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
448 instead of run-time. Global variables may optionally be initialized, may
449 have an explicit section to be placed in, and may have an optional
450 explicit alignment specified.
452 A variable may be defined as ``thread_local``, which means that it will
453 not be shared by threads (each thread will have a separated copy of the
454 variable). Not all targets support thread-local variables. Optionally, a
455 TLS model may be specified:
458 For variables that are only used within the current shared library.
460 For variables in modules that will not be loaded dynamically.
462 For variables defined in the executable and only used within it.
464 The models correspond to the ELF TLS models; see `ELF Handling For
465 Thread-Local Storage <http://people.redhat.com/drepper/tls.pdf>`_ for
466 more information on under which circumstances the different models may
467 be used. The target may choose a different TLS model if the specified
468 model is not supported, or if a better choice of model can be made.
470 A variable may be defined as a global ``constant``, which indicates that
471 the contents of the variable will **never** be modified (enabling better
472 optimization, allowing the global data to be placed in the read-only
473 section of an executable, etc). Note that variables that need runtime
474 initialization cannot be marked ``constant`` as there is a store to the
477 LLVM explicitly allows *declarations* of global variables to be marked
478 constant, even if the final definition of the global is not. This
479 capability can be used to enable slightly better optimization of the
480 program, but requires the language definition to guarantee that
481 optimizations based on the 'constantness' are valid for the translation
482 units that do not include the definition.
484 As SSA values, global variables define pointer values that are in scope
485 (i.e. they dominate) all basic blocks in the program. Global variables
486 always define a pointer to their "content" type because they describe a
487 region of memory, and all memory objects in LLVM are accessed through
490 Global variables can be marked with ``unnamed_addr`` which indicates
491 that the address is not significant, only the content. Constants marked
492 like this can be merged with other constants if they have the same
493 initializer. Note that a constant with significant address *can* be
494 merged with a ``unnamed_addr`` constant, the result being a constant
495 whose address is significant.
497 A global variable may be declared to reside in a target-specific
498 numbered address space. For targets that support them, address spaces
499 may affect how optimizations are performed and/or what target
500 instructions are used to access the variable. The default address space
501 is zero. The address space qualifier must precede any other attributes.
503 LLVM allows an explicit section to be specified for globals. If the
504 target supports it, it will emit globals to the section specified.
506 By default, global initializers are optimized by assuming that global
507 variables defined within the module are not modified from their
508 initial values before the start of the global initializer. This is
509 true even for variables potentially accessible from outside the
510 module, including those with external linkage or appearing in
511 ``@llvm.used``. This assumption may be suppressed by marking the
512 variable with ``externally_initialized``.
514 An explicit alignment may be specified for a global, which must be a
515 power of 2. If not present, or if the alignment is set to zero, the
516 alignment of the global is set by the target to whatever it feels
517 convenient. If an explicit alignment is specified, the global is forced
518 to have exactly that alignment. Targets and optimizers are not allowed
519 to over-align the global if the global has an assigned section. In this
520 case, the extra alignment could be observable: for example, code could
521 assume that the globals are densely packed in their section and try to
522 iterate over them as an array, alignment padding would break this
525 For example, the following defines a global in a numbered address space
526 with an initializer, section, and alignment:
530 @G = addrspace(5) constant float 1.0, section "foo", align 4
532 The following example defines a thread-local global with the
533 ``initialexec`` TLS model:
537 @G = thread_local(initialexec) global i32 0, align 4
539 .. _functionstructure:
544 LLVM function definitions consist of the "``define``" keyword, an
545 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
546 style <visibility>`, an optional :ref:`calling convention <callingconv>`,
547 an optional ``unnamed_addr`` attribute, a return type, an optional
548 :ref:`parameter attribute <paramattrs>` for the return type, a function
549 name, a (possibly empty) argument list (each with optional :ref:`parameter
550 attributes <paramattrs>`), optional :ref:`function attributes <fnattrs>`,
551 an optional section, an optional alignment, an optional :ref:`garbage
552 collector name <gc>`, an optional :ref:`prefix <prefixdata>`, an opening
553 curly brace, a list of basic blocks, and a closing curly brace.
555 LLVM function declarations consist of the "``declare``" keyword, an
556 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
557 style <visibility>`, an optional :ref:`calling convention <callingconv>`,
558 an optional ``unnamed_addr`` attribute, a return type, an optional
559 :ref:`parameter attribute <paramattrs>` for the return type, a function
560 name, a possibly empty list of arguments, an optional alignment, an optional
561 :ref:`garbage collector name <gc>` and an optional :ref:`prefix <prefixdata>`.
563 A function definition contains a list of basic blocks, forming the CFG
564 (Control Flow Graph) for the function. Each basic block may optionally
565 start with a label (giving the basic block a symbol table entry),
566 contains a list of instructions, and ends with a
567 :ref:`terminator <terminators>` instruction (such as a branch or function
568 return). If explicit label is not provided, a block is assigned an
569 implicit numbered label, using a next value from the same counter as used
570 for unnamed temporaries (:ref:`see above<identifiers>`). For example, if a
571 function entry block does not have explicit label, it will be assigned
572 label "%0", then first unnamed temporary in that block will be "%1", etc.
574 The first basic block in a function is special in two ways: it is
575 immediately executed on entrance to the function, and it is not allowed
576 to have predecessor basic blocks (i.e. there can not be any branches to
577 the entry block of a function). Because the block can have no
578 predecessors, it also cannot have any :ref:`PHI nodes <i_phi>`.
580 LLVM allows an explicit section to be specified for functions. If the
581 target supports it, it will emit functions to the section specified.
583 An explicit alignment may be specified for a function. If not present,
584 or if the alignment is set to zero, the alignment of the function is set
585 by the target to whatever it feels convenient. If an explicit alignment
586 is specified, the function is forced to have at least that much
587 alignment. All alignments must be a power of 2.
589 If the ``unnamed_addr`` attribute is given, the address is know to not
590 be significant and two identical functions can be merged.
594 define [linkage] [visibility]
596 <ResultType> @<FunctionName> ([argument list])
597 [fn Attrs] [section "name"] [align N]
598 [gc] [prefix Constant] { ... }
605 Aliases act as "second name" for the aliasee value (which can be either
606 function, global variable, another alias or bitcast of global value).
607 Aliases may have an optional :ref:`linkage type <linkage>`, and an optional
608 :ref:`visibility style <visibility>`.
612 @<Name> = alias [Linkage] [Visibility] <AliaseeTy> @<Aliasee>
614 The linkage must be one of ``private``, ``linker_private``,
615 ``linker_private_weak``, ``internal``, ``linkonce``, ``weak``,
616 ``linkonce_odr``, ``weak_odr``, ``linkonce_odr_auto_hide``, ``external``. Note
617 that some system linkers might not correctly handle dropping a weak symbol that
618 is aliased by a non weak alias.
620 .. _namedmetadatastructure:
625 Named metadata is a collection of metadata. :ref:`Metadata
626 nodes <metadata>` (but not metadata strings) are the only valid
627 operands for a named metadata.
631 ; Some unnamed metadata nodes, which are referenced by the named metadata.
632 !0 = metadata !{metadata !"zero"}
633 !1 = metadata !{metadata !"one"}
634 !2 = metadata !{metadata !"two"}
636 !name = !{!0, !1, !2}
643 The return type and each parameter of a function type may have a set of
644 *parameter attributes* associated with them. Parameter attributes are
645 used to communicate additional information about the result or
646 parameters of a function. Parameter attributes are considered to be part
647 of the function, not of the function type, so functions with different
648 parameter attributes can have the same function type.
650 Parameter attributes are simple keywords that follow the type specified.
651 If multiple parameter attributes are needed, they are space separated.
656 declare i32 @printf(i8* noalias nocapture, ...)
657 declare i32 @atoi(i8 zeroext)
658 declare signext i8 @returns_signed_char()
660 Note that any attributes for the function result (``nounwind``,
661 ``readonly``) come immediately after the argument list.
663 Currently, only the following parameter attributes are defined:
666 This indicates to the code generator that the parameter or return
667 value should be zero-extended to the extent required by the target's
668 ABI (which is usually 32-bits, but is 8-bits for a i1 on x86-64) by
669 the caller (for a parameter) or the callee (for a return value).
671 This indicates to the code generator that the parameter or return
672 value should be sign-extended to the extent required by the target's
673 ABI (which is usually 32-bits) by the caller (for a parameter) or
674 the callee (for a return value).
676 This indicates that this parameter or return value should be treated
677 in a special target-dependent fashion during while emitting code for
678 a function call or return (usually, by putting it in a register as
679 opposed to memory, though some targets use it to distinguish between
680 two different kinds of registers). Use of this attribute is
683 This indicates that the pointer parameter should really be passed by
684 value to the function. The attribute implies that a hidden copy of
685 the pointee is made between the caller and the callee, so the callee
686 is unable to modify the value in the caller. This attribute is only
687 valid on LLVM pointer arguments. It is generally used to pass
688 structs and arrays by value, but is also valid on pointers to
689 scalars. The copy is considered to belong to the caller not the
690 callee (for example, ``readonly`` functions should not write to
691 ``byval`` parameters). This is not a valid attribute for return
694 The byval attribute also supports specifying an alignment with the
695 align attribute. It indicates the alignment of the stack slot to
696 form and the known alignment of the pointer specified to the call
697 site. If the alignment is not specified, then the code generator
698 makes a target-specific assumption.
701 This indicates that the pointer parameter specifies the address of a
702 structure that is the return value of the function in the source
703 program. This pointer must be guaranteed by the caller to be valid:
704 loads and stores to the structure may be assumed by the callee
705 not to trap and to be properly aligned. This may only be applied to
706 the first parameter. This is not a valid attribute for return
709 This indicates that pointer values :ref:`based <pointeraliasing>` on
710 the argument or return value do not alias pointer values which are
711 not *based* on it, ignoring certain "irrelevant" dependencies. For a
712 call to the parent function, dependencies between memory references
713 from before or after the call and from those during the call are
714 "irrelevant" to the ``noalias`` keyword for the arguments and return
715 value used in that call. The caller shares the responsibility with
716 the callee for ensuring that these requirements are met. For further
717 details, please see the discussion of the NoAlias response in `alias
718 analysis <AliasAnalysis.html#MustMayNo>`_.
720 Note that this definition of ``noalias`` is intentionally similar
721 to the definition of ``restrict`` in C99 for function arguments,
722 though it is slightly weaker.
724 For function return values, C99's ``restrict`` is not meaningful,
725 while LLVM's ``noalias`` is.
727 This indicates that the callee does not make any copies of the
728 pointer that outlive the callee itself. This is not a valid
729 attribute for return values.
734 This indicates that the pointer parameter can be excised using the
735 :ref:`trampoline intrinsics <int_trampoline>`. This is not a valid
736 attribute for return values and can only be applied to one parameter.
739 This indicates that the function always returns the argument as its return
740 value. This is an optimization hint to the code generator when generating
741 the caller, allowing tail call optimization and omission of register saves
742 and restores in some cases; it is not checked or enforced when generating
743 the callee. The parameter and the function return type must be valid
744 operands for the :ref:`bitcast instruction <i_bitcast>`. This is not a
745 valid attribute for return values and can only be applied to one parameter.
749 Garbage Collector Names
750 -----------------------
752 Each function may specify a garbage collector name, which is simply a
757 define void @f() gc "name" { ... }
759 The compiler declares the supported values of *name*. Specifying a
760 collector which will cause the compiler to alter its output in order to
761 support the named garbage collection algorithm.
768 Prefix data is data associated with a function which the code generator
769 will emit immediately before the function body. The purpose of this feature
770 is to allow frontends to associate language-specific runtime metadata with
771 specific functions and make it available through the function pointer while
772 still allowing the function pointer to be called. To access the data for a
773 given function, a program may bitcast the function pointer to a pointer to
774 the constant's type. This implies that the IR symbol points to the start
777 To maintain the semantics of ordinary function calls, the prefix data must
778 have a particular format. Specifically, it must begin with a sequence of
779 bytes which decode to a sequence of machine instructions, valid for the
780 module's target, which transfer control to the point immediately succeeding
781 the prefix data, without performing any other visible action. This allows
782 the inliner and other passes to reason about the semantics of the function
783 definition without needing to reason about the prefix data. Obviously this
784 makes the format of the prefix data highly target dependent.
786 Prefix data is laid out as if it were an initializer for a global variable
787 of the prefix data's type. No padding is automatically placed between the
788 prefix data and the function body. If padding is required, it must be part
791 A trivial example of valid prefix data for the x86 architecture is ``i8 144``,
792 which encodes the ``nop`` instruction:
796 define void @f() prefix i8 144 { ... }
798 Generally prefix data can be formed by encoding a relative branch instruction
799 which skips the metadata, as in this example of valid prefix data for the
800 x86_64 architecture, where the first two bytes encode ``jmp .+10``:
804 %0 = type <{ i8, i8, i8* }>
806 define void @f() prefix %0 <{ i8 235, i8 8, i8* @md}> { ... }
808 A function may have prefix data but no body. This has similar semantics
809 to the ``available_externally`` linkage in that the data may be used by the
810 optimizers but will not be emitted in the object file.
817 Attribute groups are groups of attributes that are referenced by objects within
818 the IR. They are important for keeping ``.ll`` files readable, because a lot of
819 functions will use the same set of attributes. In the degenerative case of a
820 ``.ll`` file that corresponds to a single ``.c`` file, the single attribute
821 group will capture the important command line flags used to build that file.
823 An attribute group is a module-level object. To use an attribute group, an
824 object references the attribute group's ID (e.g. ``#37``). An object may refer
825 to more than one attribute group. In that situation, the attributes from the
826 different groups are merged.
828 Here is an example of attribute groups for a function that should always be
829 inlined, has a stack alignment of 4, and which shouldn't use SSE instructions:
833 ; Target-independent attributes:
834 attributes #0 = { alwaysinline alignstack=4 }
836 ; Target-dependent attributes:
837 attributes #1 = { "no-sse" }
839 ; Function @f has attributes: alwaysinline, alignstack=4, and "no-sse".
840 define void @f() #0 #1 { ... }
847 Function attributes are set to communicate additional information about
848 a function. Function attributes are considered to be part of the
849 function, not of the function type, so functions with different function
850 attributes can have the same function type.
852 Function attributes are simple keywords that follow the type specified.
853 If multiple attributes are needed, they are space separated. For
858 define void @f() noinline { ... }
859 define void @f() alwaysinline { ... }
860 define void @f() alwaysinline optsize { ... }
861 define void @f() optsize { ... }
864 This attribute indicates that, when emitting the prologue and
865 epilogue, the backend should forcibly align the stack pointer.
866 Specify the desired alignment, which must be a power of two, in
869 This attribute indicates that the inliner should attempt to inline
870 this function into callers whenever possible, ignoring any active
871 inlining size threshold for this caller.
873 This indicates that the callee function at a call site should be
874 recognized as a built-in function, even though the function's declaration
875 uses the ``nobuiltin`` attribute. This is only valid at call sites for
876 direct calls to functions which are declared with the ``nobuiltin``
879 This attribute indicates that this function is rarely called. When
880 computing edge weights, basic blocks post-dominated by a cold
881 function call are also considered to be cold; and, thus, given low
884 This attribute indicates that the source code contained a hint that
885 inlining this function is desirable (such as the "inline" keyword in
886 C/C++). It is just a hint; it imposes no requirements on the
889 This attribute suggests that optimization passes and code generator
890 passes make choices that keep the code size of this function as small
891 as possible and perform optimizations that may sacrifice runtime
892 performance in order to minimize the size of the generated code.
894 This attribute disables prologue / epilogue emission for the
895 function. This can have very system-specific consequences.
897 This indicates that the callee function at a call site is not recognized as
898 a built-in function. LLVM will retain the original call and not replace it
899 with equivalent code based on the semantics of the built-in function, unless
900 the call site uses the ``builtin`` attribute. This is valid at call sites
901 and on function declarations and definitions.
903 This attribute indicates that calls to the function cannot be
904 duplicated. A call to a ``noduplicate`` function may be moved
905 within its parent function, but may not be duplicated within
908 A function containing a ``noduplicate`` call may still
909 be an inlining candidate, provided that the call is not
910 duplicated by inlining. That implies that the function has
911 internal linkage and only has one call site, so the original
912 call is dead after inlining.
914 This attributes disables implicit floating point instructions.
916 This attribute indicates that the inliner should never inline this
917 function in any situation. This attribute may not be used together
918 with the ``alwaysinline`` attribute.
920 This attribute suppresses lazy symbol binding for the function. This
921 may make calls to the function faster, at the cost of extra program
922 startup time if the function is not called during program startup.
924 This attribute indicates that the code generator should not use a
925 red zone, even if the target-specific ABI normally permits it.
927 This function attribute indicates that the function never returns
928 normally. This produces undefined behavior at runtime if the
929 function ever does dynamically return.
931 This function attribute indicates that the function never returns
932 with an unwind or exceptional control flow. If the function does
933 unwind, its runtime behavior is undefined.
935 This function attribute indicates that the function is not optimized
936 by any optimization or code generator passes with the
937 exception of interprocedural optimization passes.
938 This attribute cannot be used together with the ``alwaysinline``
939 attribute; this attribute is also incompatible
940 with the ``minsize`` attribute and the ``optsize`` attribute.
942 The inliner should never inline this function in any situation.
943 Only functions with the ``alwaysinline`` attribute are valid
944 candidates for inlining inside the body of this function.
946 This attribute suggests that optimization passes and code generator
947 passes make choices that keep the code size of this function low,
948 and otherwise do optimizations specifically to reduce code size as
949 long as they do not significantly impact runtime performance.
951 On a function, this attribute indicates that the function computes its
952 result (or decides to unwind an exception) based strictly on its arguments,
953 without dereferencing any pointer arguments or otherwise accessing
954 any mutable state (e.g. memory, control registers, etc) visible to
955 caller functions. It does not write through any pointer arguments
956 (including ``byval`` arguments) and never changes any state visible
957 to callers. This means that it cannot unwind exceptions by calling
958 the ``C++`` exception throwing methods.
960 On an argument, this attribute indicates that the function does not
961 dereference that pointer argument, even though it may read or write the
962 memory that the pointer points to if accessed through other pointers.
964 On a function, this attribute indicates that the function does not write
965 through any pointer arguments (including ``byval`` arguments) or otherwise
966 modify any state (e.g. memory, control registers, etc) visible to
967 caller functions. It may dereference pointer arguments and read
968 state that may be set in the caller. A readonly function always
969 returns the same value (or unwinds an exception identically) when
970 called with the same set of arguments and global state. It cannot
971 unwind an exception by calling the ``C++`` exception throwing
974 On an argument, this attribute indicates that the function does not write
975 through this pointer argument, even though it may write to the memory that
976 the pointer points to.
978 This attribute indicates that this function can return twice. The C
979 ``setjmp`` is an example of such a function. The compiler disables
980 some optimizations (like tail calls) in the caller of these
983 This attribute indicates that AddressSanitizer checks
984 (dynamic address safety analysis) are enabled for this function.
986 This attribute indicates that MemorySanitizer checks (dynamic detection
987 of accesses to uninitialized memory) are enabled for this function.
989 This attribute indicates that ThreadSanitizer checks
990 (dynamic thread safety analysis) are enabled for this function.
992 This attribute indicates that the function should emit a stack
993 smashing protector. It is in the form of a "canary" --- a random value
994 placed on the stack before the local variables that's checked upon
995 return from the function to see if it has been overwritten. A
996 heuristic is used to determine if a function needs stack protectors
997 or not. The heuristic used will enable protectors for functions with:
999 - Character arrays larger than ``ssp-buffer-size`` (default 8).
1000 - Aggregates containing character arrays larger than ``ssp-buffer-size``.
1001 - Calls to alloca() with variable sizes or constant sizes greater than
1002 ``ssp-buffer-size``.
1004 If a function that has an ``ssp`` attribute is inlined into a
1005 function that doesn't have an ``ssp`` attribute, then the resulting
1006 function will have an ``ssp`` attribute.
1008 This attribute indicates that the function should *always* emit a
1009 stack smashing protector. This overrides the ``ssp`` function
1012 If a function that has an ``sspreq`` attribute is inlined into a
1013 function that doesn't have an ``sspreq`` attribute or which has an
1014 ``ssp`` or ``sspstrong`` attribute, then the resulting function will have
1015 an ``sspreq`` attribute.
1017 This attribute indicates that the function should emit a stack smashing
1018 protector. This attribute causes a strong heuristic to be used when
1019 determining if a function needs stack protectors. The strong heuristic
1020 will enable protectors for functions with:
1022 - Arrays of any size and type
1023 - Aggregates containing an array of any size and type.
1024 - Calls to alloca().
1025 - Local variables that have had their address taken.
1027 This overrides the ``ssp`` function attribute.
1029 If a function that has an ``sspstrong`` attribute is inlined into a
1030 function that doesn't have an ``sspstrong`` attribute, then the
1031 resulting function will have an ``sspstrong`` attribute.
1033 This attribute indicates that the ABI being targeted requires that
1034 an unwind table entry be produce for this function even if we can
1035 show that no exceptions passes by it. This is normally the case for
1036 the ELF x86-64 abi, but it can be disabled for some compilation
1041 Module-Level Inline Assembly
1042 ----------------------------
1044 Modules may contain "module-level inline asm" blocks, which corresponds
1045 to the GCC "file scope inline asm" blocks. These blocks are internally
1046 concatenated by LLVM and treated as a single unit, but may be separated
1047 in the ``.ll`` file if desired. The syntax is very simple:
1049 .. code-block:: llvm
1051 module asm "inline asm code goes here"
1052 module asm "more can go here"
1054 The strings can contain any character by escaping non-printable
1055 characters. The escape sequence used is simply "\\xx" where "xx" is the
1056 two digit hex code for the number.
1058 The inline asm code is simply printed to the machine code .s file when
1059 assembly code is generated.
1061 .. _langref_datalayout:
1066 A module may specify a target specific data layout string that specifies
1067 how data is to be laid out in memory. The syntax for the data layout is
1070 .. code-block:: llvm
1072 target datalayout = "layout specification"
1074 The *layout specification* consists of a list of specifications
1075 separated by the minus sign character ('-'). Each specification starts
1076 with a letter and may include other information after the letter to
1077 define some aspect of the data layout. The specifications accepted are
1081 Specifies that the target lays out data in big-endian form. That is,
1082 the bits with the most significance have the lowest address
1085 Specifies that the target lays out data in little-endian form. That
1086 is, the bits with the least significance have the lowest address
1089 Specifies the natural alignment of the stack in bits. Alignment
1090 promotion of stack variables is limited to the natural stack
1091 alignment to avoid dynamic stack realignment. The stack alignment
1092 must be a multiple of 8-bits. If omitted, the natural stack
1093 alignment defaults to "unspecified", which does not prevent any
1094 alignment promotions.
1095 ``p[n]:<size>:<abi>:<pref>``
1096 This specifies the *size* of a pointer and its ``<abi>`` and
1097 ``<pref>``\erred alignments for address space ``n``. All sizes are in
1098 bits. Specifying the ``<pref>`` alignment is optional. If omitted, the
1099 preceding ``:`` should be omitted too. The address space, ``n`` is
1100 optional, and if not specified, denotes the default address space 0.
1101 The value of ``n`` must be in the range [1,2^23).
1102 ``i<size>:<abi>:<pref>``
1103 This specifies the alignment for an integer type of a given bit
1104 ``<size>``. The value of ``<size>`` must be in the range [1,2^23).
1105 ``v<size>:<abi>:<pref>``
1106 This specifies the alignment for a vector type of a given bit
1108 ``f<size>:<abi>:<pref>``
1109 This specifies the alignment for a floating point type of a given bit
1110 ``<size>``. Only values of ``<size>`` that are supported by the target
1111 will work. 32 (float) and 64 (double) are supported on all targets; 80
1112 or 128 (different flavors of long double) are also supported on some
1114 ``a<size>:<abi>:<pref>``
1115 This specifies the alignment for an aggregate type of a given bit
1117 ``s<size>:<abi>:<pref>``
1118 This specifies the alignment for a stack object of a given bit
1120 ``n<size1>:<size2>:<size3>...``
1121 This specifies a set of native integer widths for the target CPU in
1122 bits. For example, it might contain ``n32`` for 32-bit PowerPC,
1123 ``n32:64`` for PowerPC 64, or ``n8:16:32:64`` for X86-64. Elements of
1124 this set are considered to support most general arithmetic operations
1127 When constructing the data layout for a given target, LLVM starts with a
1128 default set of specifications which are then (possibly) overridden by
1129 the specifications in the ``datalayout`` keyword. The default
1130 specifications are given in this list:
1132 - ``E`` - big endian
1133 - ``p:64:64:64`` - 64-bit pointers with 64-bit alignment.
1134 - ``p[n]:64:64:64`` - Other address spaces are assumed to be the
1135 same as the default address space.
1136 - ``S0`` - natural stack alignment is unspecified
1137 - ``i1:8:8`` - i1 is 8-bit (byte) aligned
1138 - ``i8:8:8`` - i8 is 8-bit (byte) aligned
1139 - ``i16:16:16`` - i16 is 16-bit aligned
1140 - ``i32:32:32`` - i32 is 32-bit aligned
1141 - ``i64:32:64`` - i64 has ABI alignment of 32-bits but preferred
1142 alignment of 64-bits
1143 - ``f16:16:16`` - half is 16-bit aligned
1144 - ``f32:32:32`` - float is 32-bit aligned
1145 - ``f64:64:64`` - double is 64-bit aligned
1146 - ``f128:128:128`` - quad is 128-bit aligned
1147 - ``v64:64:64`` - 64-bit vector is 64-bit aligned
1148 - ``v128:128:128`` - 128-bit vector is 128-bit aligned
1149 - ``a0:0:64`` - aggregates are 64-bit aligned
1151 When LLVM is determining the alignment for a given type, it uses the
1154 #. If the type sought is an exact match for one of the specifications,
1155 that specification is used.
1156 #. If no match is found, and the type sought is an integer type, then
1157 the smallest integer type that is larger than the bitwidth of the
1158 sought type is used. If none of the specifications are larger than
1159 the bitwidth then the largest integer type is used. For example,
1160 given the default specifications above, the i7 type will use the
1161 alignment of i8 (next largest) while both i65 and i256 will use the
1162 alignment of i64 (largest specified).
1163 #. If no match is found, and the type sought is a vector type, then the
1164 largest vector type that is smaller than the sought vector type will
1165 be used as a fall back. This happens because <128 x double> can be
1166 implemented in terms of 64 <2 x double>, for example.
1168 The function of the data layout string may not be what you expect.
1169 Notably, this is not a specification from the frontend of what alignment
1170 the code generator should use.
1172 Instead, if specified, the target data layout is required to match what
1173 the ultimate *code generator* expects. This string is used by the
1174 mid-level optimizers to improve code, and this only works if it matches
1175 what the ultimate code generator uses. If you would like to generate IR
1176 that does not embed this target-specific detail into the IR, then you
1177 don't have to specify the string. This will disable some optimizations
1178 that require precise layout information, but this also prevents those
1179 optimizations from introducing target specificity into the IR.
1181 .. _pointeraliasing:
1183 Pointer Aliasing Rules
1184 ----------------------
1186 Any memory access must be done through a pointer value associated with
1187 an address range of the memory access, otherwise the behavior is
1188 undefined. Pointer values are associated with address ranges according
1189 to the following rules:
1191 - A pointer value is associated with the addresses associated with any
1192 value it is *based* on.
1193 - An address of a global variable is associated with the address range
1194 of the variable's storage.
1195 - The result value of an allocation instruction is associated with the
1196 address range of the allocated storage.
1197 - A null pointer in the default address-space is associated with no
1199 - An integer constant other than zero or a pointer value returned from
1200 a function not defined within LLVM may be associated with address
1201 ranges allocated through mechanisms other than those provided by
1202 LLVM. Such ranges shall not overlap with any ranges of addresses
1203 allocated by mechanisms provided by LLVM.
1205 A pointer value is *based* on another pointer value according to the
1208 - A pointer value formed from a ``getelementptr`` operation is *based*
1209 on the first operand of the ``getelementptr``.
1210 - The result value of a ``bitcast`` is *based* on the operand of the
1212 - A pointer value formed by an ``inttoptr`` is *based* on all pointer
1213 values that contribute (directly or indirectly) to the computation of
1214 the pointer's value.
1215 - The "*based* on" relationship is transitive.
1217 Note that this definition of *"based"* is intentionally similar to the
1218 definition of *"based"* in C99, though it is slightly weaker.
1220 LLVM IR does not associate types with memory. The result type of a
1221 ``load`` merely indicates the size and alignment of the memory from
1222 which to load, as well as the interpretation of the value. The first
1223 operand type of a ``store`` similarly only indicates the size and
1224 alignment of the store.
1226 Consequently, type-based alias analysis, aka TBAA, aka
1227 ``-fstrict-aliasing``, is not applicable to general unadorned LLVM IR.
1228 :ref:`Metadata <metadata>` may be used to encode additional information
1229 which specialized optimization passes may use to implement type-based
1234 Volatile Memory Accesses
1235 ------------------------
1237 Certain memory accesses, such as :ref:`load <i_load>`'s,
1238 :ref:`store <i_store>`'s, and :ref:`llvm.memcpy <int_memcpy>`'s may be
1239 marked ``volatile``. The optimizers must not change the number of
1240 volatile operations or change their order of execution relative to other
1241 volatile operations. The optimizers *may* change the order of volatile
1242 operations relative to non-volatile operations. This is not Java's
1243 "volatile" and has no cross-thread synchronization behavior.
1245 IR-level volatile loads and stores cannot safely be optimized into
1246 llvm.memcpy or llvm.memmove intrinsics even when those intrinsics are
1247 flagged volatile. Likewise, the backend should never split or merge
1248 target-legal volatile load/store instructions.
1250 .. admonition:: Rationale
1252 Platforms may rely on volatile loads and stores of natively supported
1253 data width to be executed as single instruction. For example, in C
1254 this holds for an l-value of volatile primitive type with native
1255 hardware support, but not necessarily for aggregate types. The
1256 frontend upholds these expectations, which are intentionally
1257 unspecified in the IR. The rules above ensure that IR transformation
1258 do not violate the frontend's contract with the language.
1262 Memory Model for Concurrent Operations
1263 --------------------------------------
1265 The LLVM IR does not define any way to start parallel threads of
1266 execution or to register signal handlers. Nonetheless, there are
1267 platform-specific ways to create them, and we define LLVM IR's behavior
1268 in their presence. This model is inspired by the C++0x memory model.
