1 ==============================
2 LLVM Language Reference Manual
3 ==============================
12 This document is a reference manual for the LLVM assembly language. LLVM
13 is a Static Single Assignment (SSA) based representation that provides
14 type safety, low-level operations, flexibility, and the capability of
15 representing 'all' high-level languages cleanly. It is the common code
16 representation used throughout all phases of the LLVM compilation
22 The LLVM code representation is designed to be used in three different
23 forms: as an in-memory compiler IR, as an on-disk bitcode representation
24 (suitable for fast loading by a Just-In-Time compiler), and as a human
25 readable assembly language representation. This allows LLVM to provide a
26 powerful intermediate representation for efficient compiler
27 transformations and analysis, while providing a natural means to debug
28 and visualize the transformations. The three different forms of LLVM are
29 all equivalent. This document describes the human readable
30 representation and notation.
32 The LLVM representation aims to be light-weight and low-level while
33 being expressive, typed, and extensible at the same time. It aims to be
34 a "universal IR" of sorts, by being at a low enough level that
35 high-level ideas may be cleanly mapped to it (similar to how
36 microprocessors are "universal IR's", allowing many source languages to
37 be mapped to them). By providing type information, LLVM can be used as
38 the target of optimizations: for example, through pointer analysis, it
39 can be proven that a C automatic variable is never accessed outside of
40 the current function, allowing it to be promoted to a simple SSA value
41 instead of a memory location.
48 It is important to note that this document describes 'well formed' LLVM
49 assembly language. There is a difference between what the parser accepts
50 and what is considered 'well formed'. For example, the following
51 instruction is syntactically okay, but not well formed:
57 because the definition of ``%x`` does not dominate all of its uses. The
58 LLVM infrastructure provides a verification pass that may be used to
59 verify that an LLVM module is well formed. This pass is automatically
60 run by the parser after parsing input assembly and by the optimizer
61 before it outputs bitcode. The violations pointed out by the verifier
62 pass indicate bugs in transformation passes or input to the parser.
69 LLVM identifiers come in two basic types: global and local. Global
70 identifiers (functions, global variables) begin with the ``'@'``
71 character. Local identifiers (register names, types) begin with the
72 ``'%'`` character. Additionally, there are three different formats for
73 identifiers, for different purposes:
75 #. Named values are represented as a string of characters with their
76 prefix. For example, ``%foo``, ``@DivisionByZero``,
77 ``%a.really.long.identifier``. The actual regular expression used is
78 '``[%@][a-zA-Z$._][a-zA-Z$._0-9]*``'. Identifiers which require other
79 characters in their names can be surrounded with quotes. Special
80 characters may be escaped using ``"\xx"`` where ``xx`` is the ASCII
81 code for the character in hexadecimal. In this way, any character can
82 be used in a name value, even quotes themselves.
83 #. Unnamed values are represented as an unsigned numeric value with
84 their prefix. For example, ``%12``, ``@2``, ``%44``.
85 #. Constants, which are described in the section Constants_ below.
87 LLVM requires that values start with a prefix for two reasons: Compilers
88 don't need to worry about name clashes with reserved words, and the set
89 of reserved words may be expanded in the future without penalty.
90 Additionally, unnamed identifiers allow a compiler to quickly come up
91 with a temporary variable without having to avoid symbol table
94 Reserved words in LLVM are very similar to reserved words in other
95 languages. There are keywords for different opcodes ('``add``',
96 '``bitcast``', '``ret``', etc...), for primitive type names ('``void``',
97 '``i32``', etc...), and others. These reserved words cannot conflict
98 with variable names, because none of them start with a prefix character
101 Here is an example of LLVM code to multiply the integer variable
108 %result = mul i32 %X, 8
110 After strength reduction:
114 %result = shl i32 %X, 3
120 %0 = add i32 %X, %X ; yields {i32}:%0
121 %1 = add i32 %0, %0 ; yields {i32}:%1
122 %result = add i32 %1, %1
124 This last way of multiplying ``%X`` by 8 illustrates several important
125 lexical features of LLVM:
127 #. Comments are delimited with a '``;``' and go until the end of line.
128 #. Unnamed temporaries are created when the result of a computation is
129 not assigned to a named value.
130 #. Unnamed temporaries are numbered sequentially (using a per-function
131 incrementing counter, starting with 0). Note that basic blocks are
132 included in this numbering. For example, if the entry basic block is not
133 given a label name, then it will get number 0.
135 It also shows a convention that we follow in this document. When
136 demonstrating instructions, we will follow an instruction with a comment
137 that defines the type and name of value produced.
145 LLVM programs are composed of ``Module``'s, each of which is a
146 translation unit of the input programs. Each module consists of
147 functions, global variables, and symbol table entries. Modules may be
148 combined together with the LLVM linker, which merges function (and
149 global variable) definitions, resolves forward declarations, and merges
150 symbol table entries. Here is an example of the "hello world" module:
154 ; Declare the string constant as a global constant.
155 @.str = private unnamed_addr constant [13 x i8] c"hello world\0A\00"
157 ; External declaration of the puts function
158 declare i32 @puts(i8* nocapture) nounwind
160 ; Definition of main function
161 define i32 @main() { ; i32()*
162 ; Convert [13 x i8]* to i8 *...
163 %cast210 = getelementptr [13 x i8]* @.str, i64 0, i64 0
165 ; Call puts function to write out the string to stdout.
166 call i32 @puts(i8* %cast210)
171 !1 = metadata !{i32 42}
174 This example is made up of a :ref:`global variable <globalvars>` named
175 "``.str``", an external declaration of the "``puts``" function, a
176 :ref:`function definition <functionstructure>` for "``main``" and
177 :ref:`named metadata <namedmetadatastructure>` "``foo``".
179 In general, a module is made up of a list of global values (where both
180 functions and global variables are global values). Global values are
181 represented by a pointer to a memory location (in this case, a pointer
182 to an array of char, and a pointer to a function), and have one of the
183 following :ref:`linkage types <linkage>`.
190 All Global Variables and Functions have one of the following types of
194 Global values with "``private``" linkage are only directly
195 accessible by objects in the current module. In particular, linking
196 code into a module with an private global value may cause the
197 private to be renamed as necessary to avoid collisions. Because the
198 symbol is private to the module, all references can be updated. This
199 doesn't show up in any symbol table in the object file.
201 Similar to ``private``, but the symbol is passed through the
202 assembler and evaluated by the linker. Unlike normal strong symbols,
203 they are removed by the linker from the final linked image
204 (executable or dynamic library).
205 ``linker_private_weak``
206 Similar to "``linker_private``", but the symbol is weak. Note that
207 ``linker_private_weak`` symbols are subject to coalescing by the
208 linker. The symbols are removed by the linker from the final linked
209 image (executable or dynamic library).
211 Similar to private, but the value shows as a local symbol
212 (``STB_LOCAL`` in the case of ELF) in the object file. This
213 corresponds to the notion of the '``static``' keyword in C.
214 ``available_externally``
215 Globals with "``available_externally``" linkage are never emitted
216 into the object file corresponding to the LLVM module. They exist to
217 allow inlining and other optimizations to take place given knowledge
218 of the definition of the global, which is known to be somewhere
219 outside the module. Globals with ``available_externally`` linkage
220 are allowed to be discarded at will, and are otherwise the same as
221 ``linkonce_odr``. This linkage type is only allowed on definitions,
224 Globals with "``linkonce``" linkage are merged with other globals of
225 the same name when linkage occurs. This can be used to implement
226 some forms of inline functions, templates, or other code which must
227 be generated in each translation unit that uses it, but where the
228 body may be overridden with a more definitive definition later.
229 Unreferenced ``linkonce`` globals are allowed to be discarded. Note
230 that ``linkonce`` linkage does not actually allow the optimizer to
231 inline the body of this function into callers because it doesn't
232 know if this definition of the function is the definitive definition
233 within the program or whether it will be overridden by a stronger
234 definition. To enable inlining and other optimizations, use
235 "``linkonce_odr``" linkage.
237 "``weak``" linkage has the same merging semantics as ``linkonce``
238 linkage, except that unreferenced globals with ``weak`` linkage may
239 not be discarded. This is used for globals that are declared "weak"
242 "``common``" linkage is most similar to "``weak``" linkage, but they
243 are used for tentative definitions in C, such as "``int X;``" at
244 global scope. Symbols with "``common``" linkage are merged in the
245 same way as ``weak symbols``, and they may not be deleted if
246 unreferenced. ``common`` symbols may not have an explicit section,
247 must have a zero initializer, and may not be marked
248 ':ref:`constant <globalvars>`'. Functions and aliases may not have
251 .. _linkage_appending:
254 "``appending``" linkage may only be applied to global variables of
255 pointer to array type. When two global variables with appending
256 linkage are linked together, the two global arrays are appended
257 together. This is the LLVM, typesafe, equivalent of having the
258 system linker append together "sections" with identical names when
261 The semantics of this linkage follow the ELF object file model: the
262 symbol is weak until linked, if not linked, the symbol becomes null
263 instead of being an undefined reference.
264 ``linkonce_odr``, ``weak_odr``
265 Some languages allow differing globals to be merged, such as two
266 functions with different semantics. Other languages, such as
267 ``C++``, ensure that only equivalent globals are ever merged (the
268 "one definition rule" --- "ODR"). Such languages can use the
269 ``linkonce_odr`` and ``weak_odr`` linkage types to indicate that the
270 global will only be merged with equivalent globals. These linkage
271 types are otherwise the same as their non-``odr`` versions.
273 If none of the above identifiers are used, the global is externally
274 visible, meaning that it participates in linkage and can be used to
275 resolve external symbol references.
277 It is illegal for a function *declaration* to have any linkage type
278 other than ``external`` or ``extern_weak``.
285 LLVM :ref:`functions <functionstructure>`, :ref:`calls <i_call>` and
286 :ref:`invokes <i_invoke>` can all have an optional calling convention
287 specified for the call. The calling convention of any pair of dynamic
288 caller/callee must match, or the behavior of the program is undefined.
289 The following calling conventions are supported by LLVM, and more may be
292 "``ccc``" - The C calling convention
293 This calling convention (the default if no other calling convention
294 is specified) matches the target C calling conventions. This calling
295 convention supports varargs function calls and tolerates some
296 mismatch in the declared prototype and implemented declaration of
297 the function (as does normal C).
298 "``fastcc``" - The fast calling convention
299 This calling convention attempts to make calls as fast as possible
300 (e.g. by passing things in registers). This calling convention
301 allows the target to use whatever tricks it wants to produce fast
302 code for the target, without having to conform to an externally
303 specified ABI (Application Binary Interface). `Tail calls can only
304 be optimized when this, the GHC or the HiPE convention is
305 used. <CodeGenerator.html#id80>`_ This calling convention does not
306 support varargs and requires the prototype of all callees to exactly
307 match the prototype of the function definition.
308 "``coldcc``" - The cold calling convention
309 This calling convention attempts to make code in the caller as
310 efficient as possible under the assumption that the call is not
311 commonly executed. As such, these calls often preserve all registers
312 so that the call does not break any live ranges in the caller side.
313 This calling convention does not support varargs and requires the
314 prototype of all callees to exactly match the prototype of the
316 "``cc 10``" - GHC convention
317 This calling convention has been implemented specifically for use by
318 the `Glasgow Haskell Compiler (GHC) <http://www.haskell.org/ghc>`_.
319 It passes everything in registers, going to extremes to achieve this
320 by disabling callee save registers. This calling convention should
321 not be used lightly but only for specific situations such as an
322 alternative to the *register pinning* performance technique often
323 used when implementing functional programming languages. At the
324 moment only X86 supports this convention and it has the following
327 - On *X86-32* only supports up to 4 bit type parameters. No
328 floating point types are supported.
329 - On *X86-64* only supports up to 10 bit type parameters and 6
330 floating point parameters.
332 This calling convention supports `tail call
333 optimization <CodeGenerator.html#id80>`_ but requires both the
334 caller and callee are using it.
335 "``cc 11``" - The HiPE calling convention
336 This calling convention has been implemented specifically for use by
337 the `High-Performance Erlang
338 (HiPE) <http://www.it.uu.se/research/group/hipe/>`_ compiler, *the*
339 native code compiler of the `Ericsson's Open Source Erlang/OTP
340 system <http://www.erlang.org/download.shtml>`_. It uses more
341 registers for argument passing than the ordinary C calling
342 convention and defines no callee-saved registers. The calling
343 convention properly supports `tail call
344 optimization <CodeGenerator.html#id80>`_ but requires that both the
345 caller and the callee use it. It uses a *register pinning*
346 mechanism, similar to GHC's convention, for keeping frequently
347 accessed runtime components pinned to specific hardware registers.
348 At the moment only X86 supports this convention (both 32 and 64
350 "``webkit_jscc``" - WebKit's JavaScript calling convention
351 This calling convention has been implemented for `WebKit FTL JIT
352 <https://trac.webkit.org/wiki/FTLJIT>`_. It passes arguments on the
353 stack right to left (as cdecl does), and returns a value in the
354 platform's customary return register.
355 "``anyregcc``" - Dynamic calling convention for code patching
356 This is a special convention that supports patching an arbitrary code
357 sequence in place of a call site. This convention forces the call
358 arguments into registers but allows them to be dynamcially
359 allocated. This can currently only be used with calls to
360 llvm.experimental.patchpoint because only this intrinsic records
361 the location of its arguments in a side table. See :doc:`StackMaps`.
362 "``cc <n>``" - Numbered convention
363 Any calling convention may be specified by number, allowing
364 target-specific calling conventions to be used. Target specific
365 calling conventions start at 64.
367 More calling conventions can be added/defined on an as-needed basis, to
368 support Pascal conventions or any other well-known target-independent
371 .. _visibilitystyles:
376 All Global Variables and Functions have one of the following visibility
379 "``default``" - Default style
380 On targets that use the ELF object file format, default visibility
381 means that the declaration is visible to other modules and, in
382 shared libraries, means that the declared entity may be overridden.
383 On Darwin, default visibility means that the declaration is visible
384 to other modules. Default visibility corresponds to "external
385 linkage" in the language.
386 "``hidden``" - Hidden style
387 Two declarations of an object with hidden visibility refer to the
388 same object if they are in the same shared object. Usually, hidden
389 visibility indicates that the symbol will not be placed into the
390 dynamic symbol table, so no other module (executable or shared
391 library) can reference it directly.
392 "``protected``" - Protected style
393 On ELF, protected visibility indicates that the symbol will be
394 placed in the dynamic symbol table, but that references within the
395 defining module will bind to the local symbol. That is, the symbol
396 cannot be overridden by another module.
403 All Global Variables, Functions and Aliases can have one of the following
407 "``dllimport``" causes the compiler to reference a function or variable via
408 a global pointer to a pointer that is set up by the DLL exporting the
409 symbol. On Microsoft Windows targets, the pointer name is formed by
410 combining ``__imp_`` and the function or variable name.
412 "``dllexport``" causes the compiler to provide a global pointer to a pointer
413 in a DLL, so that it can be referenced with the ``dllimport`` attribute. On
414 Microsoft Windows targets, the pointer name is formed by combining
415 ``__imp_`` and the function or variable name. Since this storage class
416 exists for defining a dll interface, the compiler, assembler and linker know
417 it is externally referenced and must refrain from deleting the symbol.
422 LLVM IR allows you to specify name aliases for certain types. This can
423 make it easier to read the IR and make the IR more condensed
424 (particularly when recursive types are involved). An example of a name
429 %mytype = type { %mytype*, i32 }
431 You may give a name to any :ref:`type <typesystem>` except
432 ":ref:`void <t_void>`". Type name aliases may be used anywhere a type is
433 expected with the syntax "%mytype".
