub``' instruction is used to represent the '``fneg``'
3702 instruction present in most other intermediate representations.
3707 The two arguments to the '``fsub``' instruction must be :ref:`floating
3708 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3709 Both arguments must have identical types.
3714 The value produced is the floating point difference of the two operands.
3715 This instruction can also take any number of :ref:`fast-math
3716 flags <fastmath>`, which are optimization hints to enable otherwise
3717 unsafe floating point optimizations:
3722 .. code-block:: llvm
3724 <result> = fsub float 4.0, %var ; yields {float}:result = 4.0 - %var
3725 <result> = fsub float -0.0, %val ; yields {float}:result = -%var
3727 '``mul``' Instruction
3728 ^^^^^^^^^^^^^^^^^^^^^
3735 <result> = mul <ty> <op1>, <op2> ; yields {ty}:result
3736 <result> = mul nuw <ty> <op1>, <op2> ; yields {ty}:result
3737 <result> = mul nsw <ty> <op1>, <op2> ; yields {ty}:result
3738 <result> = mul nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3743 The '``mul``' instruction returns the product of its two operands.
3748 The two arguments to the '``mul``' instruction must be
3749 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3750 arguments must have identical types.
3755 The value produced is the integer product of the two operands.
3757 If the result of the multiplication has unsigned overflow, the result
3758 returned is the mathematical result modulo 2\ :sup:`n`\ , where n is the
3759 bit width of the result.
3761 Because LLVM integers use a two's complement representation, and the
3762 result is the same width as the operands, this instruction returns the
3763 correct result for both signed and unsigned integers. If a full product
3764 (e.g. ``i32`` * ``i32`` -> ``i64``) is needed, the operands should be
3765 sign-extended or zero-extended as appropriate to the width of the full
3768 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3769 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3770 result value of the ``mul`` is a :ref:`poison value <poisonvalues>` if
3771 unsigned and/or signed overflow, respectively, occurs.
3776 .. code-block:: llvm
3778 <result> = mul i32 4, %var ; yields {i32}:result = 4 * %var
3782 '``fmul``' Instruction
3783 ^^^^^^^^^^^^^^^^^^^^^^
3790 <result> = fmul [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3795 The '``fmul``' instruction returns the product of its two operands.
3800 The two arguments to the '``fmul``' instruction must be :ref:`floating
3801 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3802 Both arguments must have identical types.
3807 The value produced is the floating point product of the two operands.
3808 This instruction can also take any number of :ref:`fast-math
3809 flags <fastmath>`, which are optimization hints to enable otherwise
3810 unsafe floating point optimizations:
3815 .. code-block:: llvm
3817 <result> = fmul float 4.0, %var ; yields {float}:result = 4.0 * %var
3819 '``udiv``' Instruction
3820 ^^^^^^^^^^^^^^^^^^^^^^
3827 <result> = udiv <ty> <op1>, <op2> ; yields {ty}:result
3828 <result> = udiv exact <ty> <op1>, <op2> ; yields {ty}:result
3833 The '``udiv``' instruction returns the quotient of its two operands.
3838 The two arguments to the '``udiv``' instruction must be
3839 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3840 arguments must have identical types.
3845 The value produced is the unsigned integer quotient of the two operands.
3847 Note that unsigned integer division and signed integer division are
3848 distinct operations; for signed integer division, use '``sdiv``'.
3850 Division by zero leads to undefined behavior.
3852 If the ``exact`` keyword is present, the result value of the ``udiv`` is
3853 a :ref:`poison value <poisonvalues>` if %op1 is not a multiple of %op2 (as
3854 such, "((a udiv exact b) mul b) == a").
3859 .. code-block:: llvm
3861 <result> = udiv i32 4, %var ; yields {i32}:result = 4 / %var
3863 '``sdiv``' Instruction
3864 ^^^^^^^^^^^^^^^^^^^^^^
3871 <result> = sdiv <ty> <op1>, <op2> ; yields {ty}:result
3872 <result> = sdiv exact <ty> <op1>, <op2> ; yields {ty}:result
3877 The '``sdiv``' instruction returns the quotient of its two operands.
3882 The two arguments to the '``sdiv``' instruction must be
3883 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3884 arguments must have identical types.
3889 The value produced is the signed integer quotient of the two operands
3890 rounded towards zero.
3892 Note that signed integer division and unsigned integer division are
3893 distinct operations; for unsigned integer division, use '``udiv``'.
3895 Division by zero leads to undefined behavior. Overflow also leads to
3896 undefined behavior; this is a rare case, but can occur, for example, by
3897 doing a 32-bit division of -2147483648 by -1.
3899 If the ``exact`` keyword is present, the result value of the ``sdiv`` is
3900 a :ref:`poison value <poisonvalues>` if the result would be rounded.
3905 .. code-block:: llvm
3907 <result> = sdiv i32 4, %var ; yields {i32}:result = 4 / %var
3911 '``fdiv``' Instruction
3912 ^^^^^^^^^^^^^^^^^^^^^^
3919 <result> = fdiv [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3924 The '``fdiv``' instruction returns the quotient of its two operands.
3929 The two arguments to the '``fdiv``' instruction must be :ref:`floating
3930 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3931 Both arguments must have identical types.
3936 The value produced is the floating point quotient of the two operands.
3937 This instruction can also take any number of :ref:`fast-math
3938 flags <fastmath>`, which are optimization hints to enable otherwise
3939 unsafe floating point optimizations:
3944 .. code-block:: llvm
3946 <result> = fdiv float 4.0, %var ; yields {float}:result = 4.0 / %var
3948 '``urem``' Instruction
3949 ^^^^^^^^^^^^^^^^^^^^^^
3956 <result> = urem <ty> <op1>, <op2> ; yields {ty}:result
3961 The '``urem``' instruction returns the remainder from the unsigned
3962 division of its two arguments.
3967 The two arguments to the '``urem``' instruction must be
3968 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3969 arguments must have identical types.
3974 This instruction returns the unsigned integer *remainder* of a division.
3975 This instruction always performs an unsigned division to get the
3978 Note that unsigned integer remainder and signed integer remainder are
3979 distinct operations; for signed integer remainder, use '``srem``'.
3981 Taking the remainder of a division by zero leads to undefined behavior.
3986 .. code-block:: llvm
3988 <result> = urem i32 4, %var ; yields {i32}:result = 4 % %var
3990 '``srem``' Instruction
3991 ^^^^^^^^^^^^^^^^^^^^^^
3998 <result> = srem <ty> <op1>, <op2> ; yields {ty}:result
4003 The '``srem``' instruction returns the remainder from the signed
4004 division of its two operands. This instruction can also take
4005 :ref:`vector <t_vector>` versions of the values in which case the elements
4011 The two arguments to the '``srem``' instruction must be
4012 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4013 arguments must have identical types.
4018 This instruction returns the *remainder* of a division (where the result
4019 is either zero or has the same sign as the dividend, ``op1``), not the
4020 *modulo* operator (where the result is either zero or has the same sign
4021 as the divisor, ``op2``) of a value. For more information about the
4022 difference, see `The Math
4023 Forum <http://mathforum.org/dr.math/problems/anne.4.28.99.html>`_. For a
4024 table of how this is implemented in various languages, please see
4026 operation <http://en.wikipedia.org/wiki/Modulo_operation>`_.
4028 Note that signed integer remainder and unsigned integer remainder are
4029 distinct operations; for unsigned integer remainder, use '``urem``'.
4031 Taking the remainder of a division by zero leads to undefined behavior.
4032 Overflow also leads to undefined behavior; this is a rare case, but can
4033 occur, for example, by taking the remainder of a 32-bit division of
4034 -2147483648 by -1. (The remainder doesn't actually overflow, but this
4035 rule lets srem be implemented using instructions that return both the
4036 result of the division and the remainder.)
4041 .. code-block:: llvm
4043 <result> = srem i32 4, %var ; yields {i32}:result = 4 % %var
4047 '``frem``' Instruction
4048 ^^^^^^^^^^^^^^^^^^^^^^
4055 <result> = frem [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
4060 The '``frem``' instruction returns the remainder from the division of
4066 The two arguments to the '``frem``' instruction must be :ref:`floating
4067 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
4068 Both arguments must have identical types.
4073 This instruction returns the *remainder* of a division. The remainder
4074 has the same sign as the dividend. This instruction can also take any
4075 number of :ref:`fast-math flags <fastmath>`, which are optimization hints
4076 to enable otherwise unsafe floating point optimizations:
4081 .. code-block:: llvm
4083 <result> = frem float 4.0, %var ; yields {float}:result = 4.0 % %var
4087 Bitwise Binary Operations
4088 -------------------------
4090 Bitwise binary operators are used to do various forms of bit-twiddling
4091 in a program. They are generally very efficient instructions and can
4092 commonly be strength reduced from other instructions. They require two
4093 operands of the same type, execute an operation on them, and produce a
4094 single value. The resulting value is the same type as its operands.
4096 '``shl``' Instruction
4097 ^^^^^^^^^^^^^^^^^^^^^
4104 <result> = shl <ty> <op1>, <op2> ; yields {ty}:result
4105 <result> = shl nuw <ty> <op1>, <op2> ; yields {ty}:result
4106 <result> = shl nsw <ty> <op1>, <op2> ; yields {ty}:result
4107 <result> = shl nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
4112 The '``shl``' instruction returns the first operand shifted to the left
4113 a specified number of bits.
4118 Both arguments to the '``shl``' instruction must be the same
4119 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4120 '``op2``' is treated as an unsigned value.
4125 The value produced is ``op1`` \* 2\ :sup:`op2` mod 2\ :sup:`n`,
4126 where ``n`` is the width of the result. If ``op2`` is (statically or
4127 dynamically) negative or equal to or larger than the number of bits in
4128 ``op1``, the result is undefined. If the arguments are vectors, each
4129 vector element of ``op1`` is shifted by the corresponding shift amount
4132 If the ``nuw`` keyword is present, then the shift produces a :ref:`poison
4133 value <poisonvalues>` if it shifts out any non-zero bits. If the
4134 ``nsw`` keyword is present, then the shift produces a :ref:`poison
4135 value <poisonvalues>` if it shifts out any bits that disagree with the
4136 resultant sign bit. As such, NUW/NSW have the same semantics as they
4137 would if the shift were expressed as a mul instruction with the same
4138 nsw/nuw bits in (mul %op1, (shl 1, %op2)).
4143 .. code-block:: llvm
4145 <result> = shl i32 4, %var ; yields {i32}: 4 << %var
4146 <result> = shl i32 4, 2 ; yields {i32}: 16
4147 <result> = shl i32 1, 10 ; yields {i32}: 1024
4148 <result> = shl i32 1, 32 ; undefined
4149 <result> = shl <2 x i32> < i32 1, i32 1>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 2, i32 4>
4151 '``lshr``' Instruction
4152 ^^^^^^^^^^^^^^^^^^^^^^
4159 <result> = lshr <ty> <op1>, <op2> ; yields {ty}:result
4160 <result> = lshr exact <ty> <op1>, <op2> ; yields {ty}:result
4165 The '``lshr``' instruction (logical shift right) returns the first
4166 operand shifted to the right a specified number of bits with zero fill.
4171 Both arguments to the '``lshr``' instruction must be the same
4172 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4173 '``op2``' is treated as an unsigned value.
4178 This instruction always performs a logical shift right operation. The
4179 most significant bits of the result will be filled with zero bits after
4180 the shift. If ``op2`` is (statically or dynamically) equal to or larger
4181 than the number of bits in ``op1``, the result is undefined. If the
4182 arguments are vectors, each vector element of ``op1`` is shifted by the
4183 corresponding shift amount in ``op2``.
4185 If the ``exact`` keyword is present, the result value of the ``lshr`` is
4186 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
4192 .. code-block:: llvm
4194 <result> = lshr i32 4, 1 ; yields {i32}:result = 2
4195 <result> = lshr i32 4, 2 ; yields {i32}:result = 1
4196 <result> = lshr i8 4, 3 ; yields {i8}:result = 0
4197 <result> = lshr i8 -2, 1 ; yields {i8}:result = 0x7F
4198 <result> = lshr i32 1, 32 ; undefined
4199 <result> = lshr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 0x7FFFFFFF, i32 1>
4201 '``ashr``' Instruction
4202 ^^^^^^^^^^^^^^^^^^^^^^
4209 <result> = ashr <ty> <op1>, <op2> ; yields {ty}:result
4210 <result> = ashr exact <ty> <op1>, <op2> ; yields {ty}:result
4215 The '``ashr``' instruction (arithmetic shift right) returns the first
4216 operand shifted to the right a specified number of bits with sign
4222 Both arguments to the '``ashr``' instruction must be the same
4223 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4224 '``op2``' is treated as an unsigned value.
4229 This instruction always performs an arithmetic shift right operation,
4230 The most significant bits of the result will be filled with the sign bit
4231 of ``op1``. If ``op2`` is (statically or dynamically) equal to or larger
4232 than the number of bits in ``op1``, the result is undefined. If the
4233 arguments are vectors, each vector element of ``op1`` is shifted by the
4234 corresponding shift amount in ``op2``.
4236 If the ``exact`` keyword is present, the result value of the ``ashr`` is
4237 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
4243 .. code-block:: llvm
4245 <result> = ashr i32 4, 1 ; yields {i32}:result = 2
4246 <result> = ashr i32 4, 2 ; yields {i32}:result = 1
4247 <result> = ashr i8 4, 3 ; yields {i8}:result = 0
4248 <result> = ashr i8 -2, 1 ; yields {i8}:result = -1
4249 <result> = ashr i32 1, 32 ; undefined
4250 <result> = ashr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 3> ; yields: result=<2 x i32> < i32 -1, i32 0>
4252 '``and``' Instruction
4253 ^^^^^^^^^^^^^^^^^^^^^
4260 <result> = and <ty> <op1>, <op2> ; yields {ty}:result
4265 The '``and``' instruction returns the bitwise logical and of its two
4271 The two arguments to the '``and``' instruction must be
4272 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4273 arguments must have identical types.
4278 The truth table used for the '``and``' instruction is:
4295 .. code-block:: llvm
4297 <result> = and i32 4, %var ; yields {i32}:result = 4 & %var
4298 <result> = and i32 15, 40 ; yields {i32}:result = 8
4299 <result> = and i32 4, 8 ; yields {i32}:result = 0
4301 '``or``' Instruction
4302 ^^^^^^^^^^^^^^^^^^^^
4309 <result> = or <ty> <op1>, <op2> ; yields {ty}:result
4314 The '``or``' instruction returns the bitwise logical inclusive or of its
4320 The two arguments to the '``or``' instruction must be
4321 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4322 arguments must have identical types.
4327 The truth table used for the '``or``' instruction is:
4346 <result> = or i32 4, %var ; yields {i32}:result = 4 | %var
4347 <result> = or i32 15, 40 ; yields {i32}:result = 47
4348 <result> = or i32 4, 8 ; yields {i32}:result = 12
4350 '``xor``' Instruction
4351 ^^^^^^^^^^^^^^^^^^^^^
4358 <result> = xor <ty> <op1>, <op2> ; yields {ty}:result
4363 The '``xor``' instruction returns the bitwise logical exclusive or of
4364 its two operands. The ``xor`` is used to implement the "one's
4365 complement" operation, which is the "~" operator in C.
4370 The two arguments to the '``xor``' instruction must be
4371 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4372 arguments must have identical types.
4377 The truth table used for the '``xor``' instruction is:
4394 .. code-block:: llvm
4396 <result> = xor i32 4, %var ; yields {i32}:result = 4 ^ %var
4397 <result> = xor i32 15, 40 ; yields {i32}:result = 39
4398 <result> = xor i32 4, 8 ; yields {i32}:result = 12
4399 <result> = xor i32 %V, -1 ; yields {i32}:result = ~%V
4404 LLVM supports several instructions to represent vector operations in a
4405 target-independent manner. These instructions cover the element-access
4406 and vector-specific operations needed to process vectors effectively.
4407 While LLVM does directly support these vector operations, many
4408 sophisticated algorithms will want to use target-specific intrinsics to
4409 take full advantage of a specific target.
4411 .. _i_extractelement:
4413 '``extractelement``' Instruction
4414 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4421 <result> = extractelement <n x <ty>> <val>, i32 <idx> ; yields <ty>
4426 The '``extractelement``' instruction extracts a single scalar element
4427 from a vector at a specified index.
4432 The first operand of an '``extractelement``' instruction is a value of
4433 :ref:`vector <t_vector>` type. The second operand is an index indicating
4434 the position from which to extract the element. The index may be a
4440 The result is a scalar of the same type as the element type of ``val``.
4441 Its value is the value at position ``idx`` of ``val``. If ``idx``
4442 exceeds the length of ``val``, the results are undefined.
4447 .. code-block:: llvm
4449 <result> = extractelement <4 x i32> %vec, i32 0 ; yields i32
4451 .. _i_insertelement:
4453 '``insertelement``' Instruction
4454 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4461 <result> = insertelement <n x <ty>> <val>, <ty> <elt>, i32 <idx> ; yields <n x <ty>>
4466 The '``insertelement``' instruction inserts a scalar element into a
4467 vector at a specified index.
4472 The first operand of an '``insertelement``' instruction is a value of
4473 :ref:`vector <t_vector>` type. The second operand is a scalar value whose
4474 type must equal the element type of the first operand. The third operand
4475 is an index indicating the position at which to insert the value. The
4476 index may be a variable.