1270 For a more informal introduction to this model, see the :doc:`Atomics`.
1272 We define a *happens-before* partial order as the least partial order
1275 - Is a superset of single-thread program order, and
1276 - When a *synchronizes-with* ``b``, includes an edge from ``a`` to
1277 ``b``. *Synchronizes-with* pairs are introduced by platform-specific
1278 techniques, like pthread locks, thread creation, thread joining,
1279 etc., and by atomic instructions. (See also :ref:`Atomic Memory Ordering
1280 Constraints <ordering>`).
1282 Note that program order does not introduce *happens-before* edges
1283 between a thread and signals executing inside that thread.
1285 Every (defined) read operation (load instructions, memcpy, atomic
1286 loads/read-modify-writes, etc.) R reads a series of bytes written by
1287 (defined) write operations (store instructions, atomic
1288 stores/read-modify-writes, memcpy, etc.). For the purposes of this
1289 section, initialized globals are considered to have a write of the
1290 initializer which is atomic and happens before any other read or write
1291 of the memory in question. For each byte of a read R, R\ :sub:`byte`
1292 may see any write to the same byte, except:
1294 - If write\ :sub:`1` happens before write\ :sub:`2`, and
1295 write\ :sub:`2` happens before R\ :sub:`byte`, then
1296 R\ :sub:`byte` does not see write\ :sub:`1`.
1297 - If R\ :sub:`byte` happens before write\ :sub:`3`, then
1298 R\ :sub:`byte` does not see write\ :sub:`3`.
1300 Given that definition, R\ :sub:`byte` is defined as follows:
1302 - If R is volatile, the result is target-dependent. (Volatile is
1303 supposed to give guarantees which can support ``sig_atomic_t`` in
1304 C/C++, and may be used for accesses to addresses which do not behave
1305 like normal memory. It does not generally provide cross-thread
1307 - Otherwise, if there is no write to the same byte that happens before
1308 R\ :sub:`byte`, R\ :sub:`byte` returns ``undef`` for that byte.
1309 - Otherwise, if R\ :sub:`byte` may see exactly one write,
1310 R\ :sub:`byte` returns the value written by that write.
1311 - Otherwise, if R is atomic, and all the writes R\ :sub:`byte` may
1312 see are atomic, it chooses one of the values written. See the :ref:`Atomic
1313 Memory Ordering Constraints <ordering>` section for additional
1314 constraints on how the choice is made.
1315 - Otherwise R\ :sub:`byte` returns ``undef``.
1317 R returns the value composed of the series of bytes it read. This
1318 implies that some bytes within the value may be ``undef`` **without**
1319 the entire value being ``undef``. Note that this only defines the
1320 semantics of the operation; it doesn't mean that targets will emit more
1321 than one instruction to read the series of bytes.
1323 Note that in cases where none of the atomic intrinsics are used, this
1324 model places only one restriction on IR transformations on top of what
1325 is required for single-threaded execution: introducing a store to a byte
1326 which might not otherwise be stored is not allowed in general.
1327 (Specifically, in the case where another thread might write to and read
1328 from an address, introducing a store can change a load that may see
1329 exactly one write into a load that may see multiple writes.)
1333 Atomic Memory Ordering Constraints
1334 ----------------------------------
1336 Atomic instructions (:ref:`cmpxchg <i_cmpxchg>`,
1337 :ref:`atomicrmw <i_atomicrmw>`, :ref:`fence <i_fence>`,
1338 :ref:`atomic load <i_load>`, and :ref:`atomic store <i_store>`) take
1339 an ordering parameter that determines which other atomic instructions on
1340 the same address they *synchronize with*. These semantics are borrowed
1341 from Java and C++0x, but are somewhat more colloquial. If these
1342 descriptions aren't precise enough, check those specs (see spec
1343 references in the :doc:`atomics guide <Atomics>`).
1344 :ref:`fence <i_fence>` instructions treat these orderings somewhat
1345 differently since they don't take an address. See that instruction's
1346 documentation for details.
1348 For a simpler introduction to the ordering constraints, see the
1352 The set of values that can be read is governed by the happens-before
1353 partial order. A value cannot be read unless some operation wrote
1354 it. This is intended to provide a guarantee strong enough to model
1355 Java's non-volatile shared variables. This ordering cannot be
1356 specified for read-modify-write operations; it is not strong enough
1357 to make them atomic in any interesting way.
1359 In addition to the guarantees of ``unordered``, there is a single
1360 total order for modifications by ``monotonic`` operations on each
1361 address. All modification orders must be compatible with the
1362 happens-before order. There is no guarantee that the modification
1363 orders can be combined to a global total order for the whole program
1364 (and this often will not be possible). The read in an atomic
1365 read-modify-write operation (:ref:`cmpxchg <i_cmpxchg>` and
1366 :ref:`atomicrmw <i_atomicrmw>`) reads the value in the modification
1367 order immediately before the value it writes. If one atomic read
1368 happens before another atomic read of the same address, the later
1369 read must see the same value or a later value in the address's
1370 modification order. This disallows reordering of ``monotonic`` (or
1371 stronger) operations on the same address. If an address is written
1372 ``monotonic``-ally by one thread, and other threads ``monotonic``-ally
1373 read that address repeatedly, the other threads must eventually see
1374 the write. This corresponds to the C++0x/C1x
1375 ``memory_order_relaxed``.
1377 In addition to the guarantees of ``monotonic``, a
1378 *synchronizes-with* edge may be formed with a ``release`` operation.
1379 This is intended to model C++'s ``memory_order_acquire``.
1381 In addition to the guarantees of ``monotonic``, if this operation
1382 writes a value which is subsequently read by an ``acquire``
1383 operation, it *synchronizes-with* that operation. (This isn't a
1384 complete description; see the C++0x definition of a release
1385 sequence.) This corresponds to the C++0x/C1x
1386 ``memory_order_release``.
1387 ``acq_rel`` (acquire+release)
1388 Acts as both an ``acquire`` and ``release`` operation on its
1389 address. This corresponds to the C++0x/C1x ``memory_order_acq_rel``.
1390 ``seq_cst`` (sequentially consistent)
1391 In addition to the guarantees of ``acq_rel`` (``acquire`` for an
1392 operation which only reads, ``release`` for an operation which only
1393 writes), there is a global total order on all
1394 sequentially-consistent operations on all addresses, which is
1395 consistent with the *happens-before* partial order and with the
1396 modification orders of all the affected addresses. Each
1397 sequentially-consistent read sees the last preceding write to the
1398 same address in this global order. This corresponds to the C++0x/C1x
1399 ``memory_order_seq_cst`` and Java volatile.
1403 If an atomic operation is marked ``singlethread``, it only *synchronizes
1404 with* or participates in modification and seq\_cst total orderings with
1405 other operations running in the same thread (for example, in signal
1413 LLVM IR floating-point binary ops (:ref:`fadd <i_fadd>`,
1414 :ref:`fsub <i_fsub>`, :ref:`fmul <i_fmul>`, :ref:`fdiv <i_fdiv>`,
1415 :ref:`frem <i_frem>`) have the following flags that can set to enable
1416 otherwise unsafe floating point operations
1419 No NaNs - Allow optimizations to assume the arguments and result are not
1420 NaN. Such optimizations are required to retain defined behavior over
1421 NaNs, but the value of the result is undefined.
1424 No Infs - Allow optimizations to assume the arguments and result are not
1425 +/-Inf. Such optimizations are required to retain defined behavior over
1426 +/-Inf, but the value of the result is undefined.
1429 No Signed Zeros - Allow optimizations to treat the sign of a zero
1430 argument or result as insignificant.
1433 Allow Reciprocal - Allow optimizations to use the reciprocal of an
1434 argument rather than perform division.
1437 Fast - Allow algebraically equivalent transformations that may
1438 dramatically change results in floating point (e.g. reassociate). This
1439 flag implies all the others.
1446 The LLVM type system is one of the most important features of the
1447 intermediate representation. Being typed enables a number of
1448 optimizations to be performed on the intermediate representation
1449 directly, without having to do extra analyses on the side before the
1450 transformation. A strong type system makes it easier to read the
1451 generated code and enables novel analyses and transformations that are
1452 not feasible to perform on normal three address code representations.
1454 .. _typeclassifications:
1456 Type Classifications
1457 --------------------
1459 The types fall into a few useful classifications:
1468 * - :ref:`integer <t_integer>`
1469 - ``i1``, ``i2``, ``i3``, ... ``i8``, ... ``i16``, ... ``i32``, ...
1472 * - :ref:`floating point <t_floating>`
1473 - ``half``, ``float``, ``double``, ``x86_fp80``, ``fp128``,
1481 - :ref:`integer <t_integer>`, :ref:`floating point <t_floating>`,
1482 :ref:`pointer <t_pointer>`, :ref:`vector <t_vector>`,
1483 :ref:`structure <t_struct>`, :ref:`array <t_array>`,
1484 :ref:`label <t_label>`, :ref:`metadata <t_metadata>`.
1486 * - :ref:`primitive <t_primitive>`
1487 - :ref:`label <t_label>`,
1488 :ref:`void <t_void>`,
1489 :ref:`integer <t_integer>`,
1490 :ref:`floating point <t_floating>`,
1491 :ref:`x86mmx <t_x86mmx>`,
1492 :ref:`metadata <t_metadata>`.
1494 * - :ref:`derived <t_derived>`
1495 - :ref:`array <t_array>`,
1496 :ref:`function <t_function>`,
1497 :ref:`pointer <t_pointer>`,
1498 :ref:`structure <t_struct>`,
1499 :ref:`vector <t_vector>`,
1500 :ref:`opaque <t_opaque>`.
1502 The :ref:`first class <t_firstclass>` types are perhaps the most important.
1503 Values of these types are the only ones which can be produced by
1511 The primitive types are the fundamental building blocks of the LLVM
1522 The integer type is a very simple type that simply specifies an
1523 arbitrary bit width for the integer type desired. Any bit width from 1
1524 bit to 2\ :sup:`23`\ -1 (about 8 million) can be specified.
1533 The number of bits the integer will occupy is specified by the ``N``
1539 +----------------+------------------------------------------------+
1540 | ``i1`` | a single-bit integer. |
1541 +----------------+------------------------------------------------+
1542 | ``i32`` | a 32-bit integer. |
1543 +----------------+------------------------------------------------+
1544 | ``i1942652`` | a really big integer of over 1 million bits. |
1545 +----------------+------------------------------------------------+
1549 Floating Point Types
1550 ^^^^^^^^^^^^^^^^^^^^
1559 - 16-bit floating point value
1562 - 32-bit floating point value
1565 - 64-bit floating point value
1568 - 128-bit floating point value (112-bit mantissa)
1571 - 80-bit floating point value (X87)
1574 - 128-bit floating point value (two 64-bits)
1584 The x86mmx type represents a value held in an MMX register on an x86
1585 machine. The operations allowed on it are quite limited: parameters and
1586 return values, load and store, and bitcast. User-specified MMX
1587 instructions are represented as intrinsic or asm calls with arguments
1588 and/or results of this type. There are no arrays, vectors or constants
1606 The void type does not represent any value and has no size.
1623 The label type represents code labels.
1640 The metadata type represents embedded metadata. No derived types may be
1641 created from metadata except for :ref:`function <t_function>` arguments.
1655 The real power in LLVM comes from the derived types in the system. This
1656 is what allows a programmer to represent arrays, functions, pointers,
1657 and other useful types. Each of these types contain one or more element
1658 types which may be a primitive type, or another derived type. For
1659 example, it is possible to have a two dimensional array, using an array
1660 as the element type of another array.
1667 Aggregate Types are a subset of derived types that can contain multiple
1668 member types. :ref:`Arrays <t_array>` and :ref:`structs <t_struct>` are
1669 aggregate types. :ref:`Vectors <t_vector>` are not considered to be
1680 The array type is a very simple derived type that arranges elements
1681 sequentially in memory. The array type requires a size (number of
1682 elements) and an underlying data type.
1689 [<# elements> x <elementtype>]
1691 The number of elements is a constant integer value; ``elementtype`` may
1692 be any type with a size.
1697 +------------------+--------------------------------------+
1698 | ``[40 x i32]`` | Array of 40 32-bit integer values. |
1699 +------------------+--------------------------------------+
1700 | ``[41 x i32]`` | Array of 41 32-bit integer values. |
1701 +------------------+--------------------------------------+
1702 | ``[4 x i8]`` | Array of 4 8-bit integer values. |
1703 +------------------+--------------------------------------+
1705 Here are some examples of multidimensional arrays:
1707 +-----------------------------+----------------------------------------------------------+
1708 | ``[3 x [4 x i32]]`` | 3x4 array of 32-bit integer values. |
1709 +-----------------------------+----------------------------------------------------------+
1710 | ``[12 x [10 x float]]`` | 12x10 array of single precision floating point values. |
1711 +-----------------------------+----------------------------------------------------------+
1712 | ``[2 x [3 x [4 x i16]]]`` | 2x3x4 array of 16-bit integer values. |
1713 +-----------------------------+----------------------------------------------------------+
1715 There is no restriction on indexing beyond the end of the array implied
1716 by a static type (though there are restrictions on indexing beyond the
1717 bounds of an allocated object in some cases). This means that
1718 single-dimension 'variable sized array' addressing can be implemented in
1719 LLVM with a zero length array type. An implementation of 'pascal style
1720 arrays' in LLVM could use the type "``{ i32, [0 x float]}``", for
1731 The function type can be thought of as a function signature. It consists
1732 of a return type and a list of formal parameter types. The return type
1733 of a function type is a first class type or a void type.
1740 <returntype> (<parameter list>)
1742 ...where '``<parameter list>``' is a comma-separated list of type
1743 specifiers. Optionally, the parameter list may include a type ``...``,
1744 which indicates that the function takes a variable number of arguments.
1745 Variable argument functions can access their arguments with the
1746 :ref:`variable argument handling intrinsic <int_varargs>` functions.
1747 '``<returntype>``' is any type except :ref:`label <t_label>`.
1752 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1753 | ``i32 (i32)`` | function taking an ``i32``, returning an ``i32`` |
1754 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1755 | ``float (i16, i32 *) *`` | :ref:`Pointer <t_pointer>` to a function that takes an ``i16`` and a :ref:`pointer <t_pointer>` to ``i32``, returning ``float``. |
1756 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1757 | ``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. |
1758 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1759 | ``{i32, i32} (i32)`` | A function taking an ``i32``, returning a :ref:`structure <t_struct>` containing two ``i32`` values |
1760 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1770 The structure type is used to represent a collection of data members
1771 together in memory. The elements of a structure may be any type that has
1774 Structures in memory are accessed using '``load``' and '``store``' by
1775 getting a pointer to a field with the '``getelementptr``' instruction.
1776 Structures in registers are accessed using the '``extractvalue``' and
1777 '``insertvalue``' instructions.
1779 Structures may optionally be "packed" structures, which indicate that
1780 the alignment of the struct is one byte, and that there is no padding
1781 between the elements. In non-packed structs, padding between field types
1782 is inserted as defined by the DataLayout string in the module, which is
1783 required to match what the underlying code generator expects.
1785 Structures can either be "literal" or "identified". A literal structure
1786 is defined inline with other types (e.g. ``{i32, i32}*``) whereas
1787 identified types are always defined at the top level with a name.
1788 Literal types are uniqued by their contents and can never be recursive
1789 or opaque since there is no way to write one. Identified types can be
1790 recursive, can be opaqued, and are never uniqued.
1797 %T1 = type { <type list> } ; Identified normal struct type
1798 %T2 = type <{ <type list> }> ; Identified packed struct type
1803 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1804 | ``{ i32, i32, i32 }`` | A triple of three ``i32`` values |
1805 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1806 | ``{ 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``. |
1807 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1808 | ``<{ i8, i32 }>`` | A packed struct known to be 5 bytes in size. |
1809 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1813 Opaque Structure Types
1814 ^^^^^^^^^^^^^^^^^^^^^^
1819 Opaque structure types are used to represent named structure types that
1820 do not have a body specified. This corresponds (for example) to the C
1821 notion of a forward declared structure.
1834 +--------------+-------------------+
1835 | ``opaque`` | An opaque type. |
1836 +--------------+-------------------+
1846 The pointer type is used to specify memory locations. Pointers are
1847 commonly used to reference objects in memory.
1849 Pointer types may have an optional address space attribute defining the
1850 numbered address space where the pointed-to object resides. The default
1851 address space is number zero. The semantics of non-zero address spaces
1852 are target-specific.
1854 Note that LLVM does not permit pointers to void (``void*``) nor does it
1855 permit pointers to labels (``label*``). Use ``i8*`` instead.
1867 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1868 | ``[4 x i32]*`` | A :ref:`pointer <t_pointer>` to :ref:`array <t_array>` of four ``i32`` values. |
1869 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1870 | ``i32 (i32*) *`` | A :ref:`pointer <t_pointer>` to a :ref:`function <t_function>` that takes an ``i32*``, returning an ``i32``. |
1871 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1872 | ``i32 addrspace(5)*`` | A :ref:`pointer <t_pointer>` to an ``i32`` value that resides in address space #5. |
1873 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1883 A vector type is a simple derived type that represents a vector of
1884 elements. Vector types are used when multiple primitive data are
1885 operated in parallel using a single instruction (SIMD). A vector type
1886 requires a size (number of elements) and an underlying primitive data
1887 type. Vector types are considered :ref:`first class <t_firstclass>`.
1894 < <# elements> x <elementtype> >
1896 The number of elements is a constant integer value larger than 0;
1897 elementtype may be any integer or floating point type, or a pointer to
1898 these types. Vectors of size zero are not allowed.
1903 +-------------------+--------------------------------------------------+
1904 | ``<4 x i32>`` | Vector of 4 32-bit integer values. |
1905 +-------------------+--------------------------------------------------+
1906 | ``<8 x float>`` | Vector of 8 32-bit floating-point values. |
1907 +-------------------+--------------------------------------------------+
1908 | ``<2 x i64>`` | Vector of 2 64-bit integer values. |
1909 +-------------------+--------------------------------------------------+
1910 | ``<4 x i64*>`` | Vector of 4 pointers to 64-bit integer values. |
1911 +-------------------+--------------------------------------------------+
1916 LLVM has several different basic types of constants. This section
1917 describes them all and their syntax.
1922 **Boolean constants**
1923 The two strings '``true``' and '``false``' are both valid constants
1925 **Integer constants**
1926 Standard integers (such as '4') are constants of the
1927 :ref:`integer <t_integer>` type. Negative numbers may be used with
1929 **Floating point constants**
1930 Floating point constants use standard decimal notation (e.g.
1931 123.421), exponential notation (e.g. 1.23421e+2), or a more precise
1932 hexadecimal notation (see below). The assembler requires the exact
1933 decimal value of a floating-point constant. For example, the
1934 assembler accepts 1.25 but rejects 1.3 because 1.3 is a repeating
1935 decimal in binary. Floating point constants must have a :ref:`floating
1936 point <t_floating>` type.
1937 **Null pointer constants**
1938 The identifier '``null``' is recognized as a null pointer constant
1939 and must be of :ref:`pointer type <t_pointer>`.
1941 The one non-intuitive notation for constants is the hexadecimal form of
1942 floating point constants. For example, the form
1943 '``double 0x432ff973cafa8000``' is equivalent to (but harder to read
1944 than) '``double 4.5e+15``'. The only time hexadecimal floating point
1945 constants are required (and the only time that they are generated by the
1946 disassembler) is when a floating point constant must be emitted but it
1947 cannot be represented as a decimal floating point number in a reasonable
1948 number of digits. For example, NaN's, infinities, and other special
1949 values are represented in their IEEE hexadecimal format so that assembly
1950 and disassembly do not cause any bits to change in the constants.
1952 When using the hexadecimal form, constants of types half, float, and
1953 double are represented using the 16-digit form shown above (which
1954 matches the IEEE754 representation for double); half and float values
1955 must, however, be exactly representable as IEEE 754 half and single
1956 precision, respectively. Hexadecimal format is always used for long
1957 double, and there are three forms of long double. The 80-bit format used
1958 by x86 is represented as ``0xK`` followed by 20 hexadecimal digits. The
1959 128-bit format used by PowerPC (two adjacent doubles) is represented by
1960 ``0xM`` followed by 32 hexadecimal digits. The IEEE 128-bit format is
1961 represented by ``0xL`` followed by 32 hexadecimal digits. Long doubles
1962 will only work if they match the long double format on your target.
1963 The IEEE 16-bit format (half precision) is represented by ``0xH``
1964 followed by 4 hexadecimal digits. All hexadecimal formats are big-endian
1965 (sign bit at the left).
1967 There are no constants of type x86mmx.
1969 .. _complexconstants:
1974 Complex constants are a (potentially recursive) combination of simple
1975 constants and smaller complex constants.
1977 **Structure constants**
1978 Structure constants are represented with notation similar to
1979 structure type definitions (a comma separated list of elements,
1980 surrounded by braces (``{}``)). For example:
1981 "``{ i32 4, float 17.0, i32* @G }``", where "``@G``" is declared as
1982 "``@G = external global i32``". Structure constants must have
1983 :ref:`structure type <t_struct>`, and the number and types of elements
1984 must match those specified by the type.
1986 Array constants are represented with notation similar to array type
1987 definitions (a comma separated list of elements, surrounded by
1988 square brackets (``[]``)). For example:
1989 "``[ i32 42, i32 11, i32 74 ]``". Array constants must have
1990 :ref:`array type <t_array>`, and the number and types of elements must
1991 match those specified by the type.
1992 **Vector constants**
1993 Vector constants are represented with notation similar to vector
1994 type definitions (a comma separated list of elements, surrounded by
1995 less-than/greater-than's (``<>``)). For example:
1996 "``< i32 42, i32 11, i32 74, i32 100 >``". Vector constants
1997 must have :ref:`vector type <t_vector>`, and the number and types of
1998 elements must match those specified by the type.
1999 **Zero initialization**
2000 The string '``zeroinitializer``' can be used to zero initialize a
2001 value to zero of *any* type, including scalar and
2002 :ref:`aggregate <t_aggregate>` types. This is often used to avoid
2003 having to print large zero initializers (e.g. for large arrays) and
2004 is always exactly equivalent to using explicit zero initializers.
2006 A metadata node is a structure-like constant with :ref:`metadata
2007 type <t_metadata>`. For example:
2008 "``metadata !{ i32 0, metadata !"test" }``". Unlike other
2009 constants that are meant to be interpreted as part of the
2010 instruction stream, metadata is a place to attach additional
2011 information such as debug info.
2013 Global Variable and Function Addresses
2014 --------------------------------------
2016 The addresses of :ref:`global variables <globalvars>` and
2017 :ref:`functions <functionstructure>` are always implicitly valid
2018 (link-time) constants. These constants are explicitly referenced when
2019 the :ref:`identifier for the global <identifiers>` is used and always have
2020 :ref:`pointer <t_pointer>` type. For example, the following is a legal LLVM
2023 .. code-block:: llvm
2027 @Z = global [2 x i32*] [ i32* @X, i32* @Y ]
2034 The string '``undef``' can be used anywhere a constant is expected, and
2035 indicates that the user of the value may receive an unspecified
2036 bit-pattern. Undefined values may be of any type (other than '``label``'
2037 or '``void``') and be used anywhere a constant is permitted.
2039 Undefined values are useful because they indicate to the compiler that
2040 the program is well defined no matter what value is used. This gives the
2041 compiler more freedom to optimize. Here are some examples of
2042 (potentially surprising) transformations that are valid (in pseudo IR):
2044 .. code-block:: llvm
2054 This is safe because all of the output bits are affected by the undef
2055 bits. Any output bit can have a zero or one depending on the input bits.
2057 .. code-block:: llvm
2068 These logical operations have bits that are not always affected by the
2069 input. For example, if ``%X`` has a zero bit, then the output of the
2070 '``and``' operation will always be a zero for that bit, no matter what
2071 the corresponding bit from the '``undef``' is. As such, it is unsafe to
2072 optimize or assume that the result of the '``and``' is '``undef``'.
2073 However, it is safe to assume that all bits of the '``undef``' could be
2074 0, and optimize the '``and``' to 0. Likewise, it is safe to assume that
2075 all the bits of the '``undef``' operand to the '``or``' could be set,
2076 allowing the '``or``' to be folded to -1.
2078 .. code-block:: llvm
2080 %A = select undef, %X, %Y
2081 %B = select undef, 42, %Y
2082 %C = select %X, %Y, undef
2092 This set of examples shows that undefined '``select``' (and conditional
2093 branch) conditions can go *either way*, but they have to come from one
2094 of the two operands. In the ``%A`` example, if ``%X`` and ``%Y`` were
2095 both known to have a clear low bit, then ``%A`` would have to have a
2096 cleared low bit. However, in the ``%C`` example, the optimizer is
2097 allowed to assume that the '``undef``' operand could be the same as
2098 ``%Y``, allowing the whole '``select``' to be eliminated.
2100 .. code-block:: llvm
2102 %A = xor undef, undef
2119 This example points out that two '``undef``' operands are not
2120 necessarily the same. This can be surprising to people (and also matches
2121 C semantics) where they assume that "``X^X``" is always zero, even if
2122 ``X`` is undefined. This isn't true for a number of reasons, but the
2123 short answer is that an '``undef``' "variable" can arbitrarily change
2124 its value over its "live range". This is true because the variable
2125 doesn't actually *have a live range*. Instead, the value is logically
2126 read from arbitrary registers that happen to be around when needed, so
2127 the value is not necessarily consistent over time. In fact, ``%A`` and
2128 ``%C`` need to have the same semantics or the core LLVM "replace all
2129 uses with" concept would not hold.
2131 .. code-block:: llvm
2139 These examples show the crucial difference between an *undefined value*
2140 and *undefined behavior*. An undefined value (like '``undef``') is
2141 allowed to have an arbitrary bit-pattern. This means that the ``%A``
2142 operation can be constant folded to '``undef``', because the '``undef``'
2143 could be an SNaN, and ``fdiv`` is not (currently) defined on SNaN's.
2144 However, in the second example, we can make a more aggressive
2145 assumption: because the ``undef`` is allowed to be an arbitrary value,
2146 we are allowed to assume that it could be zero. Since a divide by zero
2147 has *undefined behavior*, we are allowed to assume that the operation
2148 does not execute at all. This allows us to delete the divide and all
2149 code after it. Because the undefined operation "can't happen", the
2150 optimizer can assume that it occurs in dead code.
2152 .. code-block:: llvm
2154 a: store undef -> %X
2155 b: store %X -> undef
2160 These examples reiterate the ``fdiv`` example: a store *of* an undefined
2161 value can be assumed to not have any effect; we can assume that the
2162 value is overwritten with bits that happen to match what was already
2163 there. However, a store *to* an undefined location could clobber
2164 arbitrary memory, therefore, it has undefined behavior.
2171 Poison values are similar to :ref:`undef values <undefvalues>`, however
2172 they also represent the fact that an instruction or constant expression
2173 which cannot evoke side effects has nevertheless detected a condition
2174 which results in undefined behavior.
2176 There is currently no way of representing a poison value in the IR; they
2177 only exist when produced by operations such as :ref:`add <i_add>` with
2180 Poison value behavior is defined in terms of value *dependence*:
2182 - Values other than :ref:`phi <i_phi>` nodes depend on their operands.
2183 - :ref:`Phi <i_phi>` nodes depend on the operand corresponding to
2184 their dynamic predecessor basic block.
2185 - Function arguments depend on the corresponding actual argument values
2186 in the dynamic callers of their functions.
2187 - :ref:`Call <i_call>` instructions depend on the :ref:`ret <i_ret>`
2188 instructions that dynamically transfer control back to them.
2189 - :ref:`Invoke <i_invoke>` instructions depend on the
2190 :ref:`ret <i_ret>`, :ref:`resume <i_resume>`, or exception-throwing
2191 call instructions that dynamically transfer control back to them.
2192 - Non-volatile loads and stores depend on the most recent stores to all
2193 of the referenced memory addresses, following the order in the IR
2194 (including loads and stores implied by intrinsics such as
2195 :ref:`@llvm.memcpy <int_memcpy>`.)
2196 - An instruction with externally visible side effects depends on the
2197 most recent preceding instruction with externally visible side
2198 effects, following the order in the IR. (This includes :ref:`volatile
2199 operations <volatile>`.)
2200 - An instruction *control-depends* on a :ref:`terminator
2201 instruction <terminators>` if the terminator instruction has
2202 multiple successors and the instruction is always executed when
2203 control transfers to one of the successors, and may not be executed
2204 when control is transferred to another.
2205 - Additionally, an instruction also *control-depends* on a terminator
2206 instruction if the set of instructions it otherwise depends on would
2207 be different if the terminator had transferred control to a different
2209 - Dependence is transitive.
2211 Poison Values have the same behavior as :ref:`undef values <undefvalues>`,
2212 with the additional affect that any instruction which has a *dependence*
2213 on a poison value has undefined behavior.
2215 Here are some examples:
2217 .. code-block:: llvm
2220 %poison = sub nuw i32 0, 1 ; Results in a poison value.
2221 %still_poison = and i32 %poison, 0 ; 0, but also poison.
2222 %poison_yet_again = getelementptr i32* @h, i32 %still_poison
2223 store i32 0, i32* %poison_yet_again ; memory at @h[0] is poisoned
2225 store i32 %poison, i32* @g ; Poison value stored to memory.
2226 %poison2 = load i32* @g ; Poison value loaded back from memory.
2228 store volatile i32 %poison, i32* @g ; External observation; undefined behavior.
2230 %narrowaddr = bitcast i32* @g to i16*
2231 %wideaddr = bitcast i32* @g to i64*
2232 %poison3 = load i16* %narrowaddr ; Returns a poison value.
2233 %poison4 = load i64* %wideaddr ; Returns a poison value.
2235 %cmp = icmp slt i32 %poison, 0 ; Returns a poison value.
2236 br i1 %cmp, label %true, label %end ; Branch to either destination.
2239 store volatile i32 0, i32* @g ; This is control-dependent on %cmp, so
2240 ; it has undefined behavior.
2244 %p = phi i32 [ 0, %entry ], [ 1, %true ]
2245 ; Both edges into this PHI are
2246 ; control-dependent on %cmp, so this
2247 ; always results in a poison value.
2249 store volatile i32 0, i32* @g ; This would depend on the store in %true
2250 ; if %cmp is true, or the store in %entry
2251 ; otherwise, so this is undefined behavior.
2253 br i1 %cmp, label %second_true, label %second_end
2254 ; The same branch again, but this time the
2255 ; true block doesn't have side effects.
2262 store volatile i32 0, i32* @g ; This time, the instruction always depends
2263 ; on the store in %end. Also, it is
2264 ; control-equivalent to %end, so this is
2265 ; well-defined (ignoring earlier undefined
2266 ; behavior in this example).