435 Note that type names are aliases for the structural type that they
436 indicate, and that you can therefore specify multiple names for the same
437 type. This often leads to confusing behavior when dumping out a .ll
438 file. Since LLVM IR uses structural typing, the name is not part of the
439 type. When printing out LLVM IR, the printer will pick *one name* to
440 render all types of a particular shape. This means that if you have code
441 where two different source types end up having the same LLVM type, that
442 the dumper will sometimes print the "wrong" or unexpected type. This is
443 an important design point and isn't going to change.
450 Global variables define regions of memory allocated at compilation time
453 Global variables definitions must be initialized, may have an explicit section
454 to be placed in, and may have an optional explicit alignment specified.
456 Global variables in other translation units can also be declared, in which
457 case they don't have an initializer.
459 A variable may be defined as ``thread_local``, which means that it will
460 not be shared by threads (each thread will have a separated copy of the
461 variable). Not all targets support thread-local variables. Optionally, a
462 TLS model may be specified:
465 For variables that are only used within the current shared library.
467 For variables in modules that will not be loaded dynamically.
469 For variables defined in the executable and only used within it.
471 The models correspond to the ELF TLS models; see `ELF Handling For
472 Thread-Local Storage <http://people.redhat.com/drepper/tls.pdf>`_ for
473 more information on under which circumstances the different models may
474 be used. The target may choose a different TLS model if the specified
475 model is not supported, or if a better choice of model can be made.
477 A variable may be defined as a global ``constant``, which indicates that
478 the contents of the variable will **never** be modified (enabling better
479 optimization, allowing the global data to be placed in the read-only
480 section of an executable, etc). Note that variables that need runtime
481 initialization cannot be marked ``constant`` as there is a store to the
484 LLVM explicitly allows *declarations* of global variables to be marked
485 constant, even if the final definition of the global is not. This
486 capability can be used to enable slightly better optimization of the
487 program, but requires the language definition to guarantee that
488 optimizations based on the 'constantness' are valid for the translation
489 units that do not include the definition.
491 As SSA values, global variables define pointer values that are in scope
492 (i.e. they dominate) all basic blocks in the program. Global variables
493 always define a pointer to their "content" type because they describe a
494 region of memory, and all memory objects in LLVM are accessed through
497 Global variables can be marked with ``unnamed_addr`` which indicates
498 that the address is not significant, only the content. Constants marked
499 like this can be merged with other constants if they have the same
500 initializer. Note that a constant with significant address *can* be
501 merged with a ``unnamed_addr`` constant, the result being a constant
502 whose address is significant.
504 A global variable may be declared to reside in a target-specific
505 numbered address space. For targets that support them, address spaces
506 may affect how optimizations are performed and/or what target
507 instructions are used to access the variable. The default address space
508 is zero. The address space qualifier must precede any other attributes.
510 LLVM allows an explicit section to be specified for globals. If the
511 target supports it, it will emit globals to the section specified.
513 By default, global initializers are optimized by assuming that global
514 variables defined within the module are not modified from their
515 initial values before the start of the global initializer. This is
516 true even for variables potentially accessible from outside the
517 module, including those with external linkage or appearing in
518 ``@llvm.used`` or dllexported variables. This assumption may be suppressed
519 by marking the variable with ``externally_initialized``.
521 An explicit alignment may be specified for a global, which must be a
522 power of 2. If not present, or if the alignment is set to zero, the
523 alignment of the global is set by the target to whatever it feels
524 convenient. If an explicit alignment is specified, the global is forced
525 to have exactly that alignment. Targets and optimizers are not allowed
526 to over-align the global if the global has an assigned section. In this
527 case, the extra alignment could be observable: for example, code could
528 assume that the globals are densely packed in their section and try to
529 iterate over them as an array, alignment padding would break this
532 Globals can also have a :ref:`DLL storage class <dllstorageclass>`.
536 [@<GlobalVarName> =] [Linkage] [Visibility] [DLLStorageClass] [ThreadLocal]
537 [AddrSpace] [unnamed_addr] [ExternallyInitialized]
538 <global | constant> <Type>
539 [, section "name"] [, align <Alignment>]
541 For example, the following defines a global in a numbered address space
542 with an initializer, section, and alignment:
546 @G = addrspace(5) constant float 1.0, section "foo", align 4
548 The following example just declares a global variable
552 @G = external global i32
554 The following example defines a thread-local global with the
555 ``initialexec`` TLS model:
559 @G = thread_local(initialexec) global i32 0, align 4
561 .. _functionstructure:
566 LLVM function definitions consist of the "``define``" keyword, an
567 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
568 style <visibility>`, an optional :ref:`DLL storage class <dllstorageclass>`,
569 an optional :ref:`calling convention <callingconv>`,
570 an optional ``unnamed_addr`` attribute, a return type, an optional
571 :ref:`parameter attribute <paramattrs>` for the return type, a function
572 name, a (possibly empty) argument list (each with optional :ref:`parameter
573 attributes <paramattrs>`), optional :ref:`function attributes <fnattrs>`,
574 an optional section, an optional alignment, an optional :ref:`garbage
575 collector name <gc>`, an optional :ref:`prefix <prefixdata>`, an opening
576 curly brace, a list of basic blocks, and a closing curly brace.
578 LLVM function declarations consist of the "``declare``" keyword, an
579 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
580 style <visibility>`, an optional :ref:`DLL storage class <dllstorageclass>`,
581 an optional :ref:`calling convention <callingconv>`,
582 an optional ``unnamed_addr`` attribute, a return type, an optional
583 :ref:`parameter attribute <paramattrs>` for the return type, a function
584 name, a possibly empty list of arguments, an optional alignment, an optional
585 :ref:`garbage collector name <gc>` and an optional :ref:`prefix <prefixdata>`.
587 A function definition contains a list of basic blocks, forming the CFG (Control
588 Flow Graph) for the function. Each basic block may optionally start with a label
589 (giving the basic block a symbol table entry), contains a list of instructions,
590 and ends with a :ref:`terminator <terminators>` instruction (such as a branch or
591 function return). If an explicit label is not provided, a block is assigned an
592 implicit numbered label, using the next value from the same counter as used for
593 unnamed temporaries (:ref:`see above<identifiers>`). For example, if a function
594 entry block does not have an explicit label, it will be assigned label "%0",
595 then the first unnamed temporary in that block will be "%1", etc.
597 The first basic block in a function is special in two ways: it is
598 immediately executed on entrance to the function, and it is not allowed
599 to have predecessor basic blocks (i.e. there can not be any branches to
600 the entry block of a function). Because the block can have no
601 predecessors, it also cannot have any :ref:`PHI nodes <i_phi>`.
603 LLVM allows an explicit section to be specified for functions. If the
604 target supports it, it will emit functions to the section specified.
606 An explicit alignment may be specified for a function. If not present,
607 or if the alignment is set to zero, the alignment of the function is set
608 by the target to whatever it feels convenient. If an explicit alignment
609 is specified, the function is forced to have at least that much
610 alignment. All alignments must be a power of 2.
612 If the ``unnamed_addr`` attribute is given, the address is know to not
613 be significant and two identical functions can be merged.
617 define [linkage] [visibility] [DLLStorageClass]
619 <ResultType> @<FunctionName> ([argument list])
620 [fn Attrs] [section "name"] [align N]
621 [gc] [prefix Constant] { ... }
628 Aliases act as "second name" for the aliasee value (which can be either
629 function, global variable, another alias or bitcast of global value).
630 Aliases may have an optional :ref:`linkage type <linkage>`, an optional
631 :ref:`visibility style <visibility>`, and an optional :ref:`DLL storage class
636 @<Name> = [Visibility] [DLLStorageClass] alias [Linkage] <AliaseeTy> @<Aliasee>
638 The linkage must be one of ``private``, ``linker_private``,
639 ``linker_private_weak``, ``internal``, ``linkonce``, ``weak``,
640 ``linkonce_odr``, ``weak_odr``, ``external``. Note that some system linkers
641 might not correctly handle dropping a weak symbol that is aliased by a non-weak
644 .. _namedmetadatastructure:
649 Named metadata is a collection of metadata. :ref:`Metadata
650 nodes <metadata>` (but not metadata strings) are the only valid
651 operands for a named metadata.
655 ; Some unnamed metadata nodes, which are referenced by the named metadata.
656 !0 = metadata !{metadata !"zero"}
657 !1 = metadata !{metadata !"one"}
658 !2 = metadata !{metadata !"two"}
660 !name = !{!0, !1, !2}
667 The return type and each parameter of a function type may have a set of
668 *parameter attributes* associated with them. Parameter attributes are
669 used to communicate additional information about the result or
670 parameters of a function. Parameter attributes are considered to be part
671 of the function, not of the function type, so functions with different
672 parameter attributes can have the same function type.
674 Parameter attributes are simple keywords that follow the type specified.
675 If multiple parameter attributes are needed, they are space separated.
680 declare i32 @printf(i8* noalias nocapture, ...)
681 declare i32 @atoi(i8 zeroext)
682 declare signext i8 @returns_signed_char()
684 Note that any attributes for the function result (``nounwind``,
685 ``readonly``) come immediately after the argument list.
687 Currently, only the following parameter attributes are defined:
690 This indicates to the code generator that the parameter or return
691 value should be zero-extended to the extent required by the target's
692 ABI (which is usually 32-bits, but is 8-bits for a i1 on x86-64) by
693 the caller (for a parameter) or the callee (for a return value).
695 This indicates to the code generator that the parameter or return
696 value should be sign-extended to the extent required by the target's
697 ABI (which is usually 32-bits) by the caller (for a parameter) or
698 the callee (for a return value).
700 This indicates that this parameter or return value should be treated
701 in a special target-dependent fashion during while emitting code for
702 a function call or return (usually, by putting it in a register as
703 opposed to memory, though some targets use it to distinguish between
704 two different kinds of registers). Use of this attribute is
707 This indicates that the pointer parameter should really be passed by
708 value to the function. The attribute implies that a hidden copy of
709 the pointee is made between the caller and the callee, so the callee
710 is unable to modify the value in the caller. This attribute is only
711 valid on LLVM pointer arguments. It is generally used to pass
712 structs and arrays by value, but is also valid on pointers to
713 scalars. The copy is considered to belong to the caller not the
714 callee (for example, ``readonly`` functions should not write to
715 ``byval`` parameters). This is not a valid attribute for return
718 The byval attribute also supports specifying an alignment with the
719 align attribute. It indicates the alignment of the stack slot to
720 form and the known alignment of the pointer specified to the call
721 site. If the alignment is not specified, then the code generator
722 makes a target-specific assumption.
728 .. Warning:: This feature is unstable and not fully implemented.
730 The ``inalloca`` argument attribute allows the caller to get the
731 address of an outgoing argument to a ``call`` or ``invoke`` before
732 it executes. It is similar to ``byval`` in that it is used to pass
733 arguments by value, but it guarantees that the argument will not be
736 To be :ref:`well formed <wellformed>`, the caller must pass in an
737 alloca value into an ``inalloca`` parameter, and an alloca may be
738 used as an ``inalloca`` argument at most once. The attribute can
739 only be applied to parameters that would be passed in memory and not
740 registers. The ``inalloca`` attribute cannot be used in conjunction
741 with other attributes that affect argument storage, like ``inreg``,
742 ``nest``, ``sret``, or ``byval``. The ``inalloca`` stack space is
743 considered to be clobbered by any call that uses it, so any
744 ``inalloca`` parameters cannot be marked ``readonly``.
746 Allocas passed with ``inalloca`` to a call must be in the opposite
747 order of the parameter list, meaning that the rightmost argument
748 must be allocated first. If a call has inalloca arguments, no other
749 allocas can occur between the first alloca used by the call and the
750 call site, unless they are are cleared by calls to
751 :ref:`llvm.stackrestore <int_stackrestore>`. Violating these rules
752 results in undefined behavior at runtime.
754 See :doc:`InAlloca` for more information on how to use this
758 This indicates that the pointer parameter specifies the address of a
759 structure that is the return value of the function in the source
760 program. This pointer must be guaranteed by the caller to be valid:
761 loads and stores to the structure may be assumed by the callee
762 not to trap and to be properly aligned. This may only be applied to
763 the first parameter. This is not a valid attribute for return
766 This indicates that pointer values :ref:`based <pointeraliasing>` on
767 the argument or return value do not alias pointer values which are
768 not *based* on it, ignoring certain "irrelevant" dependencies. For a
769 call to the parent function, dependencies between memory references
770 from before or after the call and from those during the call are
771 "irrelevant" to the ``noalias`` keyword for the arguments and return
772 value used in that call. The caller shares the responsibility with
773 the callee for ensuring that these requirements are met. For further
774 details, please see the discussion of the NoAlias response in `alias
775 analysis <AliasAnalysis.html#MustMayNo>`_.
777 Note that this definition of ``noalias`` is intentionally similar
778 to the definition of ``restrict`` in C99 for function arguments,
779 though it is slightly weaker.
781 For function return values, C99's ``restrict`` is not meaningful,
782 while LLVM's ``noalias`` is.
784 This indicates that the callee does not make any copies of the
785 pointer that outlive the callee itself. This is not a valid
786 attribute for return values.
791 This indicates that the pointer parameter can be excised using the
792 :ref:`trampoline intrinsics <int_trampoline>`. This is not a valid
793 attribute for return values and can only be applied to one parameter.
796 This indicates that the function always returns the argument as its return
797 value. This is an optimization hint to the code generator when generating
798 the caller, allowing tail call optimization and omission of register saves
799 and restores in some cases; it is not checked or enforced when generating
800 the callee. The parameter and the function return type must be valid
801 operands for the :ref:`bitcast instruction <i_bitcast>`. This is not a
802 valid attribute for return values and can only be applied to one parameter.
806 Garbage Collector Names
807 -----------------------
809 Each function may specify a garbage collector name, which is simply a
814 define void @f() gc "name" { ... }
816 The compiler declares the supported values of *name*. Specifying a
817 collector which will cause the compiler to alter its output in order to
818 support the named garbage collection algorithm.
825 Prefix data is data associated with a function which the code generator
826 will emit immediately before the function body. The purpose of this feature
827 is to allow frontends to associate language-specific runtime metadata with
828 specific functions and make it available through the function pointer while
829 still allowing the function pointer to be called. To access the data for a
830 given function, a program may bitcast the function pointer to a pointer to
831 the constant's type. This implies that the IR symbol points to the start
834 To maintain the semantics of ordinary function calls, the prefix data must
835 have a particular format. Specifically, it must begin with a sequence of
836 bytes which decode to a sequence of machine instructions, valid for the
837 module's target, which transfer control to the point immediately succeeding
838 the prefix data, without performing any other visible action. This allows
839 the inliner and other passes to reason about the semantics of the function
840 definition without needing to reason about the prefix data. Obviously this
841 makes the format of the prefix data highly target dependent.
843 Prefix data is laid out as if it were an initializer for a global variable
844 of the prefix data's type. No padding is automatically placed between the
845 prefix data and the function body. If padding is required, it must be part
848 A trivial example of valid prefix data for the x86 architecture is ``i8 144``,
849 which encodes the ``nop`` instruction:
853 define void @f() prefix i8 144 { ... }
855 Generally prefix data can be formed by encoding a relative branch instruction
856 which skips the metadata, as in this example of valid prefix data for the
857 x86_64 architecture, where the first two bytes encode ``jmp .+10``:
861 %0 = type <{ i8, i8, i8* }>
863 define void @f() prefix %0 <{ i8 235, i8 8, i8* @md}> { ... }
865 A function may have prefix data but no body. This has similar semantics
866 to the ``available_externally`` linkage in that the data may be used by the
867 optimizers but will not be emitted in the object file.