4481 The result is a vector of the same type as ``val``. Its element values
4482 are those of ``val`` except at position ``idx``, where it gets the value
4483 ``elt``. If ``idx`` exceeds the length of ``val``, the results are
4489 .. code-block:: llvm
4491 <result> = insertelement <4 x i32> %vec, i32 1, i32 0 ; yields <4 x i32>
4493 .. _i_shufflevector:
4495 '``shufflevector``' Instruction
4496 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4503 <result> = shufflevector <n x <ty>> <v1>, <n x <ty>> <v2>, <m x i32> <mask> ; yields <m x <ty>>
4508 The '``shufflevector``' instruction constructs a permutation of elements
4509 from two input vectors, returning a vector with the same element type as
4510 the input and length that is the same as the shuffle mask.
4515 The first two operands of a '``shufflevector``' instruction are vectors
4516 with the same type. The third argument is a shuffle mask whose element
4517 type is always 'i32'. The result of the instruction is a vector whose
4518 length is the same as the shuffle mask and whose element type is the
4519 same as the element type of the first two operands.
4521 The shuffle mask operand is required to be a constant vector with either
4522 constant integer or undef values.
4527 The elements of the two input vectors are numbered from left to right
4528 across both of the vectors. The shuffle mask operand specifies, for each
4529 element of the result vector, which element of the two input vectors the
4530 result element gets. The element selector may be undef (meaning "don't
4531 care") and the second operand may be undef if performing a shuffle from
4537 .. code-block:: llvm
4539 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
4540 <4 x i32> <i32 0, i32 4, i32 1, i32 5> ; yields <4 x i32>
4541 <result> = shufflevector <4 x i32> %v1, <4 x i32> undef,
4542 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32> - Identity shuffle.
4543 <result> = shufflevector <8 x i32> %v1, <8 x i32> undef,
4544 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32>
4545 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
4546 <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7 > ; yields <8 x i32>
4548 Aggregate Operations
4549 --------------------
4551 LLVM supports several instructions for working with
4552 :ref:`aggregate <t_aggregate>` values.
4556 '``extractvalue``' Instruction
4557 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4564 <result> = extractvalue <aggregate type> <val>, <idx>{, <idx>}*
4569 The '``extractvalue``' instruction extracts the value of a member field
4570 from an :ref:`aggregate <t_aggregate>` value.
4575 The first operand of an '``extractvalue``' instruction is a value of
4576 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The operands are
4577 constant indices to specify which value to extract in a similar manner
4578 as indices in a '``getelementptr``' instruction.
4580 The major differences to ``getelementptr`` indexing are:
4582 - Since the value being indexed is not a pointer, the first index is
4583 omitted and assumed to be zero.
4584 - At least one index must be specified.
4585 - Not only struct indices but also array indices must be in bounds.
4590 The result is the value at the position in the aggregate specified by
4596 .. code-block:: llvm
4598 <result> = extractvalue {i32, float} %agg, 0 ; yields i32
4602 '``insertvalue``' Instruction
4603 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4610 <result> = insertvalue <aggregate type> <val>, <ty> <elt>, <idx>{, <idx>}* ; yields <aggregate type>
4615 The '``insertvalue``' instruction inserts a value into a member field in
4616 an :ref:`aggregate <t_aggregate>` value.
4621 The first operand of an '``insertvalue``' instruction is a value of
4622 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The second operand is
4623 a first-class value to insert. The following operands are constant
4624 indices indicating the position at which to insert the value in a
4625 similar manner as indices in a '``extractvalue``' instruction. The value
4626 to insert must have the same type as the value identified by the
4632 The result is an aggregate of the same type as ``val``. Its value is
4633 that of ``val`` except that the value at the position specified by the
4634 indices is that of ``elt``.
4639 .. code-block:: llvm
4641 %agg1 = insertvalue {i32, float} undef, i32 1, 0 ; yields {i32 1, float undef}
4642 %agg2 = insertvalue {i32, float} %agg1, float %val, 1 ; yields {i32 1, float %val}
4643 %agg3 = insertvalue {i32, {float}} %agg1, float %val, 1, 0 ; yields {i32 1, float %val}
4647 Memory Access and Addressing Operations
4648 ---------------------------------------
4650 A key design point of an SSA-based representation is how it represents
4651 memory. In LLVM, no memory locations are in SSA form, which makes things
4652 very simple. This section describes how to read, write, and allocate
4657 '``alloca``' Instruction
4658 ^^^^^^^^^^^^^^^^^^^^^^^^
4665 <result> = alloca <type>[, <ty> <NumElements>][, align <alignment>] ; yields {type*}:result
4670 The '``alloca``' instruction allocates memory on the stack frame of the
4671 currently executing function, to be automatically released when this
4672 function returns to its caller. The object is always allocated in the
4673 generic address space (address space zero).
4678 The '``alloca``' instruction allocates ``sizeof(<type>)*NumElements``
4679 bytes of memory on the runtime stack, returning a pointer of the
4680 appropriate type to the program. If "NumElements" is specified, it is
4681 the number of elements allocated, otherwise "NumElements" is defaulted
4682 to be one. If a constant alignment is specified, the value result of the
4683 allocation is guaranteed to be aligned to at least that boundary. If not
4684 specified, or if zero, the target can choose to align the allocation on
4685 any convenient boundary compatible with the type.
4687 '``type``' may be any sized type.
4692 Memory is allocated; a pointer is returned. The operation is undefined
4693 if there is insufficient stack space for the allocation. '``alloca``'d
4694 memory is automatically released when the function returns. The
4695 '``alloca``' instruction is commonly used to represent automatic
4696 variables that must have an address available. When the function returns
4697 (either with the ``ret`` or ``resume`` instructions), the memory is
4698 reclaimed. Allocating zero bytes is legal, but the result is undefined.
4699 The order in which memory is allocated (ie., which way the stack grows)
4705 .. code-block:: llvm
4707 %ptr = alloca i32 ; yields {i32*}:ptr
4708 %ptr = alloca i32, i32 4 ; yields {i32*}:ptr
4709 %ptr = alloca i32, i32 4, align 1024 ; yields {i32*}:ptr
4710 %ptr = alloca i32, align 1024 ; yields {i32*}:ptr
4714 '``load``' Instruction
4715 ^^^^^^^^^^^^^^^^^^^^^^
4722 <result> = load [volatile] <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>][, !invariant.load !<index>]
4723 <result> = load atomic [volatile] <ty>* <pointer> [singlethread] <ordering>, align <alignment>
4724 !<index> = !{ i32 1 }
4729 The '``load``' instruction is used to read from memory.
4734 The argument to the ``load`` instruction specifies the memory address
4735 from which to load. The pointer must point to a :ref:`first
4736 class <t_firstclass>` type. If the ``load`` is marked as ``volatile``,
4737 then the optimizer is not allowed to modify the number or order of
4738 execution of this ``load`` with other :ref:`volatile
4739 operations <volatile>`.
4741 If the ``load`` is marked as ``atomic``, it takes an extra
4742 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
4743 ``release`` and ``acq_rel`` orderings are not valid on ``load``
4744 instructions. Atomic loads produce :ref:`defined <memmodel>` results
4745 when they may see multiple atomic stores. The type of the pointee must
4746 be an integer type whose bit width is a power of two greater than or
4747 equal to eight and less than or equal to a target-specific size limit.
4748 ``align`` must be explicitly specified on atomic loads, and the load has
4749 undefined behavior if the alignment is not set to a value which is at
4750 least the size in bytes of the pointee. ``!nontemporal`` does not have
4751 any defined semantics for atomic loads.
4753 The optional constant ``align`` argument specifies the alignment of the
4754 operation (that is, the alignment of the memory address). A value of 0
4755 or an omitted ``align`` argument means that the operation has the ABI
4756 alignment for the target. It is the responsibility of the code emitter
4757 to ensure that the alignment information is correct. Overestimating the
4758 alignment results in undefined behavior. Underestimating the alignment
4759 may produce less efficient code. An alignment of 1 is always safe.
4761 The optional ``!nontemporal`` metadata must reference a single
4762 metadata name ``<index>`` corresponding to a metadata node with one
4763 ``i32`` entry of value 1. The existence of the ``!nontemporal``
4764 metadata on the instruction tells the optimizer and code generator
4765 that this load is not expected to be reused in the cache. The code
4766 generator may select special instructions to save cache bandwidth, such
4767 as the ``MOVNT`` instruction on x86.
4769 The optional ``!invariant.load`` metadata must reference a single
4770 metadata name ``<index>`` corresponding to a metadata node with no
4771 entries. The existence of the ``!invariant.load`` metadata on the
4772 instruction tells the optimizer and code generator that this load
4773 address points to memory which does not change value during program
4774 execution. The optimizer may then move this load around, for example, by
4775 hoisting it out of loops using loop invariant code motion.
4780 The location of memory pointed to is loaded. If the value being loaded
4781 is of scalar type then the number of bytes read does not exceed the
4782 minimum number of bytes needed to hold all bits of the type. For
4783 example, loading an ``i24`` reads at most three bytes. When loading a
4784 value of a type like ``i20`` with a size that is not an integral number
4785 of bytes, the result is undefined if the value was not originally
4786 written using a store of the same type.
4791 .. code-block:: llvm
4793 %ptr = alloca i32 ; yields {i32*}:ptr
4794 store i32 3, i32* %ptr ; yields {void}
4795 %val = load i32* %ptr ; yields {i32}:val = i32 3
4799 '``store``' Instruction
4800 ^^^^^^^^^^^^^^^^^^^^^^^
4807 store [volatile] <ty> <value>, <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>] ; yields {void}
4808 store atomic [volatile] <ty> <value>, <ty>* <pointer> [singlethread] <ordering>, align <alignment> ; yields {void}
4813 The '``store``' instruction is used to write to memory.
4818 There are two arguments to the ``store`` instruction: a value to store
4819 and an address at which to store it. The type of the ``<pointer>``
4820 operand must be a pointer to the :ref:`first class <t_firstclass>` type of
4821 the ``<value>`` operand. If the ``store`` is marked as ``volatile``,
4822 then the optimizer is not allowed to modify the number or order of
4823 execution of this ``store`` with other :ref:`volatile
4824 operations <volatile>`.
4826 If the ``store`` is marked as ``atomic``, it takes an extra
4827 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
4828 ``acquire`` and ``acq_rel`` orderings aren't valid on ``store``
4829 instructions. Atomic loads produce :ref:`defined <memmodel>` results
4830 when they may see multiple atomic stores. The type of the pointee must
4831 be an integer type whose bit width is a power of two greater than or
4832 equal to eight and less than or equal to a target-specific size limit.
4833 ``align`` must be explicitly specified on atomic stores, and the store
4834 has undefined behavior if the alignment is not set to a value which is
4835 at least the size in bytes of the pointee. ``!nontemporal`` does not
4836 have any defined semantics for atomic stores.
4838 The optional constant ``align`` argument specifies the alignment of the
4839 operation (that is, the alignment of the memory address). A value of 0
4840 or an omitted ``align`` argument means that the operation has the ABI
4841 alignment for the target. It is the responsibility of the code emitter
4842 to ensure that the alignment information is correct. Overestimating the
4843 alignment results in undefined behavior. Underestimating the
4844 alignment may produce less efficient code. An alignment of 1 is always
4847 The optional ``!nontemporal`` metadata must reference a single metadata
4848 name ``<index>`` corresponding to a metadata node with one ``i32`` entry of
4849 value 1. The existence of the ``!nontemporal`` metadata on the instruction
4850 tells the optimizer and code generator that this load is not expected to
4851 be reused in the cache. The code generator may select special
4852 instructions to save cache bandwidth, such as the MOVNT instruction on
4858 The contents of memory are updated to contain ``<value>`` at the
4859 location specified by the ``<pointer>`` operand. If ``<value>`` is
4860 of scalar type then the number of bytes written does not exceed the
4861 minimum number of bytes needed to hold all bits of the type. For
4862 example, storing an ``i24`` writes at most three bytes. When writing a
4863 value of a type like ``i20`` with a size that is not an integral number
4864 of bytes, it is unspecified what happens to the extra bits that do not
4865 belong to the type, but they will typically be overwritten.
4870 .. code-block:: llvm
4872 %ptr = alloca i32 ; yields {i32*}:ptr
4873 store i32 3, i32* %ptr ; yields {void}
4874 %val = load i32* %ptr ; yields {i32}:val = i32 3
4878 '``fence``' Instruction
4879 ^^^^^^^^^^^^^^^^^^^^^^^
4886 fence [singlethread] <ordering> ; yields {void}
4891 The '``fence``' instruction is used to introduce happens-before edges
4897 '``fence``' instructions take an :ref:`ordering <ordering>` argument which
4898 defines what *synchronizes-with* edges they add. They can only be given
4899 ``acquire``, ``release``, ``acq_rel``, and ``seq_cst`` orderings.
4904 A fence A which has (at least) ``release`` ordering semantics
4905 *synchronizes with* a fence B with (at least) ``acquire`` ordering
4906 semantics if and only if there exist atomic operations X and Y, both
4907 operating on some atomic object M, such that A is sequenced before X, X
4908 modifies M (either directly or through some side effect of a sequence
4909 headed by X), Y is sequenced before B, and Y observes M. This provides a
4910 *happens-before* dependency between A and B. Rather than an explicit
4911 ``fence``, one (but not both) of the atomic operations X or Y might
4912 provide a ``release`` or ``acquire`` (resp.) ordering constraint and
4913 still *synchronize-with* the explicit ``fence`` and establish the
4914 *happens-before* edge.
4916 A ``fence`` which has ``seq_cst`` ordering, in addition to having both
4917 ``acquire`` and ``release`` semantics specified above, participates in
4918 the global program order of other ``seq_cst`` operations and/or fences.
4920 The optional ":ref:`singlethread <singlethread>`" argument specifies
4921 that the fence only synchronizes with other fences in the same thread.
4922 (This is useful for interacting with signal handlers.)
4927 .. code-block:: llvm
4929 fence acquire ; yields {void}
4930 fence singlethread seq_cst ; yields {void}
4934 '``cmpxchg``' Instruction
4935 ^^^^^^^^^^^^^^^^^^^^^^^^^
4942 cmpxchg [volatile] <ty>* <pointer>, <ty> <cmp>, <ty> <new> [singlethread] <ordering> ; yields {ty}
4947 The '``cmpxchg``' instruction is used to atomically modify memory. It
4948 loads a value in memory and compares it to a given value. If they are
4949 equal, it stores a new value into the memory.
4954 There are three arguments to the '``cmpxchg``' instruction: an address
4955 to operate on, a value to compare to the value currently be at that
4956 address, and a new value to place at that address if the compared values
4957 are equal. The type of '<cmp>' must be an integer type whose bit width
4958 is a power of two greater than or equal to eight and less than or equal
4959 to a target-specific size limit. '<cmp>' and '<new>' must have the same
4960 type, and the type of '<pointer>' must be a pointer to that type. If the
4961 ``cmpxchg`` is marked as ``volatile``, then the optimizer is not allowed
4962 to modify the number or order of execution of this ``cmpxchg`` with
4963 other :ref:`volatile operations <volatile>`.
4965 The :ref:`ordering <ordering>` argument specifies how this ``cmpxchg``
4966 synchronizes with other atomic operations.
4968 The optional "``singlethread``" argument declares that the ``cmpxchg``
4969 is only atomic with respect to code (usually signal handlers) running in
4970 the same thread as the ``cmpxchg``. Otherwise the cmpxchg is atomic with
4971 respect to all other code in the system.
4973 The pointer passed into cmpxchg must have alignment greater than or
4974 equal to the size in memory of the operand.
4979 The contents of memory at the location specified by the '``<pointer>``'
4980 operand is read and compared to '``<cmp>``'; if the read value is the
4981 equal, '``<new>``' is written. The original value at the location is
4984 A successful ``cmpxchg`` is a read-modify-write instruction for the purpose
4985 of identifying release sequences. A failed ``cmpxchg`` is equivalent to an
4986 atomic load with an ordering parameter determined by dropping any
4987 ``release`` part of the ``cmpxchg``'s ordering.
4992 .. code-block:: llvm
4995 %orig = atomic load i32* %ptr unordered ; yields {i32}
4999 %cmp = phi i32 [ %orig, %entry ], [%old, %loop]
5000 %squared = mul i32 %cmp, %cmp
5001 %old = cmpxchg i32* %ptr, i32 %cmp, i32 %squared ; yields {i32}
5002 %success = icmp eq i32 %cmp, %old
5003 br i1 %success, label %done, label %loop
5010 '``atomicrmw``' Instruction
5011 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
5018 atomicrmw [volatile] <operation> <ty>* <pointer>, <ty> <value> [singlethread] <ordering> ; yields {ty}
5023 The '``atomicrmw``' instruction is used to atomically modify memory.
5028 There are three arguments to the '``atomicrmw``' instruction: an
5029 operation to apply, an address whose value to modify, an argument to the
5030 operation. The operation must be one of the following keywords:
5044 The type of '<value>' must be an integer type whose bit width is a power
5045 of two greater than or equal to eight and less than or equal to a
5046 target-specific size limit. The type of the '``<pointer>``' operand must
5047 be a pointer to that type. If the ``atomicrmw`` is marked as
5048 ``volatile``, then the optimizer is not allowed to modify the number or
5049 order of execution of this ``atomicrmw`` with other :ref:`volatile
5050 operations <volatile>`.