2270 Addresses of Basic Blocks
2271 -------------------------
2273 ``blockaddress(@function, %block)``
2275 The '``blockaddress``' constant computes the address of the specified
2276 basic block in the specified function, and always has an ``i8*`` type.
2277 Taking the address of the entry block is illegal.
2279 This value only has defined behavior when used as an operand to the
2280 ':ref:`indirectbr <i_indirectbr>`' instruction, or for comparisons
2281 against null. Pointer equality tests between labels addresses results in
2282 undefined behavior --- though, again, comparison against null is ok, and
2283 no label is equal to the null pointer. This may be passed around as an
2284 opaque pointer sized value as long as the bits are not inspected. This
2285 allows ``ptrtoint`` and arithmetic to be performed on these values so
2286 long as the original value is reconstituted before the ``indirectbr``
2289 Finally, some targets may provide defined semantics when using the value
2290 as the operand to an inline assembly, but that is target specific.
2294 Constant Expressions
2295 --------------------
2297 Constant expressions are used to allow expressions involving other
2298 constants to be used as constants. Constant expressions may be of any
2299 :ref:`first class <t_firstclass>` type and may involve any LLVM operation
2300 that does not have side effects (e.g. load and call are not supported).
2301 The following is the syntax for constant expressions:
2303 ``trunc (CST to TYPE)``
2304 Truncate a constant to another type. The bit size of CST must be
2305 larger than the bit size of TYPE. Both types must be integers.
2306 ``zext (CST to TYPE)``
2307 Zero extend a constant to another type. The bit size of CST must be
2308 smaller than the bit size of TYPE. Both types must be integers.
2309 ``sext (CST to TYPE)``
2310 Sign extend a constant to another type. The bit size of CST must be
2311 smaller than the bit size of TYPE. Both types must be integers.
2312 ``fptrunc (CST to TYPE)``
2313 Truncate a floating point constant to another floating point type.
2314 The size of CST must be larger than the size of TYPE. Both types
2315 must be floating point.
2316 ``fpext (CST to TYPE)``
2317 Floating point extend a constant to another type. The size of CST
2318 must be smaller or equal to the size of TYPE. Both types must be
2320 ``fptoui (CST to TYPE)``
2321 Convert a floating point constant to the corresponding unsigned
2322 integer constant. TYPE must be a scalar or vector integer type. CST
2323 must be of scalar or vector floating point type. Both CST and TYPE
2324 must be scalars, or vectors of the same number of elements. If the
2325 value won't fit in the integer type, the results are undefined.
2326 ``fptosi (CST to TYPE)``
2327 Convert a floating point constant to the corresponding signed
2328 integer constant. TYPE must be a scalar or vector integer type. CST
2329 must be of scalar or vector floating point type. Both CST and TYPE
2330 must be scalars, or vectors of the same number of elements. If the
2331 value won't fit in the integer type, the results are undefined.
2332 ``uitofp (CST to TYPE)``
2333 Convert an unsigned integer constant to the corresponding floating
2334 point constant. TYPE must be a scalar or vector floating point type.
2335 CST must be of scalar or vector integer type. Both CST and TYPE must
2336 be scalars, or vectors of the same number of elements. If the value
2337 won't fit in the floating point type, the results are undefined.
2338 ``sitofp (CST to TYPE)``
2339 Convert a signed integer constant to the corresponding floating
2340 point constant. TYPE must be a scalar or vector floating point type.
2341 CST must be of scalar or vector integer type. Both CST and TYPE must
2342 be scalars, or vectors of the same number of elements. If the value
2343 won't fit in the floating point type, the results are undefined.
2344 ``ptrtoint (CST to TYPE)``
2345 Convert a pointer typed constant to the corresponding integer
2346 constant. ``TYPE`` must be an integer type. ``CST`` must be of
2347 pointer type. The ``CST`` value is zero extended, truncated, or
2348 unchanged to make it fit in ``TYPE``.
2349 ``inttoptr (CST to TYPE)``
2350 Convert an integer constant to a pointer constant. TYPE must be a
2351 pointer type. CST must be of integer type. The CST value is zero
2352 extended, truncated, or unchanged to make it fit in a pointer size.
2353 This one is *really* dangerous!
2354 ``bitcast (CST to TYPE)``
2355 Convert a constant, CST, to another TYPE. The constraints of the
2356 operands are the same as those for the :ref:`bitcast
2357 instruction <i_bitcast>`.
2358 ``getelementptr (CSTPTR, IDX0, IDX1, ...)``, ``getelementptr inbounds (CSTPTR, IDX0, IDX1, ...)``
2359 Perform the :ref:`getelementptr operation <i_getelementptr>` on
2360 constants. As with the :ref:`getelementptr <i_getelementptr>`
2361 instruction, the index list may have zero or more indexes, which are
2362 required to make sense for the type of "CSTPTR".
2363 ``select (COND, VAL1, VAL2)``
2364 Perform the :ref:`select operation <i_select>` on constants.
2365 ``icmp COND (VAL1, VAL2)``
2366 Performs the :ref:`icmp operation <i_icmp>` on constants.
2367 ``fcmp COND (VAL1, VAL2)``
2368 Performs the :ref:`fcmp operation <i_fcmp>` on constants.
2369 ``extractelement (VAL, IDX)``
2370 Perform the :ref:`extractelement operation <i_extractelement>` on
2372 ``insertelement (VAL, ELT, IDX)``
2373 Perform the :ref:`insertelement operation <i_insertelement>` on
2375 ``shufflevector (VEC1, VEC2, IDXMASK)``
2376 Perform the :ref:`shufflevector operation <i_shufflevector>` on
2378 ``extractvalue (VAL, IDX0, IDX1, ...)``
2379 Perform the :ref:`extractvalue operation <i_extractvalue>` on
2380 constants. The index list is interpreted in a similar manner as
2381 indices in a ':ref:`getelementptr <i_getelementptr>`' operation. At
2382 least one index value must be specified.
2383 ``insertvalue (VAL, ELT, IDX0, IDX1, ...)``
2384 Perform the :ref:`insertvalue operation <i_insertvalue>` on constants.
2385 The index list is interpreted in a similar manner as indices in a
2386 ':ref:`getelementptr <i_getelementptr>`' operation. At least one index
2387 value must be specified.
2388 ``OPCODE (LHS, RHS)``
2389 Perform the specified operation of the LHS and RHS constants. OPCODE
2390 may be any of the :ref:`binary <binaryops>` or :ref:`bitwise
2391 binary <bitwiseops>` operations. The constraints on operands are
2392 the same as those for the corresponding instruction (e.g. no bitwise
2393 operations on floating point values are allowed).
2400 Inline Assembler Expressions
2401 ----------------------------
2403 LLVM supports inline assembler expressions (as opposed to :ref:`Module-Level
2404 Inline Assembly <moduleasm>`) through the use of a special value. This
2405 value represents the inline assembler as a string (containing the
2406 instructions to emit), a list of operand constraints (stored as a
2407 string), a flag that indicates whether or not the inline asm expression
2408 has side effects, and a flag indicating whether the function containing
2409 the asm needs to align its stack conservatively. An example inline
2410 assembler expression is:
2412 .. code-block:: llvm
2414 i32 (i32) asm "bswap $0", "=r,r"
2416 Inline assembler expressions may **only** be used as the callee operand
2417 of a :ref:`call <i_call>` or an :ref:`invoke <i_invoke>` instruction.
2418 Thus, typically we have:
2420 .. code-block:: llvm
2422 %X = call i32 asm "bswap $0", "=r,r"(i32 %Y)
2424 Inline asms with side effects not visible in the constraint list must be
2425 marked as having side effects. This is done through the use of the
2426 '``sideeffect``' keyword, like so:
2428 .. code-block:: llvm
2430 call void asm sideeffect "eieio", ""()
2432 In some cases inline asms will contain code that will not work unless
2433 the stack is aligned in some way, such as calls or SSE instructions on
2434 x86, yet will not contain code that does that alignment within the asm.
2435 The compiler should make conservative assumptions about what the asm
2436 might contain and should generate its usual stack alignment code in the
2437 prologue if the '``alignstack``' keyword is present:
2439 .. code-block:: llvm
2441 call void asm alignstack "eieio", ""()
2443 Inline asms also support using non-standard assembly dialects. The
2444 assumed dialect is ATT. When the '``inteldialect``' keyword is present,
2445 the inline asm is using the Intel dialect. Currently, ATT and Intel are
2446 the only supported dialects. An example is:
2448 .. code-block:: llvm
2450 call void asm inteldialect "eieio", ""()
2452 If multiple keywords appear the '``sideeffect``' keyword must come
2453 first, the '``alignstack``' keyword second and the '``inteldialect``'
2459 The call instructions that wrap inline asm nodes may have a
2460 "``!srcloc``" MDNode attached to it that contains a list of constant
2461 integers. If present, the code generator will use the integer as the
2462 location cookie value when report errors through the ``LLVMContext``
2463 error reporting mechanisms. This allows a front-end to correlate backend
2464 errors that occur with inline asm back to the source code that produced
2467 .. code-block:: llvm
2469 call void asm sideeffect "something bad", ""(), !srcloc !42
2471 !42 = !{ i32 1234567 }
2473 It is up to the front-end to make sense of the magic numbers it places
2474 in the IR. If the MDNode contains multiple constants, the code generator
2475 will use the one that corresponds to the line of the asm that the error
2480 Metadata Nodes and Metadata Strings
2481 -----------------------------------
2483 LLVM IR allows metadata to be attached to instructions in the program
2484 that can convey extra information about the code to the optimizers and
2485 code generator. One example application of metadata is source-level
2486 debug information. There are two metadata primitives: strings and nodes.
2487 All metadata has the ``metadata`` type and is identified in syntax by a
2488 preceding exclamation point ('``!``').
2490 A metadata string is a string surrounded by double quotes. It can
2491 contain any character by escaping non-printable characters with
2492 "``\xx``" where "``xx``" is the two digit hex code. For example:
2495 Metadata nodes are represented with notation similar to structure
2496 constants (a comma separated list of elements, surrounded by braces and
2497 preceded by an exclamation point). Metadata nodes can have any values as
2498 their operand. For example:
2500 .. code-block:: llvm
2502 !{ metadata !"test\00", i32 10}
2504 A :ref:`named metadata <namedmetadatastructure>` is a collection of
2505 metadata nodes, which can be looked up in the module symbol table. For
2508 .. code-block:: llvm
2510 !foo = metadata !{!4, !3}
2512 Metadata can be used as function arguments. Here ``llvm.dbg.value``
2513 function is using two metadata arguments:
2515 .. code-block:: llvm
2517 call void @llvm.dbg.value(metadata !24, i64 0, metadata !25)
2519 Metadata can be attached with an instruction. Here metadata ``!21`` is
2520 attached to the ``add`` instruction using the ``!dbg`` identifier:
2522 .. code-block:: llvm
2524 %indvar.next = add i64 %indvar, 1, !dbg !21
2526 More information about specific metadata nodes recognized by the
2527 optimizers and code generator is found below.
2532 In LLVM IR, memory does not have types, so LLVM's own type system is not
2533 suitable for doing TBAA. Instead, metadata is added to the IR to
2534 describe a type system of a higher level language. This can be used to
2535 implement typical C/C++ TBAA, but it can also be used to implement
2536 custom alias analysis behavior for other languages.
2538 The current metadata format is very simple. TBAA metadata nodes have up
2539 to three fields, e.g.:
2541 .. code-block:: llvm
2543 !0 = metadata !{ metadata !"an example type tree" }
2544 !1 = metadata !{ metadata !"int", metadata !0 }
2545 !2 = metadata !{ metadata !"float", metadata !0 }
2546 !3 = metadata !{ metadata !"const float", metadata !2, i64 1 }
2548 The first field is an identity field. It can be any value, usually a
2549 metadata string, which uniquely identifies the type. The most important
2550 name in the tree is the name of the root node. Two trees with different
2551 root node names are entirely disjoint, even if they have leaves with
2554 The second field identifies the type's parent node in the tree, or is
2555 null or omitted for a root node. A type is considered to alias all of
2556 its descendants and all of its ancestors in the tree. Also, a type is
2557 considered to alias all types in other trees, so that bitcode produced
2558 from multiple front-ends is handled conservatively.
2560 If the third field is present, it's an integer which if equal to 1
2561 indicates that the type is "constant" (meaning
2562 ``pointsToConstantMemory`` should return true; see `other useful
2563 AliasAnalysis methods <AliasAnalysis.html#OtherItfs>`_).
2565 '``tbaa.struct``' Metadata
2566 ^^^^^^^^^^^^^^^^^^^^^^^^^^
2568 The :ref:`llvm.memcpy <int_memcpy>` is often used to implement
2569 aggregate assignment operations in C and similar languages, however it
2570 is defined to copy a contiguous region of memory, which is more than
2571 strictly necessary for aggregate types which contain holes due to
2572 padding. Also, it doesn't contain any TBAA information about the fields
2575 ``!tbaa.struct`` metadata can describe which memory subregions in a
2576 memcpy are padding and what the TBAA tags of the struct are.
2578 The current metadata format is very simple. ``!tbaa.struct`` metadata
2579 nodes are a list of operands which are in conceptual groups of three.
2580 For each group of three, the first operand gives the byte offset of a
2581 field in bytes, the second gives its size in bytes, and the third gives
2584 .. code-block:: llvm
2586 !4 = metadata !{ i64 0, i64 4, metadata !1, i64 8, i64 4, metadata !2 }
2588 This describes a struct with two fields. The first is at offset 0 bytes
2589 with size 4 bytes, and has tbaa tag !1. The second is at offset 8 bytes
2590 and has size 4 bytes and has tbaa tag !2.
2592 Note that the fields need not be contiguous. In this example, there is a
2593 4 byte gap between the two fields. This gap represents padding which
2594 does not carry useful data and need not be preserved.
2596 '``fpmath``' Metadata
2597 ^^^^^^^^^^^^^^^^^^^^^
2599 ``fpmath`` metadata may be attached to any instruction of floating point
2600 type. It can be used to express the maximum acceptable error in the
2601 result of that instruction, in ULPs, thus potentially allowing the
2602 compiler to use a more efficient but less accurate method of computing
2603 it. ULP is defined as follows:
2605 If ``x`` is a real number that lies between two finite consecutive
2606 floating-point numbers ``a`` and ``b``, without being equal to one
2607 of them, then ``ulp(x) = |b - a|``, otherwise ``ulp(x)`` is the
2608 distance between the two non-equal finite floating-point numbers
2609 nearest ``x``. Moreover, ``ulp(NaN)`` is ``NaN``.
2611 The metadata node shall consist of a single positive floating point
2612 number representing the maximum relative error, for example:
2614 .. code-block:: llvm
2616 !0 = metadata !{ float 2.5 } ; maximum acceptable inaccuracy is 2.5 ULPs
2618 '``range``' Metadata
2619 ^^^^^^^^^^^^^^^^^^^^
2621 ``range`` metadata may be attached only to loads of integer types. It
2622 expresses the possible ranges the loaded value is in. The ranges are
2623 represented with a flattened list of integers. The loaded value is known
2624 to be in the union of the ranges defined by each consecutive pair. Each
2625 pair has the following properties:
2627 - The type must match the type loaded by the instruction.
2628 - The pair ``a,b`` represents the range ``[a,b)``.
2629 - Both ``a`` and ``b`` are constants.
2630 - The range is allowed to wrap.
2631 - The range should not represent the full or empty set. That is,
2634 In addition, the pairs must be in signed order of the lower bound and
2635 they must be non-contiguous.
2639 .. code-block:: llvm
2641 %a = load i8* %x, align 1, !range !0 ; Can only be 0 or 1
2642 %b = load i8* %y, align 1, !range !1 ; Can only be 255 (-1), 0 or 1
2643 %c = load i8* %z, align 1, !range !2 ; Can only be 0, 1, 3, 4 or 5
2644 %d = load i8* %z, align 1, !range !3 ; Can only be -2, -1, 3, 4 or 5
2646 !0 = metadata !{ i8 0, i8 2 }
2647 !1 = metadata !{ i8 255, i8 2 }
2648 !2 = metadata !{ i8 0, i8 2, i8 3, i8 6 }
2649 !3 = metadata !{ i8 -2, i8 0, i8 3, i8 6 }
2654 It is sometimes useful to attach information to loop constructs. Currently,
2655 loop metadata is implemented as metadata attached to the branch instruction
2656 in the loop latch block. This type of metadata refer to a metadata node that is
2657 guaranteed to be separate for each loop. The loop identifier metadata is
2658 specified with the name ``llvm.loop``.
2660 The loop identifier metadata is implemented using a metadata that refers to
2661 itself to avoid merging it with any other identifier metadata, e.g.,
2662 during module linkage or function inlining. That is, each loop should refer
2663 to their own identification metadata even if they reside in separate functions.
2664 The following example contains loop identifier metadata for two separate loop
2667 .. code-block:: llvm
2669 !0 = metadata !{ metadata !0 }
2670 !1 = metadata !{ metadata !1 }
2672 The loop identifier metadata can be used to specify additional per-loop
2673 metadata. Any operands after the first operand can be treated as user-defined
2674 metadata. For example the ``llvm.vectorizer.unroll`` metadata is understood
2675 by the loop vectorizer to indicate how many times to unroll the loop:
2677 .. code-block:: llvm
2679 br i1 %exitcond, label %._crit_edge, label %.lr.ph, !llvm.loop !0
2681 !0 = metadata !{ metadata !0, metadata !1 }
2682 !1 = metadata !{ metadata !"llvm.vectorizer.unroll", i32 2 }
2687 Metadata types used to annotate memory accesses with information helpful
2688 for optimizations are prefixed with ``llvm.mem``.
2690 '``llvm.mem.parallel_loop_access``' Metadata
2691 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2693 For a loop to be parallel, in addition to using
2694 the ``llvm.loop`` metadata to mark the loop latch branch instruction,
2695 also all of the memory accessing instructions in the loop body need to be
2696 marked with the ``llvm.mem.parallel_loop_access`` metadata. If there
2697 is at least one memory accessing instruction not marked with the metadata,
2698 the loop must be considered a sequential loop. This causes parallel loops to be
2699 converted to sequential loops due to optimization passes that are unaware of
2700 the parallel semantics and that insert new memory instructions to the loop
2703 Example of a loop that is considered parallel due to its correct use of
2704 both ``llvm.loop`` and ``llvm.mem.parallel_loop_access``
2705 metadata types that refer to the same loop identifier metadata.
2707 .. code-block:: llvm
2711 %0 = load i32* %arrayidx, align 4, !llvm.mem.parallel_loop_access !0
2713 store i32 %0, i32* %arrayidx4, align 4, !llvm.mem.parallel_loop_access !0
2715 br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !0
2719 !0 = metadata !{ metadata !0 }
2721 It is also possible to have nested parallel loops. In that case the
2722 memory accesses refer to a list of loop identifier metadata nodes instead of
2723 the loop identifier metadata node directly:
2725 .. code-block:: llvm
2732 %0 = load i32* %arrayidx, align 4, !llvm.mem.parallel_loop_access !0
2734 store i32 %0, i32* %arrayidx4, align 4, !llvm.mem.parallel_loop_access !0
2736 br i1 %exitcond, label %inner.for.end, label %inner.for.body, !llvm.loop !1
2740 %0 = load i32* %arrayidx, align 4, !llvm.mem.parallel_loop_access !0
2742 store i32 %0, i32* %arrayidx4, align 4, !llvm.mem.parallel_loop_access !0
2744 br i1 %exitcond, label %outer.for.end, label %outer.for.body, !llvm.loop !2
2746 outer.for.end: ; preds = %for.body
2748 !0 = metadata !{ metadata !1, metadata !2 } ; a list of loop identifiers
2749 !1 = metadata !{ metadata !1 } ; an identifier for the inner loop
2750 !2 = metadata !{ metadata !2 } ; an identifier for the outer loop
2752 '``llvm.vectorizer``'
2753 ^^^^^^^^^^^^^^^^^^^^^
2755 Metadata prefixed with ``llvm.vectorizer`` is used to control per-loop
2756 vectorization parameters such as vectorization factor and unroll factor.
2758 ``llvm.vectorizer`` metadata should be used in conjunction with ``llvm.loop``
2759 loop identification metadata.
2761 '``llvm.vectorizer.unroll``' Metadata
2762 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2764 This metadata instructs the loop vectorizer to unroll the specified
2765 loop exactly ``N`` times.
2767 The first operand is the string ``llvm.vectorizer.unroll`` and the second
2768 operand is an integer specifying the unroll factor. For example:
2770 .. code-block:: llvm
2772 !0 = metadata !{ metadata !"llvm.vectorizer.unroll", i32 4 }
2774 Note that setting ``llvm.vectorizer.unroll`` to 1 disables unrolling of the
2777 If ``llvm.vectorizer.unroll`` is set to 0 then the amount of unrolling will be
2778 determined automatically.
2780 '``llvm.vectorizer.width``' Metadata
2781 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2783 This metadata sets the target width of the vectorizer to ``N``. Without
2784 this metadata, the vectorizer will choose a width automatically.
2785 Regardless of this metadata, the vectorizer will only vectorize loops if
2786 it believes it is valid to do so.
2788 The first operand is the string ``llvm.vectorizer.width`` and the second
2789 operand is an integer specifying the width. For example:
2791 .. code-block:: llvm
2793 !0 = metadata !{ metadata !"llvm.vectorizer.width", i32 4 }
2795 Note that setting ``llvm.vectorizer.width`` to 1 disables vectorization of the
2798 If ``llvm.vectorizer.width`` is set to 0 then the width will be determined
2801 Module Flags Metadata
2802 =====================
2804 Information about the module as a whole is difficult to convey to LLVM's
2805 subsystems. The LLVM IR isn't sufficient to transmit this information.
2806 The ``llvm.module.flags`` named metadata exists in order to facilitate
2807 this. These flags are in the form of key / value pairs --- much like a
2808 dictionary --- making it easy for any subsystem who cares about a flag to
2811 The ``llvm.module.flags`` metadata contains a list of metadata triplets.
2812 Each triplet has the following form:
2814 - The first element is a *behavior* flag, which specifies the behavior
2815 when two (or more) modules are merged together, and it encounters two
2816 (or more) metadata with the same ID. The supported behaviors are
2818 - The second element is a metadata string that is a unique ID for the
2819 metadata. Each module may only have one flag entry for each unique ID (not
2820 including entries with the **Require** behavior).
2821 - The third element is the value of the flag.
2823 When two (or more) modules are merged together, the resulting
2824 ``llvm.module.flags`` metadata is the union of the modules' flags. That is, for
2825 each unique metadata ID string, there will be exactly one entry in the merged
2826 modules ``llvm.module.flags`` metadata table, and the value for that entry will
2827 be determined by the merge behavior flag, as described below. The only exception
2828 is that entries with the *Require* behavior are always preserved.
2830 The following behaviors are supported:
2841 Emits an error if two values disagree, otherwise the resulting value
2842 is that of the operands.
2846 Emits a warning if two values disagree. The result value will be the
2847 operand for the flag from the first module being linked.
2851 Adds a requirement that another module flag be present and have a
2852 specified value after linking is performed. The value must be a
2853 metadata pair, where the first element of the pair is the ID of the
2854 module flag to be restricted, and the second element of the pair is
2855 the value the module flag should be restricted to. This behavior can
2856 be used to restrict the allowable results (via triggering of an
2857 error) of linking IDs with the **Override** behavior.
2861 Uses the specified value, regardless of the behavior or value of the
2862 other module. If both modules specify **Override**, but the values
2863 differ, an error will be emitted.
2867 Appends the two values, which are required to be metadata nodes.
2871 Appends the two values, which are required to be metadata
2872 nodes. However, duplicate entries in the second list are dropped
2873 during the append operation.
2875 It is an error for a particular unique flag ID to have multiple behaviors,
2876 except in the case of **Require** (which adds restrictions on another metadata
2877 value) or **Override**.
2879 An example of module flags:
2881 .. code-block:: llvm
2883 !0 = metadata !{ i32 1, metadata !"foo", i32 1 }
2884 !1 = metadata !{ i32 4, metadata !"bar", i32 37 }
2885 !2 = metadata !{ i32 2, metadata !"qux", i32 42 }
2886 !3 = metadata !{ i32 3, metadata !"qux",
2888 metadata !"foo", i32 1
2891 !llvm.module.flags = !{ !0, !1, !2, !3 }
2893 - Metadata ``!0`` has the ID ``!"foo"`` and the value '1'. The behavior
2894 if two or more ``!"foo"`` flags are seen is to emit an error if their
2895 values are not equal.
2897 - Metadata ``!1`` has the ID ``!"bar"`` and the value '37'. The
2898 behavior if two or more ``!"bar"`` flags are seen is to use the value
2901 - Metadata ``!2`` has the ID ``!"qux"`` and the value '42'. The
2902 behavior if two or more ``!"qux"`` flags are seen is to emit a
2903 warning if their values are not equal.
2905 - Metadata ``!3`` has the ID ``!"qux"`` and the value:
2909 metadata !{ metadata !"foo", i32 1 }
2911 The behavior is to emit an error if the ``llvm.module.flags`` does not
2912 contain a flag with the ID ``!"foo"`` that has the value '1' after linking is
2915 Objective-C Garbage Collection Module Flags Metadata
2916 ----------------------------------------------------
2918 On the Mach-O platform, Objective-C stores metadata about garbage
2919 collection in a special section called "image info". The metadata
2920 consists of a version number and a bitmask specifying what types of
2921 garbage collection are supported (if any) by the file. If two or more
2922 modules are linked together their garbage collection metadata needs to
2923 be merged rather than appended together.
2925 The Objective-C garbage collection module flags metadata consists of the
2926 following key-value pairs:
2935 * - ``Objective-C Version``
2936 - **[Required]** --- The Objective-C ABI version. Valid values are 1 and 2.
2938 * - ``Objective-C Image Info Version``
2939 - **[Required]** --- The version of the image info section. Currently
2942 * - ``Objective-C Image Info Section``
2943 - **[Required]** --- The section to place the metadata. Valid values are
2944 ``"__OBJC, __image_info, regular"`` for Objective-C ABI version 1, and
2945 ``"__DATA,__objc_imageinfo, regular, no_dead_strip"`` for
2946 Objective-C ABI version 2.
2948 * - ``Objective-C Garbage Collection``
2949 - **[Required]** --- Specifies whether garbage collection is supported or
2950 not. Valid values are 0, for no garbage collection, and 2, for garbage
2951 collection supported.
2953 * - ``Objective-C GC Only``
2954 - **[Optional]** --- Specifies that only garbage collection is supported.
2955 If present, its value must be 6. This flag requires that the
2956 ``Objective-C Garbage Collection`` flag have the value 2.
2958 Some important flag interactions:
2960 - If a module with ``Objective-C Garbage Collection`` set to 0 is
2961 merged with a module with ``Objective-C Garbage Collection`` set to
2962 2, then the resulting module has the
2963 ``Objective-C Garbage Collection`` flag set to 0.
2964 - A module with ``Objective-C Garbage Collection`` set to 0 cannot be
2965 merged with a module with ``Objective-C GC Only`` set to 6.
2967 Automatic Linker Flags Module Flags Metadata
2968 --------------------------------------------
2970 Some targets support embedding flags to the linker inside individual object
2971 files. Typically this is used in conjunction with language extensions which
2972 allow source files to explicitly declare the libraries they depend on, and have
2973 these automatically be transmitted to the linker via object files.
2975 These flags are encoded in the IR using metadata in the module flags section,
2976 using the ``Linker Options`` key. The merge behavior for this flag is required
2977 to be ``AppendUnique``, and the value for the key is expected to be a metadata
2978 node which should be a list of other metadata nodes, each of which should be a
2979 list of metadata strings defining linker options.
2981 For example, the following metadata section specifies two separate sets of
2982 linker options, presumably to link against ``libz`` and the ``Cocoa``
2985 !0 = metadata !{ i32 6, metadata !"Linker Options",
2987 metadata !{ metadata !"-lz" },
2988 metadata !{ metadata !"-framework", metadata !"Cocoa" } } }
2989 !llvm.module.flags = !{ !0 }
2991 The metadata encoding as lists of lists of options, as opposed to a collapsed
2992 list of options, is chosen so that the IR encoding can use multiple option
2993 strings to specify e.g., a single library, while still having that specifier be
2994 preserved as an atomic element that can be recognized by a target specific
2995 assembly writer or object file emitter.
2997 Each individual option is required to be either a valid option for the target's
2998 linker, or an option that is reserved by the target specific assembly writer or
2999 object file emitter. No other aspect of these options is defined by the IR.
3001 .. _intrinsicglobalvariables:
3003 Intrinsic Global Variables
3004 ==========================
3006 LLVM has a number of "magic" global variables that contain data that
3007 affect code generation or other IR semantics. These are documented here.
3008 All globals of this sort should have a section specified as
3009 "``llvm.metadata``". This section and all globals that start with
3010 "``llvm.``" are reserved for use by LLVM.
3014 The '``llvm.used``' Global Variable
3015 -----------------------------------
3017 The ``@llvm.used`` global is an array which has
3018 :ref:`appending linkage <linkage_appending>`. This array contains a list of
3019 pointers to named global variables, functions and aliases which may optionally
3020 have a pointer cast formed of bitcast or getelementptr. For example, a legal
3023 .. code-block:: llvm
3028 @llvm.used = appending global [2 x i8*] [
3030 i8* bitcast (i32* @Y to i8*)
3031 ], section "llvm.metadata"
3033 If a symbol appears in the ``@llvm.used`` list, then the compiler, assembler,
3034 and linker are required to treat the symbol as if there is a reference to the
3035 symbol that it cannot see (which is why they have to be named). For example, if
3036 a variable has internal linkage and no references other than that from the
3037 ``@llvm.used`` list, it cannot be deleted. This is commonly used to represent
3038 references from inline asms and other things the compiler cannot "see", and
3039 corresponds to "``attribute((used))``" in GNU C.
3041 On some targets, the code generator must emit a directive to the
3042 assembler or object file to prevent the assembler and linker from
3043 molesting the symbol.
3045 .. _gv_llvmcompilerused:
3047 The '``llvm.compiler.used``' Global Variable
3048 --------------------------------------------
3050 The ``@llvm.compiler.used`` directive is the same as the ``@llvm.used``
3051 directive, except that it only prevents the compiler from touching the
3052 symbol. On targets that support it, this allows an intelligent linker to
3053 optimize references to the symbol without being impeded as it would be
3056 This is a rare construct that should only be used in rare circumstances,
3057 and should not be exposed to source languages.