874 Attribute groups are groups of attributes that are referenced by objects within
875 the IR. They are important for keeping ``.ll`` files readable, because a lot of
876 functions will use the same set of attributes. In the degenerative case of a
877 ``.ll`` file that corresponds to a single ``.c`` file, the single attribute
878 group will capture the important command line flags used to build that file.
880 An attribute group is a module-level object. To use an attribute group, an
881 object references the attribute group's ID (e.g. ``#37``). An object may refer
882 to more than one attribute group. In that situation, the attributes from the
883 different groups are merged.
885 Here is an example of attribute groups for a function that should always be
886 inlined, has a stack alignment of 4, and which shouldn't use SSE instructions:
890 ; Target-independent attributes:
891 attributes #0 = { alwaysinline alignstack=4 }
893 ; Target-dependent attributes:
894 attributes #1 = { "no-sse" }
896 ; Function @f has attributes: alwaysinline, alignstack=4, and "no-sse".
897 define void @f() #0 #1 { ... }
904 Function attributes are set to communicate additional information about
905 a function. Function attributes are considered to be part of the
906 function, not of the function type, so functions with different function
907 attributes can have the same function type.
909 Function attributes are simple keywords that follow the type specified.
910 If multiple attributes are needed, they are space separated. For
915 define void @f() noinline { ... }
916 define void @f() alwaysinline { ... }
917 define void @f() alwaysinline optsize { ... }
918 define void @f() optsize { ... }
921 This attribute indicates that, when emitting the prologue and
922 epilogue, the backend should forcibly align the stack pointer.
923 Specify the desired alignment, which must be a power of two, in
926 This attribute indicates that the inliner should attempt to inline
927 this function into callers whenever possible, ignoring any active
928 inlining size threshold for this caller.
930 This indicates that the callee function at a call site should be
931 recognized as a built-in function, even though the function's declaration
932 uses the ``nobuiltin`` attribute. This is only valid at call sites for
933 direct calls to functions which are declared with the ``nobuiltin``
936 This attribute indicates that this function is rarely called. When
937 computing edge weights, basic blocks post-dominated by a cold
938 function call are also considered to be cold; and, thus, given low
941 This attribute indicates that the source code contained a hint that
942 inlining this function is desirable (such as the "inline" keyword in
943 C/C++). It is just a hint; it imposes no requirements on the
946 This attribute suggests that optimization passes and code generator
947 passes make choices that keep the code size of this function as small
948 as possible and perform optimizations that may sacrifice runtime
949 performance in order to minimize the size of the generated code.
951 This attribute disables prologue / epilogue emission for the
952 function. This can have very system-specific consequences.
954 This indicates that the callee function at a call site is not recognized as
955 a built-in function. LLVM will retain the original call and not replace it
956 with equivalent code based on the semantics of the built-in function, unless
957 the call site uses the ``builtin`` attribute. This is valid at call sites
958 and on function declarations and definitions.
960 This attribute indicates that calls to the function cannot be
961 duplicated. A call to a ``noduplicate`` function may be moved
962 within its parent function, but may not be duplicated within
965 A function containing a ``noduplicate`` call may still
966 be an inlining candidate, provided that the call is not
967 duplicated by inlining. That implies that the function has
968 internal linkage and only has one call site, so the original
969 call is dead after inlining.
971 This attributes disables implicit floating point instructions.
973 This attribute indicates that the inliner should never inline this
974 function in any situation. This attribute may not be used together
975 with the ``alwaysinline`` attribute.
977 This attribute suppresses lazy symbol binding for the function. This
978 may make calls to the function faster, at the cost of extra program
979 startup time if the function is not called during program startup.
981 This attribute indicates that the code generator should not use a
982 red zone, even if the target-specific ABI normally permits it.
984 This function attribute indicates that the function never returns
985 normally. This produces undefined behavior at runtime if the
986 function ever does dynamically return.
988 This function attribute indicates that the function never returns
989 with an unwind or exceptional control flow. If the function does
990 unwind, its runtime behavior is undefined.
992 This function attribute indicates that the function is not optimized
993 by any optimization or code generator passes with the
994 exception of interprocedural optimization passes.
995 This attribute cannot be used together with the ``alwaysinline``
996 attribute; this attribute is also incompatible
997 with the ``minsize`` attribute and the ``optsize`` attribute.
999 This attribute requires the ``noinline`` attribute to be specified on
1000 the function as well, so the function is never inlined into any caller.
1001 Only functions with the ``alwaysinline`` attribute are valid
1002 candidates for inlining into the body of this function.
1004 This attribute suggests that optimization passes and code generator
1005 passes make choices that keep the code size of this function low,
1006 and otherwise do optimizations specifically to reduce code size as
1007 long as they do not significantly impact runtime performance.
1009 On a function, this attribute indicates that the function computes its
1010 result (or decides to unwind an exception) based strictly on its arguments,
1011 without dereferencing any pointer arguments or otherwise accessing
1012 any mutable state (e.g. memory, control registers, etc) visible to
1013 caller functions. It does not write through any pointer arguments
1014 (including ``byval`` arguments) and never changes any state visible
1015 to callers. This means that it cannot unwind exceptions by calling
1016 the ``C++`` exception throwing methods.
1018 On an argument, this attribute indicates that the function does not
1019 dereference that pointer argument, even though it may read or write the
1020 memory that the pointer points to if accessed through other pointers.
1022 On a function, this attribute indicates that the function does not write
1023 through any pointer arguments (including ``byval`` arguments) or otherwise
1024 modify any state (e.g. memory, control registers, etc) visible to
1025 caller functions. It may dereference pointer arguments and read
1026 state that may be set in the caller. A readonly function always
1027 returns the same value (or unwinds an exception identically) when
1028 called with the same set of arguments and global state. It cannot
1029 unwind an exception by calling the ``C++`` exception throwing
1032 On an argument, this attribute indicates that the function does not write
1033 through this pointer argument, even though it may write to the memory that
1034 the pointer points to.
1036 This attribute indicates that this function can return twice. The C
1037 ``setjmp`` is an example of such a function. The compiler disables
1038 some optimizations (like tail calls) in the caller of these
1040 ``sanitize_address``
1041 This attribute indicates that AddressSanitizer checks
1042 (dynamic address safety analysis) are enabled for this function.
1044 This attribute indicates that MemorySanitizer checks (dynamic detection
1045 of accesses to uninitialized memory) are enabled for this function.
1047 This attribute indicates that ThreadSanitizer checks
1048 (dynamic thread safety analysis) are enabled for this function.
1050 This attribute indicates that the function should emit a stack
1051 smashing protector. It is in the form of a "canary" --- a random value
1052 placed on the stack before the local variables that's checked upon
1053 return from the function to see if it has been overwritten. A
1054 heuristic is used to determine if a function needs stack protectors
1055 or not. The heuristic used will enable protectors for functions with:
1057 - Character arrays larger than ``ssp-buffer-size`` (default 8).
1058 - Aggregates containing character arrays larger than ``ssp-buffer-size``.
1059 - Calls to alloca() with variable sizes or constant sizes greater than
1060 ``ssp-buffer-size``.
1062 If a function that has an ``ssp`` attribute is inlined into a
1063 function that doesn't have an ``ssp`` attribute, then the resulting
1064 function will have an ``ssp`` attribute.
1066 This attribute indicates that the function should *always* emit a
1067 stack smashing protector. This overrides the ``ssp`` function
1070 If a function that has an ``sspreq`` attribute is inlined into a
1071 function that doesn't have an ``sspreq`` attribute or which has an
1072 ``ssp`` or ``sspstrong`` attribute, then the resulting function will have
1073 an ``sspreq`` attribute.
1075 This attribute indicates that the function should emit a stack smashing
1076 protector. This attribute causes a strong heuristic to be used when
1077 determining if a function needs stack protectors. The strong heuristic
1078 will enable protectors for functions with:
1080 - Arrays of any size and type
1081 - Aggregates containing an array of any size and type.
1082 - Calls to alloca().
1083 - Local variables that have had their address taken.
1085 This overrides the ``ssp`` function attribute.
1087 If a function that has an ``sspstrong`` attribute is inlined into a
1088 function that doesn't have an ``sspstrong`` attribute, then the
1089 resulting function will have an ``sspstrong`` attribute.
1091 This attribute indicates that the ABI being targeted requires that
1092 an unwind table entry be produce for this function even if we can
1093 show that no exceptions passes by it. This is normally the case for
1094 the ELF x86-64 abi, but it can be disabled for some compilation
1099 Module-Level Inline Assembly
1100 ----------------------------
1102 Modules may contain "module-level inline asm" blocks, which corresponds
1103 to the GCC "file scope inline asm" blocks. These blocks are internally
1104 concatenated by LLVM and treated as a single unit, but may be separated
1105 in the ``.ll`` file if desired. The syntax is very simple:
1107 .. code-block:: llvm
1109 module asm "inline asm code goes here"
1110 module asm "more can go here"
1112 The strings can contain any character by escaping non-printable
1113 characters. The escape sequence used is simply "\\xx" where "xx" is the
1114 two digit hex code for the number.
1116 The inline asm code is simply printed to the machine code .s file when
1117 assembly code is generated.
1119 .. _langref_datalayout:
1124 A module may specify a target specific data layout string that specifies
1125 how data is to be laid out in memory. The syntax for the data layout is
1128 .. code-block:: llvm
1130 target datalayout = "layout specification"
1132 The *layout specification* consists of a list of specifications
1133 separated by the minus sign character ('-'). Each specification starts
1134 with a letter and may include other information after the letter to
1135 define some aspect of the data layout. The specifications accepted are
1139 Specifies that the target lays out data in big-endian form. That is,
1140 the bits with the most significance have the lowest address
1143 Specifies that the target lays out data in little-endian form. That
1144 is, the bits with the least significance have the lowest address
1147 Specifies the natural alignment of the stack in bits. Alignment
1148 promotion of stack variables is limited to the natural stack
1149 alignment to avoid dynamic stack realignment. The stack alignment
1150 must be a multiple of 8-bits. If omitted, the natural stack
1151 alignment defaults to "unspecified", which does not prevent any
1152 alignment promotions.
1153 ``p[n]:<size>:<abi>:<pref>``
1154 This specifies the *size* of a pointer and its ``<abi>`` and
1155 ``<pref>``\erred alignments for address space ``n``. All sizes are in
1156 bits. The address space, ``n`` is optional, and if not specified,
1157 denotes the default address space 0. The value of ``n`` must be
1158 in the range [1,2^23).
1159 ``i<size>:<abi>:<pref>``
1160 This specifies the alignment for an integer type of a given bit
1161 ``<size>``. The value of ``<size>`` must be in the range [1,2^23).
1162 ``v<size>:<abi>:<pref>``
1163 This specifies the alignment for a vector type of a given bit
1165 ``f<size>:<abi>:<pref>``
1166 This specifies the alignment for a floating point type of a given bit
1167 ``<size>``. Only values of ``<size>`` that are supported by the target
1168 will work. 32 (float) and 64 (double) are supported on all targets; 80
1169 or 128 (different flavors of long double) are also supported on some
1172 This specifies the alignment for an object of aggregate type.
1174 If present, specifies that llvm names are mangled in the output. The
1177 * ``e``: ELF mangling: Private symbols get a ``.L`` prefix.
1178 * ``m``: Mips mangling: Private symbols get a ``$`` prefix.
1179 * ``o``: Mach-O mangling: Private symbols get ``L`` prefix. Other
1180 symbols get a ``_`` prefix.
1181 * ``w``: Windows COFF prefix: Similar to Mach-O, but stdcall and fastcall
1182 functions also get a suffix based on the frame size.
1183 ``n<size1>:<size2>:<size3>...``
1184 This specifies a set of native integer widths for the target CPU in
1185 bits. For example, it might contain ``n32`` for 32-bit PowerPC,
1186 ``n32:64`` for PowerPC 64, or ``n8:16:32:64`` for X86-64. Elements of
1187 this set are considered to support most general arithmetic operations
1190 On every specification that takes a ``<abi>:<pref>``, specifying the
1191 ``<pref>`` alignment is optional. If omitted, the preceding ``:``
1192 should be omitted too and ``<pref>`` will be equal to ``<abi>``.
1194 When constructing the data layout for a given target, LLVM starts with a
1195 default set of specifications which are then (possibly) overridden by
1196 the specifications in the ``datalayout`` keyword. The default
1197 specifications are given in this list:
1199 - ``E`` - big endian
1200 - ``p:64:64:64`` - 64-bit pointers with 64-bit alignment.
1201 - ``p[n]:64:64:64`` - Other address spaces are assumed to be the
1202 same as the default address space.
1203 - ``S0`` - natural stack alignment is unspecified
1204 - ``i1:8:8`` - i1 is 8-bit (byte) aligned
1205 - ``i8:8:8`` - i8 is 8-bit (byte) aligned
1206 - ``i16:16:16`` - i16 is 16-bit aligned
1207 - ``i32:32:32`` - i32 is 32-bit aligned
1208 - ``i64:32:64`` - i64 has ABI alignment of 32-bits but preferred
1209 alignment of 64-bits
1210 - ``f16:16:16`` - half is 16-bit aligned
1211 - ``f32:32:32`` - float is 32-bit aligned
1212 - ``f64:64:64`` - double is 64-bit aligned
1213 - ``f128:128:128`` - quad is 128-bit aligned
1214 - ``v64:64:64`` - 64-bit vector is 64-bit aligned
1215 - ``v128:128:128`` - 128-bit vector is 128-bit aligned
1216 - ``a:0:64`` - aggregates are 64-bit aligned
1218 When LLVM is determining the alignment for a given type, it uses the
1221 #. If the type sought is an exact match for one of the specifications,
1222 that specification is used.
1223 #. If no match is found, and the type sought is an integer type, then
1224 the smallest integer type that is larger than the bitwidth of the
1225 sought type is used. If none of the specifications are larger than
1226 the bitwidth then the largest integer type is used. For example,
1227 given the default specifications above, the i7 type will use the
1228 alignment of i8 (next largest) while both i65 and i256 will use the
1229 alignment of i64 (largest specified).
1230 #. If no match is found, and the type sought is a vector type, then the
1231 largest vector type that is smaller than the sought vector type will
1232 be used as a fall back. This happens because <128 x double> can be
1233 implemented in terms of 64 <2 x double>, for example.
1235 The function of the data layout string may not be what you expect.
1236 Notably, this is not a specification from the frontend of what alignment
1237 the code generator should use.
1239 Instead, if specified, the target data layout is required to match what
1240 the ultimate *code generator* expects. This string is used by the
1241 mid-level optimizers to improve code, and this only works if it matches
1242 what the ultimate code generator uses. If you would like to generate IR
1243 that does not embed this target-specific detail into the IR, then you
1244 don't have to specify the string. This will disable some optimizations
1245 that require precise layout information, but this also prevents those
1246 optimizations from introducing target specificity into the IR.
1253 A module may specify a target triple string that describes the target
1254 host. The syntax for the target triple is simply:
1256 .. code-block:: llvm
1258 target triple = "x86_64-apple-macosx10.7.0"
1260 The *target triple* string consists of a series of identifiers delimited
1261 by the minus sign character ('-'). The canonical forms are:
1265 ARCHITECTURE-VENDOR-OPERATING_SYSTEM
1266 ARCHITECTURE-VENDOR-OPERATING_SYSTEM-ENVIRONMENT
1268 This information is passed along to the backend so that it generates
1269 code for the proper architecture. It's possible to override this on the
1270 command line with the ``-mtriple`` command line option.