5055 The contents of memory at the location specified by the '``<pointer>``'
5056 operand are atomically read, modified, and written back. The original
5057 value at the location is returned. The modification is specified by the
5060 - xchg: ``*ptr = val``
5061 - add: ``*ptr = *ptr + val``
5062 - sub: ``*ptr = *ptr - val``
5063 - and: ``*ptr = *ptr & val``
5064 - nand: ``*ptr = ~(*ptr & val)``
5065 - or: ``*ptr = *ptr | val``
5066 - xor: ``*ptr = *ptr ^ val``
5067 - max: ``*ptr = *ptr > val ? *ptr : val`` (using a signed comparison)
5068 - min: ``*ptr = *ptr < val ? *ptr : val`` (using a signed comparison)
5069 - umax: ``*ptr = *ptr > val ? *ptr : val`` (using an unsigned
5071 - umin: ``*ptr = *ptr < val ? *ptr : val`` (using an unsigned
5077 .. code-block:: llvm
5079 %old = atomicrmw add i32* %ptr, i32 1 acquire ; yields {i32}
5081 .. _i_getelementptr:
5083 '``getelementptr``' Instruction
5084 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5091 <result> = getelementptr <pty>* <ptrval>{, <ty> <idx>}*
5092 <result> = getelementptr inbounds <pty>* <ptrval>{, <ty> <idx>}*
5093 <result> = getelementptr <ptr vector> ptrval, <vector index type> idx
5098 The '``getelementptr``' instruction is used to get the address of a
5099 subelement of an :ref:`aggregate <t_aggregate>` data structure. It performs
5100 address calculation only and does not access memory.
5105 The first argument is always a pointer or a vector of pointers, and
5106 forms the basis of the calculation. The remaining arguments are indices
5107 that indicate which of the elements of the aggregate object are indexed.
5108 The interpretation of each index is dependent on the type being indexed
5109 into. The first index always indexes the pointer value given as the
5110 first argument, the second index indexes a value of the type pointed to
5111 (not necessarily the value directly pointed to, since the first index
5112 can be non-zero), etc. The first type indexed into must be a pointer
5113 value, subsequent types can be arrays, vectors, and structs. Note that
5114 subsequent types being indexed into can never be pointers, since that
5115 would require loading the pointer before continuing calculation.
5117 The type of each index argument depends on the type it is indexing into.
5118 When indexing into a (optionally packed) structure, only ``i32`` integer
5119 **constants** are allowed (when using a vector of indices they must all
5120 be the **same** ``i32`` integer constant). When indexing into an array,
5121 pointer or vector, integers of any width are allowed, and they are not
5122 required to be constant. These integers are treated as signed values
5125 For example, let's consider a C code fragment and how it gets compiled
5141 int *foo(struct ST *s) {
5142 return &s[1].Z.B[5][13];
5145 The LLVM code generated by Clang is:
5147 .. code-block:: llvm
5149 %struct.RT = type { i8, [10 x [20 x i32]], i8 }
5150 %struct.ST = type { i32, double, %struct.RT }
5152 define i32* @foo(%struct.ST* %s) nounwind uwtable readnone optsize ssp {
5154 %arrayidx = getelementptr inbounds %struct.ST* %s, i64 1, i32 2, i32 1, i64 5, i64 13
5161 In the example above, the first index is indexing into the
5162 '``%struct.ST*``' type, which is a pointer, yielding a '``%struct.ST``'
5163 = '``{ i32, double, %struct.RT }``' type, a structure. The second index
5164 indexes into the third element of the structure, yielding a
5165 '``%struct.RT``' = '``{ i8 , [10 x [20 x i32]], i8 }``' type, another
5166 structure. The third index indexes into the second element of the
5167 structure, yielding a '``[10 x [20 x i32]]``' type, an array. The two
5168 dimensions of the array are subscripted into, yielding an '``i32``'
5169 type. The '``getelementptr``' instruction returns a pointer to this
5170 element, thus computing a value of '``i32*``' type.
5172 Note that it is perfectly legal to index partially through a structure,
5173 returning a pointer to an inner element. Because of this, the LLVM code
5174 for the given testcase is equivalent to:
5176 .. code-block:: llvm
5178 define i32* @foo(%struct.ST* %s) {
5179 %t1 = getelementptr %struct.ST* %s, i32 1 ; yields %struct.ST*:%t1
5180 %t2 = getelementptr %struct.ST* %t1, i32 0, i32 2 ; yields %struct.RT*:%t2
5181 %t3 = getelementptr %struct.RT* %t2, i32 0, i32 1 ; yields [10 x [20 x i32]]*:%t3
5182 %t4 = getelementptr [10 x [20 x i32]]* %t3, i32 0, i32 5 ; yields [20 x i32]*:%t4
5183 %t5 = getelementptr [20 x i32]* %t4, i32 0, i32 13 ; yields i32*:%t5
5187 If the ``inbounds`` keyword is present, the result value of the
5188 ``getelementptr`` is a :ref:`poison value <poisonvalues>` if the base
5189 pointer is not an *in bounds* address of an allocated object, or if any
5190 of the addresses that would be formed by successive addition of the
5191 offsets implied by the indices to the base address with infinitely
5192 precise signed arithmetic are not an *in bounds* address of that
5193 allocated object. The *in bounds* addresses for an allocated object are
5194 all the addresses that point into the object, plus the address one byte
5195 past the end. In cases where the base is a vector of pointers the
5196 ``inbounds`` keyword applies to each of the computations element-wise.
5198 If the ``inbounds`` keyword is not present, the offsets are added to the
5199 base address with silently-wrapping two's complement arithmetic. If the
5200 offsets have a different width from the pointer, they are sign-extended
5201 or truncated to the width of the pointer. The result value of the
5202 ``getelementptr`` may be outside the object pointed to by the base
5203 pointer. The result value may not necessarily be used to access memory
5204 though, even if it happens to point into allocated storage. See the
5205 :ref:`Pointer Aliasing Rules <pointeraliasing>` section for more
5208 The getelementptr instruction is often confusing. For some more insight
5209 into how it works, see :doc:`the getelementptr FAQ <GetElementPtr>`.
5214 .. code-block:: llvm
5216 ; yields [12 x i8]*:aptr
5217 %aptr = getelementptr {i32, [12 x i8]}* %saptr, i64 0, i32 1
5219 %vptr = getelementptr {i32, <2 x i8>}* %svptr, i64 0, i32 1, i32 1
5221 %eptr = getelementptr [12 x i8]* %aptr, i64 0, i32 1
5223 %iptr = getelementptr [10 x i32]* @arr, i16 0, i16 0
5225 In cases where the pointer argument is a vector of pointers, each index
5226 must be a vector with the same number of elements. For example:
5228 .. code-block:: llvm
5230 %A = getelementptr <4 x i8*> %ptrs, <4 x i64> %offsets,
5232 Conversion Operations
5233 ---------------------
5235 The instructions in this category are the conversion instructions
5236 (casting) which all take a single operand and a type. They perform
5237 various bit conversions on the operand.
5239 '``trunc .. to``' Instruction
5240 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5247 <result> = trunc <ty> <value> to <ty2> ; yields ty2
5252 The '``trunc``' instruction truncates its operand to the type ``ty2``.
5257 The '``trunc``' instruction takes a value to trunc, and a type to trunc
5258 it to. Both types must be of :ref:`integer <t_integer>` types, or vectors
5259 of the same number of integers. The bit size of the ``value`` must be
5260 larger than the bit size of the destination type, ``ty2``. Equal sized
5261 types are not allowed.
5266 The '``trunc``' instruction truncates the high order bits in ``value``
5267 and converts the remaining bits to ``ty2``. Since the source size must
5268 be larger than the destination size, ``trunc`` cannot be a *no-op cast*.
5269 It will always truncate bits.
5274 .. code-block:: llvm
5276 %X = trunc i32 257 to i8 ; yields i8:1
5277 %Y = trunc i32 123 to i1 ; yields i1:true
5278 %Z = trunc i32 122 to i1 ; yields i1:false
5279 %W = trunc <2 x i16> <i16 8, i16 7> to <2 x i8> ; yields <i8 8, i8 7>
5281 '``zext .. to``' Instruction
5282 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5289 <result> = zext <ty> <value> to <ty2> ; yields ty2
5294 The '``zext``' instruction zero extends its operand to type ``ty2``.
5299 The '``zext``' instruction takes a value to cast, and a type to cast it
5300 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
5301 the same number of integers. The bit size of the ``value`` must be
5302 smaller than the bit size of the destination type, ``ty2``.
5307 The ``zext`` fills the high order bits of the ``value`` with zero bits
5308 until it reaches the size of the destination type, ``ty2``.
5310 When zero extending from i1, the result will always be either 0 or 1.
5315 .. code-block:: llvm
5317 %X = zext i32 257 to i64 ; yields i64:257
5318 %Y = zext i1 true to i32 ; yields i32:1
5319 %Z = zext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
5321 '``sext .. to``' Instruction
5322 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5329 <result> = sext <ty> <value> to <ty2> ; yields ty2
5334 The '``sext``' sign extends ``value`` to the type ``ty2``.
5339 The '``sext``' instruction takes a value to cast, and a type to cast it
5340 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
5341 the same number of integers. The bit size of the ``value`` must be
5342 smaller than the bit size of the destination type, ``ty2``.
5347 The '``sext``' instruction performs a sign extension by copying the sign
5348 bit (highest order bit) of the ``value`` until it reaches the bit size
5349 of the type ``ty2``.
5351 When sign extending from i1, the extension always results in -1 or 0.
5356 .. code-block:: llvm
5358 %X = sext i8 -1 to i16 ; yields i16 :65535
5359 %Y = sext i1 true to i32 ; yields i32:-1
5360 %Z = sext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
5362 '``fptrunc .. to``' Instruction
5363 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5370 <result> = fptrunc <ty> <value> to <ty2> ; yields ty2
5375 The '``fptrunc``' instruction truncates ``value`` to type ``ty2``.
5380 The '``fptrunc``' instruction takes a :ref:`floating point <t_floating>`
5381 value to cast and a :ref:`floating point <t_floating>` type to cast it to.
5382 The size of ``value`` must be larger than the size of ``ty2``. This
5383 implies that ``fptrunc`` cannot be used to make a *no-op cast*.
5388 The '``fptrunc``' instruction truncates a ``value`` from a larger
5389 :ref:`floating point <t_floating>` type to a smaller :ref:`floating
5390 point <t_floating>` type. If the value cannot fit within the
5391 destination type, ``ty2``, then the results are undefined.
5396 .. code-block:: llvm
5398 %X = fptrunc double 123.0 to float ; yields float:123.0
5399 %Y = fptrunc double 1.0E+300 to float ; yields undefined
5401 '``fpext .. to``' Instruction
5402 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5409 <result> = fpext <ty> <value> to <ty2> ; yields ty2
5414 The '``fpext``' extends a floating point ``value`` to a larger floating
5420 The '``fpext``' instruction takes a :ref:`floating point <t_floating>`
5421 ``value`` to cast, and a :ref:`floating point <t_floating>` type to cast it
5422 to. The source type must be smaller than the destination type.
5427 The '``fpext``' instruction extends the ``value`` from a smaller
5428 :ref:`floating point <t_floating>` type to a larger :ref:`floating
5429 point <t_floating>` type. The ``fpext`` cannot be used to make a
5430 *no-op cast* because it always changes bits. Use ``bitcast`` to make a
5431 *no-op cast* for a floating point cast.
5436 .. code-block:: llvm
5438 %X = fpext float 3.125 to double ; yields double:3.125000e+00
5439 %Y = fpext double %X to fp128 ; yields fp128:0xL00000000000000004000900000000000
5441 '``fptoui .. to``' Instruction
5442 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5449 <result> = fptoui <ty> <value> to <ty2> ; yields ty2
5454 The '``fptoui``' converts a floating point ``value`` to its unsigned
5455 integer equivalent of type ``ty2``.
5460 The '``fptoui``' instruction takes a value to cast, which must be a
5461 scalar or vector :ref:`floating point <t_floating>` value, and a type to
5462 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
5463 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
5464 type with the same number of elements as ``ty``
5469 The '``fptoui``' instruction converts its :ref:`floating
5470 point <t_floating>` operand into the nearest (rounding towards zero)
5471 unsigned integer value. If the value cannot fit in ``ty2``, the results
5477 .. code-block:: llvm
5479 %X = fptoui double 123.0 to i32 ; yields i32:123
5480 %Y = fptoui float 1.0E+300 to i1 ; yields undefined:1
5481 %Z = fptoui float 1.04E+17 to i8 ; yields undefined:1
5483 '``fptosi .. to``' Instruction
5484 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5491 <result> = fptosi <ty> <value> to <ty2> ; yields ty2
5496 The '``fptosi``' instruction converts :ref:`floating point <t_floating>`
5497 ``value`` to type ``ty2``.
5502 The '``fptosi``' instruction takes a value to cast, which must be a
5503 scalar or vector :ref:`floating point <t_floating>` value, and a type to
5504 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
5505 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
5506 type with the same number of elements as ``ty``
5511 The '``fptosi``' instruction converts its :ref:`floating
5512 point <t_floating>` operand into the nearest (rounding towards zero)
5513 signed integer value. If the value cannot fit in ``ty2``, the results
5519 .. code-block:: llvm
5521 %X = fptosi double -123.0 to i32 ; yields i32:-123
5522 %Y = fptosi float 1.0E-247 to i1 ; yields undefined:1
5523 %Z = fptosi float 1.04E+17 to i8 ; yields undefined:1
5525 '``uitofp .. to``' Instruction
5526 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5533 <result> = uitofp <ty> <value> to <ty2> ; yields ty2
5538 The '``uitofp``' instruction regards ``value`` as an unsigned integer
5539 and converts that value to the ``ty2`` type.
5544 The '``uitofp``' instruction takes a value to cast, which must be a
5545 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
5546 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
5547 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
5548 type with the same number of elements as ``ty``
5553 The '``uitofp``' instruction interprets its operand as an unsigned
5554 integer quantity and converts it to the corresponding floating point
5555 value. If the value cannot fit in the floating point value, the results
5561 .. code-block:: llvm
5563 %X = uitofp i32 257 to float ; yields float:257.0
5564 %Y = uitofp i8 -1 to double ; yields double:255.0
5566 '``sitofp .. to``' Instruction
5567 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5574 <result> = sitofp <ty> <value> to <ty2> ; yields ty2
5579 The '``sitofp``' instruction regards ``value`` as a signed integer and
5580 converts that value to the ``ty2`` type.
5585 The '``sitofp``' instruction takes a value to cast, which must be a
5586 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
5587 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
5588 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
5589 type with the same number of elements as ``ty``
5594 The '``sitofp``' instruction interprets its operand as a signed integer
5595 quantity and converts it to the corresponding floating point value. If
5596 the value cannot fit in the floating point value, the results are
5602 .. code-block:: llvm
5604 %X = sitofp i32 257 to float ; yields float:257.0
5605 %Y = sitofp i8 -1 to double ; yields double:-1.0
5609 '``ptrtoint .. to``' Instruction
5610 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5617 <result> = ptrtoint <ty> <value> to <ty2> ; yields ty2
5622 The '``ptrtoint``' instruction converts the pointer or a vector of
5623 pointers ``value`` to the integer (or vector of integers) type ``ty2``.
5628 The '``ptrtoint``' instruction takes a ``value`` to cast, which must be
5629 a a value of type :ref:`pointer <t_pointer>` or a vector of pointers, and a
5630 type to cast it to ``ty2``, which must be an :ref:`integer <t_integer>` or
5631 a vector of integers type.
5636 The '``ptrtoint``' instruction converts ``value`` to integer type
5637 ``ty2`` by interpreting the pointer value as an integer and either
5638 truncating or zero extending that value to the size of the integer type.
5639 If ``value`` is smaller than ``ty2`` then a zero extension is done. If
5640 ``value`` is larger than ``ty2`` then a truncation is done. If they are
5641 the same size, then nothing is done (*no-op cast*) other than a type
5647 .. code-block:: llvm
5649 %X = ptrtoint i32* %P to i8 ; yields truncation on 32-bit architecture
5650 %Y = ptrtoint i32* %P to i64 ; yields zero extension on 32-bit architecture
5651 %Z = ptrtoint <4 x i32*> %P to <4 x i64>; yields vector zero extension for a vector of addresses on 32-bit architecture
5655 '``inttoptr .. to``' Instruction
5656 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5663 <result> = inttoptr <ty> <value> to <ty2> ; yields ty2
5668 The '``inttoptr``' instruction converts an integer ``value`` to a
5669 pointer type, ``ty2``.
5674 The '``inttoptr``' instruction takes an :ref:`integer <t_integer>` value to
5675 cast, and a type to cast it to, which must be a :ref:`pointer <t_pointer>`
5681 The '``inttoptr``' instruction converts ``value`` to type ``ty2`` by
5682 applying either a zero extension or a truncation depending on the size
5683 of the integer ``value``. If ``value`` is larger than the size of a
5684 pointer then a truncation is done. If ``value`` is smaller than the size
5685 of a pointer then a zero extension is done. If they are the same size,
5686 nothing is done (*no-op cast*).