3059 .. _gv_llvmglobalctors:
3061 The '``llvm.global_ctors``' Global Variable
3062 -------------------------------------------
3064 .. code-block:: llvm
3066 %0 = type { i32, void ()* }
3067 @llvm.global_ctors = appending global [1 x %0] [%0 { i32 65535, void ()* @ctor }]
3069 The ``@llvm.global_ctors`` array contains a list of constructor
3070 functions and associated priorities. The functions referenced by this
3071 array will be called in ascending order of priority (i.e. lowest first)
3072 when the module is loaded. The order of functions with the same priority
3075 .. _llvmglobaldtors:
3077 The '``llvm.global_dtors``' Global Variable
3078 -------------------------------------------
3080 .. code-block:: llvm
3082 %0 = type { i32, void ()* }
3083 @llvm.global_dtors = appending global [1 x %0] [%0 { i32 65535, void ()* @dtor }]
3085 The ``@llvm.global_dtors`` array contains a list of destructor functions
3086 and associated priorities. The functions referenced by this array will
3087 be called in descending order of priority (i.e. highest first) when the
3088 module is loaded. The order of functions with the same priority is not
3091 Instruction Reference
3092 =====================
3094 The LLVM instruction set consists of several different classifications
3095 of instructions: :ref:`terminator instructions <terminators>`, :ref:`binary
3096 instructions <binaryops>`, :ref:`bitwise binary
3097 instructions <bitwiseops>`, :ref:`memory instructions <memoryops>`, and
3098 :ref:`other instructions <otherops>`.
3102 Terminator Instructions
3103 -----------------------
3105 As mentioned :ref:`previously <functionstructure>`, every basic block in a
3106 program ends with a "Terminator" instruction, which indicates which
3107 block should be executed after the current block is finished. These
3108 terminator instructions typically yield a '``void``' value: they produce
3109 control flow, not values (the one exception being the
3110 ':ref:`invoke <i_invoke>`' instruction).
3112 The terminator instructions are: ':ref:`ret <i_ret>`',
3113 ':ref:`br <i_br>`', ':ref:`switch <i_switch>`',
3114 ':ref:`indirectbr <i_indirectbr>`', ':ref:`invoke <i_invoke>`',
3115 ':ref:`resume <i_resume>`', and ':ref:`unreachable <i_unreachable>`'.
3119 '``ret``' Instruction
3120 ^^^^^^^^^^^^^^^^^^^^^
3127 ret <type> <value> ; Return a value from a non-void function
3128 ret void ; Return from void function
3133 The '``ret``' instruction is used to return control flow (and optionally
3134 a value) from a function back to the caller.
3136 There are two forms of the '``ret``' instruction: one that returns a
3137 value and then causes control flow, and one that just causes control
3143 The '``ret``' instruction optionally accepts a single argument, the
3144 return value. The type of the return value must be a ':ref:`first
3145 class <t_firstclass>`' type.
3147 A function is not :ref:`well formed <wellformed>` if it it has a non-void
3148 return type and contains a '``ret``' instruction with no return value or
3149 a return value with a type that does not match its type, or if it has a
3150 void return type and contains a '``ret``' instruction with a return
3156 When the '``ret``' instruction is executed, control flow returns back to
3157 the calling function's context. If the caller is a
3158 ":ref:`call <i_call>`" instruction, execution continues at the
3159 instruction after the call. If the caller was an
3160 ":ref:`invoke <i_invoke>`" instruction, execution continues at the
3161 beginning of the "normal" destination block. If the instruction returns
3162 a value, that value shall set the call or invoke instruction's return
3168 .. code-block:: llvm
3170 ret i32 5 ; Return an integer value of 5
3171 ret void ; Return from a void function
3172 ret { i32, i8 } { i32 4, i8 2 } ; Return a struct of values 4 and 2
3176 '``br``' Instruction
3177 ^^^^^^^^^^^^^^^^^^^^
3184 br i1 <cond>, label <iftrue>, label <iffalse>
3185 br label <dest> ; Unconditional branch
3190 The '``br``' instruction is used to cause control flow to transfer to a
3191 different basic block in the current function. There are two forms of
3192 this instruction, corresponding to a conditional branch and an
3193 unconditional branch.
3198 The conditional branch form of the '``br``' instruction takes a single
3199 '``i1``' value and two '``label``' values. The unconditional form of the
3200 '``br``' instruction takes a single '``label``' value as a target.
3205 Upon execution of a conditional '``br``' instruction, the '``i1``'
3206 argument is evaluated. If the value is ``true``, control flows to the
3207 '``iftrue``' ``label`` argument. If "cond" is ``false``, control flows
3208 to the '``iffalse``' ``label`` argument.
3213 .. code-block:: llvm
3216 %cond = icmp eq i32 %a, %b
3217 br i1 %cond, label %IfEqual, label %IfUnequal
3225 '``switch``' Instruction
3226 ^^^^^^^^^^^^^^^^^^^^^^^^
3233 switch <intty> <value>, label <defaultdest> [ <intty> <val>, label <dest> ... ]
3238 The '``switch``' instruction is used to transfer control flow to one of
3239 several different places. It is a generalization of the '``br``'
3240 instruction, allowing a branch to occur to one of many possible
3246 The '``switch``' instruction uses three parameters: an integer
3247 comparison value '``value``', a default '``label``' destination, and an
3248 array of pairs of comparison value constants and '``label``'s. The table
3249 is not allowed to contain duplicate constant entries.
3254 The ``switch`` instruction specifies a table of values and destinations.
3255 When the '``switch``' instruction is executed, this table is searched
3256 for the given value. If the value is found, control flow is transferred
3257 to the corresponding destination; otherwise, control flow is transferred
3258 to the default destination.
3263 Depending on properties of the target machine and the particular
3264 ``switch`` instruction, this instruction may be code generated in
3265 different ways. For example, it could be generated as a series of
3266 chained conditional branches or with a lookup table.
3271 .. code-block:: llvm
3273 ; Emulate a conditional br instruction
3274 %Val = zext i1 %value to i32
3275 switch i32 %Val, label %truedest [ i32 0, label %falsedest ]
3277 ; Emulate an unconditional br instruction
3278 switch i32 0, label %dest [ ]
3280 ; Implement a jump table:
3281 switch i32 %val, label %otherwise [ i32 0, label %onzero
3283 i32 2, label %ontwo ]
3287 '``indirectbr``' Instruction
3288 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3295 indirectbr <somety>* <address>, [ label <dest1>, label <dest2>, ... ]
3300 The '``indirectbr``' instruction implements an indirect branch to a
3301 label within the current function, whose address is specified by
3302 "``address``". Address must be derived from a
3303 :ref:`blockaddress <blockaddress>` constant.
3308 The '``address``' argument is the address of the label to jump to. The
3309 rest of the arguments indicate the full set of possible destinations
3310 that the address may point to. Blocks are allowed to occur multiple
3311 times in the destination list, though this isn't particularly useful.
3313 This destination list is required so that dataflow analysis has an
3314 accurate understanding of the CFG.
3319 Control transfers to the block specified in the address argument. All
3320 possible destination blocks must be listed in the label list, otherwise
3321 this instruction has undefined behavior. This implies that jumps to
3322 labels defined in other functions have undefined behavior as well.
3327 This is typically implemented with a jump through a register.
3332 .. code-block:: llvm
3334 indirectbr i8* %Addr, [ label %bb1, label %bb2, label %bb3 ]
3338 '``invoke``' Instruction
3339 ^^^^^^^^^^^^^^^^^^^^^^^^
3346 <result> = invoke [cconv] [ret attrs] <ptr to function ty> <function ptr val>(<function args>) [fn attrs]
3347 to label <normal label> unwind label <exception label>
3352 The '``invoke``' instruction causes control to transfer to a specified
3353 function, with the possibility of control flow transfer to either the
3354 '``normal``' label or the '``exception``' label. If the callee function
3355 returns with the "``ret``" instruction, control flow will return to the
3356 "normal" label. If the callee (or any indirect callees) returns via the
3357 ":ref:`resume <i_resume>`" instruction or other exception handling
3358 mechanism, control is interrupted and continued at the dynamically
3359 nearest "exception" label.
3361 The '``exception``' label is a `landing
3362 pad <ExceptionHandling.html#overview>`_ for the exception. As such,
3363 '``exception``' label is required to have the
3364 ":ref:`landingpad <i_landingpad>`" instruction, which contains the
3365 information about the behavior of the program after unwinding happens,
3366 as its first non-PHI instruction. The restrictions on the
3367 "``landingpad``" instruction's tightly couples it to the "``invoke``"
3368 instruction, so that the important information contained within the
3369 "``landingpad``" instruction can't be lost through normal code motion.
3374 This instruction requires several arguments:
3376 #. The optional "cconv" marker indicates which :ref:`calling
3377 convention <callingconv>` the call should use. If none is
3378 specified, the call defaults to using C calling conventions.
3379 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
3380 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
3382 #. '``ptr to function ty``': shall be the signature of the pointer to
3383 function value being invoked. In most cases, this is a direct
3384 function invocation, but indirect ``invoke``'s are just as possible,
3385 branching off an arbitrary pointer to function value.
3386 #. '``function ptr val``': An LLVM value containing a pointer to a
3387 function to be invoked.
3388 #. '``function args``': argument list whose types match the function
3389 signature argument types and parameter attributes. All arguments must
3390 be of :ref:`first class <t_firstclass>` type. If the function signature
3391 indicates the function accepts a variable number of arguments, the
3392 extra arguments can be specified.
3393 #. '``normal label``': the label reached when the called function
3394 executes a '``ret``' instruction.
3395 #. '``exception label``': the label reached when a callee returns via
3396 the :ref:`resume <i_resume>` instruction or other exception handling
3398 #. The optional :ref:`function attributes <fnattrs>` list. Only
3399 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
3400 attributes are valid here.
3405 This instruction is designed to operate as a standard '``call``'
3406 instruction in most regards. The primary difference is that it
3407 establishes an association with a label, which is used by the runtime
3408 library to unwind the stack.
3410 This instruction is used in languages with destructors to ensure that
3411 proper cleanup is performed in the case of either a ``longjmp`` or a
3412 thrown exception. Additionally, this is important for implementation of
3413 '``catch``' clauses in high-level languages that support them.
3415 For the purposes of the SSA form, the definition of the value returned
3416 by the '``invoke``' instruction is deemed to occur on the edge from the
3417 current block to the "normal" label. If the callee unwinds then no
3418 return value is available.
3423 .. code-block:: llvm
3425 %retval = invoke i32 @Test(i32 15) to label %Continue
3426 unwind label %TestCleanup ; {i32}:retval set
3427 %retval = invoke coldcc i32 %Testfnptr(i32 15) to label %Continue
3428 unwind label %TestCleanup ; {i32}:retval set
3432 '``resume``' Instruction
3433 ^^^^^^^^^^^^^^^^^^^^^^^^
3440 resume <type> <value>
3445 The '``resume``' instruction is a terminator instruction that has no
3451 The '``resume``' instruction requires one argument, which must have the
3452 same type as the result of any '``landingpad``' instruction in the same
3458 The '``resume``' instruction resumes propagation of an existing
3459 (in-flight) exception whose unwinding was interrupted with a
3460 :ref:`landingpad <i_landingpad>` instruction.
3465 .. code-block:: llvm
3467 resume { i8*, i32 } %exn
3471 '``unreachable``' Instruction
3472 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3484 The '``unreachable``' instruction has no defined semantics. This
3485 instruction is used to inform the optimizer that a particular portion of
3486 the code is not reachable. This can be used to indicate that the code
3487 after a no-return function cannot be reached, and other facts.
3492 The '``unreachable``' instruction has no defined semantics.
3499 Binary operators are used to do most of the computation in a program.
3500 They require two operands of the same type, execute an operation on
3501 them, and produce a single value. The operands might represent multiple
3502 data, as is the case with the :ref:`vector <t_vector>` data type. The
3503 result value has the same type as its operands.
3505 There are several different binary operators:
3509 '``add``' Instruction
3510 ^^^^^^^^^^^^^^^^^^^^^
3517 <result> = add <ty> <op1>, <op2> ; yields {ty}:result
3518 <result> = add nuw <ty> <op1>, <op2> ; yields {ty}:result
3519 <result> = add nsw <ty> <op1>, <op2> ; yields {ty}:result
3520 <result> = add nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3525 The '``add``' instruction returns the sum of its two operands.
3530 The two arguments to the '``add``' instruction must be
3531 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3532 arguments must have identical types.
3537 The value produced is the integer sum of the two operands.
3539 If the sum has unsigned overflow, the result returned is the
3540 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
3543 Because LLVM integers use a two's complement representation, this
3544 instruction is appropriate for both signed and unsigned integers.
3546 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3547 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3548 result value of the ``add`` is a :ref:`poison value <poisonvalues>` if
3549 unsigned and/or signed overflow, respectively, occurs.
3554 .. code-block:: llvm
3556 <result> = add i32 4, %var ; yields {i32}:result = 4 + %var
3560 '``fadd``' Instruction
3561 ^^^^^^^^^^^^^^^^^^^^^^
3568 <result> = fadd [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3573 The '``fadd``' instruction returns the sum of its two operands.
3578 The two arguments to the '``fadd``' instruction must be :ref:`floating
3579 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3580 Both arguments must have identical types.
3585 The value produced is the floating point sum of the two operands. This
3586 instruction can also take any number of :ref:`fast-math flags <fastmath>`,
3587 which are optimization hints to enable otherwise unsafe floating point
3593 .. code-block:: llvm
3595 <result> = fadd float 4.0, %var ; yields {float}:result = 4.0 + %var
3597 '``sub``' Instruction
3598 ^^^^^^^^^^^^^^^^^^^^^
3605 <result> = sub <ty> <op1>, <op2> ; yields {ty}:result
3606 <result> = sub nuw <ty> <op1>, <op2> ; yields {ty}:result
3607 <result> = sub nsw <ty> <op1>, <op2> ; yields {ty}:result
3608 <result> = sub nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3613 The '``sub``' instruction returns the difference of its two operands.
3615 Note that the '``sub``' instruction is used to represent the '``neg``'
3616 instruction present in most other intermediate representations.
3621 The two arguments to the '``sub``' instruction must be
3622 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3623 arguments must have identical types.
3628 The value produced is the integer difference of the two operands.
3630 If the difference has unsigned overflow, the result returned is the
3631 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
3634 Because LLVM integers use a two's complement representation, this
3635 instruction is appropriate for both signed and unsigned integers.
3637 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3638 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3639 result value of the ``sub`` is a :ref:`poison value <poisonvalues>` if
3640 unsigned and/or signed overflow, respectively, occurs.
3645 .. code-block:: llvm
3647 <result> = sub i32 4, %var ; yields {i32}:result = 4 - %var
3648 <result> = sub i32 0, %val ; yields {i32}:result = -%var
3652 '``fsub``' Instruction
3653 ^^^^^^^^^^^^^^^^^^^^^^
3660 <result> = fsub [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3665 The '``fsub``' instruction returns the difference of its two operands.
3667 Note that the '``fsub``' instruction is used to represent the '``fneg``'
3668 instruction present in most other intermediate representations.
3673 The two arguments to the '``fsub``' instruction must be :ref:`floating
3674 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3675 Both arguments must have identical types.
3680 The value produced is the floating point difference of the two operands.
3681 This instruction can also take any number of :ref:`fast-math
3682 flags <fastmath>`, which are optimization hints to enable otherwise
3683 unsafe floating point optimizations:
3688 .. code-block:: llvm
3690 <result> = fsub float 4.0, %var ; yields {float}:result = 4.0 - %var
3691 <result> = fsub float -0.0, %val ; yields {float}:result = -%var
3693 '``mul``' Instruction
3694 ^^^^^^^^^^^^^^^^^^^^^
3701 <result> = mul <ty> <op1>, <op2> ; yields {ty}:result
3702 <result> = mul nuw <ty> <op1>, <op2> ; yields {ty}:result
3703 <result> = mul nsw <ty> <op1>, <op2> ; yields {ty}:result
3704 <result> = mul nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3709 The '``mul``' instruction returns the product of its two operands.
3714 The two arguments to the '``mul``' instruction must be
3715 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3716 arguments must have identical types.
3721 The value produced is the integer product of the two operands.
3723 If the result of the multiplication has unsigned overflow, the result
3724 returned is the mathematical result modulo 2\ :sup:`n`\ , where n is the
3725 bit width of the result.
3727 Because LLVM integers use a two's complement representation, and the
3728 result is the same width as the operands, this instruction returns the
3729 correct result for both signed and unsigned integers. If a full product
3730 (e.g. ``i32`` * ``i32`` -> ``i64``) is needed, the operands should be
3731 sign-extended or zero-extended as appropriate to the width of the full
3734 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3735 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3736 result value of the ``mul`` is a :ref:`poison value <poisonvalues>` if
3737 unsigned and/or signed overflow, respectively, occurs.
3742 .. code-block:: llvm
3744 <result> = mul i32 4, %var ; yields {i32}:result = 4 * %var
3748 '``fmul``' Instruction
3749 ^^^^^^^^^^^^^^^^^^^^^^
3756 <result> = fmul [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3761 The '``fmul``' instruction returns the product of its two operands.
3766 The two arguments to the '``fmul``' instruction must be :ref:`floating
3767 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3768 Both arguments must have identical types.
3773 The value produced is the floating point product of the two operands.
3774 This instruction can also take any number of :ref:`fast-math
3775 flags <fastmath>`, which are optimization hints to enable otherwise
3776 unsafe floating point optimizations:
3781 .. code-block:: llvm
3783 <result> = fmul float 4.0, %var ; yields {float}:result = 4.0 * %var
3785 '``udiv``' Instruction
3786 ^^^^^^^^^^^^^^^^^^^^^^
3793 <result> = udiv <ty> <op1>, <op2> ; yields {ty}:result
3794 <result> = udiv exact <ty> <op1>, <op2> ; yields {ty}:result
3799 The '``udiv``' instruction returns the quotient of its two operands.
3804 The two arguments to the '``udiv``' instruction must be
3805 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3806 arguments must have identical types.
3811 The value produced is the unsigned integer quotient of the two operands.
3813 Note that unsigned integer division and signed integer division are
3814 distinct operations; for signed integer division, use '``sdiv``'.
3816 Division by zero leads to undefined behavior.
3818 If the ``exact`` keyword is present, the result value of the ``udiv`` is
3819 a :ref:`poison value <poisonvalues>` if %op1 is not a multiple of %op2 (as
3820 such, "((a udiv exact b) mul b) == a").
3825 .. code-block:: llvm
3827 <result> = udiv i32 4, %var ; yields {i32}:result = 4 / %var
3829 '``sdiv``' Instruction
3830 ^^^^^^^^^^^^^^^^^^^^^^
3837 <result> = sdiv <ty> <op1>, <op2> ; yields {ty}:result
3838 <result> = sdiv exact <ty> <op1>, <op2> ; yields {ty}:result
3843 The '``sdiv``' instruction returns the quotient of its two operands.
3848 The two arguments to the '``sdiv``' instruction must be
3849 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3850 arguments must have identical types.
3855 The value produced is the signed integer quotient of the two operands
3856 rounded towards zero.
3858 Note that signed integer division and unsigned integer division are
3859 distinct operations; for unsigned integer division, use '``udiv``'.
3861 Division by zero leads to undefined behavior. Overflow also leads to
3862 undefined behavior; this is a rare case, but can occur, for example, by
3863 doing a 32-bit division of -2147483648 by -1.
3865 If the ``exact`` keyword is present, the result value of the ``sdiv`` is
3866 a :ref:`poison value <poisonvalues>` if the result would be rounded.
3871 .. code-block:: llvm
3873 <result> = sdiv i32 4, %var ; yields {i32}:result = 4 / %var
3877 '``fdiv``' Instruction
3878 ^^^^^^^^^^^^^^^^^^^^^^
3885 <result> = fdiv [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3890 The '``fdiv``' instruction returns the quotient of its two operands.
3895 The two arguments to the '``fdiv``' instruction must be :ref:`floating
3896 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3897 Both arguments must have identical types.
3902 The value produced is the floating point quotient of the two operands.
3903 This instruction can also take any number of :ref:`fast-math
3904 flags <fastmath>`, which are optimization hints to enable otherwise
3905 unsafe floating point optimizations:
3910 .. code-block:: llvm
3912 <result> = fdiv float 4.0, %var ; yields {float}:result = 4.0 / %var
3914 '``urem``' Instruction
3915 ^^^^^^^^^^^^^^^^^^^^^^
3922 <result> = urem <ty> <op1>, <op2> ; yields {ty}:result
3927 The '``urem``' instruction returns the remainder from the unsigned
3928 division of its two arguments.
3933 The two arguments to the '``urem``' instruction must be
3934 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3935 arguments must have identical types.
3940 This instruction returns the unsigned integer *remainder* of a division.
3941 This instruction always performs an unsigned division to get the
3944 Note that unsigned integer remainder and signed integer remainder are
3945 distinct operations; for signed integer remainder, use '``srem``'.
3947 Taking the remainder of a division by zero leads to undefined behavior.
3952 .. code-block:: llvm
3954 <result> = urem i32 4, %var ; yields {i32}:result = 4 % %var
3956 '``srem``' Instruction
3957 ^^^^^^^^^^^^^^^^^^^^^^
3964 <result> = srem <ty> <op1>, <op2> ; yields {ty}:result
3969 The '``srem``' instruction returns the remainder from the signed
3970 division of its two operands. This instruction can also take
3971 :ref:`vector <t_vector>` versions of the values in which case the elements
3977 The two arguments to the '``srem``' instruction must be
3978 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3979 arguments must have identical types.
3984 This instruction returns the *remainder* of a division (where the result
3985 is either zero or has the same sign as the dividend, ``op1``), not the
3986 *modulo* operator (where the result is either zero or has the same sign
3987 as the divisor, ``op2``) of a value. For more information about the
3988 difference, see `The Math
3989 Forum <http://mathforum.org/dr.math/problems/anne.4.28.99.html>`_. For a
3990 table of how this is implemented in various languages, please see
3992 operation <http://en.wikipedia.org/wiki/Modulo_operation>`_.
3994 Note that signed integer remainder and unsigned integer remainder are
3995 distinct operations; for unsigned integer remainder, use '``urem``'.
3997 Taking the remainder of a division by zero leads to undefined behavior.
3998 Overflow also leads to undefined behavior; this is a rare case, but can
3999 occur, for example, by taking the remainder of a 32-bit division of
4000 -2147483648 by -1. (The remainder doesn't actually overflow, but this
4001 rule lets srem be implemented using instructions that return both the
4002 result of the division and the remainder.)
4007 .. code-block:: llvm
4009 <result> = srem i32 4, %var ; yields {i32}:result = 4 % %var
4013 '``frem``' Instruction
4014 ^^^^^^^^^^^^^^^^^^^^^^
4021 <result> = frem [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
4026 The '``frem``' instruction returns the remainder from the division of
4032 The two arguments to the '``frem``' instruction must be :ref:`floating
4033 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
4034 Both arguments must have identical types.
4039 This instruction returns the *remainder* of a division. The remainder
4040 has the same sign as the dividend. This instruction can also take any
4041 number of :ref:`fast-math flags <fastmath>`, which are optimization hints
4042 to enable otherwise unsafe floating point optimizations:
4047 .. code-block:: llvm
4049 <result> = frem float 4.0, %var ; yields {float}:result = 4.0 % %var
4053 Bitwise Binary Operations
4054 -------------------------
4056 Bitwise binary operators are used to do various forms of bit-twiddling
4057 in a program. They are generally very efficient instructions and can
4058 commonly be strength reduced from other instructions. They require two
4059 operands of the same type, execute an operation on them, and produce a
4060 single value. The resulting value is the same type as its operands.
4062 '``shl``' Instruction
4063 ^^^^^^^^^^^^^^^^^^^^^
4070 <result> = shl <ty> <op1>, <op2> ; yields {ty}:result
4071 <result> = shl nuw <ty> <op1>, <op2> ; yields {ty}:result
4072 <result> = shl nsw <ty> <op1>, <op2> ; yields {ty}:result
4073 <result> = shl nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
4078 The '``shl``' instruction returns the first operand shifted to the left
4079 a specified number of bits.
4084 Both arguments to the '``shl``' instruction must be the same
4085 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4086 '``op2``' is treated as an unsigned value.
4091 The value produced is ``op1`` \* 2\ :sup:`op2` mod 2\ :sup:`n`,
4092 where ``n`` is the width of the result. If ``op2`` is (statically or
4093 dynamically) negative or equal to or larger than the number of bits in
4094 ``op1``, the result is undefined. If the arguments are vectors, each
4095 vector element of ``op1`` is shifted by the corresponding shift amount
4098 If the ``nuw`` keyword is present, then the shift produces a :ref:`poison
4099 value <poisonvalues>` if it shifts out any non-zero bits. If the
4100 ``nsw`` keyword is present, then the shift produces a :ref:`poison
4101 value <poisonvalues>` if it shifts out any bits that disagree with the
4102 resultant sign bit. As such, NUW/NSW have the same semantics as they
4103 would if the shift were expressed as a mul instruction with the same
4104 nsw/nuw bits in (mul %op1, (shl 1, %op2)).
4109 .. code-block:: llvm
4111 <result> = shl i32 4, %var ; yields {i32}: 4 << %var
4112 <result> = shl i32 4, 2 ; yields {i32}: 16
4113 <result> = shl i32 1, 10 ; yields {i32}: 1024
4114 <result> = shl i32 1, 32 ; undefined
4115 <result> = shl <2 x i32> < i32 1, i32 1>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 2, i32 4>
4117 '``lshr``' Instruction
4118 ^^^^^^^^^^^^^^^^^^^^^^
4125 <result> = lshr <ty> <op1>, <op2> ; yields {ty}:result
4126 <result> = lshr exact <ty> <op1>, <op2> ; yields {ty}:result
4131 The '``lshr``' instruction (logical shift right) returns the first
4132 operand shifted to the right a specified number of bits with zero fill.
4137 Both arguments to the '``lshr``' instruction must be the same
4138 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4139 '``op2``' is treated as an unsigned value.
4144 This instruction always performs a logical shift right operation. The
4145 most significant bits of the result will be filled with zero bits after
4146 the shift. If ``op2`` is (statically or dynamically) equal to or larger
4147 than the number of bits in ``op1``, the result is undefined. If the
4148 arguments are vectors, each vector element of ``op1`` is shifted by the
4149 corresponding shift amount in ``op2``.
4151 If the ``exact`` keyword is present, the result value of the ``lshr`` is
4152 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
4158 .. code-block:: llvm
4160 <result> = lshr i32 4, 1 ; yields {i32}:result = 2
4161 <result> = lshr i32 4, 2 ; yields {i32}:result = 1
4162 <result> = lshr i8 4, 3 ; yields {i8}:result = 0
4163 <result> = lshr i8 -2, 1 ; yields {i8}:result = 0x7F
4164 <result> = lshr i32 1, 32 ; undefined
4165 <result> = lshr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 0x7FFFFFFF, i32 1>
4167 '``ashr``' Instruction
4168 ^^^^^^^^^^^^^^^^^^^^^^
4175 <result> = ashr <ty> <op1>, <op2> ; yields {ty}:result
4176 <result> = ashr exact <ty> <op1>, <op2> ; yields {ty}:result
4181 The '``ashr``' instruction (arithmetic shift right) returns the first
4182 operand shifted to the right a specified number of bits with sign
4188 Both arguments to the '``ashr``' instruction must be the same
4189 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4190 '``op2``' is treated as an unsigned value.
4195 This instruction always performs an arithmetic shift right operation,
4196 The most significant bits of the result will be filled with the sign bit
4197 of ``op1``. If ``op2`` is (statically or dynamically) equal to or larger
4198 than the number of bits in ``op1``, the result is undefined. If the
4199 arguments are vectors, each vector element of ``op1`` is shifted by the
4200 corresponding shift amount in ``op2``.
4202 If the ``exact`` keyword is present, the result value of the ``ashr`` is
4203 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
4209 .. code-block:: llvm
4211 <result> = ashr i32 4, 1 ; yields {i32}:result = 2
4212 <result> = ashr i32 4, 2 ; yields {i32}:result = 1
4213 <result> = ashr i8 4, 3 ; yields {i8}:result = 0
4214 <result> = ashr i8 -2, 1 ; yields {i8}:result = -1
4215 <result> = ashr i32 1, 32 ; undefined
4216 <result> = ashr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 3> ; yields: result=<2 x i32> < i32 -1, i32 0>
4218 '``and``' Instruction
4219 ^^^^^^^^^^^^^^^^^^^^^
4226 <result> = and <ty> <op1>, <op2> ; yields {ty}:result
4231 The '``and``' instruction returns the bitwise logical and of its two
4237 The two arguments to the '``and``' instruction must be
4238 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4239 arguments must have identical types.
4244 The truth table used for the '``and``' instruction is:
4261 .. code-block:: llvm
4263 <result> = and i32 4, %var ; yields {i32}:result = 4 & %var
4264 <result> = and i32 15, 40 ; yields {i32}:result = 8
4265 <result> = and i32 4, 8 ; yields {i32}:result = 0
4267 '``or``' Instruction
4268 ^^^^^^^^^^^^^^^^^^^^
4275 <result> = or <ty> <op1>, <op2> ; yields {ty}:result
4280 The '``or``' instruction returns the bitwise logical inclusive or of its
4286 The two arguments to the '``or``' instruction must be
4287 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4288 arguments must have identical types.
4293 The truth table used for the '``or``' instruction is:
4312 <result> = or i32 4, %var ; yields {i32}:result = 4 | %var
4313 <result> = or i32 15, 40 ; yields {i32}:result = 47
4314 <result> = or i32 4, 8 ; yields {i32}:result = 12
4316 '``xor``' Instruction
4317 ^^^^^^^^^^^^^^^^^^^^^
4324 <result> = xor <ty> <op1>, <op2> ; yields {ty}:result
4329 The '``xor``' instruction returns the bitwise logical exclusive or of
4330 its two operands. The ``xor`` is used to implement the "one's
4331 complement" operation, which is the "~" operator in C.
4336 The two arguments to the '``xor``' instruction must be
4337 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4338 arguments must have identical types.
4343 The truth table used for the '``xor``' instruction is:
4360 .. code-block:: llvm
4362 <result> = xor i32 4, %var ; yields {i32}:result = 4 ^ %var
4363 <result> = xor i32 15, 40 ; yields {i32}:result = 39
4364 <result> = xor i32 4, 8 ; yields {i32}:result = 12
4365 <result> = xor i32 %V, -1 ; yields {i32}:result = ~%V
4370 LLVM supports several instructions to represent vector operations in a
4371 target-independent manner. These instructions cover the element-access
4372 and vector-specific operations needed to process vectors effectively.