1272 .. _pointeraliasing:
1274 Pointer Aliasing Rules
1275 ----------------------
1277 Any memory access must be done through a pointer value associated with
1278 an address range of the memory access, otherwise the behavior is
1279 undefined. Pointer values are associated with address ranges according
1280 to the following rules:
1282 - A pointer value is associated with the addresses associated with any
1283 value it is *based* on.
1284 - An address of a global variable is associated with the address range
1285 of the variable's storage.
1286 - The result value of an allocation instruction is associated with the
1287 address range of the allocated storage.
1288 - A null pointer in the default address-space is associated with no
1290 - An integer constant other than zero or a pointer value returned from
1291 a function not defined within LLVM may be associated with address
1292 ranges allocated through mechanisms other than those provided by
1293 LLVM. Such ranges shall not overlap with any ranges of addresses
1294 allocated by mechanisms provided by LLVM.
1296 A pointer value is *based* on another pointer value according to the
1299 - A pointer value formed from a ``getelementptr`` operation is *based*
1300 on the first operand of the ``getelementptr``.
1301 - The result value of a ``bitcast`` is *based* on the operand of the
1303 - A pointer value formed by an ``inttoptr`` is *based* on all pointer
1304 values that contribute (directly or indirectly) to the computation of
1305 the pointer's value.
1306 - The "*based* on" relationship is transitive.
1308 Note that this definition of *"based"* is intentionally similar to the
1309 definition of *"based"* in C99, though it is slightly weaker.
1311 LLVM IR does not associate types with memory. The result type of a
1312 ``load`` merely indicates the size and alignment of the memory from
1313 which to load, as well as the interpretation of the value. The first
1314 operand type of a ``store`` similarly only indicates the size and
1315 alignment of the store.
1317 Consequently, type-based alias analysis, aka TBAA, aka
1318 ``-fstrict-aliasing``, is not applicable to general unadorned LLVM IR.
1319 :ref:`Metadata <metadata>` may be used to encode additional information
1320 which specialized optimization passes may use to implement type-based
1325 Volatile Memory Accesses
1326 ------------------------
1328 Certain memory accesses, such as :ref:`load <i_load>`'s,
1329 :ref:`store <i_store>`'s, and :ref:`llvm.memcpy <int_memcpy>`'s may be
1330 marked ``volatile``. The optimizers must not change the number of
1331 volatile operations or change their order of execution relative to other
1332 volatile operations. The optimizers *may* change the order of volatile
1333 operations relative to non-volatile operations. This is not Java's
1334 "volatile" and has no cross-thread synchronization behavior.
1336 IR-level volatile loads and stores cannot safely be optimized into
1337 llvm.memcpy or llvm.memmove intrinsics even when those intrinsics are
1338 flagged volatile. Likewise, the backend should never split or merge
1339 target-legal volatile load/store instructions.
1341 .. admonition:: Rationale
1343 Platforms may rely on volatile loads and stores of natively supported
1344 data width to be executed as single instruction. For example, in C
1345 this holds for an l-value of volatile primitive type with native
1346 hardware support, but not necessarily for aggregate types. The
1347 frontend upholds these expectations, which are intentionally
1348 unspecified in the IR. The rules above ensure that IR transformation
1349 do not violate the frontend's contract with the language.
1353 Memory Model for Concurrent Operations
1354 --------------------------------------
1356 The LLVM IR does not define any way to start parallel threads of
1357 execution or to register signal handlers. Nonetheless, there are
1358 platform-specific ways to create them, and we define LLVM IR's behavior
1359 in their presence. This model is inspired by the C++0x memory model.
1361 For a more informal introduction to this model, see the :doc:`Atomics`.
1363 We define a *happens-before* partial order as the least partial order
1366 - Is a superset of single-thread program order, and
1367 - When a *synchronizes-with* ``b``, includes an edge from ``a`` to
1368 ``b``. *Synchronizes-with* pairs are introduced by platform-specific
1369 techniques, like pthread locks, thread creation, thread joining,
1370 etc., and by atomic instructions. (See also :ref:`Atomic Memory Ordering
1371 Constraints <ordering>`).
1373 Note that program order does not introduce *happens-before* edges
1374 between a thread and signals executing inside that thread.
1376 Every (defined) read operation (load instructions, memcpy, atomic
1377 loads/read-modify-writes, etc.) R reads a series of bytes written by
1378 (defined) write operations (store instructions, atomic
1379 stores/read-modify-writes, memcpy, etc.). For the purposes of this
1380 section, initialized globals are considered to have a write of the
1381 initializer which is atomic and happens before any other read or write
1382 of the memory in question. For each byte of a read R, R\ :sub:`byte`
1383 may see any write to the same byte, except:
1385 - If write\ :sub:`1` happens before write\ :sub:`2`, and
1386 write\ :sub:`2` happens before R\ :sub:`byte`, then
1387 R\ :sub:`byte` does not see write\ :sub:`1`.
1388 - If R\ :sub:`byte` happens before write\ :sub:`3`, then
1389 R\ :sub:`byte` does not see write\ :sub:`3`.
1391 Given that definition, R\ :sub:`byte` is defined as follows:
1393 - If R is volatile, the result is target-dependent. (Volatile is
1394 supposed to give guarantees which can support ``sig_atomic_t`` in
1395 C/C++, and may be used for accesses to addresses which do not behave
1396 like normal memory. It does not generally provide cross-thread
1398 - Otherwise, if there is no write to the same byte that happens before
1399 R\ :sub:`byte`, R\ :sub:`byte` returns ``undef`` for that byte.
1400 - Otherwise, if R\ :sub:`byte` may see exactly one write,
1401 R\ :sub:`byte` returns the value written by that write.
1402 - Otherwise, if R is atomic, and all the writes R\ :sub:`byte` may
1403 see are atomic, it chooses one of the values written. See the :ref:`Atomic
1404 Memory Ordering Constraints <ordering>` section for additional
1405 constraints on how the choice is made.
1406 - Otherwise R\ :sub:`byte` returns ``undef``.
1408 R returns the value composed of the series of bytes it read. This
1409 implies that some bytes within the value may be ``undef`` **without**
1410 the entire value being ``undef``. Note that this only defines the
1411 semantics of the operation; it doesn't mean that targets will emit more
1412 than one instruction to read the series of bytes.
1414 Note that in cases where none of the atomic intrinsics are used, this
1415 model places only one restriction on IR transformations on top of what
1416 is required for single-threaded execution: introducing a store to a byte
1417 which might not otherwise be stored is not allowed in general.
1418 (Specifically, in the case where another thread might write to and read
1419 from an address, introducing a store can change a load that may see
1420 exactly one write into a load that may see multiple writes.)
1424 Atomic Memory Ordering Constraints
1425 ----------------------------------
1427 Atomic instructions (:ref:`cmpxchg <i_cmpxchg>`,
1428 :ref:`atomicrmw <i_atomicrmw>`, :ref:`fence <i_fence>`,
1429 :ref:`atomic load <i_load>`, and :ref:`atomic store <i_store>`) take
1430 an ordering parameter that determines which other atomic instructions on
1431 the same address they *synchronize with*. These semantics are borrowed
1432 from Java and C++0x, but are somewhat more colloquial. If these
1433 descriptions aren't precise enough, check those specs (see spec
1434 references in the :doc:`atomics guide <Atomics>`).
1435 :ref:`fence <i_fence>` instructions treat these orderings somewhat
1436 differently since they don't take an address. See that instruction's
1437 documentation for details.
1439 For a simpler introduction to the ordering constraints, see the
1443 The set of values that can be read is governed by the happens-before
1444 partial order. A value cannot be read unless some operation wrote
1445 it. This is intended to provide a guarantee strong enough to model
1446 Java's non-volatile shared variables. This ordering cannot be
1447 specified for read-modify-write operations; it is not strong enough
1448 to make them atomic in any interesting way.
1450 In addition to the guarantees of ``unordered``, there is a single
1451 total order for modifications by ``monotonic`` operations on each
1452 address. All modification orders must be compatible with the
1453 happens-before order. There is no guarantee that the modification
1454 orders can be combined to a global total order for the whole program
1455 (and this often will not be possible). The read in an atomic
1456 read-modify-write operation (:ref:`cmpxchg <i_cmpxchg>` and
1457 :ref:`atomicrmw <i_atomicrmw>`) reads the value in the modification
1458 order immediately before the value it writes. If one atomic read
1459 happens before another atomic read of the same address, the later
1460 read must see the same value or a later value in the address's
1461 modification order. This disallows reordering of ``monotonic`` (or
1462 stronger) operations on the same address. If an address is written
1463 ``monotonic``-ally by one thread, and other threads ``monotonic``-ally
1464 read that address repeatedly, the other threads must eventually see
1465 the write. This corresponds to the C++0x/C1x
1466 ``memory_order_relaxed``.
1468 In addition to the guarantees of ``monotonic``, a
1469 *synchronizes-with* edge may be formed with a ``release`` operation.
1470 This is intended to model C++'s ``memory_order_acquire``.
1472 In addition to the guarantees of ``monotonic``, if this operation
1473 writes a value which is subsequently read by an ``acquire``
1474 operation, it *synchronizes-with* that operation. (This isn't a
1475 complete description; see the C++0x definition of a release
1476 sequence.) This corresponds to the C++0x/C1x
1477 ``memory_order_release``.
1478 ``acq_rel`` (acquire+release)
1479 Acts as both an ``acquire`` and ``release`` operation on its
1480 address. This corresponds to the C++0x/C1x ``memory_order_acq_rel``.
1481 ``seq_cst`` (sequentially consistent)
1482 In addition to the guarantees of ``acq_rel`` (``acquire`` for an
1483 operation which only reads, ``release`` for an operation which only
1484 writes), there is a global total order on all
1485 sequentially-consistent operations on all addresses, which is
1486 consistent with the *happens-before* partial order and with the
1487 modification orders of all the affected addresses. Each
1488 sequentially-consistent read sees the last preceding write to the
1489 same address in this global order. This corresponds to the C++0x/C1x
1490 ``memory_order_seq_cst`` and Java volatile.
1494 If an atomic operation is marked ``singlethread``, it only *synchronizes
1495 with* or participates in modification and seq\_cst total orderings with
1496 other operations running in the same thread (for example, in signal
1504 LLVM IR floating-point binary ops (:ref:`fadd <i_fadd>`,
1505 :ref:`fsub <i_fsub>`, :ref:`fmul <i_fmul>`, :ref:`fdiv <i_fdiv>`,
1506 :ref:`frem <i_frem>`) have the following flags that can set to enable
1507 otherwise unsafe floating point operations
1510 No NaNs - Allow optimizations to assume the arguments and result are not
1511 NaN. Such optimizations are required to retain defined behavior over
1512 NaNs, but the value of the result is undefined.
1515 No Infs - Allow optimizations to assume the arguments and result are not
1516 +/-Inf. Such optimizations are required to retain defined behavior over
1517 +/-Inf, but the value of the result is undefined.
1520 No Signed Zeros - Allow optimizations to treat the sign of a zero
1521 argument or result as insignificant.
1524 Allow Reciprocal - Allow optimizations to use the reciprocal of an
1525 argument rather than perform division.
1528 Fast - Allow algebraically equivalent transformations that may
1529 dramatically change results in floating point (e.g. reassociate). This
1530 flag implies all the others.
1537 The LLVM type system is one of the most important features of the
1538 intermediate representation. Being typed enables a number of
1539 optimizations to be performed on the intermediate representation
1540 directly, without having to do extra analyses on the side before the
1541 transformation. A strong type system makes it easier to read the
1542 generated code and enables novel analyses and transformations that are
1543 not feasible to perform on normal three address code representations.
1553 The void type does not represent any value and has no size.
1571 The function type can be thought of as a function signature. It consists of a
1572 return type and a list of formal parameter types. The return type of a function
1573 type is a void type or first class type --- except for :ref:`label <t_label>`
1574 and :ref:`metadata <t_metadata>` types.
1580 <returntype> (<parameter list>)
1582 ...where '``<parameter list>``' is a comma-separated list of type
1583 specifiers. Optionally, the parameter list may include a type ``...``, which
1584 indicates that the function takes a variable number of arguments. Variable
1585 argument functions can access their arguments with the :ref:`variable argument
1586 handling intrinsic <int_varargs>` functions. '``<returntype>``' is any type
1587 except :ref:`label <t_label>` and :ref:`metadata <t_metadata>`.
1591 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1592 | ``i32 (i32)`` | function taking an ``i32``, returning an ``i32`` |
1593 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1594 | ``float (i16, i32 *) *`` | :ref:`Pointer <t_pointer>` to a function that takes an ``i16`` and a :ref:`pointer <t_pointer>` to ``i32``, returning ``float``. |
1595 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1596 | ``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. |
1597 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1598 | ``{i32, i32} (i32)`` | A function taking an ``i32``, returning a :ref:`structure <t_struct>` containing two ``i32`` values |
1599 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1606 The :ref:`first class <t_firstclass>` types are perhaps the most important.
1607 Values of these types are the only ones which can be produced by
1615 These are the types that are valid in registers from CodeGen's perspective.
1624 The integer type is a very simple type that simply specifies an
1625 arbitrary bit width for the integer type desired. Any bit width from 1
1626 bit to 2\ :sup:`23`\ -1 (about 8 million) can be specified.
1634 The number of bits the integer will occupy is specified by the ``N``
1640 +----------------+------------------------------------------------+
1641 | ``i1`` | a single-bit integer. |
1642 +----------------+------------------------------------------------+
1643 | ``i32`` | a 32-bit integer. |
1644 +----------------+------------------------------------------------+
1645 | ``i1942652`` | a really big integer of over 1 million bits. |
1646 +----------------+------------------------------------------------+
1650 Floating Point Types
1651 """"""""""""""""""""
1660 - 16-bit floating point value
1663 - 32-bit floating point value
1666 - 64-bit floating point value
1669 - 128-bit floating point value (112-bit mantissa)
1672 - 80-bit floating point value (X87)
1675 - 128-bit floating point value (two 64-bits)
1684 The x86mmx type represents a value held in an MMX register on an x86
1685 machine. The operations allowed on it are quite limited: parameters and
1686 return values, load and store, and bitcast. User-specified MMX
1687 instructions are represented as intrinsic or asm calls with arguments
1688 and/or results of this type. There are no arrays, vectors or constants
1705 The pointer type is used to specify memory locations. Pointers are
1706 commonly used to reference objects in memory.
1708 Pointer types may have an optional address space attribute defining the
1709 numbered address space where the pointed-to object resides. The default
1710 address space is number zero. The semantics of non-zero address spaces
1711 are target-specific.
1713 Note that LLVM does not permit pointers to void (``void*``) nor does it
1714 permit pointers to labels (``label*``). Use ``i8*`` instead.
1724 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1725 | ``[4 x i32]*`` | A :ref:`pointer <t_pointer>` to :ref:`array <t_array>` of four ``i32`` values. |
1726 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1727 | ``i32 (i32*) *`` | A :ref:`pointer <t_pointer>` to a :ref:`function <t_function>` that takes an ``i32*``, returning an ``i32``. |
1728 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1729 | ``i32 addrspace(5)*`` | A :ref:`pointer <t_pointer>` to an ``i32`` value that resides in address space #5. |
1730 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1739 A vector type is a simple derived type that represents a vector of
1740 elements. Vector types are used when multiple primitive data are
1741 operated in parallel using a single instruction (SIMD). A vector type
1742 requires a size (number of elements) and an underlying primitive data
1743 type. Vector types are considered :ref:`first class <t_firstclass>`.
1749 < <# elements> x <elementtype> >
1751 The number of elements is a constant integer value larger than 0;
1752 elementtype may be any integer or floating point type, or a pointer to
1753 these types. Vectors of size zero are not allowed.