5691 .. code-block:: llvm
5693 %X = inttoptr i32 255 to i32* ; yields zero extension on 64-bit architecture
5694 %Y = inttoptr i32 255 to i32* ; yields no-op on 32-bit architecture
5695 %Z = inttoptr i64 0 to i32* ; yields truncation on 32-bit architecture
5696 %Z = inttoptr <4 x i32> %G to <4 x i8*>; yields truncation of vector G to four pointers
5700 '``bitcast .. to``' Instruction
5701 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5708 <result> = bitcast <ty> <value> to <ty2> ; yields ty2
5713 The '``bitcast``' instruction converts ``value`` to type ``ty2`` without
5719 The '``bitcast``' instruction takes a value to cast, which must be a
5720 non-aggregate first class value, and a type to cast it to, which must
5721 also be a non-aggregate :ref:`first class <t_firstclass>` type. The
5722 bit sizes of ``value`` and the destination type, ``ty2``, must be
5723 identical. If the source type is a pointer, the destination type must
5724 also be a pointer of the same size. This instruction supports bitwise
5725 conversion of vectors to integers and to vectors of other types (as
5726 long as they have the same size).
5731 The '``bitcast``' instruction converts ``value`` to type ``ty2``. It
5732 is always a *no-op cast* because no bits change with this
5733 conversion. The conversion is done as if the ``value`` had been stored
5734 to memory and read back as type ``ty2``. Pointer (or vector of
5735 pointers) types may only be converted to other pointer (or vector of
5736 pointers) types with this instruction if the pointer sizes are
5737 equal. To convert pointers to other types, use the :ref:`inttoptr
5738 <i_inttoptr>` or :ref:`ptrtoint <i_ptrtoint>` instructions first.
5743 .. code-block:: llvm
5745 %X = bitcast i8 255 to i8 ; yields i8 :-1
5746 %Y = bitcast i32* %x to sint* ; yields sint*:%x
5747 %Z = bitcast <2 x int> %V to i64; ; yields i64: %V
5748 %Z = bitcast <2 x i32*> %V to <2 x i64*> ; yields <2 x i64*>
5755 The instructions in this category are the "miscellaneous" instructions,
5756 which defy better classification.
5760 '``icmp``' Instruction
5761 ^^^^^^^^^^^^^^^^^^^^^^
5768 <result> = icmp <cond> <ty> <op1>, <op2> ; yields {i1} or {<N x i1>}:result
5773 The '``icmp``' instruction returns a boolean value or a vector of
5774 boolean values based on comparison of its two integer, integer vector,
5775 pointer, or pointer vector operands.
5780 The '``icmp``' instruction takes three operands. The first operand is
5781 the condition code indicating the kind of comparison to perform. It is
5782 not a value, just a keyword. The possible condition code are:
5785 #. ``ne``: not equal
5786 #. ``ugt``: unsigned greater than
5787 #. ``uge``: unsigned greater or equal
5788 #. ``ult``: unsigned less than
5789 #. ``ule``: unsigned less or equal
5790 #. ``sgt``: signed greater than
5791 #. ``sge``: signed greater or equal
5792 #. ``slt``: signed less than
5793 #. ``sle``: signed less or equal
5795 The remaining two arguments must be :ref:`integer <t_integer>` or
5796 :ref:`pointer <t_pointer>` or integer :ref:`vector <t_vector>` typed. They
5797 must also be identical types.
5802 The '``icmp``' compares ``op1`` and ``op2`` according to the condition
5803 code given as ``cond``. The comparison performed always yields either an
5804 :ref:`i1 <t_integer>` or vector of ``i1`` result, as follows:
5806 #. ``eq``: yields ``true`` if the operands are equal, ``false``
5807 otherwise. No sign interpretation is necessary or performed.
5808 #. ``ne``: yields ``true`` if the operands are unequal, ``false``
5809 otherwise. No sign interpretation is necessary or performed.
5810 #. ``ugt``: interprets the operands as unsigned values and yields
5811 ``true`` if ``op1`` is greater than ``op2``.
5812 #. ``uge``: interprets the operands as unsigned values and yields
5813 ``true`` if ``op1`` is greater than or equal to ``op2``.
5814 #. ``ult``: interprets the operands as unsigned values and yields
5815 ``true`` if ``op1`` is less than ``op2``.
5816 #. ``ule``: interprets the operands as unsigned values and yields
5817 ``true`` if ``op1`` is less than or equal to ``op2``.
5818 #. ``sgt``: interprets the operands as signed values and yields ``true``
5819 if ``op1`` is greater than ``op2``.
5820 #. ``sge``: interprets the operands as signed values and yields ``true``
5821 if ``op1`` is greater than or equal to ``op2``.
5822 #. ``slt``: interprets the operands as signed values and yields ``true``
5823 if ``op1`` is less than ``op2``.
5824 #. ``sle``: interprets the operands as signed values and yields ``true``
5825 if ``op1`` is less than or equal to ``op2``.
5827 If the operands are :ref:`pointer <t_pointer>` typed, the pointer values
5828 are compared as if they were integers.
5830 If the operands are integer vectors, then they are compared element by
5831 element. The result is an ``i1`` vector with the same number of elements
5832 as the values being compared. Otherwise, the result is an ``i1``.
5837 .. code-block:: llvm
5839 <result> = icmp eq i32 4, 5 ; yields: result=false
5840 <result> = icmp ne float* %X, %X ; yields: result=false
5841 <result> = icmp ult i16 4, 5 ; yields: result=true
5842 <result> = icmp sgt i16 4, 5 ; yields: result=false
5843 <result> = icmp ule i16 -4, 5 ; yields: result=false
5844 <result> = icmp sge i16 4, 5 ; yields: result=false
5846 Note that the code generator does not yet support vector types with the
5847 ``icmp`` instruction.
5851 '``fcmp``' Instruction
5852 ^^^^^^^^^^^^^^^^^^^^^^
5859 <result> = fcmp <cond> <ty> <op1>, <op2> ; yields {i1} or {<N x i1>}:result
5864 The '``fcmp``' instruction returns a boolean value or vector of boolean
5865 values based on comparison of its operands.
5867 If the operands are floating point scalars, then the result type is a
5868 boolean (:ref:`i1 <t_integer>`).
5870 If the operands are floating point vectors, then the result type is a
5871 vector of boolean with the same number of elements as the operands being
5877 The '``fcmp``' instruction takes three operands. The first operand is
5878 the condition code indicating the kind of comparison to perform. It is
5879 not a value, just a keyword. The possible condition code are:
5881 #. ``false``: no comparison, always returns false
5882 #. ``oeq``: ordered and equal
5883 #. ``ogt``: ordered and greater than
5884 #. ``oge``: ordered and greater than or equal
5885 #. ``olt``: ordered and less than
5886 #. ``ole``: ordered and less than or equal
5887 #. ``one``: ordered and not equal
5888 #. ``ord``: ordered (no nans)
5889 #. ``ueq``: unordered or equal
5890 #. ``ugt``: unordered or greater than
5891 #. ``uge``: unordered or greater than or equal
5892 #. ``ult``: unordered or less than
5893 #. ``ule``: unordered or less than or equal
5894 #. ``une``: unordered or not equal
5895 #. ``uno``: unordered (either nans)
5896 #. ``true``: no comparison, always returns true
5898 *Ordered* means that neither operand is a QNAN while *unordered* means
5899 that either operand may be a QNAN.
5901 Each of ``val1`` and ``val2`` arguments must be either a :ref:`floating
5902 point <t_floating>` type or a :ref:`vector <t_vector>` of floating point
5903 type. They must have identical types.
5908 The '``fcmp``' instruction compares ``op1`` and ``op2`` according to the
5909 condition code given as ``cond``. If the operands are vectors, then the
5910 vectors are compared element by element. Each comparison performed
5911 always yields an :ref:`i1 <t_integer>` result, as follows:
5913 #. ``false``: always yields ``false``, regardless of operands.
5914 #. ``oeq``: yields ``true`` if both operands are not a QNAN and ``op1``
5915 is equal to ``op2``.
5916 #. ``ogt``: yields ``true`` if both operands are not a QNAN and ``op1``
5917 is greater than ``op2``.
5918 #. ``oge``: yields ``true`` if both operands are not a QNAN and ``op1``
5919 is greater than or equal to ``op2``.
5920 #. ``olt``: yields ``true`` if both operands are not a QNAN and ``op1``
5921 is less than ``op2``.
5922 #. ``ole``: yields ``true`` if both operands are not a QNAN and ``op1``
5923 is less than or equal to ``op2``.
5924 #. ``one``: yields ``true`` if both operands are not a QNAN and ``op1``
5925 is not equal to ``op2``.
5926 #. ``ord``: yields ``true`` if both operands are not a QNAN.
5927 #. ``ueq``: yields ``true`` if either operand is a QNAN or ``op1`` is
5929 #. ``ugt``: yields ``true`` if either operand is a QNAN or ``op1`` is
5930 greater than ``op2``.
5931 #. ``uge``: yields ``true`` if either operand is a QNAN or ``op1`` is
5932 greater than or equal to ``op2``.
5933 #. ``ult``: yields ``true`` if either operand is a QNAN or ``op1`` is
5935 #. ``ule``: yields ``true`` if either operand is a QNAN or ``op1`` is
5936 less than or equal to ``op2``.
5937 #. ``une``: yields ``true`` if either operand is a QNAN or ``op1`` is
5938 not equal to ``op2``.
5939 #. ``uno``: yields ``true`` if either operand is a QNAN.
5940 #. ``true``: always yields ``true``, regardless of operands.
5945 .. code-block:: llvm
5947 <result> = fcmp oeq float 4.0, 5.0 ; yields: result=false
5948 <result> = fcmp one float 4.0, 5.0 ; yields: result=true
5949 <result> = fcmp olt float 4.0, 5.0 ; yields: result=true
5950 <result> = fcmp ueq double 1.0, 2.0 ; yields: result=false
5952 Note that the code generator does not yet support vector types with the
5953 ``fcmp`` instruction.
5957 '``phi``' Instruction
5958 ^^^^^^^^^^^^^^^^^^^^^
5965 <result> = phi <ty> [ <val0>, <label0>], ...
5970 The '``phi``' instruction is used to implement the φ node in the SSA
5971 graph representing the function.
5976 The type of the incoming values is specified with the first type field.
5977 After this, the '``phi``' instruction takes a list of pairs as
5978 arguments, with one pair for each predecessor basic block of the current
5979 block. Only values of :ref:`first class <t_firstclass>` type may be used as
5980 the value arguments to the PHI node. Only labels may be used as the
5983 There must be no non-phi instructions between the start of a basic block
5984 and the PHI instructions: i.e. PHI instructions must be first in a basic
5987 For the purposes of the SSA form, the use of each incoming value is
5988 deemed to occur on the edge from the corresponding predecessor block to
5989 the current block (but after any definition of an '``invoke``'
5990 instruction's return value on the same edge).
5995 At runtime, the '``phi``' instruction logically takes on the value
5996 specified by the pair corresponding to the predecessor basic block that
5997 executed just prior to the current block.
6002 .. code-block:: llvm
6004 Loop: ; Infinite loop that counts from 0 on up...
6005 %indvar = phi i32 [ 0, %LoopHeader ], [ %nextindvar, %Loop ]
6006 %nextindvar = add i32 %indvar, 1
6011 '``select``' Instruction
6012 ^^^^^^^^^^^^^^^^^^^^^^^^
6019 <result> = select selty <cond>, <ty> <val1>, <ty> <val2> ; yields ty
6021 selty is either i1 or {<N x i1>}
6026 The '``select``' instruction is used to choose one value based on a
6027 condition, without branching.
6032 The '``select``' instruction requires an 'i1' value or a vector of 'i1'
6033 values indicating the condition, and two values of the same :ref:`first
6034 class <t_firstclass>` type. If the val1/val2 are vectors and the
6035 condition is a scalar, then entire vectors are selected, not individual
6041 If the condition is an i1 and it evaluates to 1, the instruction returns
6042 the first value argument; otherwise, it returns the second value
6045 If the condition is a vector of i1, then the value arguments must be
6046 vectors of the same size, and the selection is done element by element.
6051 .. code-block:: llvm
6053 %X = select i1 true, i8 17, i8 42 ; yields i8:17
6057 '``call``' Instruction
6058 ^^^^^^^^^^^^^^^^^^^^^^
6065 <result> = [tail] call [cconv] [ret attrs] <ty> [<fnty>*] <fnptrval>(<function args>) [fn attrs]
6070 The '``call``' instruction represents a simple function call.
6075 This instruction requires several arguments:
6077 #. The optional "tail" marker indicates that the callee function does
6078 not access any allocas or varargs in the caller. Note that calls may
6079 be marked "tail" even if they do not occur before a
6080 :ref:`ret <i_ret>` instruction. If the "tail" marker is present, the
6081 function call is eligible for tail call optimization, but `might not
6082 in fact be optimized into a jump <CodeGenerator.html#tailcallopt>`_.
6083 The code generator may optimize calls marked "tail" with either 1)
6084 automatic `sibling call
6085 optimization <CodeGenerator.html#sibcallopt>`_ when the caller and
6086 callee have matching signatures, or 2) forced tail call optimization
6087 when the following extra requirements are met:
6089 - Caller and callee both have the calling convention ``fastcc``.
6090 - The call is in tail position (ret immediately follows call and ret
6091 uses value of call or is void).
6092 - Option ``-tailcallopt`` is enabled, or
6093 ``llvm::GuaranteedTailCallOpt`` is ``true``.
6094 - `Platform specific constraints are
6095 met. <CodeGenerator.html#tailcallopt>`_
6097 #. The optional "cconv" marker indicates which :ref:`calling
6098 convention <callingconv>` the call should use. If none is
6099 specified, the call defaults to using C calling conventions. The
6100 calling convention of the call must match the calling convention of
6101 the target function, or else the behavior is undefined.
6102 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
6103 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
6105 #. '``ty``': the type of the call instruction itself which is also the
6106 type of the return value. Functions that return no value are marked
6108 #. '``fnty``': shall be the signature of the pointer to function value
6109 being invoked. The argument types must match the types implied by
6110 this signature. This type can be omitted if the function is not
6111 varargs and if the function type does not return a pointer to a
6113 #. '``fnptrval``': An LLVM value containing a pointer to a function to
6114 be invoked. In most cases, this is a direct function invocation, but
6115 indirect ``call``'s are just as possible, calling an arbitrary pointer
6117 #. '``function args``': argument list whose types match the function
6118 signature argument types and parameter attributes. All arguments must
6119 be of :ref:`first class <t_firstclass>` type. If the function signature
6120 indicates the function accepts a variable number of arguments, the
6121 extra arguments can be specified.
6122 #. The optional :ref:`function attributes <fnattrs>` list. Only
6123 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
6124 attributes are valid here.
6129 The '``call``' instruction is used to cause control flow to transfer to
6130 a specified function, with its incoming arguments bound to the specified
6131 values. Upon a '``ret``' instruction in the called function, control
6132 flow continues with the instruction after the function call, and the
6133 return value of the function is bound to the result argument.
6138 .. code-block:: llvm
6140 %retval = call i32 @test(i32 %argc)
6141 call i32 (i8*, ...)* @printf(i8* %msg, i32 12, i8 42) ; yields i32
6142 %X = tail call i32 @foo() ; yields i32
6143 %Y = tail call fastcc i32 @foo() ; yields i32
6144 call void %foo(i8 97 signext)
6146 %struct.A = type { i32, i8 }
6147 %r = call %struct.A @foo() ; yields { 32, i8 }
6148 %gr = extractvalue %struct.A %r, 0 ; yields i32
6149 %gr1 = extractvalue %struct.A %r, 1 ; yields i8
6150 %Z = call void @foo() noreturn ; indicates that %foo never returns normally
6151 %ZZ = call zeroext i32 @bar() ; Return value is %zero extended
6153 llvm treats calls to some functions with names and arguments that match
6154 the standard C99 library as being the C99 library functions, and may
6155 perform optimizations or generate code for them under that assumption.
6156 This is something we'd like to change in the future to provide better
6157 support for freestanding environments and non-C-based languages.
6161 '``va_arg``' Instruction
6162 ^^^^^^^^^^^^^^^^^^^^^^^^
6169 <resultval> = va_arg <va_list*> <arglist>, <argty>
6174 The '``va_arg``' instruction is used to access arguments passed through
6175 the "variable argument" area of a function call. It is used to implement
6176 the ``va_arg`` macro in C.
6181 This instruction takes a ``va_list*`` value and the type of the
6182 argument. It returns a value of the specified argument type and
6183 increments the ``va_list`` to point to the next argument. The actual
6184 type of ``va_list`` is target specific.
6189 The '``va_arg``' instruction loads an argument of the specified type
6190 from the specified ``va_list`` and causes the ``va_list`` to point to
6191 the next argument. For more information, see the variable argument
6192 handling :ref:`Intrinsic Functions <int_varargs>`.
6194 It is legal for this instruction to be called in a function which does
6195 not take a variable number of arguments, for example, the ``vfprintf``
6198 ``va_arg`` is an LLVM instruction instead of an :ref:`intrinsic
6199 function <intrinsics>` because it takes a type as an argument.
6204 See the :ref:`variable argument processing <int_varargs>` section.
6206 Note that the code generator does not yet fully support va\_arg on many
6207 targets. Also, it does not currently support va\_arg with aggregate
6208 types on any target.
6212 '``landingpad``' Instruction
6213 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6220 <resultval> = landingpad <resultty> personality <type> <pers_fn> <clause>+
6221 <resultval> = landingpad <resultty> personality <type> <pers_fn> cleanup <clause>*
6223 <clause> := catch <type> <value>
6224 <clause> := filter <array constant type> <array constant>
6229 The '``landingpad``' instruction is used by `LLVM's exception handling
6230 system <ExceptionHandling.html#overview>`_ to specify that a basic block
6231 is a landing pad --- one where the exception lands, and corresponds to the
6232 code found in the ``catch`` portion of a ``try``/``catch`` sequence. It
6233 defines values supplied by the personality function (``pers_fn``) upon
6234 re-entry to the function. The ``resultval`` has the type ``resultty``.