4373 While LLVM does directly support these vector operations, many
4374 sophisticated algorithms will want to use target-specific intrinsics to
4375 take full advantage of a specific target.
4377 .. _i_extractelement:
4379 '``extractelement``' Instruction
4380 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4387 <result> = extractelement <n x <ty>> <val>, i32 <idx> ; yields <ty>
4392 The '``extractelement``' instruction extracts a single scalar element
4393 from a vector at a specified index.
4398 The first operand of an '``extractelement``' instruction is a value of
4399 :ref:`vector <t_vector>` type. The second operand is an index indicating
4400 the position from which to extract the element. The index may be a
4406 The result is a scalar of the same type as the element type of ``val``.
4407 Its value is the value at position ``idx`` of ``val``. If ``idx``
4408 exceeds the length of ``val``, the results are undefined.
4413 .. code-block:: llvm
4415 <result> = extractelement <4 x i32> %vec, i32 0 ; yields i32
4417 .. _i_insertelement:
4419 '``insertelement``' Instruction
4420 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4427 <result> = insertelement <n x <ty>> <val>, <ty> <elt>, i32 <idx> ; yields <n x <ty>>
4432 The '``insertelement``' instruction inserts a scalar element into a
4433 vector at a specified index.
4438 The first operand of an '``insertelement``' instruction is a value of
4439 :ref:`vector <t_vector>` type. The second operand is a scalar value whose
4440 type must equal the element type of the first operand. The third operand
4441 is an index indicating the position at which to insert the value. The
4442 index may be a variable.
4447 The result is a vector of the same type as ``val``. Its element values
4448 are those of ``val`` except at position ``idx``, where it gets the value
4449 ``elt``. If ``idx`` exceeds the length of ``val``, the results are
4455 .. code-block:: llvm
4457 <result> = insertelement <4 x i32> %vec, i32 1, i32 0 ; yields <4 x i32>
4459 .. _i_shufflevector:
4461 '``shufflevector``' Instruction
4462 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4469 <result> = shufflevector <n x <ty>> <v1>, <n x <ty>> <v2>, <m x i32> <mask> ; yields <m x <ty>>
4474 The '``shufflevector``' instruction constructs a permutation of elements
4475 from two input vectors, returning a vector with the same element type as
4476 the input and length that is the same as the shuffle mask.
4481 The first two operands of a '``shufflevector``' instruction are vectors
4482 with the same type. The third argument is a shuffle mask whose element
4483 type is always 'i32'. The result of the instruction is a vector whose
4484 length is the same as the shuffle mask and whose element type is the
4485 same as the element type of the first two operands.
4487 The shuffle mask operand is required to be a constant vector with either
4488 constant integer or undef values.
4493 The elements of the two input vectors are numbered from left to right
4494 across both of the vectors. The shuffle mask operand specifies, for each
4495 element of the result vector, which element of the two input vectors the
4496 result element gets. The element selector may be undef (meaning "don't
4497 care") and the second operand may be undef if performing a shuffle from
4503 .. code-block:: llvm
4505 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
4506 <4 x i32> <i32 0, i32 4, i32 1, i32 5> ; yields <4 x i32>
4507 <result> = shufflevector <4 x i32> %v1, <4 x i32> undef,
4508 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32> - Identity shuffle.
4509 <result> = shufflevector <8 x i32> %v1, <8 x i32> undef,
4510 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32>
4511 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
4512 <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7 > ; yields <8 x i32>
4514 Aggregate Operations
4515 --------------------
4517 LLVM supports several instructions for working with
4518 :ref:`aggregate <t_aggregate>` values.
4522 '``extractvalue``' Instruction
4523 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4530 <result> = extractvalue <aggregate type> <val>, <idx>{, <idx>}*
4535 The '``extractvalue``' instruction extracts the value of a member field
4536 from an :ref:`aggregate <t_aggregate>` value.
4541 The first operand of an '``extractvalue``' instruction is a value of
4542 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The operands are
4543 constant indices to specify which value to extract in a similar manner
4544 as indices in a '``getelementptr``' instruction.
4546 The major differences to ``getelementptr`` indexing are:
4548 - Since the value being indexed is not a pointer, the first index is
4549 omitted and assumed to be zero.
4550 - At least one index must be specified.
4551 - Not only struct indices but also array indices must be in bounds.
4556 The result is the value at the position in the aggregate specified by
4562 .. code-block:: llvm
4564 <result> = extractvalue {i32, float} %agg, 0 ; yields i32
4568 '``insertvalue``' Instruction
4569 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4576 <result> = insertvalue <aggregate type> <val>, <ty> <elt>, <idx>{, <idx>}* ; yields <aggregate type>
4581 The '``insertvalue``' instruction inserts a value into a member field in
4582 an :ref:`aggregate <t_aggregate>` value.
4587 The first operand of an '``insertvalue``' instruction is a value of
4588 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The second operand is
4589 a first-class value to insert. The following operands are constant
4590 indices indicating the position at which to insert the value in a
4591 similar manner as indices in a '``extractvalue``' instruction. The value
4592 to insert must have the same type as the value identified by the
4598 The result is an aggregate of the same type as ``val``. Its value is
4599 that of ``val`` except that the value at the position specified by the
4600 indices is that of ``elt``.
4605 .. code-block:: llvm
4607 %agg1 = insertvalue {i32, float} undef, i32 1, 0 ; yields {i32 1, float undef}
4608 %agg2 = insertvalue {i32, float} %agg1, float %val, 1 ; yields {i32 1, float %val}
4609 %agg3 = insertvalue {i32, {float}} %agg1, float %val, 1, 0 ; yields {i32 1, float %val}
4613 Memory Access and Addressing Operations
4614 ---------------------------------------
4616 A key design point of an SSA-based representation is how it represents
4617 memory. In LLVM, no memory locations are in SSA form, which makes things
4618 very simple. This section describes how to read, write, and allocate
4623 '``alloca``' Instruction
4624 ^^^^^^^^^^^^^^^^^^^^^^^^
4631 <result> = alloca <type>[, <ty> <NumElements>][, align <alignment>] ; yields {type*}:result
4636 The '``alloca``' instruction allocates memory on the stack frame of the
4637 currently executing function, to be automatically released when this
4638 function returns to its caller. The object is always allocated in the
4639 generic address space (address space zero).
4644 The '``alloca``' instruction allocates ``sizeof(<type>)*NumElements``
4645 bytes of memory on the runtime stack, returning a pointer of the
4646 appropriate type to the program. If "NumElements" is specified, it is
4647 the number of elements allocated, otherwise "NumElements" is defaulted
4648 to be one. If a constant alignment is specified, the value result of the
4649 allocation is guaranteed to be aligned to at least that boundary. If not
4650 specified, or if zero, the target can choose to align the allocation on
4651 any convenient boundary compatible with the type.
4653 '``type``' may be any sized type.
4658 Memory is allocated; a pointer is returned. The operation is undefined
4659 if there is insufficient stack space for the allocation. '``alloca``'d
4660 memory is automatically released when the function returns. The
4661 '``alloca``' instruction is commonly used to represent automatic
4662 variables that must have an address available. When the function returns
4663 (either with the ``ret`` or ``resume`` instructions), the memory is
4664 reclaimed. Allocating zero bytes is legal, but the result is undefined.
4665 The order in which memory is allocated (ie., which way the stack grows)
4671 .. code-block:: llvm
4673 %ptr = alloca i32 ; yields {i32*}:ptr
4674 %ptr = alloca i32, i32 4 ; yields {i32*}:ptr
4675 %ptr = alloca i32, i32 4, align 1024 ; yields {i32*}:ptr
4676 %ptr = alloca i32, align 1024 ; yields {i32*}:ptr
4680 '``load``' Instruction
4681 ^^^^^^^^^^^^^^^^^^^^^^
4688 <result> = load [volatile] <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>][, !invariant.load !<index>]
4689 <result> = load atomic [volatile] <ty>* <pointer> [singlethread] <ordering>, align <alignment>
4690 !<index> = !{ i32 1 }
4695 The '``load``' instruction is used to read from memory.
4700 The argument to the ``load`` instruction specifies the memory address
4701 from which to load. The pointer must point to a :ref:`first
4702 class <t_firstclass>` type. If the ``load`` is marked as ``volatile``,
4703 then the optimizer is not allowed to modify the number or order of
4704 execution of this ``load`` with other :ref:`volatile
4705 operations <volatile>`.
4707 If the ``load`` is marked as ``atomic``, it takes an extra
4708 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
4709 ``release`` and ``acq_rel`` orderings are not valid on ``load``
4710 instructions. Atomic loads produce :ref:`defined <memmodel>` results
4711 when they may see multiple atomic stores. The type of the pointee must
4712 be an integer type whose bit width is a power of two greater than or
4713 equal to eight and less than or equal to a target-specific size limit.
4714 ``align`` must be explicitly specified on atomic loads, and the load has
4715 undefined behavior if the alignment is not set to a value which is at
4716 least the size in bytes of the pointee. ``!nontemporal`` does not have
4717 any defined semantics for atomic loads.
4719 The optional constant ``align`` argument specifies the alignment of the
4720 operation (that is, the alignment of the memory address). A value of 0
4721 or an omitted ``align`` argument means that the operation has the ABI
4722 alignment for the target. It is the responsibility of the code emitter
4723 to ensure that the alignment information is correct. Overestimating the
4724 alignment results in undefined behavior. Underestimating the alignment
4725 may produce less efficient code. An alignment of 1 is always safe.
4727 The optional ``!nontemporal`` metadata must reference a single
4728 metadata name ``<index>`` corresponding to a metadata node with one
4729 ``i32`` entry of value 1. The existence of the ``!nontemporal``
4730 metadata on the instruction tells the optimizer and code generator
4731 that this load is not expected to be reused in the cache. The code
4732 generator may select special instructions to save cache bandwidth, such
4733 as the ``MOVNT`` instruction on x86.
4735 The optional ``!invariant.load`` metadata must reference a single
4736 metadata name ``<index>`` corresponding to a metadata node with no
4737 entries. The existence of the ``!invariant.load`` metadata on the
4738 instruction tells the optimizer and code generator that this load
4739 address points to memory which does not change value during program
4740 execution. The optimizer may then move this load around, for example, by
4741 hoisting it out of loops using loop invariant code motion.
4746 The location of memory pointed to is loaded. If the value being loaded
4747 is of scalar type then the number of bytes read does not exceed the
4748 minimum number of bytes needed to hold all bits of the type. For
4749 example, loading an ``i24`` reads at most three bytes. When loading a
4750 value of a type like ``i20`` with a size that is not an integral number
4751 of bytes, the result is undefined if the value was not originally
4752 written using a store of the same type.
4757 .. code-block:: llvm
4759 %ptr = alloca i32 ; yields {i32*}:ptr
4760 store i32 3, i32* %ptr ; yields {void}
4761 %val = load i32* %ptr ; yields {i32}:val = i32 3
4765 '``store``' Instruction
4766 ^^^^^^^^^^^^^^^^^^^^^^^
4773 store [volatile] <ty> <value>, <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>] ; yields {void}
4774 store atomic [volatile] <ty> <value>, <ty>* <pointer> [singlethread] <ordering>, align <alignment> ; yields {void}
4779 The '``store``' instruction is used to write to memory.
4784 There are two arguments to the ``store`` instruction: a value to store
4785 and an address at which to store it. The type of the ``<pointer>``
4786 operand must be a pointer to the :ref:`first class <t_firstclass>` type of
4787 the ``<value>`` operand. If the ``store`` is marked as ``volatile``,
4788 then the optimizer is not allowed to modify the number or order of
4789 execution of this ``store`` with other :ref:`volatile
4790 operations <volatile>`.
4792 If the ``store`` is marked as ``atomic``, it takes an extra
4793 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
4794 ``acquire`` and ``acq_rel`` orderings aren't valid on ``store``
4795 instructions. Atomic loads produce :ref:`defined <memmodel>` results
4796 when they may see multiple atomic stores. The type of the pointee must
4797 be an integer type whose bit width is a power of two greater than or
4798 equal to eight and less than or equal to a target-specific size limit.
4799 ``align`` must be explicitly specified on atomic stores, and the store
4800 has undefined behavior if the alignment is not set to a value which is
4801 at least the size in bytes of the pointee. ``!nontemporal`` does not
4802 have any defined semantics for atomic stores.
4804 The optional constant ``align`` argument specifies the alignment of the
4805 operation (that is, the alignment of the memory address). A value of 0
4806 or an omitted ``align`` argument means that the operation has the ABI
4807 alignment for the target. It is the responsibility of the code emitter
4808 to ensure that the alignment information is correct. Overestimating the
4809 alignment results in undefined behavior. Underestimating the
4810 alignment may produce less efficient code. An alignment of 1 is always
4813 The optional ``!nontemporal`` metadata must reference a single metadata
4814 name ``<index>`` corresponding to a metadata node with one ``i32`` entry of
4815 value 1. The existence of the ``!nontemporal`` metadata on the instruction
4816 tells the optimizer and code generator that this load is not expected to
4817 be reused in the cache. The code generator may select special
4818 instructions to save cache bandwidth, such as the MOVNT instruction on
4824 The contents of memory are updated to contain ``<value>`` at the
4825 location specified by the ``<pointer>`` operand. If ``<value>`` is
4826 of scalar type then the number of bytes written does not exceed the
4827 minimum number of bytes needed to hold all bits of the type. For
4828 example, storing an ``i24`` writes at most three bytes. When writing a
4829 value of a type like ``i20`` with a size that is not an integral number
4830 of bytes, it is unspecified what happens to the extra bits that do not
4831 belong to the type, but they will typically be overwritten.
4836 .. code-block:: llvm
4838 %ptr = alloca i32 ; yields {i32*}:ptr
4839 store i32 3, i32* %ptr ; yields {void}
4840 %val = load i32* %ptr ; yields {i32}:val = i32 3
4844 '``fence``' Instruction
4845 ^^^^^^^^^^^^^^^^^^^^^^^
4852 fence [singlethread] <ordering> ; yields {void}
4857 The '``fence``' instruction is used to introduce happens-before edges
4863 '``fence``' instructions take an :ref:`ordering <ordering>` argument which
4864 defines what *synchronizes-with* edges they add. They can only be given
4865 ``acquire``, ``release``, ``acq_rel``, and ``seq_cst`` orderings.
4870 A fence A which has (at least) ``release`` ordering semantics
4871 *synchronizes with* a fence B with (at least) ``acquire`` ordering
4872 semantics if and only if there exist atomic operations X and Y, both
4873 operating on some atomic object M, such that A is sequenced before X, X
4874 modifies M (either directly or through some side effect of a sequence
4875 headed by X), Y is sequenced before B, and Y observes M. This provides a
4876 *happens-before* dependency between A and B. Rather than an explicit
4877 ``fence``, one (but not both) of the atomic operations X or Y might
4878 provide a ``release`` or ``acquire`` (resp.) ordering constraint and
4879 still *synchronize-with* the explicit ``fence`` and establish the
4880 *happens-before* edge.
4882 A ``fence`` which has ``seq_cst`` ordering, in addition to having both
4883 ``acquire`` and ``release`` semantics specified above, participates in
4884 the global program order of other ``seq_cst`` operations and/or fences.
4886 The optional ":ref:`singlethread <singlethread>`" argument specifies
4887 that the fence only synchronizes with other fences in the same thread.
4888 (This is useful for interacting with signal handlers.)
4893 .. code-block:: llvm
4895 fence acquire ; yields {void}
4896 fence singlethread seq_cst ; yields {void}
4900 '``cmpxchg``' Instruction
4901 ^^^^^^^^^^^^^^^^^^^^^^^^^
4908 cmpxchg [volatile] <ty>* <pointer>, <ty> <cmp>, <ty> <new> [singlethread] <ordering> ; yields {ty}
4913 The '``cmpxchg``' instruction is used to atomically modify memory. It
4914 loads a value in memory and compares it to a given value. If they are
4915 equal, it stores a new value into the memory.
4920 There are three arguments to the '``cmpxchg``' instruction: an address
4921 to operate on, a value to compare to the value currently be at that
4922 address, and a new value to place at that address if the compared values
4923 are equal. The type of '<cmp>' must be an integer type whose bit width
4924 is a power of two greater than or equal to eight and less than or equal
4925 to a target-specific size limit. '<cmp>' and '<new>' must have the same
4926 type, and the type of '<pointer>' must be a pointer to that type. If the
4927 ``cmpxchg`` is marked as ``volatile``, then the optimizer is not allowed
4928 to modify the number or order of execution of this ``cmpxchg`` with
4929 other :ref:`volatile operations <volatile>`.
4931 The :ref:`ordering <ordering>` argument specifies how this ``cmpxchg``
4932 synchronizes with other atomic operations.
4934 The optional "``singlethread``" argument declares that the ``cmpxchg``
4935 is only atomic with respect to code (usually signal handlers) running in
4936 the same thread as the ``cmpxchg``. Otherwise the cmpxchg is atomic with
4937 respect to all other code in the system.
4939 The pointer passed into cmpxchg must have alignment greater than or
4940 equal to the size in memory of the operand.
4945 The contents of memory at the location specified by the '``<pointer>``'
4946 operand is read and compared to '``<cmp>``'; if the read value is the
4947 equal, '``<new>``' is written. The original value at the location is
4950 A successful ``cmpxchg`` is a read-modify-write instruction for the purpose
4951 of identifying release sequences. A failed ``cmpxchg`` is equivalent to an
4952 atomic load with an ordering parameter determined by dropping any
4953 ``release`` part of the ``cmpxchg``'s ordering.
4958 .. code-block:: llvm
4961 %orig = atomic load i32* %ptr unordered ; yields {i32}
4965 %cmp = phi i32 [ %orig, %entry ], [%old, %loop]
4966 %squared = mul i32 %cmp, %cmp
4967 %old = cmpxchg i32* %ptr, i32 %cmp, i32 %squared ; yields {i32}
4968 %success = icmp eq i32 %cmp, %old
4969 br i1 %success, label %done, label %loop
4976 '``atomicrmw``' Instruction
4977 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
4984 atomicrmw [volatile] <operation> <ty>* <pointer>, <ty> <value> [singlethread] <ordering> ; yields {ty}
4989 The '``atomicrmw``' instruction is used to atomically modify memory.
4994 There are three arguments to the '``atomicrmw``' instruction: an
4995 operation to apply, an address whose value to modify, an argument to the
4996 operation. The operation must be one of the following keywords:
5010 The type of '<value>' must be an integer type whose bit width is a power
5011 of two greater than or equal to eight and less than or equal to a
5012 target-specific size limit. The type of the '``<pointer>``' operand must
5013 be a pointer to that type. If the ``atomicrmw`` is marked as
5014 ``volatile``, then the optimizer is not allowed to modify the number or
5015 order of execution of this ``atomicrmw`` with other :ref:`volatile
5016 operations <volatile>`.
5021 The contents of memory at the location specified by the '``<pointer>``'
5022 operand are atomically read, modified, and written back. The original
5023 value at the location is returned. The modification is specified by the
5026 - xchg: ``*ptr = val``
5027 - add: ``*ptr = *ptr + val``
5028 - sub: ``*ptr = *ptr - val``
5029 - and: ``*ptr = *ptr & val``
5030 - nand: ``*ptr = ~(*ptr & val)``
5031 - or: ``*ptr = *ptr | val``
5032 - xor: ``*ptr = *ptr ^ val``
5033 - max: ``*ptr = *ptr > val ? *ptr : val`` (using a signed comparison)
5034 - min: ``*ptr = *ptr < val ? *ptr : val`` (using a signed comparison)
5035 - umax: ``*ptr = *ptr > val ? *ptr : val`` (using an unsigned
5037 - umin: ``*ptr = *ptr < val ? *ptr : val`` (using an unsigned
5043 .. code-block:: llvm
5045 %old = atomicrmw add i32* %ptr, i32 1 acquire ; yields {i32}
5047 .. _i_getelementptr:
5049 '``getelementptr``' Instruction
5050 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5057 <result> = getelementptr <pty>* <ptrval>{, <ty> <idx>}*
5058 <result> = getelementptr inbounds <pty>* <ptrval>{, <ty> <idx>}*
5059 <result> = getelementptr <ptr vector> ptrval, <vector index type> idx
5064 The '``getelementptr``' instruction is used to get the address of a
5065 subelement of an :ref:`aggregate <t_aggregate>` data structure. It performs
5066 address calculation only and does not access memory.
5071 The first argument is always a pointer or a vector of pointers, and
5072 forms the basis of the calculation. The remaining arguments are indices
5073 that indicate which of the elements of the aggregate object are indexed.
5074 The interpretation of each index is dependent on the type being indexed
5075 into. The first index always indexes the pointer value given as the
5076 first argument, the second index indexes a value of the type pointed to
5077 (not necessarily the value directly pointed to, since the first index
5078 can be non-zero), etc. The first type indexed into must be a pointer
5079 value, subsequent types can be arrays, vectors, and structs. Note that
5080 subsequent types being indexed into can never be pointers, since that
5081 would require loading the pointer before continuing calculation.
5083 The type of each index argument depends on the type it is indexing into.
5084 When indexing into a (optionally packed) structure, only ``i32`` integer
5085 **constants** are allowed (when using a vector of indices they must all
5086 be the **same** ``i32`` integer constant). When indexing into an array,
5087 pointer or vector, integers of any width are allowed, and they are not
5088 required to be constant. These integers are treated as signed values
5091 For example, let's consider a C code fragment and how it gets compiled
5107 int *foo(struct ST *s) {
5108 return &s[1].Z.B[5][13];
5111 The LLVM code generated by Clang is:
5113 .. code-block:: llvm
5115 %struct.RT = type { i8, [10 x [20 x i32]], i8 }
5116 %struct.ST = type { i32, double, %struct.RT }
5118 define i32* @foo(%struct.ST* %s) nounwind uwtable readnone optsize ssp {
5120 %arrayidx = getelementptr inbounds %struct.ST* %s, i64 1, i32 2, i32 1, i64 5, i64 13
5127 In the example above, the first index is indexing into the
5128 '``%struct.ST*``' type, which is a pointer, yielding a '``%struct.ST``'
5129 = '``{ i32, double, %struct.RT }``' type, a structure. The second index
5130 indexes into the third element of the structure, yielding a
5131 '``%struct.RT``' = '``{ i8 , [10 x [20 x i32]], i8 }``' type, another
5132 structure. The third index indexes into the second element of the
5133 structure, yielding a '``[10 x [20 x i32]]``' type, an array. The two
5134 dimensions of the array are subscripted into, yielding an '``i32``'
5135 type. The '``getelementptr``' instruction returns a pointer to this
5136 element, thus computing a value of '``i32*``' type.
5138 Note that it is perfectly legal to index partially through a structure,
5139 returning a pointer to an inner element. Because of this, the LLVM code
5140 for the given testcase is equivalent to:
5142 .. code-block:: llvm
5144 define i32* @foo(%struct.ST* %s) {
5145 %t1 = getelementptr %struct.ST* %s, i32 1 ; yields %struct.ST*:%t1
5146 %t2 = getelementptr %struct.ST* %t1, i32 0, i32 2 ; yields %struct.RT*:%t2
5147 %t3 = getelementptr %struct.RT* %t2, i32 0, i32 1 ; yields [10 x [20 x i32]]*:%t3
5148 %t4 = getelementptr [10 x [20 x i32]]* %t3, i32 0, i32 5 ; yields [20 x i32]*:%t4
5149 %t5 = getelementptr [20 x i32]* %t4, i32 0, i32 13 ; yields i32*:%t5
5153 If the ``inbounds`` keyword is present, the result value of the
5154 ``getelementptr`` is a :ref:`poison value <poisonvalues>` if the base
5155 pointer is not an *in bounds* address of an allocated object, or if any
5156 of the addresses that would be formed by successive addition of the
5157 offsets implied by the indices to the base address with infinitely
5158 precise signed arithmetic are not an *in bounds* address of that
5159 allocated object. The *in bounds* addresses for an allocated object are
5160 all the addresses that point into the object, plus the address one byte
5161 past the end. In cases where the base is a vector of pointers the
5162 ``inbounds`` keyword applies to each of the computations element-wise.
5164 If the ``inbounds`` keyword is not present, the offsets are added to the
5165 base address with silently-wrapping two's complement arithmetic. If the
5166 offsets have a different width from the pointer, they are sign-extended
5167 or truncated to the width of the pointer. The result value of the
5168 ``getelementptr`` may be outside the object pointed to by the base
5169 pointer. The result value may not necessarily be used to access memory
5170 though, even if it happens to point into allocated storage. See the
5171 :ref:`Pointer Aliasing Rules <pointeraliasing>` section for more
5174 The getelementptr instruction is often confusing. For some more insight
5175 into how it works, see :doc:`the getelementptr FAQ <GetElementPtr>`.
5180 .. code-block:: llvm
5182 ; yields [12 x i8]*:aptr
5183 %aptr = getelementptr {i32, [12 x i8]}* %saptr, i64 0, i32 1
5185 %vptr = getelementptr {i32, <2 x i8>}* %svptr, i64 0, i32 1, i32 1
5187 %eptr = getelementptr [12 x i8]* %aptr, i64 0, i32 1
5189 %iptr = getelementptr [10 x i32]* @arr, i16 0, i16 0
5191 In cases where the pointer argument is a vector of pointers, each index
5192 must be a vector with the same number of elements. For example:
5194 .. code-block:: llvm
5196 %A = getelementptr <4 x i8*> %ptrs, <4 x i64> %offsets,
5198 Conversion Operations
5199 ---------------------
5201 The instructions in this category are the conversion instructions
5202 (casting) which all take a single operand and a type. They perform
5203 various bit conversions on the operand.
5205 '``trunc .. to``' Instruction
5206 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5213 <result> = trunc <ty> <value> to <ty2> ; yields ty2
5218 The '``trunc``' instruction truncates its operand to the type ``ty2``.
5223 The '``trunc``' instruction takes a value to trunc, and a type to trunc
5224 it to. Both types must be of :ref:`integer <t_integer>` types, or vectors
5225 of the same number of integers. The bit size of the ``value`` must be
5226 larger than the bit size of the destination type, ``ty2``. Equal sized
5227 types are not allowed.
5232 The '``trunc``' instruction truncates the high order bits in ``value``
5233 and converts the remaining bits to ``ty2``. Since the source size must
5234 be larger than the destination size, ``trunc`` cannot be a *no-op cast*.
5235 It will always truncate bits.
5240 .. code-block:: llvm
5242 %X = trunc i32 257 to i8 ; yields i8:1
5243 %Y = trunc i32 123 to i1 ; yields i1:true
5244 %Z = trunc i32 122 to i1 ; yields i1:false
5245 %W = trunc <2 x i16> <i16 8, i16 7> to <2 x i8> ; yields <i8 8, i8 7>
5247 '``zext .. to``' Instruction
5248 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5255 <result> = zext <ty> <value> to <ty2> ; yields ty2
5260 The '``zext``' instruction zero extends its operand to type ``ty2``.
5265 The '``zext``' instruction takes a value to cast, and a type to cast it
5266 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
5267 the same number of integers. The bit size of the ``value`` must be
5268 smaller than the bit size of the destination type, ``ty2``.
5273 The ``zext`` fills the high order bits of the ``value`` with zero bits
5274 until it reaches the size of the destination type, ``ty2``.
5276 When zero extending from i1, the result will always be either 0 or 1.
5281 .. code-block:: llvm
5283 %X = zext i32 257 to i64 ; yields i64:257
5284 %Y = zext i1 true to i32 ; yields i32:1
5285 %Z = zext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
5287 '``sext .. to``' Instruction
5288 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5295 <result> = sext <ty> <value> to <ty2> ; yields ty2
5300 The '``sext``' sign extends ``value`` to the type ``ty2``.
5305 The '``sext``' instruction takes a value to cast, and a type to cast it
5306 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
5307 the same number of integers. The bit size of the ``value`` must be
5308 smaller than the bit size of the destination type, ``ty2``.
5313 The '``sext``' instruction performs a sign extension by copying the sign
5314 bit (highest order bit) of the ``value`` until it reaches the bit size
5315 of the type ``ty2``.
5317 When sign extending from i1, the extension always results in -1 or 0.
5322 .. code-block:: llvm
5324 %X = sext i8 -1 to i16 ; yields i16 :65535
5325 %Y = sext i1 true to i32 ; yields i32:-1
5326 %Z = sext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
5328 '``fptrunc .. to``' Instruction
5329 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5336 <result> = fptrunc <ty> <value> to <ty2> ; yields ty2
5341 The '``fptrunc``' instruction truncates ``value`` to type ``ty2``.
5346 The '``fptrunc``' instruction takes a :ref:`floating point <t_floating>`
5347 value to cast and a :ref:`floating point <t_floating>` type to cast it to.
5348 The size of ``value`` must be larger than the size of ``ty2``. This
5349 implies that ``fptrunc`` cannot be used to make a *no-op cast*.
5354 The '``fptrunc``' instruction truncates a ``value`` from a larger
5355 :ref:`floating point <t_floating>` type to a smaller :ref:`floating
5356 point <t_floating>` type. If the value cannot fit within the
5357 destination type, ``ty2``, then the results are undefined.
5362 .. code-block:: llvm
5364 %X = fptrunc double 123.0 to float ; yields float:123.0
5365 %Y = fptrunc double 1.0E+300 to float ; yields undefined
5367 '``fpext .. to``' Instruction
5368 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5375 <result> = fpext <ty> <value> to <ty2> ; yields ty2
5380 The '``fpext``' extends a floating point ``value`` to a larger floating
5386 The '``fpext``' instruction takes a :ref:`floating point <t_floating>`
5387 ``value`` to cast, and a :ref:`floating point <t_floating>` type to cast it
5388 to. The source type must be smaller than the destination type.
5393 The '``fpext``' instruction extends the ``value`` from a smaller
5394 :ref:`floating point <t_floating>` type to a larger :ref:`floating
5395 point <t_floating>` type. The ``fpext`` cannot be used to make a
5396 *no-op cast* because it always changes bits. Use ``bitcast`` to make a
5397 *no-op cast* for a floating point cast.
5402 .. code-block:: llvm
5404 %X = fpext float 3.125 to double ; yields double:3.125000e+00
5405 %Y = fpext double %X to fp128 ; yields fp128:0xL00000000000000004000900000000000
5407 '``fptoui .. to``' Instruction
5408 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5415 <result> = fptoui <ty> <value> to <ty2> ; yields ty2
5420 The '``fptoui``' converts a floating point ``value`` to its unsigned
5421 integer equivalent of type ``ty2``.