1757 +-------------------+--------------------------------------------------+
1758 | ``<4 x i32>`` | Vector of 4 32-bit integer values. |
1759 +-------------------+--------------------------------------------------+
1760 | ``<8 x float>`` | Vector of 8 32-bit floating-point values. |
1761 +-------------------+--------------------------------------------------+
1762 | ``<2 x i64>`` | Vector of 2 64-bit integer values. |
1763 +-------------------+--------------------------------------------------+
1764 | ``<4 x i64*>`` | Vector of 4 pointers to 64-bit integer values. |
1765 +-------------------+--------------------------------------------------+
1774 The label type represents code labels.
1789 The metadata type represents embedded metadata. No derived types may be
1790 created from metadata except for :ref:`function <t_function>` arguments.
1803 Aggregate Types are a subset of derived types that can contain multiple
1804 member types. :ref:`Arrays <t_array>` and :ref:`structs <t_struct>` are
1805 aggregate types. :ref:`Vectors <t_vector>` are not considered to be
1815 The array type is a very simple derived type that arranges elements
1816 sequentially in memory. The array type requires a size (number of
1817 elements) and an underlying data type.
1823 [<# elements> x <elementtype>]
1825 The number of elements is a constant integer value; ``elementtype`` may
1826 be any type with a size.
1830 +------------------+--------------------------------------+
1831 | ``[40 x i32]`` | Array of 40 32-bit integer values. |
1832 +------------------+--------------------------------------+
1833 | ``[41 x i32]`` | Array of 41 32-bit integer values. |
1834 +------------------+--------------------------------------+
1835 | ``[4 x i8]`` | Array of 4 8-bit integer values. |
1836 +------------------+--------------------------------------+
1838 Here are some examples of multidimensional arrays:
1840 +-----------------------------+----------------------------------------------------------+
1841 | ``[3 x [4 x i32]]`` | 3x4 array of 32-bit integer values. |
1842 +-----------------------------+----------------------------------------------------------+
1843 | ``[12 x [10 x float]]`` | 12x10 array of single precision floating point values. |
1844 +-----------------------------+----------------------------------------------------------+
1845 | ``[2 x [3 x [4 x i16]]]`` | 2x3x4 array of 16-bit integer values. |
1846 +-----------------------------+----------------------------------------------------------+
1848 There is no restriction on indexing beyond the end of the array implied
1849 by a static type (though there are restrictions on indexing beyond the
1850 bounds of an allocated object in some cases). This means that
1851 single-dimension 'variable sized array' addressing can be implemented in
1852 LLVM with a zero length array type. An implementation of 'pascal style
1853 arrays' in LLVM could use the type "``{ i32, [0 x float]}``", for
1863 The structure type is used to represent a collection of data members
1864 together in memory. The elements of a structure may be any type that has
1867 Structures in memory are accessed using '``load``' and '``store``' by
1868 getting a pointer to a field with the '``getelementptr``' instruction.
1869 Structures in registers are accessed using the '``extractvalue``' and
1870 '``insertvalue``' instructions.
1872 Structures may optionally be "packed" structures, which indicate that
1873 the alignment of the struct is one byte, and that there is no padding
1874 between the elements. In non-packed structs, padding between field types
1875 is inserted as defined by the DataLayout string in the module, which is
1876 required to match what the underlying code generator expects.
1878 Structures can either be "literal" or "identified". A literal structure
1879 is defined inline with other types (e.g. ``{i32, i32}*``) whereas
1880 identified types are always defined at the top level with a name.
1881 Literal types are uniqued by their contents and can never be recursive
1882 or opaque since there is no way to write one. Identified types can be
1883 recursive, can be opaqued, and are never uniqued.
1889 %T1 = type { <type list> } ; Identified normal struct type
1890 %T2 = type <{ <type list> }> ; Identified packed struct type
1894 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1895 | ``{ i32, i32, i32 }`` | A triple of three ``i32`` values |
1896 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1897 | ``{ 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``. |
1898 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1899 | ``<{ i8, i32 }>`` | A packed struct known to be 5 bytes in size. |
1900 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1904 Opaque Structure Types
1905 """"""""""""""""""""""
1909 Opaque structure types are used to represent named structure types that
1910 do not have a body specified. This corresponds (for example) to the C
1911 notion of a forward declared structure.
1922 +--------------+-------------------+
1923 | ``opaque`` | An opaque type. |
1924 +--------------+-------------------+
1929 LLVM has several different basic types of constants. This section
1930 describes them all and their syntax.
1935 **Boolean constants**
1936 The two strings '``true``' and '``false``' are both valid constants
1938 **Integer constants**
1939 Standard integers (such as '4') are constants of the
1940 :ref:`integer <t_integer>` type. Negative numbers may be used with
1942 **Floating point constants**
1943 Floating point constants use standard decimal notation (e.g.
1944 123.421), exponential notation (e.g. 1.23421e+2), or a more precise
1945 hexadecimal notation (see below). The assembler requires the exact
1946 decimal value of a floating-point constant. For example, the
1947 assembler accepts 1.25 but rejects 1.3 because 1.3 is a repeating
1948 decimal in binary. Floating point constants must have a :ref:`floating
1949 point <t_floating>` type.
1950 **Null pointer constants**
1951 The identifier '``null``' is recognized as a null pointer constant
1952 and must be of :ref:`pointer type <t_pointer>`.
1954 The one non-intuitive notation for constants is the hexadecimal form of
1955 floating point constants. For example, the form
1956 '``double 0x432ff973cafa8000``' is equivalent to (but harder to read
1957 than) '``double 4.5e+15``'. The only time hexadecimal floating point
1958 constants are required (and the only time that they are generated by the
1959 disassembler) is when a floating point constant must be emitted but it
1960 cannot be represented as a decimal floating point number in a reasonable
1961 number of digits. For example, NaN's, infinities, and other special
1962 values are represented in their IEEE hexadecimal format so that assembly
1963 and disassembly do not cause any bits to change in the constants.
1965 When using the hexadecimal form, constants of types half, float, and
1966 double are represented using the 16-digit form shown above (which
1967 matches the IEEE754 representation for double); half and float values
1968 must, however, be exactly representable as IEEE 754 half and single
1969 precision, respectively. Hexadecimal format is always used for long
1970 double, and there are three forms of long double. The 80-bit format used
1971 by x86 is represented as ``0xK`` followed by 20 hexadecimal digits. The
1972 128-bit format used by PowerPC (two adjacent doubles) is represented by
1973 ``0xM`` followed by 32 hexadecimal digits. The IEEE 128-bit format is
1974 represented by ``0xL`` followed by 32 hexadecimal digits. Long doubles
1975 will only work if they match the long double format on your target.
1976 The IEEE 16-bit format (half precision) is represented by ``0xH``
1977 followed by 4 hexadecimal digits. All hexadecimal formats are big-endian
1978 (sign bit at the left).
1980 There are no constants of type x86mmx.
1982 .. _complexconstants:
1987 Complex constants are a (potentially recursive) combination of simple
1988 constants and smaller complex constants.
1990 **Structure constants**
1991 Structure constants are represented with notation similar to
1992 structure type definitions (a comma separated list of elements,
1993 surrounded by braces (``{}``)). For example:
1994 "``{ i32 4, float 17.0, i32* @G }``", where "``@G``" is declared as
1995 "``@G = external global i32``". Structure constants must have
1996 :ref:`structure type <t_struct>`, and the number and types of elements
1997 must match those specified by the type.
1999 Array constants are represented with notation similar to array type
2000 definitions (a comma separated list of elements, surrounded by
2001 square brackets (``[]``)). For example:
2002 "``[ i32 42, i32 11, i32 74 ]``". Array constants must have
2003 :ref:`array type <t_array>`, and the number and types of elements must
2004 match those specified by the type.
2005 **Vector constants**
2006 Vector constants are represented with notation similar to vector
2007 type definitions (a comma separated list of elements, surrounded by
2008 less-than/greater-than's (``<>``)). For example:
2009 "``< i32 42, i32 11, i32 74, i32 100 >``". Vector constants
2010 must have :ref:`vector type <t_vector>`, and the number and types of
2011 elements must match those specified by the type.
2012 **Zero initialization**
2013 The string '``zeroinitializer``' can be used to zero initialize a
2014 value to zero of *any* type, including scalar and
2015 :ref:`aggregate <t_aggregate>` types. This is often used to avoid
2016 having to print large zero initializers (e.g. for large arrays) and
2017 is always exactly equivalent to using explicit zero initializers.
2019 A metadata node is a structure-like constant with :ref:`metadata
2020 type <t_metadata>`. For example:
2021 "``metadata !{ i32 0, metadata !"test" }``". Unlike other
2022 constants that are meant to be interpreted as part of the
2023 instruction stream, metadata is a place to attach additional
2024 information such as debug info.
2026 Global Variable and Function Addresses
2027 --------------------------------------
2029 The addresses of :ref:`global variables <globalvars>` and
2030 :ref:`functions <functionstructure>` are always implicitly valid
2031 (link-time) constants. These constants are explicitly referenced when
2032 the :ref:`identifier for the global <identifiers>` is used and always have
2033 :ref:`pointer <t_pointer>` type. For example, the following is a legal LLVM
2036 .. code-block:: llvm
2040 @Z = global [2 x i32*] [ i32* @X, i32* @Y ]
2047 The string '``undef``' can be used anywhere a constant is expected, and
2048 indicates that the user of the value may receive an unspecified
2049 bit-pattern. Undefined values may be of any type (other than '``label``'
2050 or '``void``') and be used anywhere a constant is permitted.
2052 Undefined values are useful because they indicate to the compiler that
2053 the program is well defined no matter what value is used. This gives the
2054 compiler more freedom to optimize. Here are some examples of
2055 (potentially surprising) transformations that are valid (in pseudo IR):
2057 .. code-block:: llvm
2067 This is safe because all of the output bits are affected by the undef
2068 bits. Any output bit can have a zero or one depending on the input bits.
2070 .. code-block:: llvm
2081 These logical operations have bits that are not always affected by the
2082 input. For example, if ``%X`` has a zero bit, then the output of the
2083 '``and``' operation will always be a zero for that bit, no matter what
2084 the corresponding bit from the '``undef``' is. As such, it is unsafe to
2085 optimize or assume that the result of the '``and``' is '``undef``'.
2086 However, it is safe to assume that all bits of the '``undef``' could be
2087 0, and optimize the '``and``' to 0. Likewise, it is safe to assume that
2088 all the bits of the '``undef``' operand to the '``or``' could be set,
2089 allowing the '``or``' to be folded to -1.
2091 .. code-block:: llvm
2093 %A = select undef, %X, %Y
2094 %B = select undef, 42, %Y
2095 %C = select %X, %Y, undef
2105 This set of examples shows that undefined '``select``' (and conditional
2106 branch) conditions can go *either way*, but they have to come from one
2107 of the two operands. In the ``%A`` example, if ``%X`` and ``%Y`` were
2108 both known to have a clear low bit, then ``%A`` would have to have a
2109 cleared low bit. However, in the ``%C`` example, the optimizer is
2110 allowed to assume that the '``undef``' operand could be the same as
2111 ``%Y``, allowing the whole '``select``' to be eliminated.
2113 .. code-block:: llvm
2115 %A = xor undef, undef
2132 This example points out that two '``undef``' operands are not
2133 necessarily the same. This can be surprising to people (and also matches
2134 C semantics) where they assume that "``X^X``" is always zero, even if
2135 ``X`` is undefined. This isn't true for a number of reasons, but the
2136 short answer is that an '``undef``' "variable" can arbitrarily change
2137 its value over its "live range". This is true because the variable
2138 doesn't actually *have a live range*. Instead, the value is logically
2139 read from arbitrary registers that happen to be around when needed, so
2140 the value is not necessarily consistent over time. In fact, ``%A`` and
2141 ``%C`` need to have the same semantics or the core LLVM "replace all
2142 uses with" concept would not hold.
2144 .. code-block:: llvm
2152 These examples show the crucial difference between an *undefined value*
2153 and *undefined behavior*. An undefined value (like '``undef``') is
2154 allowed to have an arbitrary bit-pattern. This means that the ``%A``
2155 operation can be constant folded to '``undef``', because the '``undef``'
2156 could be an SNaN, and ``fdiv`` is not (currently) defined on SNaN's.
2157 However, in the second example, we can make a more aggressive
2158 assumption: because the ``undef`` is allowed to be an arbitrary value,
2159 we are allowed to assume that it could be zero. Since a divide by zero
2160 has *undefined behavior*, we are allowed to assume that the operation
2161 does not execute at all. This allows us to delete the divide and all
2162 code after it. Because the undefined operation "can't happen", the
2163 optimizer can assume that it occurs in dead code.
2165 .. code-block:: llvm
2167 a: store undef -> %X
2168 b: store %X -> undef
2173 These examples reiterate the ``fdiv`` example: a store *of* an undefined
2174 value can be assumed to not have any effect; we can assume that the
2175 value is overwritten with bits that happen to match what was already
2176 there. However, a store *to* an undefined location could clobber
2177 arbitrary memory, therefore, it has undefined behavior.
2184 Poison values are similar to :ref:`undef values <undefvalues>`, however
2185 they also represent the fact that an instruction or constant expression
2186 which cannot evoke side effects has nevertheless detected a condition
2187 which results in undefined behavior.
2189 There is currently no way of representing a poison value in the IR; they
2190 only exist when produced by operations such as :ref:`add <i_add>` with
2193 Poison value behavior is defined in terms of value *dependence*:
2195 - Values other than :ref:`phi <i_phi>` nodes depend on their operands.
2196 - :ref:`Phi <i_phi>` nodes depend on the operand corresponding to
2197 their dynamic predecessor basic block.
2198 - Function arguments depend on the corresponding actual argument values
2199 in the dynamic callers of their functions.
2200 - :ref:`Call <i_call>` instructions depend on the :ref:`ret <i_ret>`
2201 instructions that dynamically transfer control back to them.
2202 - :ref:`Invoke <i_invoke>` instructions depend on the
2203 :ref:`ret <i_ret>`, :ref:`resume <i_resume>`, or exception-throwing
2204 call instructions that dynamically transfer control back to them.
2205 - Non-volatile loads and stores depend on the most recent stores to all
2206 of the referenced memory addresses, following the order in the IR
2207 (including loads and stores implied by intrinsics such as
2208 :ref:`@llvm.memcpy <int_memcpy>`.)
2209 - An instruction with externally visible side effects depends on the
2210 most recent preceding instruction with externally visible side
2211 effects, following the order in the IR. (This includes :ref:`volatile
2212 operations <volatile>`.)
2213 - An instruction *control-depends* on a :ref:`terminator
2214 instruction <terminators>` if the terminator instruction has
2215 multiple successors and the instruction is always executed when
2216 control transfers to one of the successors, and may not be executed
2217 when control is transferred to another.
2218 - Additionally, an instruction also *control-depends* on a terminator
2219 instruction if the set of instructions it otherwise depends on would
2220 be different if the terminator had transferred control to a different
2222 - Dependence is transitive.
2224 Poison Values have the same behavior as :ref:`undef values <undefvalues>`,
2225 with the additional affect that any instruction which has a *dependence*
2226 on a poison value has undefined behavior.
2228 Here are some examples:
2230 .. code-block:: llvm
2233 %poison = sub nuw i32 0, 1 ; Results in a poison value.
2234 %still_poison = and i32 %poison, 0 ; 0, but also poison.
2235 %poison_yet_again = getelementptr i32* @h, i32 %still_poison
2236 store i32 0, i32* %poison_yet_again ; memory at @h[0] is poisoned
2238 store i32 %poison, i32* @g ; Poison value stored to memory.