6239 This instruction takes a ``pers_fn`` value. This is the personality
6240 function associated with the unwinding mechanism. The optional
6241 ``cleanup`` flag indicates that the landing pad block is a cleanup.
6243 A ``clause`` begins with the clause type --- ``catch`` or ``filter`` --- and
6244 contains the global variable representing the "type" that may be caught
6245 or filtered respectively. Unlike the ``catch`` clause, the ``filter``
6246 clause takes an array constant as its argument. Use
6247 "``[0 x i8**] undef``" for a filter which cannot throw. The
6248 '``landingpad``' instruction must contain *at least* one ``clause`` or
6249 the ``cleanup`` flag.
6254 The '``landingpad``' instruction defines the values which are set by the
6255 personality function (``pers_fn``) upon re-entry to the function, and
6256 therefore the "result type" of the ``landingpad`` instruction. As with
6257 calling conventions, how the personality function results are
6258 represented in LLVM IR is target specific.
6260 The clauses are applied in order from top to bottom. If two
6261 ``landingpad`` instructions are merged together through inlining, the
6262 clauses from the calling function are appended to the list of clauses.
6263 When the call stack is being unwound due to an exception being thrown,
6264 the exception is compared against each ``clause`` in turn. If it doesn't
6265 match any of the clauses, and the ``cleanup`` flag is not set, then
6266 unwinding continues further up the call stack.
6268 The ``landingpad`` instruction has several restrictions:
6270 - A landing pad block is a basic block which is the unwind destination
6271 of an '``invoke``' instruction.
6272 - A landing pad block must have a '``landingpad``' instruction as its
6273 first non-PHI instruction.
6274 - There can be only one '``landingpad``' instruction within the landing
6276 - A basic block that is not a landing pad block may not include a
6277 '``landingpad``' instruction.
6278 - All '``landingpad``' instructions in a function must have the same
6279 personality function.
6284 .. code-block:: llvm
6286 ;; A landing pad which can catch an integer.
6287 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6289 ;; A landing pad that is a cleanup.
6290 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6292 ;; A landing pad which can catch an integer and can only throw a double.
6293 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6295 filter [1 x i8**] [@_ZTId]
6302 LLVM supports the notion of an "intrinsic function". These functions
6303 have well known names and semantics and are required to follow certain
6304 restrictions. Overall, these intrinsics represent an extension mechanism
6305 for the LLVM language that does not require changing all of the
6306 transformations in LLVM when adding to the language (or the bitcode
6307 reader/writer, the parser, etc...).
6309 Intrinsic function names must all start with an "``llvm.``" prefix. This
6310 prefix is reserved in LLVM for intrinsic names; thus, function names may
6311 not begin with this prefix. Intrinsic functions must always be external
6312 functions: you cannot define the body of intrinsic functions. Intrinsic
6313 functions may only be used in call or invoke instructions: it is illegal
6314 to take the address of an intrinsic function. Additionally, because
6315 intrinsic functions are part of the LLVM language, it is required if any
6316 are added that they be documented here.
6318 Some intrinsic functions can be overloaded, i.e., the intrinsic
6319 represents a family of functions that perform the same operation but on
6320 different data types. Because LLVM can represent over 8 million
6321 different integer types, overloading is used commonly to allow an
6322 intrinsic function to operate on any integer type. One or more of the
6323 argument types or the result type can be overloaded to accept any
6324 integer type. Argument types may also be defined as exactly matching a
6325 previous argument's type or the result type. This allows an intrinsic
6326 function which accepts multiple arguments, but needs all of them to be
6327 of the same type, to only be overloaded with respect to a single
6328 argument or the result.
6330 Overloaded intrinsics will have the names of its overloaded argument
6331 types encoded into its function name, each preceded by a period. Only
6332 those types which are overloaded result in a name suffix. Arguments
6333 whose type is matched against another type do not. For example, the
6334 ``llvm.ctpop`` function can take an integer of any width and returns an
6335 integer of exactly the same integer width. This leads to a family of
6336 functions such as ``i8 @llvm.ctpop.i8(i8 %val)`` and
6337 ``i29 @llvm.ctpop.i29(i29 %val)``. Only one type, the return type, is
6338 overloaded, and only one type suffix is required. Because the argument's
6339 type is matched against the return type, it does not require its own
6342 To learn how to add an intrinsic function, please see the `Extending
6343 LLVM Guide <ExtendingLLVM.html>`_.
6347 Variable Argument Handling Intrinsics
6348 -------------------------------------
6350 Variable argument support is defined in LLVM with the
6351 :ref:`va_arg <i_va_arg>` instruction and these three intrinsic
6352 functions. These functions are related to the similarly named macros
6353 defined in the ``<stdarg.h>`` header file.
6355 All of these functions operate on arguments that use a target-specific
6356 value type "``va_list``". The LLVM assembly language reference manual
6357 does not define what this type is, so all transformations should be
6358 prepared to handle these functions regardless of the type used.
6360 This example shows how the :ref:`va_arg <i_va_arg>` instruction and the
6361 variable argument handling intrinsic functions are used.
6363 .. code-block:: llvm
6365 define i32 @test(i32 %X, ...) {
6366 ; Initialize variable argument processing
6368 %ap2 = bitcast i8** %ap to i8*
6369 call void @llvm.va_start(i8* %ap2)
6371 ; Read a single integer argument
6372 %tmp = va_arg i8** %ap, i32
6374 ; Demonstrate usage of llvm.va_copy and llvm.va_end
6376 %aq2 = bitcast i8** %aq to i8*
6377 call void @llvm.va_copy(i8* %aq2, i8* %ap2)
6378 call void @llvm.va_end(i8* %aq2)
6380 ; Stop processing of arguments.
6381 call void @llvm.va_end(i8* %ap2)
6385 declare void @llvm.va_start(i8*)
6386 declare void @llvm.va_copy(i8*, i8*)
6387 declare void @llvm.va_end(i8*)
6391 '``llvm.va_start``' Intrinsic
6392 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6399 declare void @llvm.va_start(i8* <arglist>)
6404 The '``llvm.va_start``' intrinsic initializes ``*<arglist>`` for
6405 subsequent use by ``va_arg``.
6410 The argument is a pointer to a ``va_list`` element to initialize.
6415 The '``llvm.va_start``' intrinsic works just like the ``va_start`` macro
6416 available in C. In a target-dependent way, it initializes the
6417 ``va_list`` element to which the argument points, so that the next call
6418 to ``va_arg`` will produce the first variable argument passed to the
6419 function. Unlike the C ``va_start`` macro, this intrinsic does not need
6420 to know the last argument of the function as the compiler can figure
6423 '``llvm.va_end``' Intrinsic
6424 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6431 declare void @llvm.va_end(i8* <arglist>)
6436 The '``llvm.va_end``' intrinsic destroys ``*<arglist>``, which has been
6437 initialized previously with ``llvm.va_start`` or ``llvm.va_copy``.
6442 The argument is a pointer to a ``va_list`` to destroy.
6447 The '``llvm.va_end``' intrinsic works just like the ``va_end`` macro
6448 available in C. In a target-dependent way, it destroys the ``va_list``
6449 element to which the argument points. Calls to
6450 :ref:`llvm.va_start <int_va_start>` and
6451 :ref:`llvm.va_copy <int_va_copy>` must be matched exactly with calls to
6456 '``llvm.va_copy``' Intrinsic
6457 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6464 declare void @llvm.va_copy(i8* <destarglist>, i8* <srcarglist>)
6469 The '``llvm.va_copy``' intrinsic copies the current argument position
6470 from the source argument list to the destination argument list.
6475 The first argument is a pointer to a ``va_list`` element to initialize.
6476 The second argument is a pointer to a ``va_list`` element to copy from.
6481 The '``llvm.va_copy``' intrinsic works just like the ``va_copy`` macro
6482 available in C. In a target-dependent way, it copies the source
6483 ``va_list`` element into the destination ``va_list`` element. This
6484 intrinsic is necessary because the `` llvm.va_start`` intrinsic may be
6485 arbitrarily complex and require, for example, memory allocation.
6487 Accurate Garbage Collection Intrinsics
6488 --------------------------------------
6490 LLVM support for `Accurate Garbage Collection <GarbageCollection.html>`_
6491 (GC) requires the implementation and generation of these intrinsics.
6492 These intrinsics allow identification of :ref:`GC roots on the
6493 stack <int_gcroot>`, as well as garbage collector implementations that
6494 require :ref:`read <int_gcread>` and :ref:`write <int_gcwrite>` barriers.
6495 Front-ends for type-safe garbage collected languages should generate
6496 these intrinsics to make use of the LLVM garbage collectors. For more
6497 details, see `Accurate Garbage Collection with
6498 LLVM <GarbageCollection.html>`_.
6500 The garbage collection intrinsics only operate on objects in the generic
6501 address space (address space zero).
6505 '``llvm.gcroot``' Intrinsic
6506 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6513 declare void @llvm.gcroot(i8** %ptrloc, i8* %metadata)
6518 The '``llvm.gcroot``' intrinsic declares the existence of a GC root to
6519 the code generator, and allows some metadata to be associated with it.
6524 The first argument specifies the address of a stack object that contains
6525 the root pointer. The second pointer (which must be either a constant or
6526 a global value address) contains the meta-data to be associated with the
6532 At runtime, a call to this intrinsic stores a null pointer into the
6533 "ptrloc" location. At compile-time, the code generator generates
6534 information to allow the runtime to find the pointer at GC safe points.
6535 The '``llvm.gcroot``' intrinsic may only be used in a function which
6536 :ref:`specifies a GC algorithm <gc>`.
6540 '``llvm.gcread``' Intrinsic
6541 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6548 declare i8* @llvm.gcread(i8* %ObjPtr, i8** %Ptr)
6553 The '``llvm.gcread``' intrinsic identifies reads of references from heap
6554 locations, allowing garbage collector implementations that require read
6560 The second argument is the address to read from, which should be an
6561 address allocated from the garbage collector. The first object is a
6562 pointer to the start of the referenced object, if needed by the language
6563 runtime (otherwise null).
6568 The '``llvm.gcread``' intrinsic has the same semantics as a load
6569 instruction, but may be replaced with substantially more complex code by
6570 the garbage collector runtime, as needed. The '``llvm.gcread``'
6571 intrinsic may only be used in a function which :ref:`specifies a GC
6576 '``llvm.gcwrite``' Intrinsic
6577 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6584 declare void @llvm.gcwrite(i8* %P1, i8* %Obj, i8** %P2)
6589 The '``llvm.gcwrite``' intrinsic identifies writes of references to heap
6590 locations, allowing garbage collector implementations that require write
6591 barriers (such as generational or reference counting collectors).
6596 The first argument is the reference to store, the second is the start of
6597 the object to store it to, and the third is the address of the field of
6598 Obj to store to. If the runtime does not require a pointer to the
6599 object, Obj may be null.
6604 The '``llvm.gcwrite``' intrinsic has the same semantics as a store
6605 instruction, but may be replaced with substantially more complex code by
6606 the garbage collector runtime, as needed. The '``llvm.gcwrite``'
6607 intrinsic may only be used in a function which :ref:`specifies a GC
6610 Code Generator Intrinsics
6611 -------------------------
6613 These intrinsics are provided by LLVM to expose special features that
6614 may only be implemented with code generator support.
6616 '``llvm.returnaddress``' Intrinsic
6617 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6624 declare i8 *@llvm.returnaddress(i32 <level>)
6629 The '``llvm.returnaddress``' intrinsic attempts to compute a
6630 target-specific value indicating the return address of the current
6631 function or one of its callers.
6636 The argument to this intrinsic indicates which function to return the
6637 address for. Zero indicates the calling function, one indicates its
6638 caller, etc. The argument is **required** to be a constant integer
6644 The '``llvm.returnaddress``' intrinsic either returns a pointer
6645 indicating the return address of the specified call frame, or zero if it
6646 cannot be identified. The value returned by this intrinsic is likely to
6647 be incorrect or 0 for arguments other than zero, so it should only be
6648 used for debugging purposes.
6650 Note that calling this intrinsic does not prevent function inlining or
6651 other aggressive transformations, so the value returned may not be that
6652 of the obvious source-language caller.
6654 '``llvm.frameaddress``' Intrinsic
6655 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6662 declare i8* @llvm.frameaddress(i32 <level>)
6667 The '``llvm.frameaddress``' intrinsic attempts to return the
6668 target-specific frame pointer value for the specified stack frame.
6673 The argument to this intrinsic indicates which function to return the
6674 frame pointer for. Zero indicates the calling function, one indicates
6675 its caller, etc. The argument is **required** to be a constant integer
6681 The '``llvm.frameaddress``' intrinsic either returns a pointer
6682 indicating the frame address of the specified call frame, or zero if it
6683 cannot be identified. The value returned by this intrinsic is likely to
6684 be incorrect or 0 for arguments other than zero, so it should only be
6685 used for debugging purposes.
6687 Note that calling this intrinsic does not prevent function inlining or
6688 other aggressive transformations, so the value returned may not be that
6689 of the obvious source-language caller.
6693 '``llvm.stacksave``' Intrinsic
6694 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6701 declare i8* @llvm.stacksave()
6706 The '``llvm.stacksave``' intrinsic is used to remember the current state
6707 of the function stack, for use with
6708 :ref:`llvm.stackrestore <int_stackrestore>`. This is useful for
6709 implementing language features like scoped automatic variable sized
6715 This intrinsic returns a opaque pointer value that can be passed to
6716 :ref:`llvm.stackrestore <int_stackrestore>`. When an
6717 ``llvm.stackrestore`` intrinsic is executed with a value saved from
6718 ``llvm.stacksave``, it effectively restores the state of the stack to
6719 the state it was in when the ``llvm.stacksave`` intrinsic executed. In
6720 practice, this pops any :ref:`alloca <i_alloca>` blocks from the stack that
6721 were allocated after the ``llvm.stacksave`` was executed.
6723 .. _int_stackrestore:
6725 '``llvm.stackrestore``' Intrinsic
6726 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6733 declare void @llvm.stackrestore(i8* %ptr)
6738 The '``llvm.stackrestore``' intrinsic is used to restore the state of
6739 the function stack to the state it was in when the corresponding
6740 :ref:`llvm.stacksave <int_stacksave>` intrinsic executed. This is
6741 useful for implementing language features like scoped automatic variable
6742 sized arrays in C99.
6747 See the description for :ref:`llvm.stacksave <int_stacksave>`.
6749 '``llvm.prefetch``' Intrinsic
6750 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6757 declare void @llvm.prefetch(i8* <address>, i32 <rw>, i32 <locality>, i32 <cache type>)
6762 The '``llvm.prefetch``' intrinsic is a hint to the code generator to
6763 insert a prefetch instruction if supported; otherwise, it is a noop.
6764 Prefetches have no effect on the behavior of the program but can change
6765 its performance characteristics.
6770 ``address`` is the address to be prefetched, ``rw`` is the specifier
6771 determining if the fetch should be for a read (0) or write (1), and
6772 ``locality`` is a temporal locality specifier ranging from (0) - no
6773 locality, to (3) - extremely local keep in cache. The ``cache type``
6774 specifies whether the prefetch is performed on the data (1) or
6775 instruction (0) cache. The ``rw``, ``locality`` and ``cache type``
6776 arguments must be constant integers.
6781 This intrinsic does not modify the behavior of the program. In
6782 particular, prefetches cannot trap and do not produce a value. On
6783 targets that support this intrinsic, the prefetch can provide hints to
6784 the processor cache for better performance.
6786 '``llvm.pcmarker``' Intrinsic
6787 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6794 declare void @llvm.pcmarker(i32 <id>)
6799 The '``llvm.pcmarker``' intrinsic is a method to export a Program
6800 Counter (PC) in a region of code to simulators and other tools. The
6801 method is target specific, but it is expected that the marker will use
6802 exported symbols to transmit the PC of the marker. The marker makes no
6803 guarantees that it will remain with any specific instruction after
6804 optimizations. It is possible that the presence of a marker will inhibit
6805 optimizations. The intended use is to be inserted after optimizations to
6806 allow correlations of simulation runs.
6811 ``id`` is a numerical id identifying the marker.
6816 This intrinsic does not modify the behavior of the program. Backends
6817 that do not support this intrinsic may ignore it.
6819 '``llvm.readcyclecounter``' Intrinsic
6820 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6827 declare i64 @llvm.readcyclecounter()
6832 The '``llvm.readcyclecounter``' intrinsic provides access to the cycle
6833 counter register (or similar low latency, high accuracy clocks) on those
6834 targets that support it. On X86, it should map to RDTSC. On Alpha, it
6835 should map to RPCC. As the backing counters overflow quickly (on the
6836 order of 9 seconds on alpha), this should only be used for small
6842 When directly supported, reading the cycle counter should not modify any
6843 memory. Implementations are allowed to either return a application
6844 specific value or a system wide value. On backends without support, this
6845 is lowered to a constant 0.
6847 Note that runtime support may be conditional on the privilege-level code is
6848 running at and the host platform.