5426 The '``fptoui``' instruction takes a value to cast, which must be a
5427 scalar or vector :ref:`floating point <t_floating>` value, and a type to
5428 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
5429 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
5430 type with the same number of elements as ``ty``
5435 The '``fptoui``' instruction converts its :ref:`floating
5436 point <t_floating>` operand into the nearest (rounding towards zero)
5437 unsigned integer value. If the value cannot fit in ``ty2``, the results
5443 .. code-block:: llvm
5445 %X = fptoui double 123.0 to i32 ; yields i32:123
5446 %Y = fptoui float 1.0E+300 to i1 ; yields undefined:1
5447 %Z = fptoui float 1.04E+17 to i8 ; yields undefined:1
5449 '``fptosi .. to``' Instruction
5450 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5457 <result> = fptosi <ty> <value> to <ty2> ; yields ty2
5462 The '``fptosi``' instruction converts :ref:`floating point <t_floating>`
5463 ``value`` to type ``ty2``.
5468 The '``fptosi``' instruction takes a value to cast, which must be a
5469 scalar or vector :ref:`floating point <t_floating>` value, and a type to
5470 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
5471 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
5472 type with the same number of elements as ``ty``
5477 The '``fptosi``' instruction converts its :ref:`floating
5478 point <t_floating>` operand into the nearest (rounding towards zero)
5479 signed integer value. If the value cannot fit in ``ty2``, the results
5485 .. code-block:: llvm
5487 %X = fptosi double -123.0 to i32 ; yields i32:-123
5488 %Y = fptosi float 1.0E-247 to i1 ; yields undefined:1
5489 %Z = fptosi float 1.04E+17 to i8 ; yields undefined:1
5491 '``uitofp .. to``' Instruction
5492 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5499 <result> = uitofp <ty> <value> to <ty2> ; yields ty2
5504 The '``uitofp``' instruction regards ``value`` as an unsigned integer
5505 and converts that value to the ``ty2`` type.
5510 The '``uitofp``' instruction takes a value to cast, which must be a
5511 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
5512 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
5513 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
5514 type with the same number of elements as ``ty``
5519 The '``uitofp``' instruction interprets its operand as an unsigned
5520 integer quantity and converts it to the corresponding floating point
5521 value. If the value cannot fit in the floating point value, the results
5527 .. code-block:: llvm
5529 %X = uitofp i32 257 to float ; yields float:257.0
5530 %Y = uitofp i8 -1 to double ; yields double:255.0
5532 '``sitofp .. to``' Instruction
5533 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5540 <result> = sitofp <ty> <value> to <ty2> ; yields ty2
5545 The '``sitofp``' instruction regards ``value`` as a signed integer and
5546 converts that value to the ``ty2`` type.
5551 The '``sitofp``' instruction takes a value to cast, which must be a
5552 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
5553 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
5554 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
5555 type with the same number of elements as ``ty``
5560 The '``sitofp``' instruction interprets its operand as a signed integer
5561 quantity and converts it to the corresponding floating point value. If
5562 the value cannot fit in the floating point value, the results are
5568 .. code-block:: llvm
5570 %X = sitofp i32 257 to float ; yields float:257.0
5571 %Y = sitofp i8 -1 to double ; yields double:-1.0
5575 '``ptrtoint .. to``' Instruction
5576 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5583 <result> = ptrtoint <ty> <value> to <ty2> ; yields ty2
5588 The '``ptrtoint``' instruction converts the pointer or a vector of
5589 pointers ``value`` to the integer (or vector of integers) type ``ty2``.
5594 The '``ptrtoint``' instruction takes a ``value`` to cast, which must be
5595 a a value of type :ref:`pointer <t_pointer>` or a vector of pointers, and a
5596 type to cast it to ``ty2``, which must be an :ref:`integer <t_integer>` or
5597 a vector of integers type.
5602 The '``ptrtoint``' instruction converts ``value`` to integer type
5603 ``ty2`` by interpreting the pointer value as an integer and either
5604 truncating or zero extending that value to the size of the integer type.
5605 If ``value`` is smaller than ``ty2`` then a zero extension is done. If
5606 ``value`` is larger than ``ty2`` then a truncation is done. If they are
5607 the same size, then nothing is done (*no-op cast*) other than a type
5613 .. code-block:: llvm
5615 %X = ptrtoint i32* %P to i8 ; yields truncation on 32-bit architecture
5616 %Y = ptrtoint i32* %P to i64 ; yields zero extension on 32-bit architecture
5617 %Z = ptrtoint <4 x i32*> %P to <4 x i64>; yields vector zero extension for a vector of addresses on 32-bit architecture
5621 '``inttoptr .. to``' Instruction
5622 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5629 <result> = inttoptr <ty> <value> to <ty2> ; yields ty2
5634 The '``inttoptr``' instruction converts an integer ``value`` to a
5635 pointer type, ``ty2``.
5640 The '``inttoptr``' instruction takes an :ref:`integer <t_integer>` value to
5641 cast, and a type to cast it to, which must be a :ref:`pointer <t_pointer>`
5647 The '``inttoptr``' instruction converts ``value`` to type ``ty2`` by
5648 applying either a zero extension or a truncation depending on the size
5649 of the integer ``value``. If ``value`` is larger than the size of a
5650 pointer then a truncation is done. If ``value`` is smaller than the size
5651 of a pointer then a zero extension is done. If they are the same size,
5652 nothing is done (*no-op cast*).
5657 .. code-block:: llvm
5659 %X = inttoptr i32 255 to i32* ; yields zero extension on 64-bit architecture
5660 %Y = inttoptr i32 255 to i32* ; yields no-op on 32-bit architecture
5661 %Z = inttoptr i64 0 to i32* ; yields truncation on 32-bit architecture
5662 %Z = inttoptr <4 x i32> %G to <4 x i8*>; yields truncation of vector G to four pointers
5666 '``bitcast .. to``' Instruction
5667 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5674 <result> = bitcast <ty> <value> to <ty2> ; yields ty2
5679 The '``bitcast``' instruction converts ``value`` to type ``ty2`` without
5685 The '``bitcast``' instruction takes a value to cast, which must be a
5686 non-aggregate first class value, and a type to cast it to, which must
5687 also be a non-aggregate :ref:`first class <t_firstclass>` type. The
5688 bit sizes of ``value`` and the destination type, ``ty2``, must be
5689 identical. If the source type is a pointer, the destination type must
5690 also be a pointer of the same size. This instruction supports bitwise
5691 conversion of vectors to integers and to vectors of other types (as
5692 long as they have the same size).
5697 The '``bitcast``' instruction converts ``value`` to type ``ty2``. It
5698 is always a *no-op cast* because no bits change with this
5699 conversion. The conversion is done as if the ``value`` had been stored
5700 to memory and read back as type ``ty2``. Pointer (or vector of
5701 pointers) types may only be converted to other pointer (or vector of
5702 pointers) types with this instruction if the pointer sizes are
5703 equal. To convert pointers to other types, use the :ref:`inttoptr
5704 <i_inttoptr>` or :ref:`ptrtoint <i_ptrtoint>` instructions first.
5709 .. code-block:: llvm
5711 %X = bitcast i8 255 to i8 ; yields i8 :-1
5712 %Y = bitcast i32* %x to sint* ; yields sint*:%x
5713 %Z = bitcast <2 x int> %V to i64; ; yields i64: %V
5714 %Z = bitcast <2 x i32*> %V to <2 x i64*> ; yields <2 x i64*>
5721 The instructions in this category are the "miscellaneous" instructions,
5722 which defy better classification.
5726 '``icmp``' Instruction
5727 ^^^^^^^^^^^^^^^^^^^^^^
5734 <result> = icmp <cond> <ty> <op1>, <op2> ; yields {i1} or {<N x i1>}:result
5739 The '``icmp``' instruction returns a boolean value or a vector of
5740 boolean values based on comparison of its two integer, integer vector,
5741 pointer, or pointer vector operands.
5746 The '``icmp``' instruction takes three operands. The first operand is
5747 the condition code indicating the kind of comparison to perform. It is
5748 not a value, just a keyword. The possible condition code are:
5751 #. ``ne``: not equal
5752 #. ``ugt``: unsigned greater than
5753 #. ``uge``: unsigned greater or equal
5754 #. ``ult``: unsigned less than
5755 #. ``ule``: unsigned less or equal
5756 #. ``sgt``: signed greater than
5757 #. ``sge``: signed greater or equal
5758 #. ``slt``: signed less than
5759 #. ``sle``: signed less or equal
5761 The remaining two arguments must be :ref:`integer <t_integer>` or
5762 :ref:`pointer <t_pointer>` or integer :ref:`vector <t_vector>` typed. They
5763 must also be identical types.
5768 The '``icmp``' compares ``op1`` and ``op2`` according to the condition
5769 code given as ``cond``. The comparison performed always yields either an
5770 :ref:`i1 <t_integer>` or vector of ``i1`` result, as follows:
5772 #. ``eq``: yields ``true`` if the operands are equal, ``false``
5773 otherwise. No sign interpretation is necessary or performed.
5774 #. ``ne``: yields ``true`` if the operands are unequal, ``false``
5775 otherwise. No sign interpretation is necessary or performed.
5776 #. ``ugt``: interprets the operands as unsigned values and yields
5777 ``true`` if ``op1`` is greater than ``op2``.
5778 #. ``uge``: interprets the operands as unsigned values and yields
5779 ``true`` if ``op1`` is greater than or equal to ``op2``.
5780 #. ``ult``: interprets the operands as unsigned values and yields
5781 ``true`` if ``op1`` is less than ``op2``.
5782 #. ``ule``: interprets the operands as unsigned values and yields
5783 ``true`` if ``op1`` is less than or equal to ``op2``.
5784 #. ``sgt``: interprets the operands as signed values and yields ``true``
5785 if ``op1`` is greater than ``op2``.
5786 #. ``sge``: interprets the operands as signed values and yields ``true``
5787 if ``op1`` is greater than or equal to ``op2``.
5788 #. ``slt``: interprets the operands as signed values and yields ``true``
5789 if ``op1`` is less than ``op2``.
5790 #. ``sle``: interprets the operands as signed values and yields ``true``
5791 if ``op1`` is less than or equal to ``op2``.
5793 If the operands are :ref:`pointer <t_pointer>` typed, the pointer values
5794 are compared as if they were integers.
5796 If the operands are integer vectors, then they are compared element by
5797 element. The result is an ``i1`` vector with the same number of elements
5798 as the values being compared. Otherwise, the result is an ``i1``.
5803 .. code-block:: llvm
5805 <result> = icmp eq i32 4, 5 ; yields: result=false
5806 <result> = icmp ne float* %X, %X ; yields: result=false
5807 <result> = icmp ult i16 4, 5 ; yields: result=true
5808 <result> = icmp sgt i16 4, 5 ; yields: result=false
5809 <result> = icmp ule i16 -4, 5 ; yields: result=false
5810 <result> = icmp sge i16 4, 5 ; yields: result=false
5812 Note that the code generator does not yet support vector types with the
5813 ``icmp`` instruction.
5817 '``fcmp``' Instruction
5818 ^^^^^^^^^^^^^^^^^^^^^^
5825 <result> = fcmp <cond> <ty> <op1>, <op2> ; yields {i1} or {<N x i1>}:result
5830 The '``fcmp``' instruction returns a boolean value or vector of boolean
5831 values based on comparison of its operands.
5833 If the operands are floating point scalars, then the result type is a
5834 boolean (:ref:`i1 <t_integer>`).
5836 If the operands are floating point vectors, then the result type is a
5837 vector of boolean with the same number of elements as the operands being
5843 The '``fcmp``' instruction takes three operands. The first operand is
5844 the condition code indicating the kind of comparison to perform. It is
5845 not a value, just a keyword. The possible condition code are:
5847 #. ``false``: no comparison, always returns false
5848 #. ``oeq``: ordered and equal
5849 #. ``ogt``: ordered and greater than
5850 #. ``oge``: ordered and greater than or equal
5851 #. ``olt``: ordered and less than
5852 #. ``ole``: ordered and less than or equal
5853 #. ``one``: ordered and not equal
5854 #. ``ord``: ordered (no nans)
5855 #. ``ueq``: unordered or equal
5856 #. ``ugt``: unordered or greater than
5857 #. ``uge``: unordered or greater than or equal
5858 #. ``ult``: unordered or less than
5859 #. ``ule``: unordered or less than or equal
5860 #. ``une``: unordered or not equal
5861 #. ``uno``: unordered (either nans)
5862 #. ``true``: no comparison, always returns true
5864 *Ordered* means that neither operand is a QNAN while *unordered* means
5865 that either operand may be a QNAN.
5867 Each of ``val1`` and ``val2`` arguments must be either a :ref:`floating
5868 point <t_floating>` type or a :ref:`vector <t_vector>` of floating point
5869 type. They must have identical types.
5874 The '``fcmp``' instruction compares ``op1`` and ``op2`` according to the
5875 condition code given as ``cond``. If the operands are vectors, then the
5876 vectors are compared element by element. Each comparison performed
5877 always yields an :ref:`i1 <t_integer>` result, as follows:
5879 #. ``false``: always yields ``false``, regardless of operands.
5880 #. ``oeq``: yields ``true`` if both operands are not a QNAN and ``op1``
5881 is equal to ``op2``.
5882 #. ``ogt``: yields ``true`` if both operands are not a QNAN and ``op1``
5883 is greater than ``op2``.
5884 #. ``oge``: yields ``true`` if both operands are not a QNAN and ``op1``
5885 is greater than or equal to ``op2``.
5886 #. ``olt``: yields ``true`` if both operands are not a QNAN and ``op1``
5887 is less than ``op2``.
5888 #. ``ole``: yields ``true`` if both operands are not a QNAN and ``op1``
5889 is less than or equal to ``op2``.
5890 #. ``one``: yields ``true`` if both operands are not a QNAN and ``op1``
5891 is not equal to ``op2``.
5892 #. ``ord``: yields ``true`` if both operands are not a QNAN.
5893 #. ``ueq``: yields ``true`` if either operand is a QNAN or ``op1`` is
5895 #. ``ugt``: yields ``true`` if either operand is a QNAN or ``op1`` is
5896 greater than ``op2``.
5897 #. ``uge``: yields ``true`` if either operand is a QNAN or ``op1`` is
5898 greater than or equal to ``op2``.
5899 #. ``ult``: yields ``true`` if either operand is a QNAN or ``op1`` is
5901 #. ``ule``: yields ``true`` if either operand is a QNAN or ``op1`` is
5902 less than or equal to ``op2``.
5903 #. ``une``: yields ``true`` if either operand is a QNAN or ``op1`` is
5904 not equal to ``op2``.
5905 #. ``uno``: yields ``true`` if either operand is a QNAN.
5906 #. ``true``: always yields ``true``, regardless of operands.
5911 .. code-block:: llvm
5913 <result> = fcmp oeq float 4.0, 5.0 ; yields: result=false
5914 <result> = fcmp one float 4.0, 5.0 ; yields: result=true
5915 <result> = fcmp olt float 4.0, 5.0 ; yields: result=true
5916 <result> = fcmp ueq double 1.0, 2.0 ; yields: result=false
5918 Note that the code generator does not yet support vector types with the
5919 ``fcmp`` instruction.
5923 '``phi``' Instruction
5924 ^^^^^^^^^^^^^^^^^^^^^
5931 <result> = phi <ty> [ <val0>, <label0>], ...
5936 The '``phi``' instruction is used to implement the φ node in the SSA
5937 graph representing the function.
5942 The type of the incoming values is specified with the first type field.
5943 After this, the '``phi``' instruction takes a list of pairs as
5944 arguments, with one pair for each predecessor basic block of the current
5945 block. Only values of :ref:`first class <t_firstclass>` type may be used as
5946 the value arguments to the PHI node. Only labels may be used as the
5949 There must be no non-phi instructions between the start of a basic block
5950 and the PHI instructions: i.e. PHI instructions must be first in a basic
5953 For the purposes of the SSA form, the use of each incoming value is
5954 deemed to occur on the edge from the corresponding predecessor block to
5955 the current block (but after any definition of an '``invoke``'
5956 instruction's return value on the same edge).
5961 At runtime, the '``phi``' instruction logically takes on the value
5962 specified by the pair corresponding to the predecessor basic block that
5963 executed just prior to the current block.
5968 .. code-block:: llvm
5970 Loop: ; Infinite loop that counts from 0 on up...
5971 %indvar = phi i32 [ 0, %LoopHeader ], [ %nextindvar, %Loop ]
5972 %nextindvar = add i32 %indvar, 1
5977 '``select``' Instruction
5978 ^^^^^^^^^^^^^^^^^^^^^^^^
5985 <result> = select selty <cond>, <ty> <val1>, <ty> <val2> ; yields ty
5987 selty is either i1 or {<N x i1>}
5992 The '``select``' instruction is used to choose one value based on a
5993 condition, without branching.
5998 The '``select``' instruction requires an 'i1' value or a vector of 'i1'
5999 values indicating the condition, and two values of the same :ref:`first
6000 class <t_firstclass>` type. If the val1/val2 are vectors and the
6001 condition is a scalar, then entire vectors are selected, not individual
6007 If the condition is an i1 and it evaluates to 1, the instruction returns
6008 the first value argument; otherwise, it returns the second value
6011 If the condition is a vector of i1, then the value arguments must be
6012 vectors of the same size, and the selection is done element by element.
6017 .. code-block:: llvm
6019 %X = select i1 true, i8 17, i8 42 ; yields i8:17
6023 '``call``' Instruction
6024 ^^^^^^^^^^^^^^^^^^^^^^
6031 <result> = [tail] call [cconv] [ret attrs] <ty> [<fnty>*] <fnptrval>(<function args>) [fn attrs]
6036 The '``call``' instruction represents a simple function call.
6041 This instruction requires several arguments:
6043 #. The optional "tail" marker indicates that the callee function does
6044 not access any allocas or varargs in the caller. Note that calls may
6045 be marked "tail" even if they do not occur before a
6046 :ref:`ret <i_ret>` instruction. If the "tail" marker is present, the
6047 function call is eligible for tail call optimization, but `might not
6048 in fact be optimized into a jump <CodeGenerator.html#tailcallopt>`_.
6049 The code generator may optimize calls marked "tail" with either 1)
6050 automatic `sibling call
6051 optimization <CodeGenerator.html#sibcallopt>`_ when the caller and
6052 callee have matching signatures, or 2) forced tail call optimization
6053 when the following extra requirements are met:
6055 - Caller and callee both have the calling convention ``fastcc``.
6056 - The call is in tail position (ret immediately follows call and ret
6057 uses value of call or is void).
6058 - Option ``-tailcallopt`` is enabled, or
6059 ``llvm::GuaranteedTailCallOpt`` is ``true``.
6060 - `Platform specific constraints are
6061 met. <CodeGenerator.html#tailcallopt>`_
6063 #. The optional "cconv" marker indicates which :ref:`calling
6064 convention <callingconv>` the call should use. If none is
6065 specified, the call defaults to using C calling conventions. The
6066 calling convention of the call must match the calling convention of
6067 the target function, or else the behavior is undefined.
6068 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
6069 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
6071 #. '``ty``': the type of the call instruction itself which is also the
6072 type of the return value. Functions that return no value are marked
6074 #. '``fnty``': shall be the signature of the pointer to function value
6075 being invoked. The argument types must match the types implied by
6076 this signature. This type can be omitted if the function is not
6077 varargs and if the function type does not return a pointer to a
6079 #. '``fnptrval``': An LLVM value containing a pointer to a function to
6080 be invoked. In most cases, this is a direct function invocation, but
6081 indirect ``call``'s are just as possible, calling an arbitrary pointer
6083 #. '``function args``': argument list whose types match the function
6084 signature argument types and parameter attributes. All arguments must
6085 be of :ref:`first class <t_firstclass>` type. If the function signature
6086 indicates the function accepts a variable number of arguments, the
6087 extra arguments can be specified.
6088 #. The optional :ref:`function attributes <fnattrs>` list. Only
6089 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
6090 attributes are valid here.
6095 The '``call``' instruction is used to cause control flow to transfer to
6096 a specified function, with its incoming arguments bound to the specified
6097 values. Upon a '``ret``' instruction in the called function, control
6098 flow continues with the instruction after the function call, and the
6099 return value of the function is bound to the result argument.
6104 .. code-block:: llvm
6106 %retval = call i32 @test(i32 %argc)
6107 call i32 (i8*, ...)* @printf(i8* %msg, i32 12, i8 42) ; yields i32
6108 %X = tail call i32 @foo() ; yields i32
6109 %Y = tail call fastcc i32 @foo() ; yields i32
6110 call void %foo(i8 97 signext)
6112 %struct.A = type { i32, i8 }
6113 %r = call %struct.A @foo() ; yields { 32, i8 }
6114 %gr = extractvalue %struct.A %r, 0 ; yields i32
6115 %gr1 = extractvalue %struct.A %r, 1 ; yields i8
6116 %Z = call void @foo() noreturn ; indicates that %foo never returns normally
6117 %ZZ = call zeroext i32 @bar() ; Return value is %zero extended
6119 llvm treats calls to some functions with names and arguments that match
6120 the standard C99 library as being the C99 library functions, and may
6121 perform optimizations or generate code for them under that assumption.
6122 This is something we'd like to change in the future to provide better
6123 support for freestanding environments and non-C-based languages.
6127 '``va_arg``' Instruction
6128 ^^^^^^^^^^^^^^^^^^^^^^^^
6135 <resultval> = va_arg <va_list*> <arglist>, <argty>
6140 The '``va_arg``' instruction is used to access arguments passed through
6141 the "variable argument" area of a function call. It is used to implement
6142 the ``va_arg`` macro in C.
6147 This instruction takes a ``va_list*`` value and the type of the
6148 argument. It returns a value of the specified argument type and
6149 increments the ``va_list`` to point to the next argument. The actual
6150 type of ``va_list`` is target specific.
6155 The '``va_arg``' instruction loads an argument of the specified type
6156 from the specified ``va_list`` and causes the ``va_list`` to point to
6157 the next argument. For more information, see the variable argument
6158 handling :ref:`Intrinsic Functions <int_varargs>`.
6160 It is legal for this instruction to be called in a function which does
6161 not take a variable number of arguments, for example, the ``vfprintf``
6164 ``va_arg`` is an LLVM instruction instead of an :ref:`intrinsic
6165 function <intrinsics>` because it takes a type as an argument.
6170 See the :ref:`variable argument processing <int_varargs>` section.
6172 Note that the code generator does not yet fully support va\_arg on many
6173 targets. Also, it does not currently support va\_arg with aggregate
6174 types on any target.
6178 '``landingpad``' Instruction
6179 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6186 <resultval> = landingpad <resultty> personality <type> <pers_fn> <clause>+
6187 <resultval> = landingpad <resultty> personality <type> <pers_fn> cleanup <clause>*
6189 <clause> := catch <type> <value>
6190 <clause> := filter <array constant type> <array constant>
6195 The '``landingpad``' instruction is used by `LLVM's exception handling
6196 system <ExceptionHandling.html#overview>`_ to specify that a basic block
6197 is a landing pad --- one where the exception lands, and corresponds to the
6198 code found in the ``catch`` portion of a ``try``/``catch`` sequence. It
6199 defines values supplied by the personality function (``pers_fn``) upon
6200 re-entry to the function. The ``resultval`` has the type ``resultty``.
6205 This instruction takes a ``pers_fn`` value. This is the personality
6206 function associated with the unwinding mechanism. The optional
6207 ``cleanup`` flag indicates that the landing pad block is a cleanup.
6209 A ``clause`` begins with the clause type --- ``catch`` or ``filter`` --- and
6210 contains the global variable representing the "type" that may be caught
6211 or filtered respectively. Unlike the ``catch`` clause, the ``filter``
6212 clause takes an array constant as its argument. Use
6213 "``[0 x i8**] undef``" for a filter which cannot throw. The
6214 '``landingpad``' instruction must contain *at least* one ``clause`` or
6215 the ``cleanup`` flag.
6220 The '``landingpad``' instruction defines the values which are set by the
6221 personality function (``pers_fn``) upon re-entry to the function, and
6222 therefore the "result type" of the ``landingpad`` instruction. As with
6223 calling conventions, how the personality function results are
6224 represented in LLVM IR is target specific.
6226 The clauses are applied in order from top to bottom. If two
6227 ``landingpad`` instructions are merged together through inlining, the
6228 clauses from the calling function are appended to the list of clauses.
6229 When the call stack is being unwound due to an exception being thrown,
6230 the exception is compared against each ``clause`` in turn. If it doesn't
6231 match any of the clauses, and the ``cleanup`` flag is not set, then
6232 unwinding continues further up the call stack.
6234 The ``landingpad`` instruction has several restrictions:
6236 - A landing pad block is a basic block which is the unwind destination
6237 of an '``invoke``' instruction.
6238 - A landing pad block must have a '``landingpad``' instruction as its
6239 first non-PHI instruction.
6240 - There can be only one '``landingpad``' instruction within the landing
6242 - A basic block that is not a landing pad block may not include a
6243 '``landingpad``' instruction.
6244 - All '``landingpad``' instructions in a function must have the same
6245 personality function.
6250 .. code-block:: llvm
6252 ;; A landing pad which can catch an integer.
6253 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6255 ;; A landing pad that is a cleanup.
6256 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6258 ;; A landing pad which can catch an integer and can only throw a double.
6259 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6261 filter [1 x i8**] [@_ZTId]
6268 LLVM supports the notion of an "intrinsic function". These functions
6269 have well known names and semantics and are required to follow certain
6270 restrictions. Overall, these intrinsics represent an extension mechanism
6271 for the LLVM language that does not require changing all of the
6272 transformations in LLVM when adding to the language (or the bitcode
6273 reader/writer, the parser, etc...).
6275 Intrinsic function names must all start with an "``llvm.``" prefix. This
6276 prefix is reserved in LLVM for intrinsic names; thus, function names may
6277 not begin with this prefix. Intrinsic functions must always be external
6278 functions: you cannot define the body of intrinsic functions. Intrinsic
6279 functions may only be used in call or invoke instructions: it is illegal
6280 to take the address of an intrinsic function. Additionally, because
6281 intrinsic functions are part of the LLVM language, it is required if any
6282 are added that they be documented here.
6284 Some intrinsic functions can be overloaded, i.e., the intrinsic
6285 represents a family of functions that perform the same operation but on
6286 different data types. Because LLVM can represent over 8 million
6287 different integer types, overloading is used commonly to allow an
6288 intrinsic function to operate on any integer type. One or more of the
6289 argument types or the result type can be overloaded to accept any
6290 integer type. Argument types may also be defined as exactly matching a
6291 previous argument's type or the result type. This allows an intrinsic
6292 function which accepts multiple arguments, but needs all of them to be
6293 of the same type, to only be overloaded with respect to a single
6294 argument or the result.
6296 Overloaded intrinsics will have the names of its overloaded argument
6297 types encoded into its function name, each preceded by a period. Only
6298 those types which are overloaded result in a name suffix. Arguments
6299 whose type is matched against another type do not. For example, the
6300 ``llvm.ctpop`` function can take an integer of any width and returns an
6301 integer of exactly the same integer width. This leads to a family of
6302 functions such as ``i8 @llvm.ctpop.i8(i8 %val)`` and
6303 ``i29 @llvm.ctpop.i29(i29 %val)``. Only one type, the return type, is
6304 overloaded, and only one type suffix is required. Because the argument's
6305 type is matched against the return type, it does not require its own
6308 To learn how to add an intrinsic function, please see the `Extending
6309 LLVM Guide <ExtendingLLVM.html>`_.
6313 Variable Argument Handling Intrinsics
6314 -------------------------------------
6316 Variable argument support is defined in LLVM with the
6317 :ref:`va_arg <i_va_arg>` instruction and these three intrinsic
6318 functions. These functions are related to the similarly named macros
6319 defined in the ``<stdarg.h>`` header file.
6321 All of these functions operate on arguments that use a target-specific
6322 value type "``va_list``". The LLVM assembly language reference manual
6323 does not define what this type is, so all transformations should be
6324 prepared to handle these functions regardless of the type used.
6326 This example shows how the :ref:`va_arg <i_va_arg>` instruction and the
6327 variable argument handling intrinsic functions are used.
6329 .. code-block:: llvm
6331 define i32 @test(i32 %X, ...) {
6332 ; Initialize variable argument processing
6334 %ap2 = bitcast i8** %ap to i8*
6335 call void @llvm.va_start(i8* %ap2)
6337 ; Read a single integer argument
6338 %tmp = va_arg i8** %ap, i32
6340 ; Demonstrate usage of llvm.va_copy and llvm.va_end
6342 %aq2 = bitcast i8** %aq to i8*
6343 call void @llvm.va_copy(i8* %aq2, i8* %ap2)
6344 call void @llvm.va_end(i8* %aq2)
6346 ; Stop processing of arguments.
6347 call void @llvm.va_end(i8* %ap2)
6351 declare void @llvm.va_start(i8*)
6352 declare void @llvm.va_copy(i8*, i8*)
6353 declare void @llvm.va_end(i8*)
6357 '``llvm.va_start``' Intrinsic
6358 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6365 declare void @llvm.va_start(i8* <arglist>)
6370 The '``llvm.va_start``' intrinsic initializes ``*<arglist>`` for
6371 subsequent use by ``va_arg``.
6376 The argument is a pointer to a ``va_list`` element to initialize.
6381 The '``llvm.va_start``' intrinsic works just like the ``va_start`` macro
6382 available in C. In a target-dependent way, it initializes the
6383 ``va_list`` element to which the argument points, so that the next call
6384 to ``va_arg`` will produce the first variable argument passed to the
6385 function. Unlike the C ``va_start`` macro, this intrinsic does not need
6386 to know the last argument of the function as the compiler can figure
6389 '``llvm.va_end``' Intrinsic
6390 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6397 declare void @llvm.va_end(i8* <arglist>)
6402 The '``llvm.va_end``' intrinsic destroys ``*<arglist>``, which has been
6403 initialized previously with ``llvm.va_start`` or ``llvm.va_copy``.
6408 The argument is a pointer to a ``va_list`` to destroy.
6413 The '``llvm.va_end``' intrinsic works just like the ``va_end`` macro
6414 available in C. In a target-dependent way, it destroys the ``va_list``
6415 element to which the argument points. Calls to
6416 :ref:`llvm.va_start <int_va_start>` and
6417 :ref:`llvm.va_copy <int_va_copy>` must be matched exactly with calls to
6422 '``llvm.va_copy``' Intrinsic
6423 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6430 declare void @llvm.va_copy(i8* <destarglist>, i8* <srcarglist>)
6435 The '``llvm.va_copy``' intrinsic copies the current argument position
6436 from the source argument list to the destination argument list.