2239 %poison2 = load i32* @g ; Poison value loaded back from memory.
2241 store volatile i32 %poison, i32* @g ; External observation; undefined behavior.
2243 %narrowaddr = bitcast i32* @g to i16*
2244 %wideaddr = bitcast i32* @g to i64*
2245 %poison3 = load i16* %narrowaddr ; Returns a poison value.
2246 %poison4 = load i64* %wideaddr ; Returns a poison value.
2248 %cmp = icmp slt i32 %poison, 0 ; Returns a poison value.
2249 br i1 %cmp, label %true, label %end ; Branch to either destination.
2252 store volatile i32 0, i32* @g ; This is control-dependent on %cmp, so
2253 ; it has undefined behavior.
2257 %p = phi i32 [ 0, %entry ], [ 1, %true ]
2258 ; Both edges into this PHI are
2259 ; control-dependent on %cmp, so this
2260 ; always results in a poison value.
2262 store volatile i32 0, i32* @g ; This would depend on the store in %true
2263 ; if %cmp is true, or the store in %entry
2264 ; otherwise, so this is undefined behavior.
2266 br i1 %cmp, label %second_true, label %second_end
2267 ; The same branch again, but this time the
2268 ; true block doesn't have side effects.
2275 store volatile i32 0, i32* @g ; This time, the instruction always depends
2276 ; on the store in %end. Also, it is
2277 ; control-equivalent to %end, so this is
2278 ; well-defined (ignoring earlier undefined
2279 ; behavior in this example).
2283 Addresses of Basic Blocks
2284 -------------------------
2286 ``blockaddress(@function, %block)``
2288 The '``blockaddress``' constant computes the address of the specified
2289 basic block in the specified function, and always has an ``i8*`` type.
2290 Taking the address of the entry block is illegal.
2292 This value only has defined behavior when used as an operand to the
2293 ':ref:`indirectbr <i_indirectbr>`' instruction, or for comparisons
2294 against null. Pointer equality tests between labels addresses results in
2295 undefined behavior --- though, again, comparison against null is ok, and
2296 no label is equal to the null pointer. This may be passed around as an
2297 opaque pointer sized value as long as the bits are not inspected. This
2298 allows ``ptrtoint`` and arithmetic to be performed on these values so
2299 long as the original value is reconstituted before the ``indirectbr``
2302 Finally, some targets may provide defined semantics when using the value
2303 as the operand to an inline assembly, but that is target specific.
2307 Constant Expressions
2308 --------------------
2310 Constant expressions are used to allow expressions involving other
2311 constants to be used as constants. Constant expressions may be of any
2312 :ref:`first class <t_firstclass>` type and may involve any LLVM operation
2313 that does not have side effects (e.g. load and call are not supported).
2314 The following is the syntax for constant expressions:
2316 ``trunc (CST to TYPE)``
2317 Truncate a constant to another type. The bit size of CST must be
2318 larger than the bit size of TYPE. Both types must be integers.
2319 ``zext (CST to TYPE)``
2320 Zero extend a constant to another type. The bit size of CST must be
2321 smaller than the bit size of TYPE. Both types must be integers.
2322 ``sext (CST to TYPE)``
2323 Sign extend a constant to another type. The bit size of CST must be
2324 smaller than the bit size of TYPE. Both types must be integers.
2325 ``fptrunc (CST to TYPE)``
2326 Truncate a floating point constant to another floating point type.
2327 The size of CST must be larger than the size of TYPE. Both types
2328 must be floating point.
2329 ``fpext (CST to TYPE)``
2330 Floating point extend a constant to another type. The size of CST
2331 must be smaller or equal to the size of TYPE. Both types must be
2333 ``fptoui (CST to TYPE)``
2334 Convert a floating point constant to the corresponding unsigned
2335 integer constant. TYPE must be a scalar or vector integer type. CST
2336 must be of scalar or vector floating point type. Both CST and TYPE
2337 must be scalars, or vectors of the same number of elements. If the
2338 value won't fit in the integer type, the results are undefined.
2339 ``fptosi (CST to TYPE)``
2340 Convert a floating point constant to the corresponding signed
2341 integer constant. TYPE must be a scalar or vector integer type. CST
2342 must be of scalar or vector floating point type. Both CST and TYPE
2343 must be scalars, or vectors of the same number of elements. If the
2344 value won't fit in the integer type, the results are undefined.
2345 ``uitofp (CST to TYPE)``
2346 Convert an unsigned integer constant to the corresponding floating
2347 point constant. TYPE must be a scalar or vector floating point type.
2348 CST must be of scalar or vector integer type. Both CST and TYPE must
2349 be scalars, or vectors of the same number of elements. If the value
2350 won't fit in the floating point type, the results are undefined.
2351 ``sitofp (CST to TYPE)``
2352 Convert a signed integer constant to the corresponding floating
2353 point constant. TYPE must be a scalar or vector floating point type.
2354 CST must be of scalar or vector integer type. Both CST and TYPE must
2355 be scalars, or vectors of the same number of elements. If the value
2356 won't fit in the floating point type, the results are undefined.
2357 ``ptrtoint (CST to TYPE)``
2358 Convert a pointer typed constant to the corresponding integer
2359 constant. ``TYPE`` must be an integer type. ``CST`` must be of
2360 pointer type. The ``CST`` value is zero extended, truncated, or
2361 unchanged to make it fit in ``TYPE``.
2362 ``inttoptr (CST to TYPE)``
2363 Convert an integer constant to a pointer constant. TYPE must be a
2364 pointer type. CST must be of integer type. The CST value is zero
2365 extended, truncated, or unchanged to make it fit in a pointer size.
2366 This one is *really* dangerous!
2367 ``bitcast (CST to TYPE)``
2368 Convert a constant, CST, to another TYPE. The constraints of the
2369 operands are the same as those for the :ref:`bitcast
2370 instruction <i_bitcast>`.
2371 ``addrspacecast (CST to TYPE)``
2372 Convert a constant pointer or constant vector of pointer, CST, to another
2373 TYPE in a different address space. The constraints of the operands are the
2374 same as those for the :ref:`addrspacecast instruction <i_addrspacecast>`.
2375 ``getelementptr (CSTPTR, IDX0, IDX1, ...)``, ``getelementptr inbounds (CSTPTR, IDX0, IDX1, ...)``
2376 Perform the :ref:`getelementptr operation <i_getelementptr>` on
2377 constants. As with the :ref:`getelementptr <i_getelementptr>`
2378 instruction, the index list may have zero or more indexes, which are
2379 required to make sense for the type of "CSTPTR".
2380 ``select (COND, VAL1, VAL2)``
2381 Perform the :ref:`select operation <i_select>` on constants.
2382 ``icmp COND (VAL1, VAL2)``
2383 Performs the :ref:`icmp operation <i_icmp>` on constants.
2384 ``fcmp COND (VAL1, VAL2)``
2385 Performs the :ref:`fcmp operation <i_fcmp>` on constants.
2386 ``extractelement (VAL, IDX)``
2387 Perform the :ref:`extractelement operation <i_extractelement>` on
2389 ``insertelement (VAL, ELT, IDX)``
2390 Perform the :ref:`insertelement operation <i_insertelement>` on
2392 ``shufflevector (VEC1, VEC2, IDXMASK)``
2393 Perform the :ref:`shufflevector operation <i_shufflevector>` on
2395 ``extractvalue (VAL, IDX0, IDX1, ...)``
2396 Perform the :ref:`extractvalue operation <i_extractvalue>` on
2397 constants. The index list is interpreted in a similar manner as
2398 indices in a ':ref:`getelementptr <i_getelementptr>`' operation. At
2399 least one index value must be specified.
2400 ``insertvalue (VAL, ELT, IDX0, IDX1, ...)``
2401 Perform the :ref:`insertvalue operation <i_insertvalue>` on constants.
2402 The index list is interpreted in a similar manner as indices in a
2403 ':ref:`getelementptr <i_getelementptr>`' operation. At least one index
2404 value must be specified.
2405 ``OPCODE (LHS, RHS)``
2406 Perform the specified operation of the LHS and RHS constants. OPCODE
2407 may be any of the :ref:`binary <binaryops>` or :ref:`bitwise
2408 binary <bitwiseops>` operations. The constraints on operands are
2409 the same as those for the corresponding instruction (e.g. no bitwise
2410 operations on floating point values are allowed).
2417 Inline Assembler Expressions
2418 ----------------------------
2420 LLVM supports inline assembler expressions (as opposed to :ref:`Module-Level
2421 Inline Assembly <moduleasm>`) through the use of a special value. This
2422 value represents the inline assembler as a string (containing the
2423 instructions to emit), a list of operand constraints (stored as a
2424 string), a flag that indicates whether or not the inline asm expression
2425 has side effects, and a flag indicating whether the function containing
2426 the asm needs to align its stack conservatively. An example inline
2427 assembler expression is:
2429 .. code-block:: llvm
2431 i32 (i32) asm "bswap $0", "=r,r"
2433 Inline assembler expressions may **only** be used as the callee operand
2434 of a :ref:`call <i_call>` or an :ref:`invoke <i_invoke>` instruction.
2435 Thus, typically we have:
2437 .. code-block:: llvm
2439 %X = call i32 asm "bswap $0", "=r,r"(i32 %Y)
2441 Inline asms with side effects not visible in the constraint list must be
2442 marked as having side effects. This is done through the use of the
2443 '``sideeffect``' keyword, like so:
2445 .. code-block:: llvm
2447 call void asm sideeffect "eieio", ""()
2449 In some cases inline asms will contain code that will not work unless
2450 the stack is aligned in some way, such as calls or SSE instructions on
2451 x86, yet will not contain code that does that alignment within the asm.
2452 The compiler should make conservative assumptions about what the asm
2453 might contain and should generate its usual stack alignment code in the
2454 prologue if the '``alignstack``' keyword is present:
2456 .. code-block:: llvm
2458 call void asm alignstack "eieio", ""()
2460 Inline asms also support using non-standard assembly dialects. The
2461 assumed dialect is ATT. When the '``inteldialect``' keyword is present,
2462 the inline asm is using the Intel dialect. Currently, ATT and Intel are
2463 the only supported dialects. An example is:
2465 .. code-block:: llvm
2467 call void asm inteldialect "eieio", ""()
2469 If multiple keywords appear the '``sideeffect``' keyword must come
2470 first, the '``alignstack``' keyword second and the '``inteldialect``'
2476 The call instructions that wrap inline asm nodes may have a
2477 "``!srcloc``" MDNode attached to it that contains a list of constant
2478 integers. If present, the code generator will use the integer as the
2479 location cookie value when report errors through the ``LLVMContext``
2480 error reporting mechanisms. This allows a front-end to correlate backend
2481 errors that occur with inline asm back to the source code that produced
2484 .. code-block:: llvm
2486 call void asm sideeffect "something bad", ""(), !srcloc !42
2488 !42 = !{ i32 1234567 }
2490 It is up to the front-end to make sense of the magic numbers it places
2491 in the IR. If the MDNode contains multiple constants, the code generator
2492 will use the one that corresponds to the line of the asm that the error
2497 Metadata Nodes and Metadata Strings
2498 -----------------------------------
2500 LLVM IR allows metadata to be attached to instructions in the program
2501 that can convey extra information about the code to the optimizers and
2502 code generator. One example application of metadata is source-level
2503 debug information. There are two metadata primitives: strings and nodes.
2504 All metadata has the ``metadata`` type and is identified in syntax by a
2505 preceding exclamation point ('``!``').
2507 A metadata string is a string surrounded by double quotes. It can
2508 contain any character by escaping non-printable characters with
2509 "``\xx``" where "``xx``" is the two digit hex code. For example:
2512 Metadata nodes are represented with notation similar to structure
2513 constants (a comma separated list of elements, surrounded by braces and
2514 preceded by an exclamation point). Metadata nodes can have any values as
2515 their operand. For example:
2517 .. code-block:: llvm
2519 !{ metadata !"test\00", i32 10}
2521 A :ref:`named metadata <namedmetadatastructure>` is a collection of
2522 metadata nodes, which can be looked up in the module symbol table. For
2525 .. code-block:: llvm
2527 !foo = metadata !{!4, !3}
2529 Metadata can be used as function arguments. Here ``llvm.dbg.value``
2530 function is using two metadata arguments:
2532 .. code-block:: llvm
2534 call void @llvm.dbg.value(metadata !24, i64 0, metadata !25)
2536 Metadata can be attached with an instruction. Here metadata ``!21`` is
2537 attached to the ``add`` instruction using the ``!dbg`` identifier:
2539 .. code-block:: llvm
2541 %indvar.next = add i64 %indvar, 1, !dbg !21
2543 More information about specific metadata nodes recognized by the
2544 optimizers and code generator is found below.
2549 In LLVM IR, memory does not have types, so LLVM's own type system is not
2550 suitable for doing TBAA. Instead, metadata is added to the IR to
2551 describe a type system of a higher level language. This can be used to
2552 implement typical C/C++ TBAA, but it can also be used to implement
2553 custom alias analysis behavior for other languages.
2555 The current metadata format is very simple. TBAA metadata nodes have up
2556 to three fields, e.g.:
2558 .. code-block:: llvm
2560 !0 = metadata !{ metadata !"an example type tree" }
2561 !1 = metadata !{ metadata !"int", metadata !0 }
2562 !2 = metadata !{ metadata !"float", metadata !0 }
2563 !3 = metadata !{ metadata !"const float", metadata !2, i64 1 }
2565 The first field is an identity field. It can be any value, usually a
2566 metadata string, which uniquely identifies the type. The most important
2567 name in the tree is the name of the root node. Two trees with different
2568 root node names are entirely disjoint, even if they have leaves with
2571 The second field identifies the type's parent node in the tree, or is
2572 null or omitted for a root node. A type is considered to alias all of
2573 its descendants and all of its ancestors in the tree. Also, a type is
2574 considered to alias all types in other trees, so that bitcode produced
2575 from multiple front-ends is handled conservatively.
2577 If the third field is present, it's an integer which if equal to 1
2578 indicates that the type is "constant" (meaning
2579 ``pointsToConstantMemory`` should return true; see `other useful
2580 AliasAnalysis methods <AliasAnalysis.html#OtherItfs>`_).
2582 '``tbaa.struct``' Metadata
2583 ^^^^^^^^^^^^^^^^^^^^^^^^^^
2585 The :ref:`llvm.memcpy <int_memcpy>` is often used to implement
2586 aggregate assignment operations in C and similar languages, however it
2587 is defined to copy a contiguous region of memory, which is more than
2588 strictly necessary for aggregate types which contain holes due to
2589 padding. Also, it doesn't contain any TBAA information about the fields
2592 ``!tbaa.struct`` metadata can describe which memory subregions in a
2593 memcpy are padding and what the TBAA tags of the struct are.
2595 The current metadata format is very simple. ``!tbaa.struct`` metadata
2596 nodes are a list of operands which are in conceptual groups of three.
2597 For each group of three, the first operand gives the byte offset of a
2598 field in bytes, the second gives its size in bytes, and the third gives
2601 .. code-block:: llvm
2603 !4 = metadata !{ i64 0, i64 4, metadata !1, i64 8, i64 4, metadata !2 }
2605 This describes a struct with two fields. The first is at offset 0 bytes
2606 with size 4 bytes, and has tbaa tag !1. The second is at offset 8 bytes
2607 and has size 4 bytes and has tbaa tag !2.
2609 Note that the fields need not be contiguous. In this example, there is a
2610 4 byte gap between the two fields. This gap represents padding which
2611 does not carry useful data and need not be preserved.