6850 Standard C Library Intrinsics
6851 -----------------------------
6853 LLVM provides intrinsics for a few important standard C library
6854 functions. These intrinsics allow source-language front-ends to pass
6855 information about the alignment of the pointer arguments to the code
6856 generator, providing opportunity for more efficient code generation.
6860 '``llvm.memcpy``' Intrinsic
6861 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6866 This is an overloaded intrinsic. You can use ``llvm.memcpy`` on any
6867 integer bit width and for different address spaces. Not all targets
6868 support all bit widths however.
6872 declare void @llvm.memcpy.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
6873 i32 <len>, i32 <align>, i1 <isvolatile>)
6874 declare void @llvm.memcpy.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
6875 i64 <len>, i32 <align>, i1 <isvolatile>)
6880 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
6881 source location to the destination location.
6883 Note that, unlike the standard libc function, the ``llvm.memcpy.*``
6884 intrinsics do not return a value, takes extra alignment/isvolatile
6885 arguments and the pointers can be in specified address spaces.
6890 The first argument is a pointer to the destination, the second is a
6891 pointer to the source. The third argument is an integer argument
6892 specifying the number of bytes to copy, the fourth argument is the
6893 alignment of the source and destination locations, and the fifth is a
6894 boolean indicating a volatile access.
6896 If the call to this intrinsic has an alignment value that is not 0 or 1,
6897 then the caller guarantees that both the source and destination pointers
6898 are aligned to that boundary.
6900 If the ``isvolatile`` parameter is ``true``, the ``llvm.memcpy`` call is
6901 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
6902 very cleanly specified and it is unwise to depend on it.
6907 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
6908 source location to the destination location, which are not allowed to
6909 overlap. It copies "len" bytes of memory over. If the argument is known
6910 to be aligned to some boundary, this can be specified as the fourth
6911 argument, otherwise it should be set to 0 or 1 (both meaning no alignment).
6913 '``llvm.memmove``' Intrinsic
6914 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6919 This is an overloaded intrinsic. You can use llvm.memmove on any integer
6920 bit width and for different address space. Not all targets support all
6925 declare void @llvm.memmove.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
6926 i32 <len>, i32 <align>, i1 <isvolatile>)
6927 declare void @llvm.memmove.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
6928 i64 <len>, i32 <align>, i1 <isvolatile>)
6933 The '``llvm.memmove.*``' intrinsics move a block of memory from the
6934 source location to the destination location. It is similar to the
6935 '``llvm.memcpy``' intrinsic but allows the two memory locations to
6938 Note that, unlike the standard libc function, the ``llvm.memmove.*``
6939 intrinsics do not return a value, takes extra alignment/isvolatile
6940 arguments and the pointers can be in specified address spaces.
6945 The first argument is a pointer to the destination, the second is a
6946 pointer to the source. The third argument is an integer argument
6947 specifying the number of bytes to copy, the fourth argument is the
6948 alignment of the source and destination locations, and the fifth is a
6949 boolean indicating a volatile access.
6951 If the call to this intrinsic has an alignment value that is not 0 or 1,
6952 then the caller guarantees that the source and destination pointers are
6953 aligned to that boundary.
6955 If the ``isvolatile`` parameter is ``true``, the ``llvm.memmove`` call
6956 is a :ref:`volatile operation <volatile>`. The detailed access behavior is
6957 not very cleanly specified and it is unwise to depend on it.
6962 The '``llvm.memmove.*``' intrinsics copy a block of memory from the
6963 source location to the destination location, which may overlap. It
6964 copies "len" bytes of memory over. If the argument is known to be
6965 aligned to some boundary, this can be specified as the fourth argument,
6966 otherwise it should be set to 0 or 1 (both meaning no alignment).
6968 '``llvm.memset.*``' Intrinsics
6969 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6974 This is an overloaded intrinsic. You can use llvm.memset on any integer
6975 bit width and for different address spaces. However, not all targets
6976 support all bit widths.
6980 declare void @llvm.memset.p0i8.i32(i8* <dest>, i8 <val>,
6981 i32 <len>, i32 <align>, i1 <isvolatile>)
6982 declare void @llvm.memset.p0i8.i64(i8* <dest>, i8 <val>,
6983 i64 <len>, i32 <align>, i1 <isvolatile>)
6988 The '``llvm.memset.*``' intrinsics fill a block of memory with a
6989 particular byte value.
6991 Note that, unlike the standard libc function, the ``llvm.memset``
6992 intrinsic does not return a value and takes extra alignment/volatile
6993 arguments. Also, the destination can be in an arbitrary address space.
6998 The first argument is a pointer to the destination to fill, the second
6999 is the byte value with which to fill it, the third argument is an
7000 integer argument specifying the number of bytes to fill, and the fourth
7001 argument is the known alignment of the destination location.
7003 If the call to this intrinsic has an alignment value that is not 0 or 1,
7004 then the caller guarantees that the destination pointer is aligned to
7007 If the ``isvolatile`` parameter is ``true``, the ``llvm.memset`` call is
7008 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
7009 very cleanly specified and it is unwise to depend on it.
7014 The '``llvm.memset.*``' intrinsics fill "len" bytes of memory starting
7015 at the destination location. If the argument is known to be aligned to
7016 some boundary, this can be specified as the fourth argument, otherwise
7017 it should be set to 0 or 1 (both meaning no alignment).
7019 '``llvm.sqrt.*``' Intrinsic
7020 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7025 This is an overloaded intrinsic. You can use ``llvm.sqrt`` on any
7026 floating point or vector of floating point type. Not all targets support
7031 declare float @llvm.sqrt.f32(float %Val)
7032 declare double @llvm.sqrt.f64(double %Val)
7033 declare x86_fp80 @llvm.sqrt.f80(x86_fp80 %Val)
7034 declare fp128 @llvm.sqrt.f128(fp128 %Val)
7035 declare ppc_fp128 @llvm.sqrt.ppcf128(ppc_fp128 %Val)
7040 The '``llvm.sqrt``' intrinsics return the sqrt of the specified operand,
7041 returning the same value as the libm '``sqrt``' functions would. Unlike
7042 ``sqrt`` in libm, however, ``llvm.sqrt`` has undefined behavior for
7043 negative numbers other than -0.0 (which allows for better optimization,
7044 because there is no need to worry about errno being set).
7045 ``llvm.sqrt(-0.0)`` is defined to return -0.0 like IEEE sqrt.
7050 The argument and return value are floating point numbers of the same
7056 This function returns the sqrt of the specified operand if it is a
7057 nonnegative floating point number.
7059 '``llvm.powi.*``' Intrinsic
7060 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7065 This is an overloaded intrinsic. You can use ``llvm.powi`` on any
7066 floating point or vector of floating point type. Not all targets support
7071 declare float @llvm.powi.f32(float %Val, i32 %power)
7072 declare double @llvm.powi.f64(double %Val, i32 %power)
7073 declare x86_fp80 @llvm.powi.f80(x86_fp80 %Val, i32 %power)
7074 declare fp128 @llvm.powi.f128(fp128 %Val, i32 %power)
7075 declare ppc_fp128 @llvm.powi.ppcf128(ppc_fp128 %Val, i32 %power)
7080 The '``llvm.powi.*``' intrinsics return the first operand raised to the
7081 specified (positive or negative) power. The order of evaluation of
7082 multiplications is not defined. When a vector of floating point type is
7083 used, the second argument remains a scalar integer value.
7088 The second argument is an integer power, and the first is a value to
7089 raise to that power.
7094 This function returns the first value raised to the second power with an
7095 unspecified sequence of rounding operations.
7097 '``llvm.sin.*``' Intrinsic
7098 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7103 This is an overloaded intrinsic. You can use ``llvm.sin`` on any
7104 floating point or vector of floating point type. Not all targets support
7109 declare float @llvm.sin.f32(float %Val)
7110 declare double @llvm.sin.f64(double %Val)
7111 declare x86_fp80 @llvm.sin.f80(x86_fp80 %Val)
7112 declare fp128 @llvm.sin.f128(fp128 %Val)
7113 declare ppc_fp128 @llvm.sin.ppcf128(ppc_fp128 %Val)
7118 The '``llvm.sin.*``' intrinsics return the sine of the operand.
7123 The argument and return value are floating point numbers of the same
7129 This function returns the sine of the specified operand, returning the
7130 same values as the libm ``sin`` functions would, and handles error
7131 conditions in the same way.
7133 '``llvm.cos.*``' Intrinsic
7134 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7139 This is an overloaded intrinsic. You can use ``llvm.cos`` on any
7140 floating point or vector of floating point type. Not all targets support
7145 declare float @llvm.cos.f32(float %Val)
7146 declare double @llvm.cos.f64(double %Val)
7147 declare x86_fp80 @llvm.cos.f80(x86_fp80 %Val)
7148 declare fp128 @llvm.cos.f128(fp128 %Val)
7149 declare ppc_fp128 @llvm.cos.ppcf128(ppc_fp128 %Val)
7154 The '``llvm.cos.*``' intrinsics return the cosine of the operand.
7159 The argument and return value are floating point numbers of the same
7165 This function returns the cosine of the specified operand, returning the
7166 same values as the libm ``cos`` functions would, and handles error
7167 conditions in the same way.
7169 '``llvm.pow.*``' Intrinsic
7170 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7175 This is an overloaded intrinsic. You can use ``llvm.pow`` on any
7176 floating point or vector of floating point type. Not all targets support
7181 declare float @llvm.pow.f32(float %Val, float %Power)
7182 declare double @llvm.pow.f64(double %Val, double %Power)
7183 declare x86_fp80 @llvm.pow.f80(x86_fp80 %Val, x86_fp80 %Power)
7184 declare fp128 @llvm.pow.f128(fp128 %Val, fp128 %Power)
7185 declare ppc_fp128 @llvm.pow.ppcf128(ppc_fp128 %Val, ppc_fp128 Power)
7190 The '``llvm.pow.*``' intrinsics return the first operand raised to the
7191 specified (positive or negative) power.
7196 The second argument is a floating point power, and the first is a value
7197 to raise to that power.
7202 This function returns the first value raised to the second power,
7203 returning the same values as the libm ``pow`` functions would, and
7204 handles error conditions in the same way.
7206 '``llvm.exp.*``' Intrinsic
7207 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7212 This is an overloaded intrinsic. You can use ``llvm.exp`` on any
7213 floating point or vector of floating point type. Not all targets support
7218 declare float @llvm.exp.f32(float %Val)
7219 declare double @llvm.exp.f64(double %Val)
7220 declare x86_fp80 @llvm.exp.f80(x86_fp80 %Val)
7221 declare fp128 @llvm.exp.f128(fp128 %Val)
7222 declare ppc_fp128 @llvm.exp.ppcf128(ppc_fp128 %Val)
7227 The '``llvm.exp.*``' intrinsics perform the exp function.
7232 The argument and return value are floating point numbers of the same
7238 This function returns the same values as the libm ``exp`` functions
7239 would, and handles error conditions in the same way.
7241 '``llvm.exp2.*``' Intrinsic
7242 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7247 This is an overloaded intrinsic. You can use ``llvm.exp2`` on any
7248 floating point or vector of floating point type. Not all targets support
7253 declare float @llvm.exp2.f32(float %Val)
7254 declare double @llvm.exp2.f64(double %Val)
7255 declare x86_fp80 @llvm.exp2.f80(x86_fp80 %Val)
7256 declare fp128 @llvm.exp2.f128(fp128 %Val)
7257 declare ppc_fp128 @llvm.exp2.ppcf128(ppc_fp128 %Val)
7262 The '``llvm.exp2.*``' intrinsics perform the exp2 function.
7267 The argument and return value are floating point numbers of the same
7273 This function returns the same values as the libm ``exp2`` functions
7274 would, and handles error conditions in the same way.
7276 '``llvm.log.*``' Intrinsic
7277 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7282 This is an overloaded intrinsic. You can use ``llvm.log`` on any
7283 floating point or vector of floating point type. Not all targets support
7288 declare float @llvm.log.f32(float %Val)
7289 declare double @llvm.log.f64(double %Val)
7290 declare x86_fp80 @llvm.log.f80(x86_fp80 %Val)
7291 declare fp128 @llvm.log.f128(fp128 %Val)
7292 declare ppc_fp128 @llvm.log.ppcf128(ppc_fp128 %Val)
7297 The '``llvm.log.*``' intrinsics perform the log function.
7302 The argument and return value are floating point numbers of the same
7308 This function returns the same values as the libm ``log`` functions
7309 would, and handles error conditions in the same way.
7311 '``llvm.log10.*``' Intrinsic
7312 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7317 This is an overloaded intrinsic. You can use ``llvm.log10`` on any
7318 floating point or vector of floating point type. Not all targets support
7323 declare float @llvm.log10.f32(float %Val)
7324 declare double @llvm.log10.f64(double %Val)
7325 declare x86_fp80 @llvm.log10.f80(x86_fp80 %Val)
7326 declare fp128 @llvm.log10.f128(fp128 %Val)
7327 declare ppc_fp128 @llvm.log10.ppcf128(ppc_fp128 %Val)
7332 The '``llvm.log10.*``' intrinsics perform the log10 function.
7337 The argument and return value are floating point numbers of the same
7343 This function returns the same values as the libm ``log10`` functions
7344 would, and handles error conditions in the same way.
7346 '``llvm.log2.*``' Intrinsic
7347 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7352 This is an overloaded intrinsic. You can use ``llvm.log2`` on any
7353 floating point or vector of floating point type. Not all targets support
7358 declare float @llvm.log2.f32(float %Val)
7359 declare double @llvm.log2.f64(double %Val)
7360 declare x86_fp80 @llvm.log2.f80(x86_fp80 %Val)
7361 declare fp128 @llvm.log2.f128(fp128 %Val)
7362 declare ppc_fp128 @llvm.log2.ppcf128(ppc_fp128 %Val)
7367 The '``llvm.log2.*``' intrinsics perform the log2 function.
7372 The argument and return value are floating point numbers of the same
7378 This function returns the same values as the libm ``log2`` functions
7379 would, and handles error conditions in the same way.
7381 '``llvm.fma.*``' Intrinsic
7382 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7387 This is an overloaded intrinsic. You can use ``llvm.fma`` on any
7388 floating point or vector of floating point type. Not all targets support
7393 declare float @llvm.fma.f32(float %a, float %b, float %c)
7394 declare double @llvm.fma.f64(double %a, double %b, double %c)
7395 declare x86_fp80 @llvm.fma.f80(x86_fp80 %a, x86_fp80 %b, x86_fp80 %c)
7396 declare fp128 @llvm.fma.f128(fp128 %a, fp128 %b, fp128 %c)
7397 declare ppc_fp128 @llvm.fma.ppcf128(ppc_fp128 %a, ppc_fp128 %b, ppc_fp128 %c)
7402 The '``llvm.fma.*``' intrinsics perform the fused multiply-add
7408 The argument and return value are floating point numbers of the same
7414 This function returns the same values as the libm ``fma`` functions
7417 '``llvm.fabs.*``' Intrinsic
7418 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7423 This is an overloaded intrinsic. You can use ``llvm.fabs`` on any
7424 floating point or vector of floating point type. Not all targets support
7429 declare float @llvm.fabs.f32(float %Val)
7430 declare double @llvm.fabs.f64(double %Val)
7431 declare x86_fp80 @llvm.fabs.f80(x86_fp80 %Val)
7432 declare fp128 @llvm.fabs.f128(fp128 %Val)
7433 declare ppc_fp128 @llvm.fabs.ppcf128(ppc_fp128 %Val)
7438 The '``llvm.fabs.*``' intrinsics return the absolute value of the
7444 The argument and return value are floating point numbers of the same
7450 This function returns the same values as the libm ``fabs`` functions
7451 would, and handles error conditions in the same way.
7453 '``llvm.copysign.*``' Intrinsic
7454 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7459 This is an overloaded intrinsic. You can use ``llvm.copysign`` on any
7460 floating point or vector of floating point type. Not all targets support
7465 declare float @llvm.copysign.f32(float %Mag, float %Sgn)
7466 declare double @llvm.copysign.f64(double %Mag, double %Sgn)
7467 declare x86_fp80 @llvm.copysign.f80(x86_fp80 %Mag, x86_fp80 %Sgn)
7468 declare fp128 @llvm.copysign.f128(fp128 %Mag, fp128 %Sgn)
7469 declare ppc_fp128 @llvm.copysign.ppcf128(ppc_fp128 %Mag, ppc_fp128 %Sgn)
7474 The '``llvm.copysign.*``' intrinsics return a value with the magnitude of the
7475 first operand and the sign of the second operand.
7480 The arguments and return value are floating point numbers of the same
7486 This function returns the same values as the libm ``copysign``
7487 functions would, and handles error conditions in the same way.
7489 '``llvm.floor.*``' Intrinsic
7490 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7495 This is an overloaded intrinsic. You can use ``llvm.floor`` on any
7496 floating point or vector of floating point type. Not all targets support
7501 declare float @llvm.floor.f32(float %Val)
7502 declare double @llvm.floor.f64(double %Val)
7503 declare x86_fp80 @llvm.floor.f80(x86_fp80 %Val)
7504 declare fp128 @llvm.floor.f128(fp128 %Val)
7505 declare ppc_fp128 @llvm.floor.ppcf128(ppc_fp128 %Val)
7510 The '``llvm.floor.*``' intrinsics return the floor of the operand.
7515 The argument and return value are floating point numbers of the same
7521 This function returns the same values as the libm ``floor`` functions
7522 would, and handles error conditions in the same way.