6441 The first argument is a pointer to a ``va_list`` element to initialize.
6442 The second argument is a pointer to a ``va_list`` element to copy from.
6447 The '``llvm.va_copy``' intrinsic works just like the ``va_copy`` macro
6448 available in C. In a target-dependent way, it copies the source
6449 ``va_list`` element into the destination ``va_list`` element. This
6450 intrinsic is necessary because the `` llvm.va_start`` intrinsic may be
6451 arbitrarily complex and require, for example, memory allocation.
6453 Accurate Garbage Collection Intrinsics
6454 --------------------------------------
6456 LLVM support for `Accurate Garbage Collection <GarbageCollection.html>`_
6457 (GC) requires the implementation and generation of these intrinsics.
6458 These intrinsics allow identification of :ref:`GC roots on the
6459 stack <int_gcroot>`, as well as garbage collector implementations that
6460 require :ref:`read <int_gcread>` and :ref:`write <int_gcwrite>` barriers.
6461 Front-ends for type-safe garbage collected languages should generate
6462 these intrinsics to make use of the LLVM garbage collectors. For more
6463 details, see `Accurate Garbage Collection with
6464 LLVM <GarbageCollection.html>`_.
6466 The garbage collection intrinsics only operate on objects in the generic
6467 address space (address space zero).
6471 '``llvm.gcroot``' Intrinsic
6472 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6479 declare void @llvm.gcroot(i8** %ptrloc, i8* %metadata)
6484 The '``llvm.gcroot``' intrinsic declares the existence of a GC root to
6485 the code generator, and allows some metadata to be associated with it.
6490 The first argument specifies the address of a stack object that contains
6491 the root pointer. The second pointer (which must be either a constant or
6492 a global value address) contains the meta-data to be associated with the
6498 At runtime, a call to this intrinsic stores a null pointer into the
6499 "ptrloc" location. At compile-time, the code generator generates
6500 information to allow the runtime to find the pointer at GC safe points.
6501 The '``llvm.gcroot``' intrinsic may only be used in a function which
6502 :ref:`specifies a GC algorithm <gc>`.
6506 '``llvm.gcread``' Intrinsic
6507 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6514 declare i8* @llvm.gcread(i8* %ObjPtr, i8** %Ptr)
6519 The '``llvm.gcread``' intrinsic identifies reads of references from heap
6520 locations, allowing garbage collector implementations that require read
6526 The second argument is the address to read from, which should be an
6527 address allocated from the garbage collector. The first object is a
6528 pointer to the start of the referenced object, if needed by the language
6529 runtime (otherwise null).
6534 The '``llvm.gcread``' intrinsic has the same semantics as a load
6535 instruction, but may be replaced with substantially more complex code by
6536 the garbage collector runtime, as needed. The '``llvm.gcread``'
6537 intrinsic may only be used in a function which :ref:`specifies a GC
6542 '``llvm.gcwrite``' Intrinsic
6543 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6550 declare void @llvm.gcwrite(i8* %P1, i8* %Obj, i8** %P2)
6555 The '``llvm.gcwrite``' intrinsic identifies writes of references to heap
6556 locations, allowing garbage collector implementations that require write
6557 barriers (such as generational or reference counting collectors).
6562 The first argument is the reference to store, the second is the start of
6563 the object to store it to, and the third is the address of the field of
6564 Obj to store to. If the runtime does not require a pointer to the
6565 object, Obj may be null.
6570 The '``llvm.gcwrite``' intrinsic has the same semantics as a store
6571 instruction, but may be replaced with substantially more complex code by
6572 the garbage collector runtime, as needed. The '``llvm.gcwrite``'
6573 intrinsic may only be used in a function which :ref:`specifies a GC
6576 Code Generator Intrinsics
6577 -------------------------
6579 These intrinsics are provided by LLVM to expose special features that
6580 may only be implemented with code generator support.
6582 '``llvm.returnaddress``' Intrinsic
6583 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6590 declare i8 *@llvm.returnaddress(i32 <level>)
6595 The '``llvm.returnaddress``' intrinsic attempts to compute a
6596 target-specific value indicating the return address of the current
6597 function or one of its callers.
6602 The argument to this intrinsic indicates which function to return the
6603 address for. Zero indicates the calling function, one indicates its
6604 caller, etc. The argument is **required** to be a constant integer
6610 The '``llvm.returnaddress``' intrinsic either returns a pointer
6611 indicating the return address of the specified call frame, or zero if it
6612 cannot be identified. The value returned by this intrinsic is likely to
6613 be incorrect or 0 for arguments other than zero, so it should only be
6614 used for debugging purposes.
6616 Note that calling this intrinsic does not prevent function inlining or
6617 other aggressive transformations, so the value returned may not be that
6618 of the obvious source-language caller.
6620 '``llvm.frameaddress``' Intrinsic
6621 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6628 declare i8* @llvm.frameaddress(i32 <level>)
6633 The '``llvm.frameaddress``' intrinsic attempts to return the
6634 target-specific frame pointer value for the specified stack frame.
6639 The argument to this intrinsic indicates which function to return the
6640 frame pointer for. Zero indicates the calling function, one indicates
6641 its caller, etc. The argument is **required** to be a constant integer
6647 The '``llvm.frameaddress``' intrinsic either returns a pointer
6648 indicating the frame address of the specified call frame, or zero if it
6649 cannot be identified. The value returned by this intrinsic is likely to
6650 be incorrect or 0 for arguments other than zero, so it should only be
6651 used for debugging purposes.
6653 Note that calling this intrinsic does not prevent function inlining or
6654 other aggressive transformations, so the value returned may not be that
6655 of the obvious source-language caller.
6659 '``llvm.stacksave``' Intrinsic
6660 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6667 declare i8* @llvm.stacksave()
6672 The '``llvm.stacksave``' intrinsic is used to remember the current state
6673 of the function stack, for use with
6674 :ref:`llvm.stackrestore <int_stackrestore>`. This is useful for
6675 implementing language features like scoped automatic variable sized
6681 This intrinsic returns a opaque pointer value that can be passed to
6682 :ref:`llvm.stackrestore <int_stackrestore>`. When an
6683 ``llvm.stackrestore`` intrinsic is executed with a value saved from
6684 ``llvm.stacksave``, it effectively restores the state of the stack to
6685 the state it was in when the ``llvm.stacksave`` intrinsic executed. In
6686 practice, this pops any :ref:`alloca <i_alloca>` blocks from the stack that
6687 were allocated after the ``llvm.stacksave`` was executed.
6689 .. _int_stackrestore:
6691 '``llvm.stackrestore``' Intrinsic
6692 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6699 declare void @llvm.stackrestore(i8* %ptr)
6704 The '``llvm.stackrestore``' intrinsic is used to restore the state of
6705 the function stack to the state it was in when the corresponding
6706 :ref:`llvm.stacksave <int_stacksave>` intrinsic executed. This is
6707 useful for implementing language features like scoped automatic variable
6708 sized arrays in C99.
6713 See the description for :ref:`llvm.stacksave <int_stacksave>`.
6715 '``llvm.prefetch``' Intrinsic
6716 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6723 declare void @llvm.prefetch(i8* <address>, i32 <rw>, i32 <locality>, i32 <cache type>)
6728 The '``llvm.prefetch``' intrinsic is a hint to the code generator to
6729 insert a prefetch instruction if supported; otherwise, it is a noop.
6730 Prefetches have no effect on the behavior of the program but can change
6731 its performance characteristics.
6736 ``address`` is the address to be prefetched, ``rw`` is the specifier
6737 determining if the fetch should be for a read (0) or write (1), and
6738 ``locality`` is a temporal locality specifier ranging from (0) - no
6739 locality, to (3) - extremely local keep in cache. The ``cache type``
6740 specifies whether the prefetch is performed on the data (1) or
6741 instruction (0) cache. The ``rw``, ``locality`` and ``cache type``
6742 arguments must be constant integers.
6747 This intrinsic does not modify the behavior of the program. In
6748 particular, prefetches cannot trap and do not produce a value. On
6749 targets that support this intrinsic, the prefetch can provide hints to
6750 the processor cache for better performance.
6752 '``llvm.pcmarker``' Intrinsic
6753 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6760 declare void @llvm.pcmarker(i32 <id>)
6765 The '``llvm.pcmarker``' intrinsic is a method to export a Program
6766 Counter (PC) in a region of code to simulators and other tools. The
6767 method is target specific, but it is expected that the marker will use
6768 exported symbols to transmit the PC of the marker. The marker makes no
6769 guarantees that it will remain with any specific instruction after
6770 optimizations. It is possible that the presence of a marker will inhibit
6771 optimizations. The intended use is to be inserted after optimizations to
6772 allow correlations of simulation runs.
6777 ``id`` is a numerical id identifying the marker.
6782 This intrinsic does not modify the behavior of the program. Backends
6783 that do not support this intrinsic may ignore it.
6785 '``llvm.readcyclecounter``' Intrinsic
6786 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6793 declare i64 @llvm.readcyclecounter()
6798 The '``llvm.readcyclecounter``' intrinsic provides access to the cycle
6799 counter register (or similar low latency, high accuracy clocks) on those
6800 targets that support it. On X86, it should map to RDTSC. On Alpha, it
6801 should map to RPCC. As the backing counters overflow quickly (on the
6802 order of 9 seconds on alpha), this should only be used for small
6808 When directly supported, reading the cycle counter should not modify any
6809 memory. Implementations are allowed to either return a application
6810 specific value or a system wide value. On backends without support, this
6811 is lowered to a constant 0.
6813 Note that runtime support may be conditional on the privilege-level code is
6814 running at and the host platform.
6816 Standard C Library Intrinsics
6817 -----------------------------
6819 LLVM provides intrinsics for a few important standard C library
6820 functions. These intrinsics allow source-language front-ends to pass
6821 information about the alignment of the pointer arguments to the code
6822 generator, providing opportunity for more efficient code generation.
6826 '``llvm.memcpy``' Intrinsic
6827 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6832 This is an overloaded intrinsic. You can use ``llvm.memcpy`` on any
6833 integer bit width and for different address spaces. Not all targets
6834 support all bit widths however.
6838 declare void @llvm.memcpy.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
6839 i32 <len>, i32 <align>, i1 <isvolatile>)
6840 declare void @llvm.memcpy.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
6841 i64 <len>, i32 <align>, i1 <isvolatile>)
6846 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
6847 source location to the destination location.
6849 Note that, unlike the standard libc function, the ``llvm.memcpy.*``
6850 intrinsics do not return a value, takes extra alignment/isvolatile
6851 arguments and the pointers can be in specified address spaces.
6856 The first argument is a pointer to the destination, the second is a
6857 pointer to the source. The third argument is an integer argument
6858 specifying the number of bytes to copy, the fourth argument is the
6859 alignment of the source and destination locations, and the fifth is a
6860 boolean indicating a volatile access.
6862 If the call to this intrinsic has an alignment value that is not 0 or 1,
6863 then the caller guarantees that both the source and destination pointers
6864 are aligned to that boundary.
6866 If the ``isvolatile`` parameter is ``true``, the ``llvm.memcpy`` call is
6867 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
6868 very cleanly specified and it is unwise to depend on it.
6873 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
6874 source location to the destination location, which are not allowed to
6875 overlap. It copies "len" bytes of memory over. If the argument is known
6876 to be aligned to some boundary, this can be specified as the fourth
6877 argument, otherwise it should be set to 0 or 1.
6879 '``llvm.memmove``' Intrinsic
6880 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6885 This is an overloaded intrinsic. You can use llvm.memmove on any integer
6886 bit width and for different address space. Not all targets support all
6891 declare void @llvm.memmove.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
6892 i32 <len>, i32 <align>, i1 <isvolatile>)
6893 declare void @llvm.memmove.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
6894 i64 <len>, i32 <align>, i1 <isvolatile>)
6899 The '``llvm.memmove.*``' intrinsics move a block of memory from the
6900 source location to the destination location. It is similar to the
6901 '``llvm.memcpy``' intrinsic but allows the two memory locations to
6904 Note that, unlike the standard libc function, the ``llvm.memmove.*``
6905 intrinsics do not return a value, takes extra alignment/isvolatile
6906 arguments and the pointers can be in specified address spaces.
6911 The first argument is a pointer to the destination, the second is a
6912 pointer to the source. The third argument is an integer argument
6913 specifying the number of bytes to copy, the fourth argument is the
6914 alignment of the source and destination locations, and the fifth is a
6915 boolean indicating a volatile access.
6917 If the call to this intrinsic has an alignment value that is not 0 or 1,
6918 then the caller guarantees that the source and destination pointers are
6919 aligned to that boundary.
6921 If the ``isvolatile`` parameter is ``true``, the ``llvm.memmove`` call
6922 is a :ref:`volatile operation <volatile>`. The detailed access behavior is
6923 not very cleanly specified and it is unwise to depend on it.
6928 The '``llvm.memmove.*``' intrinsics copy a block of memory from the
6929 source location to the destination location, which may overlap. It
6930 copies "len" bytes of memory over. If the argument is known to be
6931 aligned to some boundary, this can be specified as the fourth argument,
6932 otherwise it should be set to 0 or 1.
6934 '``llvm.memset.*``' Intrinsics
6935 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6940 This is an overloaded intrinsic. You can use llvm.memset on any integer
6941 bit width and for different address spaces. However, not all targets
6942 support all bit widths.
6946 declare void @llvm.memset.p0i8.i32(i8* <dest>, i8 <val>,
6947 i32 <len>, i32 <align>, i1 <isvolatile>)
6948 declare void @llvm.memset.p0i8.i64(i8* <dest>, i8 <val>,
6949 i64 <len>, i32 <align>, i1 <isvolatile>)
6954 The '``llvm.memset.*``' intrinsics fill a block of memory with a
6955 particular byte value.
6957 Note that, unlike the standard libc function, the ``llvm.memset``
6958 intrinsic does not return a value and takes extra alignment/volatile
6959 arguments. Also, the destination can be in an arbitrary address space.
6964 The first argument is a pointer to the destination to fill, the second
6965 is the byte value with which to fill it, the third argument is an
6966 integer argument specifying the number of bytes to fill, and the fourth
6967 argument is the known alignment of the destination location.
6969 If the call to this intrinsic has an alignment value that is not 0 or 1,
6970 then the caller guarantees that the destination pointer is aligned to
6973 If the ``isvolatile`` parameter is ``true``, the ``llvm.memset`` call is
6974 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
6975 very cleanly specified and it is unwise to depend on it.
6980 The '``llvm.memset.*``' intrinsics fill "len" bytes of memory starting
6981 at the destination location. If the argument is known to be aligned to
6982 some boundary, this can be specified as the fourth argument, otherwise
6983 it should be set to 0 or 1.
6985 '``llvm.sqrt.*``' Intrinsic
6986 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6991 This is an overloaded intrinsic. You can use ``llvm.sqrt`` on any
6992 floating point or vector of floating point type. Not all targets support
6997 declare float @llvm.sqrt.f32(float %Val)
6998 declare double @llvm.sqrt.f64(double %Val)
6999 declare x86_fp80 @llvm.sqrt.f80(x86_fp80 %Val)
7000 declare fp128 @llvm.sqrt.f128(fp128 %Val)
7001 declare ppc_fp128 @llvm.sqrt.ppcf128(ppc_fp128 %Val)
7006 The '``llvm.sqrt``' intrinsics return the sqrt of the specified operand,
7007 returning the same value as the libm '``sqrt``' functions would. Unlike
7008 ``sqrt`` in libm, however, ``llvm.sqrt`` has undefined behavior for
7009 negative numbers other than -0.0 (which allows for better optimization,
7010 because there is no need to worry about errno being set).
7011 ``llvm.sqrt(-0.0)`` is defined to return -0.0 like IEEE sqrt.
7016 The argument and return value are floating point numbers of the same
7022 This function returns the sqrt of the specified operand if it is a
7023 nonnegative floating point number.
7025 '``llvm.powi.*``' Intrinsic
7026 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7031 This is an overloaded intrinsic. You can use ``llvm.powi`` on any
7032 floating point or vector of floating point type. Not all targets support
7037 declare float @llvm.powi.f32(float %Val, i32 %power)
7038 declare double @llvm.powi.f64(double %Val, i32 %power)
7039 declare x86_fp80 @llvm.powi.f80(x86_fp80 %Val, i32 %power)
7040 declare fp128 @llvm.powi.f128(fp128 %Val, i32 %power)
7041 declare ppc_fp128 @llvm.powi.ppcf128(ppc_fp128 %Val, i32 %power)
7046 The '``llvm.powi.*``' intrinsics return the first operand raised to the
7047 specified (positive or negative) power. The order of evaluation of
7048 multiplications is not defined. When a vector of floating point type is
7049 used, the second argument remains a scalar integer value.
7054 The second argument is an integer power, and the first is a value to
7055 raise to that power.
7060 This function returns the first value raised to the second power with an
7061 unspecified sequence of rounding operations.
7063 '``llvm.sin.*``' Intrinsic
7064 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7069 This is an overloaded intrinsic. You can use ``llvm.sin`` on any
7070 floating point or vector of floating point type. Not all targets support
7075 declare float @llvm.sin.f32(float %Val)
7076 declare double @llvm.sin.f64(double %Val)
7077 declare x86_fp80 @llvm.sin.f80(x86_fp80 %Val)
7078 declare fp128 @llvm.sin.f128(fp128 %Val)
7079 declare ppc_fp128 @llvm.sin.ppcf128(ppc_fp128 %Val)
7084 The '``llvm.sin.*``' intrinsics return the sine of the operand.
7089 The argument and return value are floating point numbers of the same
7095 This function returns the sine of the specified operand, returning the
7096 same values as the libm ``sin`` functions would, and handles error
7097 conditions in the same way.
7099 '``llvm.cos.*``' Intrinsic
7100 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7105 This is an overloaded intrinsic. You can use ``llvm.cos`` on any
7106 floating point or vector of floating point type. Not all targets support
7111 declare float @llvm.cos.f32(float %Val)
7112 declare double @llvm.cos.f64(double %Val)
7113 declare x86_fp80 @llvm.cos.f80(x86_fp80 %Val)
7114 declare fp128 @llvm.cos.f128(fp128 %Val)
7115 declare ppc_fp128 @llvm.cos.ppcf128(ppc_fp128 %Val)
7120 The '``llvm.cos.*``' intrinsics return the cosine of the operand.
7125 The argument and return value are floating point numbers of the same
7131 This function returns the cosine of the specified operand, returning the
7132 same values as the libm ``cos`` functions would, and handles error
7133 conditions in the same way.
7135 '``llvm.pow.*``' Intrinsic
7136 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7141 This is an overloaded intrinsic. You can use ``llvm.pow`` on any
7142 floating point or vector of floating point type. Not all targets support
7147 declare float @llvm.pow.f32(float %Val, float %Power)
7148 declare double @llvm.pow.f64(double %Val, double %Power)
7149 declare x86_fp80 @llvm.pow.f80(x86_fp80 %Val, x86_fp80 %Power)
7150 declare fp128 @llvm.pow.f128(fp128 %Val, fp128 %Power)
7151 declare ppc_fp128 @llvm.pow.ppcf128(ppc_fp128 %Val, ppc_fp128 Power)
7156 The '``llvm.pow.*``' intrinsics return the first operand raised to the
7157 specified (positive or negative) power.
7162 The second argument is a floating point power, and the first is a value
7163 to raise to that power.
7168 This function returns the first value raised to the second power,
7169 returning the same values as the libm ``pow`` functions would, and
7170 handles error conditions in the same way.
7172 '``llvm.exp.*``' Intrinsic
7173 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7178 This is an overloaded intrinsic. You can use ``llvm.exp`` on any
7179 floating point or vector of floating point type. Not all targets support
7184 declare float @llvm.exp.f32(float %Val)
7185 declare double @llvm.exp.f64(double %Val)
7186 declare x86_fp80 @llvm.exp.f80(x86_fp80 %Val)
7187 declare fp128 @llvm.exp.f128(fp128 %Val)
7188 declare ppc_fp128 @llvm.exp.ppcf128(ppc_fp128 %Val)
7193 The '``llvm.exp.*``' intrinsics perform the exp function.
7198 The argument and return value are floating point numbers of the same
7204 This function returns the same values as the libm ``exp`` functions
7205 would, and handles error conditions in the same way.
7207 '``llvm.exp2.*``' Intrinsic
7208 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7213 This is an overloaded intrinsic. You can use ``llvm.exp2`` on any
7214 floating point or vector of floating point type. Not all targets support
7219 declare float @llvm.exp2.f32(float %Val)
7220 declare double @llvm.exp2.f64(double %Val)
7221 declare x86_fp80 @llvm.exp2.f80(x86_fp80 %Val)
7222 declare fp128 @llvm.exp2.f128(fp128 %Val)
7223 declare ppc_fp128 @llvm.exp2.ppcf128(ppc_fp128 %Val)
7228 The '``llvm.exp2.*``' intrinsics perform the exp2 function.
7233 The argument and return value are floating point numbers of the same
7239 This function returns the same values as the libm ``exp2`` functions
7240 would, and handles error conditions in the same way.
7242 '``llvm.log.*``' Intrinsic
7243 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7248 This is an overloaded intrinsic. You can use ``llvm.log`` on any
7249 floating point or vector of floating point type. Not all targets support
7254 declare float @llvm.log.f32(float %Val)
7255 declare double @llvm.log.f64(double %Val)
7256 declare x86_fp80 @llvm.log.f80(x86_fp80 %Val)
7257 declare fp128 @llvm.log.f128(fp128 %Val)
7258 declare ppc_fp128 @llvm.log.ppcf128(ppc_fp128 %Val)
7263 The '``llvm.log.*``' intrinsics perform the log function.
7268 The argument and return value are floating point numbers of the same
7274 This function returns the same values as the libm ``log`` functions
7275 would, and handles error conditions in the same way.
7277 '``llvm.log10.*``' Intrinsic
7278 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7283 This is an overloaded intrinsic. You can use ``llvm.log10`` on any
7284 floating point or vector of floating point type. Not all targets support
7289 declare float @llvm.log10.f32(float %Val)
7290 declare double @llvm.log10.f64(double %Val)
7291 declare x86_fp80 @llvm.log10.f80(x86_fp80 %Val)
7292 declare fp128 @llvm.log10.f128(fp128 %Val)
7293 declare ppc_fp128 @llvm.log10.ppcf128(ppc_fp128 %Val)
7298 The '``llvm.log10.*``' intrinsics perform the log10 function.
7303 The argument and return value are floating point numbers of the same
7309 This function returns the same values as the libm ``log10`` functions
7310 would, and handles error conditions in the same way.
7312 '``llvm.log2.*``' Intrinsic
7313 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7318 This is an overloaded intrinsic. You can use ``llvm.log2`` on any
7319 floating point or vector of floating point type. Not all targets support
7324 declare float @llvm.log2.f32(float %Val)
7325 declare double @llvm.log2.f64(double %Val)
7326 declare x86_fp80 @llvm.log2.f80(x86_fp80 %Val)
7327 declare fp128 @llvm.log2.f128(fp128 %Val)
7328 declare ppc_fp128 @llvm.log2.ppcf128(ppc_fp128 %Val)
7333 The '``llvm.log2.*``' intrinsics perform the log2 function.
7338 The argument and return value are floating point numbers of the same
7344 This function returns the same values as the libm ``log2`` functions
7345 would, and handles error conditions in the same way.
7347 '``llvm.fma.*``' Intrinsic
7348 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7353 This is an overloaded intrinsic. You can use ``llvm.fma`` on any
7354 floating point or vector of floating point type. Not all targets support
7359 declare float @llvm.fma.f32(float %a, float %b, float %c)
7360 declare double @llvm.fma.f64(double %a, double %b, double %c)
7361 declare x86_fp80 @llvm.fma.f80(x86_fp80 %a, x86_fp80 %b, x86_fp80 %c)
7362 declare fp128 @llvm.fma.f128(fp128 %a, fp128 %b, fp128 %c)
7363 declare ppc_fp128 @llvm.fma.ppcf128(ppc_fp128 %a, ppc_fp128 %b, ppc_fp128 %c)
7368 The '``llvm.fma.*``' intrinsics perform the fused multiply-add
7374 The argument and return value are floating point numbers of the same
7380 This function returns the same values as the libm ``fma`` functions
7383 '``llvm.fabs.*``' Intrinsic
7384 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7389 This is an overloaded intrinsic. You can use ``llvm.fabs`` on any
7390 floating point or vector of floating point type. Not all targets support
7395 declare float @llvm.fabs.f32(float %Val)
7396 declare double @llvm.fabs.f64(double %Val)
7397 declare x86_fp80 @llvm.fabs.f80(x86_fp80 %Val)
7398 declare fp128 @llvm.fabs.f128(fp128 %Val)
7399 declare ppc_fp128 @llvm.fabs.ppcf128(ppc_fp128 %Val)
7404 The '``llvm.fabs.*``' intrinsics return the absolute value of the
7410 The argument and return value are floating point numbers of the same
7416 This function returns the same values as the libm ``fabs`` functions
7417 would, and handles error conditions in the same way.
7419 '``llvm.copysign.*``' Intrinsic
7420 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7425 This is an overloaded intrinsic. You can use ``llvm.copysign`` on any
7426 floating point or vector of floating point type. Not all targets support
7431 declare float @llvm.copysign.f32(float %Mag, float %Sgn)
7432 declare double @llvm.copysign.f64(double %Mag, double %Sgn)
7433 declare x86_fp80 @llvm.copysign.f80(x86_fp80 %Mag, x86_fp80 %Sgn)
7434 declare fp128 @llvm.copysign.f128(fp128 %Mag, fp128 %Sgn)
7435 declare ppc_fp128 @llvm.copysign.ppcf128(ppc_fp128 %Mag, ppc_fp128 %Sgn)
7440 The '``llvm.copysign.*``' intrinsics return a value with the magnitude of the
7441 first operand and the sign of the second operand.
7446 The arguments and return value are floating point numbers of the same
7452 This function returns the same values as the libm ``copysign``
7453 functions would, and handles error conditions in the same way.
7455 '``llvm.floor.*``' Intrinsic
7456 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7461 This is an overloaded intrinsic. You can use ``llvm.floor`` on any
7462 floating point or vector of floating point type. Not all targets support
7467 declare float @llvm.floor.f32(float %Val)
7468 declare double @llvm.floor.f64(double %Val)
7469 declare x86_fp80 @llvm.floor.f80(x86_fp80 %Val)
7470 declare fp128 @llvm.floor.f128(fp128 %Val)
7471 declare ppc_fp128 @llvm.floor.ppcf128(ppc_fp128 %Val)
7476 The '``llvm.floor.*``' intrinsics return the floor of the operand.
7481 The argument and return value are floating point numbers of the same
7487 This function returns the same values as the libm ``floor`` functions
7488 would, and handles error conditions in the same way.
7490 '``llvm.ceil.*``' Intrinsic
7491 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7496 This is an overloaded intrinsic. You can use ``llvm.ceil`` on any
7497 floating point or vector of floating point type. Not all targets support
7502 declare float @llvm.ceil.f32(float %Val)
7503 declare double @llvm.ceil.f64(double %Val)
7504 declare x86_fp80 @llvm.ceil.f80(x86_fp80 %Val)
7505 declare fp128 @llvm.ceil.f128(fp128 %Val)
7506 declare ppc_fp128 @llvm.ceil.ppcf128(ppc_fp128 %Val)
7511 The '``llvm.ceil.*``' intrinsics return the ceiling of the operand.
7516 The argument and return value are floating point numbers of the same
7522 This function returns the same values as the libm ``ceil`` functions
7523 would, and handles error conditions in the same way.
7525 '``llvm.trunc.*``' Intrinsic
7526 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7531 This is an overloaded intrinsic. You can use ``llvm.trunc`` on any
7532 floating point or vector of floating point type. Not all targets support
7537 declare float @llvm.trunc.f32(float %Val)
7538 declare double @llvm.trunc.f64(double %Val)
7539 declare x86_fp80 @llvm.trunc.f80(x86_fp80 %Val)
7540 declare fp128 @llvm.trunc.f128(fp128 %Val)
7541 declare ppc_fp128 @llvm.trunc.ppcf128(ppc_fp128 %Val)
7546 The '``llvm.trunc.*``' intrinsics returns the operand rounded to the
7547 nearest integer not larger in magnitude than the operand.
7552 The argument and return value are floating point numbers of the same
7558 This function returns the same values as the libm ``trunc`` functions
7559 would, and handles error conditions in the same way.
7561 '``llvm.rint.*``' Intrinsic
7562 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7567 This is an overloaded intrinsic. You can use ``llvm.rint`` on any
7568 floating point or vector of floating point type. Not all targets support
7573 declare float @llvm.rint.f32(float %Val)
7574 declare double @llvm.rint.f64(double %Val)
7575 declare x86_fp80 @llvm.rint.f80(x86_fp80 %Val)
7576 declare fp128 @llvm.rint.f128(fp128 %Val)
7577 declare ppc_fp128 @llvm.rint.ppcf128(ppc_fp128 %Val)
7582 The '``llvm.rint.*``' intrinsics returns the operand rounded to the
7583 nearest integer. It may raise an inexact floating-point exception if the
7584 operand isn't an integer.
7589 The argument and return value are floating point numbers of the same
7595 This function returns the same values as the libm ``rint`` functions
7596 would, and handles error conditions in the same way.
7598 '``llvm.nearbyint.*``' Intrinsic
7599 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7604 This is an overloaded intrinsic. You can use ``llvm.nearbyint`` on any
7605 floating point or vector of floating point type. Not all targets support
7610 declare float @llvm.nearbyint.f32(float %Val)
7611 declare double @llvm.nearbyint.f64(double %Val)
7612 declare x86_fp80 @llvm.nearbyint.f80(x86_fp80 %Val)
7613 declare fp128 @llvm.nearbyint.f128(fp128 %Val)
7614 declare ppc_fp128 @llvm.nearbyint.ppcf128(ppc_fp128 %Val)
7619 The '``llvm.nearbyint.*``' intrinsics returns the operand rounded to the
7625 The argument and return value are floating point numbers of the same
7631 This function returns the same values as the libm ``nearbyint``
7632 functions would, and handles error conditions in the same way.
7634 '``llvm.round.*``' Intrinsic
7635 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7640 This is an overloaded intrinsic. You can use ``llvm.round`` on any
7641 floating point or vector of floating point type. Not all targets support
7646 declare float @llvm.round.f32(float %Val)
7647 declare double @llvm.round.f64(double %Val)
7648 declare x86_fp80 @llvm.round.f80(x86_fp80 %Val)
7649 declare fp128 @llvm.round.f128(fp128 %Val)
7650 declare ppc_fp128 @llvm.round.ppcf128(ppc_fp128 %Val)
7655 The '``llvm.round.*``' intrinsics returns the operand rounded to the
7661 The argument and return value are floating point numbers of the same
7667 This function returns the same values as the libm ``round``
7668 functions would, and handles error conditions in the same way.