2613 '``fpmath``' Metadata
2614 ^^^^^^^^^^^^^^^^^^^^^
2616 ``fpmath`` metadata may be attached to any instruction of floating point
2617 type. It can be used to express the maximum acceptable error in the
2618 result of that instruction, in ULPs, thus potentially allowing the
2619 compiler to use a more efficient but less accurate method of computing
2620 it. ULP is defined as follows:
2622 If ``x`` is a real number that lies between two finite consecutive
2623 floating-point numbers ``a`` and ``b``, without being equal to one
2624 of them, then ``ulp(x) = |b - a|``, otherwise ``ulp(x)`` is the
2625 distance between the two non-equal finite floating-point numbers
2626 nearest ``x``. Moreover, ``ulp(NaN)`` is ``NaN``.
2628 The metadata node shall consist of a single positive floating point
2629 number representing the maximum relative error, for example:
2631 .. code-block:: llvm
2633 !0 = metadata !{ float 2.5 } ; maximum acceptable inaccuracy is 2.5 ULPs
2635 '``range``' Metadata
2636 ^^^^^^^^^^^^^^^^^^^^
2638 ``range`` metadata may be attached only to loads of integer types. It
2639 expresses the possible ranges the loaded value is in. The ranges are
2640 represented with a flattened list of integers. The loaded value is known
2641 to be in the union of the ranges defined by each consecutive pair. Each
2642 pair has the following properties:
2644 - The type must match the type loaded by the instruction.
2645 - The pair ``a,b`` represents the range ``[a,b)``.
2646 - Both ``a`` and ``b`` are constants.
2647 - The range is allowed to wrap.
2648 - The range should not represent the full or empty set. That is,
2651 In addition, the pairs must be in signed order of the lower bound and
2652 they must be non-contiguous.
2656 .. code-block:: llvm
2658 %a = load i8* %x, align 1, !range !0 ; Can only be 0 or 1
2659 %b = load i8* %y, align 1, !range !1 ; Can only be 255 (-1), 0 or 1
2660 %c = load i8* %z, align 1, !range !2 ; Can only be 0, 1, 3, 4 or 5
2661 %d = load i8* %z, align 1, !range !3 ; Can only be -2, -1, 3, 4 or 5
2663 !0 = metadata !{ i8 0, i8 2 }
2664 !1 = metadata !{ i8 255, i8 2 }
2665 !2 = metadata !{ i8 0, i8 2, i8 3, i8 6 }
2666 !3 = metadata !{ i8 -2, i8 0, i8 3, i8 6 }
2671 It is sometimes useful to attach information to loop constructs. Currently,
2672 loop metadata is implemented as metadata attached to the branch instruction
2673 in the loop latch block. This type of metadata refer to a metadata node that is
2674 guaranteed to be separate for each loop. The loop identifier metadata is
2675 specified with the name ``llvm.loop``.
2677 The loop identifier metadata is implemented using a metadata that refers to
2678 itself to avoid merging it with any other identifier metadata, e.g.,
2679 during module linkage or function inlining. That is, each loop should refer
2680 to their own identification metadata even if they reside in separate functions.
2681 The following example contains loop identifier metadata for two separate loop
2684 .. code-block:: llvm
2686 !0 = metadata !{ metadata !0 }
2687 !1 = metadata !{ metadata !1 }
2689 The loop identifier metadata can be used to specify additional per-loop
2690 metadata. Any operands after the first operand can be treated as user-defined
2691 metadata. For example the ``llvm.vectorizer.unroll`` metadata is understood
2692 by the loop vectorizer to indicate how many times to unroll the loop:
2694 .. code-block:: llvm
2696 br i1 %exitcond, label %._crit_edge, label %.lr.ph, !llvm.loop !0
2698 !0 = metadata !{ metadata !0, metadata !1 }
2699 !1 = metadata !{ metadata !"llvm.vectorizer.unroll", i32 2 }
2704 Metadata types used to annotate memory accesses with information helpful
2705 for optimizations are prefixed with ``llvm.mem``.
2707 '``llvm.mem.parallel_loop_access``' Metadata
2708 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2710 For a loop to be parallel, in addition to using
2711 the ``llvm.loop`` metadata to mark the loop latch branch instruction,
2712 also all of the memory accessing instructions in the loop body need to be
2713 marked with the ``llvm.mem.parallel_loop_access`` metadata. If there
2714 is at least one memory accessing instruction not marked with the metadata,
2715 the loop must be considered a sequential loop. This causes parallel loops to be
2716 converted to sequential loops due to optimization passes that are unaware of
2717 the parallel semantics and that insert new memory instructions to the loop
2720 Example of a loop that is considered parallel due to its correct use of
2721 both ``llvm.loop`` and ``llvm.mem.parallel_loop_access``
2722 metadata types that refer to the same loop identifier metadata.
2724 .. code-block:: llvm
2728 %0 = load i32* %arrayidx, align 4, !llvm.mem.parallel_loop_access !0
2730 store i32 %0, i32* %arrayidx4, align 4, !llvm.mem.parallel_loop_access !0
2732 br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !0
2736 !0 = metadata !{ metadata !0 }
2738 It is also possible to have nested parallel loops. In that case the
2739 memory accesses refer to a list of loop identifier metadata nodes instead of
2740 the loop identifier metadata node directly:
2742 .. code-block:: llvm
2749 %0 = load i32* %arrayidx, align 4, !llvm.mem.parallel_loop_access !0
2751 store i32 %0, i32* %arrayidx4, align 4, !llvm.mem.parallel_loop_access !0
2753 br i1 %exitcond, label %inner.for.end, label %inner.for.body, !llvm.loop !1
2757 %0 = load i32* %arrayidx, align 4, !llvm.mem.parallel_loop_access !0
2759 store i32 %0, i32* %arrayidx4, align 4, !llvm.mem.parallel_loop_access !0
2761 br i1 %exitcond, label %outer.for.end, label %outer.for.body, !llvm.loop !2
2763 outer.for.end: ; preds = %for.body
2765 !0 = metadata !{ metadata !1, metadata !2 } ; a list of loop identifiers
2766 !1 = metadata !{ metadata !1 } ; an identifier for the inner loop
2767 !2 = metadata !{ metadata !2 } ; an identifier for the outer loop
2769 '``llvm.vectorizer``'
2770 ^^^^^^^^^^^^^^^^^^^^^
2772 Metadata prefixed with ``llvm.vectorizer`` is used to control per-loop
2773 vectorization parameters such as vectorization factor and unroll factor.
2775 ``llvm.vectorizer`` metadata should be used in conjunction with ``llvm.loop``
2776 loop identification metadata.
2778 '``llvm.vectorizer.unroll``' Metadata
2779 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2781 This metadata instructs the loop vectorizer to unroll the specified
2782 loop exactly ``N`` times.
2784 The first operand is the string ``llvm.vectorizer.unroll`` and the second
2785 operand is an integer specifying the unroll factor. For example:
2787 .. code-block:: llvm
2789 !0 = metadata !{ metadata !"llvm.vectorizer.unroll", i32 4 }
2791 Note that setting ``llvm.vectorizer.unroll`` to 1 disables unrolling of the
2794 If ``llvm.vectorizer.unroll`` is set to 0 then the amount of unrolling will be
2795 determined automatically.
2797 '``llvm.vectorizer.width``' Metadata
2798 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2800 This metadata sets the target width of the vectorizer to ``N``. Without
2801 this metadata, the vectorizer will choose a width automatically.
2802 Regardless of this metadata, the vectorizer will only vectorize loops if
2803 it believes it is valid to do so.
2805 The first operand is the string ``llvm.vectorizer.width`` and the second
2806 operand is an integer specifying the width. For example:
2808 .. code-block:: llvm
2810 !0 = metadata !{ metadata !"llvm.vectorizer.width", i32 4 }
2812 Note that setting ``llvm.vectorizer.width`` to 1 disables vectorization of the
2815 If ``llvm.vectorizer.width`` is set to 0 then the width will be determined
2818 Module Flags Metadata
2819 =====================
2821 Information about the module as a whole is difficult to convey to LLVM's
2822 subsystems. The LLVM IR isn't sufficient to transmit this information.
2823 The ``llvm.module.flags`` named metadata exists in order to facilitate
2824 this. These flags are in the form of key / value pairs --- much like a
2825 dictionary --- making it easy for any subsystem who cares about a flag to
2828 The ``llvm.module.flags`` metadata contains a list of metadata triplets.
2829 Each triplet has the following form:
2831 - The first element is a *behavior* flag, which specifies the behavior
2832 when two (or more) modules are merged together, and it encounters two
2833 (or more) metadata with the same ID. The supported behaviors are
2835 - The second element is a metadata string that is a unique ID for the
2836 metadata. Each module may only have one flag entry for each unique ID (not
2837 including entries with the **Require** behavior).
2838 - The third element is the value of the flag.
2840 When two (or more) modules are merged together, the resulting
2841 ``llvm.module.flags`` metadata is the union of the modules' flags. That is, for
2842 each unique metadata ID string, there will be exactly one entry in the merged
2843 modules ``llvm.module.flags`` metadata table, and the value for that entry will
2844 be determined by the merge behavior flag, as described below. The only exception
2845 is that entries with the *Require* behavior are always preserved.
2847 The following behaviors are supported:
2858 Emits an error if two values disagree, otherwise the resulting value
2859 is that of the operands.
2863 Emits a warning if two values disagree. The result value will be the
2864 operand for the flag from the first module being linked.
2868 Adds a requirement that another module flag be present and have a
2869 specified value after linking is performed. The value must be a
2870 metadata pair, where the first element of the pair is the ID of the
2871 module flag to be restricted, and the second element of the pair is
2872 the value the module flag should be restricted to. This behavior can
2873 be used to restrict the allowable results (via triggering of an
2874 error) of linking IDs with the **Override** behavior.
2878 Uses the specified value, regardless of the behavior or value of the
2879 other module. If both modules specify **Override**, but the values
2880 differ, an error will be emitted.
2884 Appends the two values, which are required to be metadata nodes.
2888 Appends the two values, which are required to be metadata
2889 nodes. However, duplicate entries in the second list are dropped
2890 during the append operation.
2892 It is an error for a particular unique flag ID to have multiple behaviors,
2893 except in the case of **Require** (which adds restrictions on another metadata
2894 value) or **Override**.
2896 An example of module flags:
2898 .. code-block:: llvm
2900 !0 = metadata !{ i32 1, metadata !"foo", i32 1 }
2901 !1 = metadata !{ i32 4, metadata !"bar", i32 37 }
2902 !2 = metadata !{ i32 2, metadata !"qux", i32 42 }
2903 !3 = metadata !{ i32 3, metadata !"qux",
2905 metadata !"foo", i32 1
2908 !llvm.module.flags = !{ !0, !1, !2, !3 }
2910 - Metadata ``!0`` has the ID ``!"foo"`` and the value '1'. The behavior
2911 if two or more ``!"foo"`` flags are seen is to emit an error if their
2912 values are not equal.
2914 - Metadata ``!1`` has the ID ``!"bar"`` and the value '37'. The
2915 behavior if two or more ``!"bar"`` flags are seen is to use the value
2918 - Metadata ``!2`` has the ID ``!"qux"`` and the value '42'. The
2919 behavior if two or more ``!"qux"`` flags are seen is to emit a
2920 warning if their values are not equal.
2922 - Metadata ``!3`` has the ID ``!"qux"`` and the value:
2926 metadata !{ metadata !"foo", i32 1 }
2928 The behavior is to emit an error if the ``llvm.module.flags`` does not
2929 contain a flag with the ID ``!"foo"`` that has the value '1' after linking is
2932 Objective-C Garbage Collection Module Flags Metadata
2933 ----------------------------------------------------
2935 On the Mach-O platform, Objective-C stores metadata about garbage
2936 collection in a special section called "image info". The metadata
2937 consists of a version number and a bitmask specifying what types of
2938 garbage collection are supported (if any) by the file. If two or more
2939 modules are linked together their garbage collection metadata needs to
2940 be merged rather than appended together.
2942 The Objective-C garbage collection module flags metadata consists of the
2943 following key-value pairs:
2952 * - ``Objective-C Version``
2953 - **[Required]** --- The Objective-C ABI version. Valid values are 1 and 2.
2955 * - ``Objective-C Image Info Version``
2956 - **[Required]** --- The version of the image info section. Currently
2959 * - ``Objective-C Image Info Section``
2960 - **[Required]** --- The section to place the metadata. Valid values are
2961 ``"__OBJC, __image_info, regular"`` for Objective-C ABI version 1, and
2962 ``"__DATA,__objc_imageinfo, regular, no_dead_strip"`` for
2963 Objective-C ABI version 2.
2965 * - ``Objective-C Garbage Collection``
2966 - **[Required]** --- Specifies whether garbage collection is supported or
2967 not. Valid values are 0, for no garbage collection, and 2, for garbage
2968 collection supported.
2970 * - ``Objective-C GC Only``
2971 - **[Optional]** --- Specifies that only garbage collection is supported.
2972 If present, its value must be 6. This flag requires that the
2973 ``Objective-C Garbage Collection`` flag have the value 2.
2975 Some important flag interactions:
2977 - If a module with ``Objective-C Garbage Collection`` set to 0 is
2978 merged with a module with ``Objective-C Garbage Collection`` set to
2979 2, then the resulting module has the
2980 ``Objective-C Garbage Collection`` flag set to 0.
2981 - A module with ``Objective-C Garbage Collection`` set to 0 cannot be
2982 merged with a module with ``Objective-C GC Only`` set to 6.
2984 Automatic Linker Flags Module Flags Metadata
2985 --------------------------------------------
2987 Some targets support embedding flags to the linker inside individual object
2988 files. Typically this is used in conjunction with language extensions which
2989 allow source files to explicitly declare the libraries they depend on, and have
2990 these automatically be transmitted to the linker via object files.
2992 These flags are encoded in the IR using metadata in the module flags section,
2993 using the ``Linker Options`` key. The merge behavior for this flag is required
2994 to be ``AppendUnique``, and the value for the key is expected to be a metadata
2995 node which should be a list of other metadata nodes, each of which should be a
2996 list of metadata strings defining linker options.
2998 For example, the following metadata section specifies two separate sets of
2999 linker options, presumably to link against ``libz`` and the ``Cocoa``
3002 !0 = metadata !{ i32 6, metadata !"Linker Options",
3004 metadata !{ metadata !"-lz" },
3005 metadata !{ metadata !"-framework", metadata !"Cocoa" } } }
3006 !llvm.module.flags = !{ !0 }
3008 The metadata encoding as lists of lists of options, as opposed to a collapsed
3009 list of options, is chosen so that the IR encoding can use multiple option
3010 strings to specify e.g., a single library, while still having that specifier be
3011 preserved as an atomic element that can be recognized by a target specific
3012 assembly writer or object file emitter.
3014 Each individual option is required to be either a valid option for the target's
3015 linker, or an option that is reserved by the target specific assembly writer or
3016 object file emitter. No other aspect of these options is defined by the IR.
3018 .. _intrinsicglobalvariables:
3020 Intrinsic Global Variables
3021 ==========================
3023 LLVM has a number of "magic" global variables that contain data that
3024 affect code generation or other IR semantics. These are documented here.
3025 All globals of this sort should have a section specified as
3026 "``llvm.metadata``". This section and all globals that start with
3027 "``llvm.``" are reserved for use by LLVM.