7524 '``llvm.ceil.*``' Intrinsic
7525 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7530 This is an overloaded intrinsic. You can use ``llvm.ceil`` on any
7531 floating point or vector of floating point type. Not all targets support
7536 declare float @llvm.ceil.f32(float %Val)
7537 declare double @llvm.ceil.f64(double %Val)
7538 declare x86_fp80 @llvm.ceil.f80(x86_fp80 %Val)
7539 declare fp128 @llvm.ceil.f128(fp128 %Val)
7540 declare ppc_fp128 @llvm.ceil.ppcf128(ppc_fp128 %Val)
7545 The '``llvm.ceil.*``' intrinsics return the ceiling of the operand.
7550 The argument and return value are floating point numbers of the same
7556 This function returns the same values as the libm ``ceil`` functions
7557 would, and handles error conditions in the same way.
7559 '``llvm.trunc.*``' Intrinsic
7560 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7565 This is an overloaded intrinsic. You can use ``llvm.trunc`` on any
7566 floating point or vector of floating point type. Not all targets support
7571 declare float @llvm.trunc.f32(float %Val)
7572 declare double @llvm.trunc.f64(double %Val)
7573 declare x86_fp80 @llvm.trunc.f80(x86_fp80 %Val)
7574 declare fp128 @llvm.trunc.f128(fp128 %Val)
7575 declare ppc_fp128 @llvm.trunc.ppcf128(ppc_fp128 %Val)
7580 The '``llvm.trunc.*``' intrinsics returns the operand rounded to the
7581 nearest integer not larger in magnitude than the operand.
7586 The argument and return value are floating point numbers of the same
7592 This function returns the same values as the libm ``trunc`` functions
7593 would, and handles error conditions in the same way.
7595 '``llvm.rint.*``' Intrinsic
7596 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7601 This is an overloaded intrinsic. You can use ``llvm.rint`` on any
7602 floating point or vector of floating point type. Not all targets support
7607 declare float @llvm.rint.f32(float %Val)
7608 declare double @llvm.rint.f64(double %Val)
7609 declare x86_fp80 @llvm.rint.f80(x86_fp80 %Val)
7610 declare fp128 @llvm.rint.f128(fp128 %Val)
7611 declare ppc_fp128 @llvm.rint.ppcf128(ppc_fp128 %Val)
7616 The '``llvm.rint.*``' intrinsics returns the operand rounded to the
7617 nearest integer. It may raise an inexact floating-point exception if the
7618 operand isn't an integer.
7623 The argument and return value are floating point numbers of the same
7629 This function returns the same values as the libm ``rint`` functions
7630 would, and handles error conditions in the same way.
7632 '``llvm.nearbyint.*``' Intrinsic
7633 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7638 This is an overloaded intrinsic. You can use ``llvm.nearbyint`` on any
7639 floating point or vector of floating point type. Not all targets support
7644 declare float @llvm.nearbyint.f32(float %Val)
7645 declare double @llvm.nearbyint.f64(double %Val)
7646 declare x86_fp80 @llvm.nearbyint.f80(x86_fp80 %Val)
7647 declare fp128 @llvm.nearbyint.f128(fp128 %Val)
7648 declare ppc_fp128 @llvm.nearbyint.ppcf128(ppc_fp128 %Val)
7653 The '``llvm.nearbyint.*``' intrinsics returns the operand rounded to the
7659 The argument and return value are floating point numbers of the same
7665 This function returns the same values as the libm ``nearbyint``
7666 functions would, and handles error conditions in the same way.
7668 '``llvm.round.*``' Intrinsic
7669 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7674 This is an overloaded intrinsic. You can use ``llvm.round`` on any
7675 floating point or vector of floating point type. Not all targets support
7680 declare float @llvm.round.f32(float %Val)
7681 declare double @llvm.round.f64(double %Val)
7682 declare x86_fp80 @llvm.round.f80(x86_fp80 %Val)
7683 declare fp128 @llvm.round.f128(fp128 %Val)
7684 declare ppc_fp128 @llvm.round.ppcf128(ppc_fp128 %Val)
7689 The '``llvm.round.*``' intrinsics returns the operand rounded to the
7695 The argument and return value are floating point numbers of the same
7701 This function returns the same values as the libm ``round``
7702 functions would, and handles error conditions in the same way.
7704 Bit Manipulation Intrinsics
7705 ---------------------------
7707 LLVM provides intrinsics for a few important bit manipulation
7708 operations. These allow efficient code generation for some algorithms.
7710 '``llvm.bswap.*``' Intrinsics
7711 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7716 This is an overloaded intrinsic function. You can use bswap on any
7717 integer type that is an even number of bytes (i.e. BitWidth % 16 == 0).
7721 declare i16 @llvm.bswap.i16(i16 <id>)
7722 declare i32 @llvm.bswap.i32(i32 <id>)
7723 declare i64 @llvm.bswap.i64(i64 <id>)
7728 The '``llvm.bswap``' family of intrinsics is used to byte swap integer
7729 values with an even number of bytes (positive multiple of 16 bits).
7730 These are useful for performing operations on data that is not in the
7731 target's native byte order.
7736 The ``llvm.bswap.i16`` intrinsic returns an i16 value that has the high
7737 and low byte of the input i16 swapped. Similarly, the ``llvm.bswap.i32``
7738 intrinsic returns an i32 value that has the four bytes of the input i32
7739 swapped, so that if the input bytes are numbered 0, 1, 2, 3 then the
7740 returned i32 will have its bytes in 3, 2, 1, 0 order. The
7741 ``llvm.bswap.i48``, ``llvm.bswap.i64`` and other intrinsics extend this
7742 concept to additional even-byte lengths (6 bytes, 8 bytes and more,
7745 '``llvm.ctpop.*``' Intrinsic
7746 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7751 This is an overloaded intrinsic. You can use llvm.ctpop on any integer
7752 bit width, or on any vector with integer elements. Not all targets
7753 support all bit widths or vector types, however.
7757 declare i8 @llvm.ctpop.i8(i8 <src>)
7758 declare i16 @llvm.ctpop.i16(i16 <src>)
7759 declare i32 @llvm.ctpop.i32(i32 <src>)
7760 declare i64 @llvm.ctpop.i64(i64 <src>)
7761 declare i256 @llvm.ctpop.i256(i256 <src>)
7762 declare <2 x i32> @llvm.ctpop.v2i32(<2 x i32> <src>)
7767 The '``llvm.ctpop``' family of intrinsics counts the number of bits set
7773 The only argument is the value to be counted. The argument may be of any
7774 integer type, or a vector with integer elements. The return type must
7775 match the argument type.
7780 The '``llvm.ctpop``' intrinsic counts the 1's in a variable, or within
7781 each element of a vector.
7783 '``llvm.ctlz.*``' Intrinsic
7784 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7789 This is an overloaded intrinsic. You can use ``llvm.ctlz`` on any
7790 integer bit width, or any vector whose elements are integers. Not all
7791 targets support all bit widths or vector types, however.
7795 declare i8 @llvm.ctlz.i8 (i8 <src>, i1 <is_zero_undef>)
7796 declare i16 @llvm.ctlz.i16 (i16 <src>, i1 <is_zero_undef>)
7797 declare i32 @llvm.ctlz.i32 (i32 <src>, i1 <is_zero_undef>)
7798 declare i64 @llvm.ctlz.i64 (i64 <src>, i1 <is_zero_undef>)
7799 declare i256 @llvm.ctlz.i256(i256 <src>, i1 <is_zero_undef>)
7800 declase <2 x i32> @llvm.ctlz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
7805 The '``llvm.ctlz``' family of intrinsic functions counts the number of
7806 leading zeros in a variable.
7811 The first argument is the value to be counted. This argument may be of
7812 any integer type, or a vectory with integer element type. The return
7813 type must match the first argument type.
7815 The second argument must be a constant and is a flag to indicate whether
7816 the intrinsic should ensure that a zero as the first argument produces a
7817 defined result. Historically some architectures did not provide a
7818 defined result for zero values as efficiently, and many algorithms are
7819 now predicated on avoiding zero-value inputs.
7824 The '``llvm.ctlz``' intrinsic counts the leading (most significant)
7825 zeros in a variable, or within each element of the vector. If
7826 ``src == 0`` then the result is the size in bits of the type of ``src``
7827 if ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
7828 ``llvm.ctlz(i32 2) = 30``.
7830 '``llvm.cttz.*``' Intrinsic
7831 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7836 This is an overloaded intrinsic. You can use ``llvm.cttz`` on any
7837 integer bit width, or any vector of integer elements. Not all targets
7838 support all bit widths or vector types, however.
7842 declare i8 @llvm.cttz.i8 (i8 <src>, i1 <is_zero_undef>)
7843 declare i16 @llvm.cttz.i16 (i16 <src>, i1 <is_zero_undef>)
7844 declare i32 @llvm.cttz.i32 (i32 <src>, i1 <is_zero_undef>)
7845 declare i64 @llvm.cttz.i64 (i64 <src>, i1 <is_zero_undef>)
7846 declare i256 @llvm.cttz.i256(i256 <src>, i1 <is_zero_undef>)
7847 declase <2 x i32> @llvm.cttz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
7852 The '``llvm.cttz``' family of intrinsic functions counts the number of
7858 The first argument is the value to be counted. This argument may be of
7859 any integer type, or a vectory with integer element type. The return
7860 type must match the first argument type.
7862 The second argument must be a constant and is a flag to indicate whether
7863 the intrinsic should ensure that a zero as the first argument produces a
7864 defined result. Historically some architectures did not provide a
7865 defined result for zero values as efficiently, and many algorithms are
7866 now predicated on avoiding zero-value inputs.
7871 The '``llvm.cttz``' intrinsic counts the trailing (least significant)
7872 zeros in a variable, or within each element of a vector. If ``src == 0``
7873 then the result is the size in bits of the type of ``src`` if
7874 ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
7875 ``llvm.cttz(2) = 1``.
7877 Arithmetic with Overflow Intrinsics
7878 -----------------------------------
7880 LLVM provides intrinsics for some arithmetic with overflow operations.
7882 '``llvm.sadd.with.overflow.*``' Intrinsics
7883 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7888 This is an overloaded intrinsic. You can use ``llvm.sadd.with.overflow``
7889 on any integer bit width.
7893 declare {i16, i1} @llvm.sadd.with.overflow.i16(i16 %a, i16 %b)
7894 declare {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
7895 declare {i64, i1} @llvm.sadd.with.overflow.i64(i64 %a, i64 %b)
7900 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
7901 a signed addition of the two arguments, and indicate whether an overflow
7902 occurred during the signed summation.
7907 The arguments (%a and %b) and the first element of the result structure
7908 may be of integer types of any bit width, but they must have the same
7909 bit width. The second element of the result structure must be of type
7910 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
7916 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
7917 a signed addition of the two variables. They return a structure --- the
7918 first element of which is the signed summation, and the second element
7919 of which is a bit specifying if the signed summation resulted in an
7925 .. code-block:: llvm
7927 %res = call {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
7928 %sum = extractvalue {i32, i1} %res, 0
7929 %obit = extractvalue {i32, i1} %res, 1
7930 br i1 %obit, label %overflow, label %normal
7932 '``llvm.uadd.with.overflow.*``' Intrinsics
7933 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7938 This is an overloaded intrinsic. You can use ``llvm.uadd.with.overflow``
7939 on any integer bit width.
7943 declare {i16, i1} @llvm.uadd.with.overflow.i16(i16 %a, i16 %b)
7944 declare {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
7945 declare {i64, i1} @llvm.uadd.with.overflow.i64(i64 %a, i64 %b)
7950 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
7951 an unsigned addition of the two arguments, and indicate whether a carry
7952 occurred during the unsigned summation.
7957 The arguments (%a and %b) and the first element of the result structure
7958 may be of integer types of any bit width, but they must have the same
7959 bit width. The second element of the result structure must be of type
7960 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
7966 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
7967 an unsigned addition of the two arguments. They return a structure --- the
7968 first element of which is the sum, and the second element of which is a
7969 bit specifying if the unsigned summation resulted in a carry.
7974 .. code-block:: llvm
7976 %res = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
7977 %sum = extractvalue {i32, i1} %res, 0
7978 %obit = extractvalue {i32, i1} %res, 1
7979 br i1 %obit, label %carry, label %normal
7981 '``llvm.ssub.with.overflow.*``' Intrinsics
7982 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7987 This is an overloaded intrinsic. You can use ``llvm.ssub.with.overflow``
7988 on any integer bit width.
7992 declare {i16, i1} @llvm.ssub.with.overflow.i16(i16 %a, i16 %b)
7993 declare {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
7994 declare {i64, i1} @llvm.ssub.with.overflow.i64(i64 %a, i64 %b)
7999 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
8000 a signed subtraction of the two arguments, and indicate whether an
8001 overflow occurred during the signed subtraction.
8006 The arguments (%a and %b) and the first element of the result structure
8007 may be of integer types of any bit width, but they must have the same
8008 bit width. The second element of the result structure must be of type
8009 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
8015 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
8016 a signed subtraction of the two arguments. They return a structure --- the
8017 first element of which is the subtraction, and the second element of
8018 which is a bit specifying if the signed subtraction resulted in an
8024 .. code-block:: llvm
8026 %res = call {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
8027 %sum = extractvalue {i32, i1} %res, 0
8028 %obit = extractvalue {i32, i1} %res, 1
8029 br i1 %obit, label %overflow, label %normal
8031 '``llvm.usub.with.overflow.*``' Intrinsics
8032 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8037 This is an overloaded intrinsic. You can use ``llvm.usub.with.overflow``
8038 on any integer bit width.
8042 declare {i16, i1} @llvm.usub.with.overflow.i16(i16 %a, i16 %b)
8043 declare {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
8044 declare {i64, i1} @llvm.usub.with.overflow.i64(i64 %a, i64 %b)
8049 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
8050 an unsigned subtraction of the two arguments, and indicate whether an
8051 overflow occurred during the unsigned subtraction.
8056 The arguments (%a and %b) and the first element of the result structure
8057 may be of integer types of any bit width, but they must have the same
8058 bit width. The second element of the result structure must be of type
8059 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
8065 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
8066 an unsigned subtraction of the two arguments. They return a structure ---
8067 the first element of which is the subtraction, and the second element of
8068 which is a bit specifying if the unsigned subtraction resulted in an
8074 .. code-block:: llvm
8076 %res = call {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
8077 %sum = extractvalue {i32, i1} %res, 0
8078 %obit = extractvalue {i32, i1} %res, 1
8079 br i1 %obit, label %overflow, label %normal
8081 '``llvm.smul.with.overflow.*``' Intrinsics
8082 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8087 This is an overloaded intrinsic. You can use ``llvm.smul.with.overflow``
8088 on any integer bit width.
8092 declare {i16, i1} @llvm.smul.with.overflow.i16(i16 %a, i16 %b)
8093 declare {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
8094 declare {i64, i1} @llvm.smul.with.overflow.i64(i64 %a, i64 %b)
8099 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
8100 a signed multiplication of the two arguments, and indicate whether an
8101 overflow occurred during the signed multiplication.
8106 The arguments (%a and %b) and the first element of the result structure
8107 may be of integer types of any bit width, but they must have the same
8108 bit width. The second element of the result structure must be of type
8109 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
8115 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
8116 a signed multiplication of the two arguments. They return a structure ---
8117 the first element of which is the multiplication, and the second element
8118 of which is a bit specifying if the signed multiplication resulted in an
8124 .. code-block:: llvm
8126 %res = call {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
8127 %sum = extractvalue {i32, i1} %res, 0
8128 %obit = extractvalue {i32, i1} %res, 1
8129 br i1 %obit, label %overflow, label %normal
8131 '``llvm.umul.with.overflow.*``' Intrinsics
8132 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8137 This is an overloaded intrinsic. You can use ``llvm.umul.with.overflow``
8138 on any integer bit width.
8142 declare {i16, i1} @llvm.umul.with.overflow.i16(i16 %a, i16 %b)
8143 declare {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
8144 declare {i64, i1} @llvm.umul.with.overflow.i64(i64 %a, i64 %b)
8149 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
8150 a unsigned multiplication of the two arguments, and indicate whether an
8151 overflow occurred during the unsigned multiplication.
8156 The arguments (%a and %b) and the first element of the result structure
8157 may be of integer types of any bit width, but they must have the same
8158 bit width. The second element of the result structure must be of type
8159 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
8165 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
8166 an unsigned multiplication of the two arguments. They return a structure ---
8167 the first element of which is the multiplication, and the second
8168 element of which is a bit specifying if the unsigned multiplication
8169 resulted in an overflow.
8174 .. code-block:: llvm
8176 %res = call {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
8177 %sum = extractvalue {i32, i1} %res, 0
8178 %obit = extractvalue {i32, i1} %res, 1
8179 br i1 %obit, label %overflow, label %normal
8181 Specialised Arithmetic Intrinsics
8182 ---------------------------------
8184 '``llvm.fmuladd.*``' Intrinsic
8185 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8192 declare float @llvm.fmuladd.f32(float %a, float %b, float %c)
8193 declare double @llvm.fmuladd.f64(double %a, double %b, double %c)
8198 The '``llvm.fmuladd.*``' intrinsic functions represent multiply-add
8199 expressions that can be fused if the code generator determines that (a) the
8200 target instruction set has support for a fused operation, and (b) that the
8201 fused operation is more efficient than the equivalent, separate pair of mul
8202 and add instructions.