7670 Bit Manipulation Intrinsics
7671 ---------------------------
7673 LLVM provides intrinsics for a few important bit manipulation
7674 operations. These allow efficient code generation for some algorithms.
7676 '``llvm.bswap.*``' Intrinsics
7677 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7682 This is an overloaded intrinsic function. You can use bswap on any
7683 integer type that is an even number of bytes (i.e. BitWidth % 16 == 0).
7687 declare i16 @llvm.bswap.i16(i16 <id>)
7688 declare i32 @llvm.bswap.i32(i32 <id>)
7689 declare i64 @llvm.bswap.i64(i64 <id>)
7694 The '``llvm.bswap``' family of intrinsics is used to byte swap integer
7695 values with an even number of bytes (positive multiple of 16 bits).
7696 These are useful for performing operations on data that is not in the
7697 target's native byte order.
7702 The ``llvm.bswap.i16`` intrinsic returns an i16 value that has the high
7703 and low byte of the input i16 swapped. Similarly, the ``llvm.bswap.i32``
7704 intrinsic returns an i32 value that has the four bytes of the input i32
7705 swapped, so that if the input bytes are numbered 0, 1, 2, 3 then the
7706 returned i32 will have its bytes in 3, 2, 1, 0 order. The
7707 ``llvm.bswap.i48``, ``llvm.bswap.i64`` and other intrinsics extend this
7708 concept to additional even-byte lengths (6 bytes, 8 bytes and more,
7711 '``llvm.ctpop.*``' Intrinsic
7712 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7717 This is an overloaded intrinsic. You can use llvm.ctpop on any integer
7718 bit width, or on any vector with integer elements. Not all targets
7719 support all bit widths or vector types, however.
7723 declare i8 @llvm.ctpop.i8(i8 <src>)
7724 declare i16 @llvm.ctpop.i16(i16 <src>)
7725 declare i32 @llvm.ctpop.i32(i32 <src>)
7726 declare i64 @llvm.ctpop.i64(i64 <src>)
7727 declare i256 @llvm.ctpop.i256(i256 <src>)
7728 declare <2 x i32> @llvm.ctpop.v2i32(<2 x i32> <src>)
7733 The '``llvm.ctpop``' family of intrinsics counts the number of bits set
7739 The only argument is the value to be counted. The argument may be of any
7740 integer type, or a vector with integer elements. The return type must
7741 match the argument type.
7746 The '``llvm.ctpop``' intrinsic counts the 1's in a variable, or within
7747 each element of a vector.
7749 '``llvm.ctlz.*``' Intrinsic
7750 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7755 This is an overloaded intrinsic. You can use ``llvm.ctlz`` on any
7756 integer bit width, or any vector whose elements are integers. Not all
7757 targets support all bit widths or vector types, however.
7761 declare i8 @llvm.ctlz.i8 (i8 <src>, i1 <is_zero_undef>)
7762 declare i16 @llvm.ctlz.i16 (i16 <src>, i1 <is_zero_undef>)
7763 declare i32 @llvm.ctlz.i32 (i32 <src>, i1 <is_zero_undef>)
7764 declare i64 @llvm.ctlz.i64 (i64 <src>, i1 <is_zero_undef>)
7765 declare i256 @llvm.ctlz.i256(i256 <src>, i1 <is_zero_undef>)
7766 declase <2 x i32> @llvm.ctlz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
7771 The '``llvm.ctlz``' family of intrinsic functions counts the number of
7772 leading zeros in a variable.
7777 The first argument is the value to be counted. This argument may be of
7778 any integer type, or a vectory with integer element type. The return
7779 type must match the first argument type.
7781 The second argument must be a constant and is a flag to indicate whether
7782 the intrinsic should ensure that a zero as the first argument produces a
7783 defined result. Historically some architectures did not provide a
7784 defined result for zero values as efficiently, and many algorithms are
7785 now predicated on avoiding zero-value inputs.
7790 The '``llvm.ctlz``' intrinsic counts the leading (most significant)
7791 zeros in a variable, or within each element of the vector. If
7792 ``src == 0`` then the result is the size in bits of the type of ``src``
7793 if ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
7794 ``llvm.ctlz(i32 2) = 30``.
7796 '``llvm.cttz.*``' Intrinsic
7797 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7802 This is an overloaded intrinsic. You can use ``llvm.cttz`` on any
7803 integer bit width, or any vector of integer elements. Not all targets
7804 support all bit widths or vector types, however.
7808 declare i8 @llvm.cttz.i8 (i8 <src>, i1 <is_zero_undef>)
7809 declare i16 @llvm.cttz.i16 (i16 <src>, i1 <is_zero_undef>)
7810 declare i32 @llvm.cttz.i32 (i32 <src>, i1 <is_zero_undef>)
7811 declare i64 @llvm.cttz.i64 (i64 <src>, i1 <is_zero_undef>)
7812 declare i256 @llvm.cttz.i256(i256 <src>, i1 <is_zero_undef>)
7813 declase <2 x i32> @llvm.cttz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
7818 The '``llvm.cttz``' family of intrinsic functions counts the number of
7824 The first argument is the value to be counted. This argument may be of
7825 any integer type, or a vectory with integer element type. The return
7826 type must match the first argument type.
7828 The second argument must be a constant and is a flag to indicate whether
7829 the intrinsic should ensure that a zero as the first argument produces a
7830 defined result. Historically some architectures did not provide a
7831 defined result for zero values as efficiently, and many algorithms are
7832 now predicated on avoiding zero-value inputs.
7837 The '``llvm.cttz``' intrinsic counts the trailing (least significant)
7838 zeros in a variable, or within each element of a vector. If ``src == 0``
7839 then the result is the size in bits of the type of ``src`` if
7840 ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
7841 ``llvm.cttz(2) = 1``.
7843 Arithmetic with Overflow Intrinsics
7844 -----------------------------------
7846 LLVM provides intrinsics for some arithmetic with overflow operations.
7848 '``llvm.sadd.with.overflow.*``' Intrinsics
7849 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7854 This is an overloaded intrinsic. You can use ``llvm.sadd.with.overflow``
7855 on any integer bit width.
7859 declare {i16, i1} @llvm.sadd.with.overflow.i16(i16 %a, i16 %b)
7860 declare {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
7861 declare {i64, i1} @llvm.sadd.with.overflow.i64(i64 %a, i64 %b)
7866 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
7867 a signed addition of the two arguments, and indicate whether an overflow
7868 occurred during the signed summation.
7873 The arguments (%a and %b) and the first element of the result structure
7874 may be of integer types of any bit width, but they must have the same
7875 bit width. The second element of the result structure must be of type
7876 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
7882 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
7883 a signed addition of the two variables. They return a structure --- the
7884 first element of which is the signed summation, and the second element
7885 of which is a bit specifying if the signed summation resulted in an
7891 .. code-block:: llvm
7893 %res = call {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
7894 %sum = extractvalue {i32, i1} %res, 0
7895 %obit = extractvalue {i32, i1} %res, 1
7896 br i1 %obit, label %overflow, label %normal
7898 '``llvm.uadd.with.overflow.*``' Intrinsics
7899 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7904 This is an overloaded intrinsic. You can use ``llvm.uadd.with.overflow``
7905 on any integer bit width.
7909 declare {i16, i1} @llvm.uadd.with.overflow.i16(i16 %a, i16 %b)
7910 declare {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
7911 declare {i64, i1} @llvm.uadd.with.overflow.i64(i64 %a, i64 %b)
7916 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
7917 an unsigned addition of the two arguments, and indicate whether a carry
7918 occurred during the unsigned summation.
7923 The arguments (%a and %b) and the first element of the result structure
7924 may be of integer types of any bit width, but they must have the same
7925 bit width. The second element of the result structure must be of type
7926 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
7932 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
7933 an unsigned addition of the two arguments. They return a structure --- the
7934 first element of which is the sum, and the second element of which is a
7935 bit specifying if the unsigned summation resulted in a carry.
7940 .. code-block:: llvm
7942 %res = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
7943 %sum = extractvalue {i32, i1} %res, 0
7944 %obit = extractvalue {i32, i1} %res, 1
7945 br i1 %obit, label %carry, label %normal
7947 '``llvm.ssub.with.overflow.*``' Intrinsics
7948 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7953 This is an overloaded intrinsic. You can use ``llvm.ssub.with.overflow``
7954 on any integer bit width.
7958 declare {i16, i1} @llvm.ssub.with.overflow.i16(i16 %a, i16 %b)
7959 declare {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
7960 declare {i64, i1} @llvm.ssub.with.overflow.i64(i64 %a, i64 %b)
7965 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
7966 a signed subtraction of the two arguments, and indicate whether an
7967 overflow occurred during the signed subtraction.
7972 The arguments (%a and %b) and the first element of the result structure
7973 may be of integer types of any bit width, but they must have the same
7974 bit width. The second element of the result structure must be of type
7975 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
7981 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
7982 a signed subtraction of the two arguments. They return a structure --- the
7983 first element of which is the subtraction, and the second element of
7984 which is a bit specifying if the signed subtraction resulted in an
7990 .. code-block:: llvm
7992 %res = call {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
7993 %sum = extractvalue {i32, i1} %res, 0
7994 %obit = extractvalue {i32, i1} %res, 1
7995 br i1 %obit, label %overflow, label %normal
7997 '``llvm.usub.with.overflow.*``' Intrinsics
7998 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8003 This is an overloaded intrinsic. You can use ``llvm.usub.with.overflow``
8004 on any integer bit width.
8008 declare {i16, i1} @llvm.usub.with.overflow.i16(i16 %a, i16 %b)
8009 declare {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
8010 declare {i64, i1} @llvm.usub.with.overflow.i64(i64 %a, i64 %b)
8015 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
8016 an unsigned subtraction of the two arguments, and indicate whether an
8017 overflow occurred during the unsigned subtraction.
8022 The arguments (%a and %b) and the first element of the result structure
8023 may be of integer types of any bit width, but they must have the same
8024 bit width. The second element of the result structure must be of type
8025 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
8031 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
8032 an unsigned subtraction of the two arguments. They return a structure ---
8033 the first element of which is the subtraction, and the second element of
8034 which is a bit specifying if the unsigned subtraction resulted in an
8040 .. code-block:: llvm
8042 %res = call {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
8043 %sum = extractvalue {i32, i1} %res, 0
8044 %obit = extractvalue {i32, i1} %res, 1
8045 br i1 %obit, label %overflow, label %normal
8047 '``llvm.smul.with.overflow.*``' Intrinsics
8048 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8053 This is an overloaded intrinsic. You can use ``llvm.smul.with.overflow``
8054 on any integer bit width.
8058 declare {i16, i1} @llvm.smul.with.overflow.i16(i16 %a, i16 %b)
8059 declare {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
8060 declare {i64, i1} @llvm.smul.with.overflow.i64(i64 %a, i64 %b)
8065 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
8066 a signed multiplication of the two arguments, and indicate whether an
8067 overflow occurred during the signed multiplication.
8072 The arguments (%a and %b) and the first element of the result structure
8073 may be of integer types of any bit width, but they must have the same
8074 bit width. The second element of the result structure must be of type
8075 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
8081 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
8082 a signed multiplication of the two arguments. They return a structure ---
8083 the first element of which is the multiplication, and the second element
8084 of which is a bit specifying if the signed multiplication resulted in an
8090 .. code-block:: llvm
8092 %res = call {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
8093 %sum = extractvalue {i32, i1} %res, 0
8094 %obit = extractvalue {i32, i1} %res, 1
8095 br i1 %obit, label %overflow, label %normal
8097 '``llvm.umul.with.overflow.*``' Intrinsics
8098 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8103 This is an overloaded intrinsic. You can use ``llvm.umul.with.overflow``
8104 on any integer bit width.
8108 declare {i16, i1} @llvm.umul.with.overflow.i16(i16 %a, i16 %b)
8109 declare {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
8110 declare {i64, i1} @llvm.umul.with.overflow.i64(i64 %a, i64 %b)
8115 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
8116 a unsigned multiplication of the two arguments, and indicate whether an
8117 overflow occurred during the unsigned multiplication.
8122 The arguments (%a and %b) and the first element of the result structure
8123 may be of integer types of any bit width, but they must have the same
8124 bit width. The second element of the result structure must be of type
8125 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
8131 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
8132 an unsigned multiplication of the two arguments. They return a structure ---
8133 the first element of which is the multiplication, and the second
8134 element of which is a bit specifying if the unsigned multiplication
8135 resulted in an overflow.
8140 .. code-block:: llvm
8142 %res = call {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
8143 %sum = extractvalue {i32, i1} %res, 0
8144 %obit = extractvalue {i32, i1} %res, 1
8145 br i1 %obit, label %overflow, label %normal
8147 Specialised Arithmetic Intrinsics
8148 ---------------------------------
8150 '``llvm.fmuladd.*``' Intrinsic
8151 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8158 declare float @llvm.fmuladd.f32(float %a, float %b, float %c)
8159 declare double @llvm.fmuladd.f64(double %a, double %b, double %c)
8164 The '``llvm.fmuladd.*``' intrinsic functions represent multiply-add
8165 expressions that can be fused if the code generator determines that (a) the
8166 target instruction set has support for a fused operation, and (b) that the
8167 fused operation is more efficient than the equivalent, separate pair of mul
8168 and add instructions.
8173 The '``llvm.fmuladd.*``' intrinsics each take three arguments: two
8174 multiplicands, a and b, and an addend c.
8183 %0 = call float @llvm.fmuladd.f32(%a, %b, %c)
8185 is equivalent to the expression a \* b + c, except that rounding will
8186 not be performed between the multiplication and addition steps if the
8187 code generator fuses the operations. Fusion is not guaranteed, even if
8188 the target platform supports it. If a fused multiply-add is required the
8189 corresponding llvm.fma.\* intrinsic function should be used instead.
8194 .. code-block:: llvm
8196 %r2 = call float @llvm.fmuladd.f32(float %a, float %b, float %c) ; yields {float}:r2 = (a * b) + c
8198 Half Precision Floating Point Intrinsics
8199 ----------------------------------------
8201 For most target platforms, half precision floating point is a
8202 storage-only format. This means that it is a dense encoding (in memory)
8203 but does not support computation in the format.
8205 This means that code must first load the half-precision floating point
8206 value as an i16, then convert it to float with
8207 :ref:`llvm.convert.from.fp16 <int_convert_from_fp16>`. Computation can
8208 then be performed on the float value (including extending to double
8209 etc). To store the value back to memory, it is first converted to float
8210 if needed, then converted to i16 with
8211 :ref:`llvm.convert.to.fp16 <int_convert_to_fp16>`, then storing as an
8214 .. _int_convert_to_fp16:
8216 '``llvm.convert.to.fp16``' Intrinsic
8217 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8224 declare i16 @llvm.convert.to.fp16(f32 %a)
8229 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion
8230 from single precision floating point format to half precision floating
8236 The intrinsic function contains single argument - the value to be
8242 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion
8243 from single precision floating point format to half precision floating
8244 point format. The return value is an ``i16`` which contains the
8250 .. code-block:: llvm
8252 %res = call i16 @llvm.convert.to.fp16(f32 %a)
8253 store i16 %res, i16* @x, align 2
8255 .. _int_convert_from_fp16:
8257 '``llvm.convert.from.fp16``' Intrinsic
8258 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8265 declare f32 @llvm.convert.from.fp16(i16 %a)
8270 The '``llvm.convert.from.fp16``' intrinsic function performs a
8271 conversion from half precision floating point format to single precision
8272 floating point format.
8277 The intrinsic function contains single argument - the value to be
8283 The '``llvm.convert.from.fp16``' intrinsic function performs a
8284 conversion from half single precision floating point format to single
8285 precision floating point format. The input half-float value is
8286 represented by an ``i16`` value.
8291 .. code-block:: llvm
8293 %a = load i16* @x, align 2
8294 %res = call f32 @llvm.convert.from.fp16(i16 %a)
8299 The LLVM debugger intrinsics (which all start with ``llvm.dbg.``
8300 prefix), are described in the `LLVM Source Level
8301 Debugging <SourceLevelDebugging.html#format_common_intrinsics>`_
8304 Exception Handling Intrinsics
8305 -----------------------------
8307 The LLVM exception handling intrinsics (which all start with
8308 ``llvm.eh.`` prefix), are described in the `LLVM Exception
8309 Handling <ExceptionHandling.html#format_common_intrinsics>`_ document.
8313 Trampoline Intrinsics
8314 ---------------------
8316 These intrinsics make it possible to excise one parameter, marked with
8317 the :ref:`nest <nest>` attribute, from a function. The result is a
8318 callable function pointer lacking the nest parameter - the caller does
8319 not need to provide a value for it. Instead, the value to use is stored
8320 in advance in a "trampoline", a block of memory usually allocated on the
8321 stack, which also contains code to splice the nest value into the
8322 argument list. This is used to implement the GCC nested function address
8325 For example, if the function is ``i32 f(i8* nest %c, i32 %x, i32 %y)``
8326 then the resulting function pointer has signature ``i32 (i32, i32)*``.
8327 It can be created as follows:
8329 .. code-block:: llvm
8331 %tramp = alloca [10 x i8], align 4 ; size and alignment only correct for X86
8332 %tramp1 = getelementptr [10 x i8]* %tramp, i32 0, i32 0
8333 call i8* @llvm.init.trampoline(i8* %tramp1, i8* bitcast (i32 (i8*, i32, i32)* @f to i8*), i8* %nval)
8334 %p = call i8* @llvm.adjust.trampoline(i8* %tramp1)
8335 %fp = bitcast i8* %p to i32 (i32, i32)*
8337 The call ``%val = call i32 %fp(i32 %x, i32 %y)`` is then equivalent to
8338 ``%val = call i32 %f(i8* %nval, i32 %x, i32 %y)``.
8342 '``llvm.init.trampoline``' Intrinsic
8343 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8350 declare void @llvm.init.trampoline(i8* <tramp>, i8* <func>, i8* <nval>)
8355 This fills the memory pointed to by ``tramp`` with executable code,
8356 turning it into a trampoline.
8361 The ``llvm.init.trampoline`` intrinsic takes three arguments, all
8362 pointers. The ``tramp`` argument must point to a sufficiently large and
8363 sufficiently aligned block of memory; this memory is written to by the
8364 intrinsic. Note that the size and the alignment are target-specific -
8365 LLVM currently provides no portable way of determining them, so a
8366 front-end that generates this intrinsic needs to have some
8367 target-specific knowledge. The ``func`` argument must hold a function
8368 bitcast to an ``i8*``.
8373 The block of memory pointed to by ``tramp`` is filled with target
8374 dependent code, turning it into a function. Then ``tramp`` needs to be
8375 passed to :ref:`llvm.adjust.trampoline <int_at>` to get a pointer which can
8376 be :ref:`bitcast (to a new function) and called <int_trampoline>`. The new
8377 function's signature is the same as that of ``func`` with any arguments
8378 marked with the ``nest`` attribute removed. At most one such ``nest``
8379 argument is allowed, and it must be of pointer type. Calling the new
8380 function is equivalent to calling ``func`` with the same argument list,
8381 but with ``nval`` used for the missing ``nest`` argument. If, after
8382 calling ``llvm.init.trampoline``, the memory pointed to by ``tramp`` is
8383 modified, then the effect of any later call to the returned function
8384 pointer is undefined.
8388 '``llvm.adjust.trampoline``' Intrinsic
8389 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8396 declare i8* @llvm.adjust.trampoline(i8* <tramp>)
8401 This performs any required machine-specific adjustment to the address of
8402 a trampoline (passed as ``tramp``).
8407 ``tramp`` must point to a block of memory which already has trampoline
8408 code filled in by a previous call to
8409 :ref:`llvm.init.trampoline <int_it>`.
8414 On some architectures the address of the code to be executed needs to be
8415 different to the address where the trampoline is actually stored. This
8416 intrinsic returns the executable address corresponding to ``tramp``
8417 after performing the required machine specific adjustments. The pointer
8418 returned can then be :ref:`bitcast and executed <int_trampoline>`.
8423 This class of intrinsics exists to information about the lifetime of
8424 memory objects and ranges where variables are immutable.
8426 '``llvm.lifetime.start``' Intrinsic
8427 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8434 declare void @llvm.lifetime.start(i64 <size>, i8* nocapture <ptr>)
8439 The '``llvm.lifetime.start``' intrinsic specifies the start of a memory
8445 The first argument is a constant integer representing the size of the
8446 object, or -1 if it is variable sized. The second argument is a pointer
8452 This intrinsic indicates that before this point in the code, the value
8453 of the memory pointed to by ``ptr`` is dead. This means that it is known
8454 to never be used and has an undefined value. A load from the pointer
8455 that precedes this intrinsic can be replaced with ``'undef'``.
8457 '``llvm.lifetime.end``' Intrinsic
8458 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8465 declare void @llvm.lifetime.end(i64 <size>, i8* nocapture <ptr>)
8470 The '``llvm.lifetime.end``' intrinsic specifies the end of a memory
8476 The first argument is a constant integer representing the size of the
8477 object, or -1 if it is variable sized. The second argument is a pointer
8483 This intrinsic indicates that after this point in the code, the value of
8484 the memory pointed to by ``ptr`` is dead. This means that it is known to
8485 never be used and has an undefined value. Any stores into the memory
8486 object following this intrinsic may be removed as dead.
8488 '``llvm.invariant.start``' Intrinsic
8489 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8496 declare {}* @llvm.invariant.start(i64 <size>, i8* nocapture <ptr>)
8501 The '``llvm.invariant.start``' intrinsic specifies that the contents of
8502 a memory object will not change.
8507 The first argument is a constant integer representing the size of the
8508 object, or -1 if it is variable sized. The second argument is a pointer
8514 This intrinsic indicates that until an ``llvm.invariant.end`` that uses
8515 the return value, the referenced memory location is constant and
8518 '``llvm.invariant.end``' Intrinsic
8519 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8526 declare void @llvm.invariant.end({}* <start>, i64 <size>, i8* nocapture <ptr>)
8531 The '``llvm.invariant.end``' intrinsic specifies that the contents of a
8532 memory object are mutable.
8537 The first argument is the matching ``llvm.invariant.start`` intrinsic.
8538 The second argument is a constant integer representing the size of the
8539 object, or -1 if it is variable sized and the third argument is a
8540 pointer to the object.
8545 This intrinsic indicates that the memory is mutable again.
8550 This class of intrinsics is designed to be generic and has no specific
8553 '``llvm.var.annotation``' Intrinsic
8554 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8561 declare void @llvm.var.annotation(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
8566 The '``llvm.var.annotation``' intrinsic.
8571 The first argument is a pointer to a value, the second is a pointer to a
8572 global string, the third is a pointer to a global string which is the
8573 source file name, and the last argument is the line number.
8578 This intrinsic allows annotation of local variables with arbitrary
8579 strings. This can be useful for special purpose optimizations that want
8580 to look for these annotations. These have no other defined use; they are
8581 ignored by code generation and optimization.
8583 '``llvm.ptr.annotation.*``' Intrinsic
8584 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8589 This is an overloaded intrinsic. You can use '``llvm.ptr.annotation``' on a
8590 pointer to an integer of any width. *NOTE* you must specify an address space for
8591 the pointer. The identifier for the default address space is the integer
8596 declare i8* @llvm.ptr.annotation.p<address space>i8(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
8597 declare i16* @llvm.ptr.annotation.p<address space>i16(i16* <val>, i8* <str>, i8* <str>, i32 <int>)
8598 declare i32* @llvm.ptr.annotation.p<address space>i32(i32* <val>, i8* <str>, i8* <str>, i32 <int>)
8599 declare i64* @llvm.ptr.annotation.p<address space>i64(i64* <val>, i8* <str>, i8* <str>, i32 <int>)
8600 declare i256* @llvm.ptr.annotation.p<address space>i256(i256* <val>, i8* <str>, i8* <str>, i32 <int>)
8605 The '``llvm.ptr.annotation``' intrinsic.
8610 The first argument is a pointer to an integer value of arbitrary bitwidth
8611 (result of some expression), the second is a pointer to a global string, the
8612 third is a pointer to a global string which is the source file name, and the
8613 last argument is the line number. It returns the value of the first argument.
8618 This intrinsic allows annotation of a pointer to an integer with arbitrary
8619 strings. This can be useful for special purpose optimizations that want to look
8620 for these annotations. These have no other defined use; they are ignored by code
8621 generation and optimization.
8623 '``llvm.annotation.*``' Intrinsic
8624 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8629 This is an overloaded intrinsic. You can use '``llvm.annotation``' on
8630 any integer bit width.
8634 declare i8 @llvm.annotation.i8(i8 <val>, i8* <str>, i8* <str>, i32 <int>)
8635 declare i16 @llvm.annotation.i16(i16 <val>, i8* <str>, i8* <str>, i32 <int>)
8636 declare i32 @llvm.annotation.i32(i32 <val>, i8* <str>, i8* <str>, i32 <int>)
8637 declare i64 @llvm.annotation.i64(i64 <val>, i8* <str>, i8* <str>, i32 <int>)
8638 declare i256 @llvm.annotation.i256(i256 <val>, i8* <str>, i8* <str>, i32 <int>)
8643 The '``llvm.annotation``' intrinsic.
8648 The first argument is an integer value (result of some expression), the
8649 second is a pointer to a global string, the third is a pointer to a
8650 global string which is the source file name, and the last argument is
8651 the line number. It returns the value of the first argument.
8656 This intrinsic allows annotations to be put on arbitrary expressions
8657 with arbitrary strings. This can be useful for special purpose
8658 optimizations that want to look for these annotations. These have no
8659 other defined use; they are ignored by code generation and optimization.
8661 '``llvm.trap``' Intrinsic
8662 ^^^^^^^^^^^^^^^^^^^^^^^^^
8669 declare void @llvm.trap() noreturn nounwind
8674 The '``llvm.trap``' intrinsic.
8684 This intrinsic is lowered to the target dependent trap instruction. If
8685 the target does not have a trap instruction, this intrinsic will be
8686 lowered to a call of the ``abort()`` function.
8688 '``llvm.debugtrap``' Intrinsic
8689 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8696 declare void @llvm.debugtrap() nounwind
8701 The '``llvm.debugtrap``' intrinsic.
8711 This intrinsic is lowered to code which is intended to cause an
8712 execution trap with the intention of requesting the attention of a
8715 '``llvm.stackprotector``' Intrinsic
8716 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8723 declare void @llvm.stackprotector(i8* <guard>, i8** <slot>)
8728 The ``llvm.stackprotector`` intrinsic takes the ``guard`` and stores it
8729 onto the stack at ``slot``. The stack slot is adjusted to ensure that it
8730 is placed on the stack before local variables.
8735 The ``llvm.stackprotector`` intrinsic requires two pointer arguments.
8736 The first argument is the value loaded from the stack guard
8737 ``@__stack_chk_guard``. The second variable is an ``alloca`` that has
8738 enough space to hold the value of the guard.
8743 This intrinsic causes the prologue/epilogue inserter to force the position of
8744 the ``AllocaInst`` stack slot to be before local variables on the stack. This is
8745 to ensure that if a local variable on the stack is overwritten, it will destroy
8746 the value of the guard. When the function exits, the guard on the stack is
8747 checked against the original guard by ``llvm.stackprotectorcheck``. If they are
8748 different, then ``llvm.stackprotectorcheck`` causes the program to abort by
8749 calling the ``__stack_chk_fail()`` function.
8751 '``llvm.stackprotectorcheck``' Intrinsic
8752 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8759 declare void @llvm.stackprotectorcheck(i8** <guard>)
8764 The ``llvm.stackprotectorcheck`` intrinsic compares ``guard`` against an already
8765 created stack protector and if they are not equal calls the
8766 ``__stack_chk_fail()`` function.
8771 The ``llvm.stackprotectorcheck`` intrinsic requires one pointer argument, the
8772 the variable ``@__stack_chk_guard``.
8777 This intrinsic is provided to perform the stack protector check by comparing
8778 ``guard`` with the stack slot created by ``llvm.stackprotector`` and if the
8779 values do not match call the ``__stack_chk_fail()`` function.
8781 The reason to provide this as an IR level intrinsic instead of implementing it
8782 via other IR operations is that in order to perform this operation at the IR
8783 level without an intrinsic, one would need to create additional basic blocks to
8784 handle the success/failure cases. This makes it difficult to stop the stack
8785 protector check from disrupting sibling tail calls in Codegen. With this
8786 intrinsic, we are able to generate the stack protector basic blocks late in
8787 codegen after the tail call decision has occured.
8789 '``llvm.objectsize``' Intrinsic
8790 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8797 declare i32 @llvm.objectsize.i32(i8* <object>, i1 <min>)
8798 declare i64 @llvm.objectsize.i64(i8* <object>, i1 <min>)
8803 The ``llvm.objectsize`` intrinsic is designed to provide information to
8804 the optimizers to determine at compile time whether a) an operation
8805 (like memcpy) will overflow a buffer that corresponds to an object, or
8806 b) that a runtime check for overflow isn't necessary. An object in this
8807 context means an allocation of a specific class, structure, array, or
8813 The ``llvm.objectsize`` intrinsic takes two arguments. The first
8814 argument is a pointer to or into the ``object``. The second argument is
8815 a boolean and determines whether ``llvm.objectsize`` returns 0 (if true)
8816 or -1 (if false) when the object size is unknown. The second argument
8817 only accepts constants.
8822 The ``llvm.objectsize`` intrinsic is lowered to a constant representing
8823 the size of the object concerned. If the size cannot be determined at
8824 compile time, ``llvm.objectsize`` returns ``i32/i64 -1 or 0`` (depending
8825 on the ``min`` argument).
8827 '``llvm.expect``' Intrinsic
8828 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8835 declare i32 @llvm.expect.i32(i32 <val>, i32 <expected_val>)
8836 declare i64 @llvm.expect.i64(i64 <val>, i64 <expected_val>)
8841 The ``llvm.expect`` intrinsic provides information about expected (the
8842 most probable) value of ``val``, which can be used by optimizers.
8847 The ``llvm.expect`` intrinsic takes two arguments. The first argument is
8848 a value. The second argument is an expected value, this needs to be a
8849 constant value, variables are not allowed.
8854 This intrinsic is lowered to the ``val``.
8856 '``llvm.donothing``' Intrinsic
8857 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8864 declare void @llvm.donothing() nounwind readnone
8869 The ``llvm.donothing`` intrinsic doesn't perform any operation. It's the
8870 only intrinsic that can be called with an invoke instruction.
8880 This intrinsic does nothing, and it's removed by optimizers and ignored