8207 The '``llvm.fmuladd.*``' intrinsics each take three arguments: two
8208 multiplicands, a and b, and an addend c.
8217 %0 = call float @llvm.fmuladd.f32(%a, %b, %c)
8219 is equivalent to the expression a \* b + c, except that rounding will
8220 not be performed between the multiplication and addition steps if the
8221 code generator fuses the operations. Fusion is not guaranteed, even if
8222 the target platform supports it. If a fused multiply-add is required the
8223 corresponding llvm.fma.\* intrinsic function should be used instead.
8228 .. code-block:: llvm
8230 %r2 = call float @llvm.fmuladd.f32(float %a, float %b, float %c) ; yields {float}:r2 = (a * b) + c
8232 Half Precision Floating Point Intrinsics
8233 ----------------------------------------
8235 For most target platforms, half precision floating point is a
8236 storage-only format. This means that it is a dense encoding (in memory)
8237 but does not support computation in the format.
8239 This means that code must first load the half-precision floating point
8240 value as an i16, then convert it to float with
8241 :ref:`llvm.convert.from.fp16 <int_convert_from_fp16>`. Computation can
8242 then be performed on the float value (including extending to double
8243 etc). To store the value back to memory, it is first converted to float
8244 if needed, then converted to i16 with
8245 :ref:`llvm.convert.to.fp16 <int_convert_to_fp16>`, then storing as an
8248 .. _int_convert_to_fp16:
8250 '``llvm.convert.to.fp16``' Intrinsic
8251 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8258 declare i16 @llvm.convert.to.fp16(f32 %a)
8263 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion
8264 from single precision floating point format to half precision floating
8270 The intrinsic function contains single argument - the value to be
8276 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion
8277 from single precision floating point format to half precision floating
8278 point format. The return value is an ``i16`` which contains the
8284 .. code-block:: llvm
8286 %res = call i16 @llvm.convert.to.fp16(f32 %a)
8287 store i16 %res, i16* @x, align 2
8289 .. _int_convert_from_fp16:
8291 '``llvm.convert.from.fp16``' Intrinsic
8292 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8299 declare f32 @llvm.convert.from.fp16(i16 %a)
8304 The '``llvm.convert.from.fp16``' intrinsic function performs a
8305 conversion from half precision floating point format to single precision
8306 floating point format.
8311 The intrinsic function contains single argument - the value to be
8317 The '``llvm.convert.from.fp16``' intrinsic function performs a
8318 conversion from half single precision floating point format to single
8319 precision floating point format. The input half-float value is
8320 represented by an ``i16`` value.
8325 .. code-block:: llvm
8327 %a = load i16* @x, align 2
8328 %res = call f32 @llvm.convert.from.fp16(i16 %a)
8333 The LLVM debugger intrinsics (which all start with ``llvm.dbg.``
8334 prefix), are described in the `LLVM Source Level
8335 Debugging <SourceLevelDebugging.html#format_common_intrinsics>`_
8338 Exception Handling Intrinsics
8339 -----------------------------
8341 The LLVM exception handling intrinsics (which all start with
8342 ``llvm.eh.`` prefix), are described in the `LLVM Exception
8343 Handling <ExceptionHandling.html#format_common_intrinsics>`_ document.
8347 Trampoline Intrinsics
8348 ---------------------
8350 These intrinsics make it possible to excise one parameter, marked with
8351 the :ref:`nest <nest>` attribute, from a function. The result is a
8352 callable function pointer lacking the nest parameter - the caller does
8353 not need to provide a value for it. Instead, the value to use is stored
8354 in advance in a "trampoline", a block of memory usually allocated on the
8355 stack, which also contains code to splice the nest value into the
8356 argument list. This is used to implement the GCC nested function address
8359 For example, if the function is ``i32 f(i8* nest %c, i32 %x, i32 %y)``
8360 then the resulting function pointer has signature ``i32 (i32, i32)*``.
8361 It can be created as follows:
8363 .. code-block:: llvm
8365 %tramp = alloca [10 x i8], align 4 ; size and alignment only correct for X86
8366 %tramp1 = getelementptr [10 x i8]* %tramp, i32 0, i32 0
8367 call i8* @llvm.init.trampoline(i8* %tramp1, i8* bitcast (i32 (i8*, i32, i32)* @f to i8*), i8* %nval)
8368 %p = call i8* @llvm.adjust.trampoline(i8* %tramp1)
8369 %fp = bitcast i8* %p to i32 (i32, i32)*
8371 The call ``%val = call i32 %fp(i32 %x, i32 %y)`` is then equivalent to
8372 ``%val = call i32 %f(i8* %nval, i32 %x, i32 %y)``.
8376 '``llvm.init.trampoline``' Intrinsic
8377 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8384 declare void @llvm.init.trampoline(i8* <tramp>, i8* <func>, i8* <nval>)
8389 This fills the memory pointed to by ``tramp`` with executable code,
8390 turning it into a trampoline.
8395 The ``llvm.init.trampoline`` intrinsic takes three arguments, all
8396 pointers. The ``tramp`` argument must point to a sufficiently large and
8397 sufficiently aligned block of memory; this memory is written to by the
8398 intrinsic. Note that the size and the alignment are target-specific -
8399 LLVM currently provides no portable way of determining them, so a
8400 front-end that generates this intrinsic needs to have some
8401 target-specific knowledge. The ``func`` argument must hold a function
8402 bitcast to an ``i8*``.
8407 The block of memory pointed to by ``tramp`` is filled with target
8408 dependent code, turning it into a function. Then ``tramp`` needs to be
8409 passed to :ref:`llvm.adjust.trampoline <int_at>` to get a pointer which can
8410 be :ref:`bitcast (to a new function) and called <int_trampoline>`. The new
8411 function's signature is the same as that of ``func`` with any arguments
8412 marked with the ``nest`` attribute removed. At most one such ``nest``
8413 argument is allowed, and it must be of pointer type. Calling the new
8414 function is equivalent to calling ``func`` with the same argument list,
8415 but with ``nval`` used for the missing ``nest`` argument. If, after
8416 calling ``llvm.init.trampoline``, the memory pointed to by ``tramp`` is
8417 modified, then the effect of any later call to the returned function
8418 pointer is undefined.
8422 '``llvm.adjust.trampoline``' Intrinsic
8423 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8430 declare i8* @llvm.adjust.trampoline(i8* <tramp>)
8435 This performs any required machine-specific adjustment to the address of
8436 a trampoline (passed as ``tramp``).
8441 ``tramp`` must point to a block of memory which already has trampoline
8442 code filled in by a previous call to
8443 :ref:`llvm.init.trampoline <int_it>`.
8448 On some architectures the address of the code to be executed needs to be
8449 different to the address where the trampoline is actually stored. This
8450 intrinsic returns the executable address corresponding to ``tramp``
8451 after performing the required machine specific adjustments. The pointer
8452 returned can then be :ref:`bitcast and executed <int_trampoline>`.
8457 This class of intrinsics exists to information about the lifetime of
8458 memory objects and ranges where variables are immutable.
8460 '``llvm.lifetime.start``' Intrinsic
8461 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8468 declare void @llvm.lifetime.start(i64 <size>, i8* nocapture <ptr>)
8473 The '``llvm.lifetime.start``' intrinsic specifies the start of a memory
8479 The first argument is a constant integer representing the size of the
8480 object, or -1 if it is variable sized. The second argument is a pointer
8486 This intrinsic indicates that before this point in the code, the value
8487 of the memory pointed to by ``ptr`` is dead. This means that it is known
8488 to never be used and has an undefined value. A load from the pointer
8489 that precedes this intrinsic can be replaced with ``'undef'``.
8491 '``llvm.lifetime.end``' Intrinsic
8492 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8499 declare void @llvm.lifetime.end(i64 <size>, i8* nocapture <ptr>)
8504 The '``llvm.lifetime.end``' intrinsic specifies the end of a memory
8510 The first argument is a constant integer representing the size of the
8511 object, or -1 if it is variable sized. The second argument is a pointer
8517 This intrinsic indicates that after this point in the code, the value of
8518 the memory pointed to by ``ptr`` is dead. This means that it is known to
8519 never be used and has an undefined value. Any stores into the memory
8520 object following this intrinsic may be removed as dead.
8522 '``llvm.invariant.start``' Intrinsic
8523 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8530 declare {}* @llvm.invariant.start(i64 <size>, i8* nocapture <ptr>)
8535 The '``llvm.invariant.start``' intrinsic specifies that the contents of
8536 a memory object will not change.
8541 The first argument is a constant integer representing the size of the
8542 object, or -1 if it is variable sized. The second argument is a pointer
8548 This intrinsic indicates that until an ``llvm.invariant.end`` that uses
8549 the return value, the referenced memory location is constant and
8552 '``llvm.invariant.end``' Intrinsic
8553 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8560 declare void @llvm.invariant.end({}* <start>, i64 <size>, i8* nocapture <ptr>)
8565 The '``llvm.invariant.end``' intrinsic specifies that the contents of a
8566 memory object are mutable.
8571 The first argument is the matching ``llvm.invariant.start`` intrinsic.
8572 The second argument is a constant integer representing the size of the
8573 object, or -1 if it is variable sized and the third argument is a
8574 pointer to the object.
8579 This intrinsic indicates that the memory is mutable again.
8584 This class of intrinsics is designed to be generic and has no specific
8587 '``llvm.var.annotation``' Intrinsic
8588 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8595 declare void @llvm.var.annotation(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
8600 The '``llvm.var.annotation``' intrinsic.
8605 The first argument is a pointer to a value, the second is a pointer to a
8606 global string, the third is a pointer to a global string which is the
8607 source file name, and the last argument is the line number.
8612 This intrinsic allows annotation of local variables with arbitrary
8613 strings. This can be useful for special purpose optimizations that want
8614 to look for these annotations. These have no other defined use; they are
8615 ignored by code generation and optimization.
8617 '``llvm.ptr.annotation.*``' Intrinsic
8618 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8623 This is an overloaded intrinsic. You can use '``llvm.ptr.annotation``' on a
8624 pointer to an integer of any width. *NOTE* you must specify an address space for
8625 the pointer. The identifier for the default address space is the integer
8630 declare i8* @llvm.ptr.annotation.p<address space>i8(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
8631 declare i16* @llvm.ptr.annotation.p<address space>i16(i16* <val>, i8* <str>, i8* <str>, i32 <int>)
8632 declare i32* @llvm.ptr.annotation.p<address space>i32(i32* <val>, i8* <str>, i8* <str>, i32 <int>)
8633 declare i64* @llvm.ptr.annotation.p<address space>i64(i64* <val>, i8* <str>, i8* <str>, i32 <int>)
8634 declare i256* @llvm.ptr.annotation.p<address space>i256(i256* <val>, i8* <str>, i8* <str>, i32 <int>)
8639 The '``llvm.ptr.annotation``' intrinsic.
8644 The first argument is a pointer to an integer value of arbitrary bitwidth
8645 (result of some expression), the second is a pointer to a global string, the
8646 third is a pointer to a global string which is the source file name, and the
8647 last argument is the line number. It returns the value of the first argument.
8652 This intrinsic allows annotation of a pointer to an integer with arbitrary
8653 strings. This can be useful for special purpose optimizations that want to look
8654 for these annotations. These have no other defined use; they are ignored by code
8655 generation and optimization.
8657 '``llvm.annotation.*``' Intrinsic
8658 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8663 This is an overloaded intrinsic. You can use '``llvm.annotation``' on
8664 any integer bit width.
8668 declare i8 @llvm.annotation.i8(i8 <val>, i8* <str>, i8* <str>, i32 <int>)
8669 declare i16 @llvm.annotation.i16(i16 <val>, i8* <str>, i8* <str>, i32 <int>)
8670 declare i32 @llvm.annotation.i32(i32 <val>, i8* <str>, i8* <str>, i32 <int>)
8671 declare i64 @llvm.annotation.i64(i64 <val>, i8* <str>, i8* <str>, i32 <int>)
8672 declare i256 @llvm.annotation.i256(i256 <val>, i8* <str>, i8* <str>, i32 <int>)
8677 The '``llvm.annotation``' intrinsic.
8682 The first argument is an integer value (result of some expression), the
8683 second is a pointer to a global string, the third is a pointer to a
8684 global string which is the source file name, and the last argument is
8685 the line number. It returns the value of the first argument.
8690 This intrinsic allows annotations to be put on arbitrary expressions
8691 with arbitrary strings. This can be useful for special purpose
8692 optimizations that want to look for these annotations. These have no
8693 other defined use; they are ignored by code generation and optimization.
8695 '``llvm.trap``' Intrinsic
8696 ^^^^^^^^^^^^^^^^^^^^^^^^^
8703 declare void @llvm.trap() noreturn nounwind
8708 The '``llvm.trap``' intrinsic.
8718 This intrinsic is lowered to the target dependent trap instruction. If
8719 the target does not have a trap instruction, this intrinsic will be
8720 lowered to a call of the ``abort()`` function.
8722 '``llvm.debugtrap``' Intrinsic
8723 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8730 declare void @llvm.debugtrap() nounwind
8735 The '``llvm.debugtrap``' intrinsic.
8745 This intrinsic is lowered to code which is intended to cause an
8746 execution trap with the intention of requesting the attention of a
8749 '``llvm.stackprotector``' Intrinsic
8750 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8757 declare void @llvm.stackprotector(i8* <guard>, i8** <slot>)
8762 The ``llvm.stackprotector`` intrinsic takes the ``guard`` and stores it
8763 onto the stack at ``slot``. The stack slot is adjusted to ensure that it
8764 is placed on the stack before local variables.
8769 The ``llvm.stackprotector`` intrinsic requires two pointer arguments.
8770 The first argument is the value loaded from the stack guard
8771 ``@__stack_chk_guard``. The second variable is an ``alloca`` that has
8772 enough space to hold the value of the guard.
8777 This intrinsic causes the prologue/epilogue inserter to force the position of
8778 the ``AllocaInst`` stack slot to be before local variables on the stack. This is
8779 to ensure that if a local variable on the stack is overwritten, it will destroy
8780 the value of the guard. When the function exits, the guard on the stack is
8781 checked against the original guard by ``llvm.stackprotectorcheck``. If they are
8782 different, then ``llvm.stackprotectorcheck`` causes the program to abort by
8783 calling the ``__stack_chk_fail()`` function.
8785 '``llvm.stackprotectorcheck``' Intrinsic
8786 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8793 declare void @llvm.stackprotectorcheck(i8** <guard>)
8798 The ``llvm.stackprotectorcheck`` intrinsic compares ``guard`` against an already
8799 created stack protector and if they are not equal calls the
8800 ``__stack_chk_fail()`` function.
8805 The ``llvm.stackprotectorcheck`` intrinsic requires one pointer argument, the
8806 the variable ``@__stack_chk_guard``.
8811 This intrinsic is provided to perform the stack protector check by comparing
8812 ``guard`` with the stack slot created by ``llvm.stackprotector`` and if the
8813 values do not match call the ``__stack_chk_fail()`` function.
8815 The reason to provide this as an IR level intrinsic instead of implementing it
8816 via other IR operations is that in order to perform this operation at the IR
8817 level without an intrinsic, one would need to create additional basic blocks to
8818 handle the success/failure cases. This makes it difficult to stop the stack
8819 protector check from disrupting sibling tail calls in Codegen. With this
8820 intrinsic, we are able to generate the stack protector basic blocks late in
8821 codegen after the tail call decision has occurred.
8823 '``llvm.objectsize``' Intrinsic
8824 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8831 declare i32 @llvm.objectsize.i32(i8* <object>, i1 <min>)
8832 declare i64 @llvm.objectsize.i64(i8* <object>, i1 <min>)
8837 The ``llvm.objectsize`` intrinsic is designed to provide information to
8838 the optimizers to determine at compile time whether a) an operation
8839 (like memcpy) will overflow a buffer that corresponds to an object, or
8840 b) that a runtime check for overflow isn't necessary. An object in this
8841 context means an allocation of a specific class, structure, array, or
8847 The ``llvm.objectsize`` intrinsic takes two arguments. The first
8848 argument is a pointer to or into the ``object``. The second argument is
8849 a boolean and determines whether ``llvm.objectsize`` returns 0 (if true)
8850 or -1 (if false) when the object size is unknown. The second argument
8851 only accepts constants.
8856 The ``llvm.objectsize`` intrinsic is lowered to a constant representing
8857 the size of the object concerned. If the size cannot be determined at
8858 compile time, ``llvm.objectsize`` returns ``i32/i64 -1 or 0`` (depending
8859 on the ``min`` argument).
8861 '``llvm.expect``' Intrinsic
8862 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8869 declare i32 @llvm.expect.i32(i32 <val>, i32 <expected_val>)
8870 declare i64 @llvm.expect.i64(i64 <val>, i64 <expected_val>)
8875 The ``llvm.expect`` intrinsic provides information about expected (the
8876 most probable) value of ``val``, which can be used by optimizers.
8881 The ``llvm.expect`` intrinsic takes two arguments. The first argument is
8882 a value. The second argument is an expected value, this needs to be a
8883 constant value, variables are not allowed.
8888 This intrinsic is lowered to the ``val``.
8890 '``llvm.donothing``' Intrinsic
8891 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8898 declare void @llvm.donothing() nounwind readnone
8903 The ``llvm.donothing`` intrinsic doesn't perform any operation. It's the
8904 only intrinsic that can be called with an invoke instruction.
8914 This intrinsic does nothing, and it's removed by optimizers and ignored