3031 The '``llvm.used``' Global Variable
3032 -----------------------------------
3034 The ``@llvm.used`` global is an array which has
3035 :ref:`appending linkage <linkage_appending>`. This array contains a list of
3036 pointers to named global variables, functions and aliases which may optionally
3037 have a pointer cast formed of bitcast or getelementptr. For example, a legal
3040 .. code-block:: llvm
3045 @llvm.used = appending global [2 x i8*] [
3047 i8* bitcast (i32* @Y to i8*)
3048 ], section "llvm.metadata"
3050 If a symbol appears in the ``@llvm.used`` list, then the compiler, assembler,
3051 and linker are required to treat the symbol as if there is a reference to the
3052 symbol that it cannot see (which is why they have to be named). For example, if
3053 a variable has internal linkage and no references other than that from the
3054 ``@llvm.used`` list, it cannot be deleted. This is commonly used to represent
3055 references from inline asms and other things the compiler cannot "see", and
3056 corresponds to "``attribute((used))``" in GNU C.
3058 On some targets, the code generator must emit a directive to the
3059 assembler or object file to prevent the assembler and linker from
3060 molesting the symbol.
3062 .. _gv_llvmcompilerused:
3064 The '``llvm.compiler.used``' Global Variable
3065 --------------------------------------------
3067 The ``@llvm.compiler.used`` directive is the same as the ``@llvm.used``
3068 directive, except that it only prevents the compiler from touching the
3069 symbol. On targets that support it, this allows an intelligent linker to
3070 optimize references to the symbol without being impeded as it would be
3073 This is a rare construct that should only be used in rare circumstances,
3074 and should not be exposed to source languages.
3076 .. _gv_llvmglobalctors:
3078 The '``llvm.global_ctors``' Global Variable
3079 -------------------------------------------
3081 .. code-block:: llvm
3083 %0 = type { i32, void ()* }
3084 @llvm.global_ctors = appending global [1 x %0] [%0 { i32 65535, void ()* @ctor }]
3086 The ``@llvm.global_ctors`` array contains a list of constructor
3087 functions and associated priorities. The functions referenced by this
3088 array will be called in ascending order of priority (i.e. lowest first)
3089 when the module is loaded. The order of functions with the same priority
3092 .. _llvmglobaldtors:
3094 The '``llvm.global_dtors``' Global Variable
3095 -------------------------------------------
3097 .. code-block:: llvm
3099 %0 = type { i32, void ()* }
3100 @llvm.global_dtors = appending global [1 x %0] [%0 { i32 65535, void ()* @dtor }]
3102 The ``@llvm.global_dtors`` array contains a list of destructor functions
3103 and associated priorities. The functions referenced by this array will
3104 be called in descending order of priority (i.e. highest first) when the
3105 module is loaded. The order of functions with the same priority is not
3108 Instruction Reference
3109 =====================
3111 The LLVM instruction set consists of several different classifications
3112 of instructions: :ref:`terminator instructions <terminators>`, :ref:`binary
3113 instructions <binaryops>`, :ref:`bitwise binary
3114 instructions <bitwiseops>`, :ref:`memory instructions <memoryops>`, and
3115 :ref:`other instructions <otherops>`.
3119 Terminator Instructions
3120 -----------------------
3122 As mentioned :ref:`previously <functionstructure>`, every basic block in a
3123 program ends with a "Terminator" instruction, which indicates which
3124 block should be executed after the current block is finished. These
3125 terminator instructions typically yield a '``void``' value: they produce
3126 control flow, not values (the one exception being the
3127 ':ref:`invoke <i_invoke>`' instruction).
3129 The terminator instructions are: ':ref:`ret <i_ret>`',
3130 ':ref:`br <i_br>`', ':ref:`switch <i_switch>`',
3131 ':ref:`indirectbr <i_indirectbr>`', ':ref:`invoke <i_invoke>`',
3132 ':ref:`resume <i_resume>`', and ':ref:`unreachable <i_unreachable>`'.
3136 '``ret``' Instruction
3137 ^^^^^^^^^^^^^^^^^^^^^
3144 ret <type> <value> ; Return a value from a non-void function
3145 ret void ; Return from void function
3150 The '``ret``' instruction is used to return control flow (and optionally
3151 a value) from a function back to the caller.
3153 There are two forms of the '``ret``' instruction: one that returns a
3154 value and then causes control flow, and one that just causes control
3160 The '``ret``' instruction optionally accepts a single argument, the
3161 return value. The type of the return value must be a ':ref:`first
3162 class <t_firstclass>`' type.
3164 A function is not :ref:`well formed <wellformed>` if it it has a non-void
3165 return type and contains a '``ret``' instruction with no return value or
3166 a return value with a type that does not match its type, or if it has a
3167 void return type and contains a '``ret``' instruction with a return
3173 When the '``ret``' instruction is executed, control flow returns back to
3174 the calling function's context. If the caller is a
3175 ":ref:`call <i_call>`" instruction, execution continues at the
3176 instruction after the call. If the caller was an
3177 ":ref:`invoke <i_invoke>`" instruction, execution continues at the
3178 beginning of the "normal" destination block. If the instruction returns
3179 a value, that value shall set the call or invoke instruction's return
3185 .. code-block:: llvm
3187 ret i32 5 ; Return an integer value of 5
3188 ret void ; Return from a void function
3189 ret { i32, i8 } { i32 4, i8 2 } ; Return a struct of values 4 and 2
3193 '``br``' Instruction
3194 ^^^^^^^^^^^^^^^^^^^^
3201 br i1 <cond>, label <iftrue>, label <iffalse>
3202 br label <dest> ; Unconditional branch
3207 The '``br``' instruction is used to cause control flow to transfer to a
3208 different basic block in the current function. There are two forms of
3209 this instruction, corresponding to a conditional branch and an
3210 unconditional branch.
3215 The conditional branch form of the '``br``' instruction takes a single
3216 '``i1``' value and two '``label``' values. The unconditional form of the
3217 '``br``' instruction takes a single '``label``' value as a target.
3222 Upon execution of a conditional '``br``' instruction, the '``i1``'
3223 argument is evaluated. If the value is ``true``, control flows to the
3224 '``iftrue``' ``label`` argument. If "cond" is ``false``, control flows
3225 to the '``iffalse``' ``label`` argument.
3230 .. code-block:: llvm
3233 %cond = icmp eq i32 %a, %b
3234 br i1 %cond, label %IfEqual, label %IfUnequal
3242 '``switch``' Instruction
3243 ^^^^^^^^^^^^^^^^^^^^^^^^
3250 switch <intty> <value>, label <defaultdest> [ <intty> <val>, label <dest> ... ]
3255 The '``switch``' instruction is used to transfer control flow to one of
3256 several different places. It is a generalization of the '``br``'
3257 instruction, allowing a branch to occur to one of many possible
3263 The '``switch``' instruction uses three parameters: an integer
3264 comparison value '``value``', a default '``label``' destination, and an
3265 array of pairs of comparison value constants and '``label``'s. The table
3266 is not allowed to contain duplicate constant entries.
3271 The ``switch`` instruction specifies a table of values and destinations.
3272 When the '``switch``' instruction is executed, this table is searched
3273 for the given value. If the value is found, control flow is transferred
3274 to the corresponding destination; otherwise, control flow is transferred
3275 to the default destination.
3280 Depending on properties of the target machine and the particular
3281 ``switch`` instruction, this instruction may be code generated in
3282 different ways. For example, it could be generated as a series of
3283 chained conditional branches or with a lookup table.
3288 .. code-block:: llvm
3290 ; Emulate a conditional br instruction
3291 %Val = zext i1 %value to i32
3292 switch i32 %Val, label %truedest [ i32 0, label %falsedest ]
3294 ; Emulate an unconditional br instruction
3295 switch i32 0, label %dest [ ]
3297 ; Implement a jump table:
3298 switch i32 %val, label %otherwise [ i32 0, label %onzero
3300 i32 2, label %ontwo ]
3304 '``indirectbr``' Instruction
3305 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3312 indirectbr <somety>* <address>, [ label <dest1>, label <dest2>, ... ]
3317 The '``indirectbr``' instruction implements an indirect branch to a
3318 label within the current function, whose address is specified by
3319 "``address``". Address must be derived from a
3320 :ref:`blockaddress <blockaddress>` constant.
3325 The '``address``' argument is the address of the label to jump to. The
3326 rest of the arguments indicate the full set of possible destinations
3327 that the address may point to. Blocks are allowed to occur multiple
3328 times in the destination list, though this isn't particularly useful.
3330 This destination list is required so that dataflow analysis has an
3331 accurate understanding of the CFG.
3336 Control transfers to the block specified in the address argument. All
3337 possible destination blocks must be listed in the label list, otherwise
3338 this instruction has undefined behavior. This implies that jumps to
3339 labels defined in other functions have undefined behavior as well.
3344 This is typically implemented with a jump through a register.
3349 .. code-block:: llvm
3351 indirectbr i8* %Addr, [ label %bb1, label %bb2, label %bb3 ]
3355 '``invoke``' Instruction
3356 ^^^^^^^^^^^^^^^^^^^^^^^^
3363 <result> = invoke [cconv] [ret attrs] <ptr to function ty> <function ptr val>(<function args>) [fn attrs]
3364 to label <normal label> unwind label <exception label>
3369 The '``invoke``' instruction causes control to transfer to a specified
3370 function, with the possibility of control flow transfer to either the
3371 '``normal``' label or the '``exception``' label. If the callee function
3372 returns with the "``ret``" instruction, control flow will return to the
3373 "normal" label. If the callee (or any indirect callees) returns via the
3374 ":ref:`resume <i_resume>`" instruction or other exception handling
3375 mechanism, control is interrupted and continued at the dynamically
3376 nearest "exception" label.
3378 The '``exception``' label is a `landing
3379 pad <ExceptionHandling.html#overview>`_ for the exception. As such,
3380 '``exception``' label is required to have the
3381 ":ref:`landingpad <i_landingpad>`" instruction, which contains the
3382 information about the behavior of the program after unwinding happens,
3383 as its first non-PHI instruction. The restrictions on the
3384 "``landingpad``" instruction's tightly couples it to the "``invoke``"
3385 instruction, so that the important information contained within the
3386 "``landingpad``" instruction can't be lost through normal code motion.
3391 This instruction requires several arguments:
3393 #. The optional "cconv" marker indicates which :ref:`calling
3394 convention <callingconv>` the call should use. If none is
3395 specified, the call defaults to using C calling conventions.
3396 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
3397 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
3399 #. '``ptr to function ty``': shall be the signature of the pointer to
3400 function value being invoked. In most cases, this is a direct
3401 function invocation, but indirect ``invoke``'s are just as possible,
3402 branching off an arbitrary pointer to function value.
3403 #. '``function ptr val``': An LLVM value containing a pointer to a
3404 function to be invoked.
3405 #. '``function args``': argument list whose types match the function
3406 signature argument types and parameter attributes. All arguments must
3407 be of :ref:`first class <t_firstclass>` type. If the function signature
3408 indicates the function accepts a variable number of arguments, the
3409 extra arguments can be specified.
3410 #. '``normal label``': the label reached when the called function
3411 executes a '``ret``' instruction.
3412 #. '``exception label``': the label reached when a callee returns via
3413 the :ref:`resume <i_resume>` instruction or other exception handling
3415 #. The optional :ref:`function attributes <fnattrs>` list. Only
3416 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
3417 attributes are valid here.
3422 This instruction is designed to operate as a standard '``call``'
3423 instruction in most regards. The primary difference is that it
3424 establishes an association with a label, which is used by the runtime
3425 library to unwind the stack.
3427 This instruction is used in languages with destructors to ensure that
3428 proper cleanup is performed in the case of either a ``longjmp`` or a
3429 thrown exception. Additionally, this is important for implementation of
3430 '``catch``' clauses in high-level languages that support them.
3432 For the purposes of the SSA form, the definition of the value returned
3433 by the '``invoke``' instruction is deemed to occur on the edge from the
3434 current block to the "normal" label. If the callee unwinds then no
3435 return value is available.
3440 .. code-block:: llvm
3442 %retval = invoke i32 @Test(i32 15) to label %Continue
3443 unwind label %TestCleanup ; {i32}:retval set
3444 %retval = invoke coldcc i32 %Testfnptr(i32 15) to label %Continue
3445 unwind label %TestCleanup ; {i32}:retval set
3449 '``resume``' Instruction
3450 ^^^^^^^^^^^^^^^^^^^^^^^^
3457 resume <type> <value>
3462 The '``resume``' instruction is a terminator instruction that has no
3468 The '``resume``' instruction requires one argument, which must have the
3469 same type as the result of any '``landingpad``' instruction in the same
3475 The '``resume``' instruction resumes propagation of an existing
3476 (in-flight) exception whose unwinding was interrupted with a
3477 :ref:`landingpad <i_landingpad>` instruction.
3482 .. code-block:: llvm
3484 resume { i8*, i32 } %exn
3488 '``unreachable``' Instruction
3489 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3501 The '``unreachable``' instruction has no defined semantics. This
3502 instruction is used to inform the optimizer that a particular portion of
3503 the code is not reachable. This can be used to indicate that the code
3504 after a no-return function cannot be reached, and other facts.
3509 The '``unreachable``' instruction has no defined semantics.
3516 Binary operators are used to do most of the computation in a program.
3517 They require two operands of the same type, execute an operation on
3518 them, and produce a single value. The operands might represent multiple
3519 data, as is the case with the :ref:`vector <t_vector>` data type. The
3520 result value has the same type as its operands.
3522 There are several different binary operators:
3526 '``add``' Instruction
3527 ^^^^^^^^^^^^^^^^^^^^^
3534 <result> = add <ty> <op1>, <op2> ; yields {ty}:result
3535 <result> = add nuw <ty> <op1>, <op2> ; yields {ty}:result
3536 <result> = add nsw <ty> <op1>, <op2> ; yields {ty}:result
3537 <result> = add nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3542 The '``add``' instruction returns the sum of its two operands.
3547 The two arguments to the '``add``' instruction must be
3548 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3549 arguments must have identical types.
3554 The value produced is the integer sum of the two operands.
3556 If the sum has unsigned overflow, the result returned is the
3557 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
3560 Because LLVM integers use a two's complement representation, this
3561 instruction is appropriate for both signed and unsigned integers.
3563 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3564 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3565 result value of the ``add`` is a :ref:`poison value <poisonvalues>` if
3566 unsigned and/or signed overflow, respectively, occurs.
3571 .. code-block:: llvm
3573 <result> = add i32 4, %var ; yields {i32}:result = 4 + %var
3577 '``fadd``' Instruction
3578 ^^^^^^^^^^^^^^^^^^^^^^
3585 <result> = fadd [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3590 The '``fadd``' instruction returns the sum of its two operands.
3595 The two arguments to the '``fadd``' instruction must be :ref:`floating
3596 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3597 Both arguments must have identical types.
3602 The value produced is the floating point sum of the two operands. This
3603 instruction can also take any number of :ref:`fast-math flags <fastmath>`,
3604 which are optimization hints to enable otherwise unsafe floating point
3610 .. code-block:: llvm
3612 <result> = fadd float 4.0, %var ; yields {float}:result = 4.0 + %var
3614 '``sub``' Instruction
3615 ^^^^^^^^^^^^^^^^^^^^^
3622 <result> = sub <ty> <op1>, <op2> ; yields {ty}:result
3623 <result> = sub nuw <ty> <op1>, <op2> ; yields {ty}:result
3624 <result> = sub nsw <ty> <op1>, <op2> ; yields {ty}:result
3625 <result> = sub nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3630 The '``sub``' instruction returns the difference of its two operands.
3632 Note that the '``sub``' instruction is used to represent the '``neg``'
3633 instruction present in most other intermediate representations.
3638 The two arguments to the '``sub``' instruction must be
3639 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3640 arguments must have identical types.
3645 The value produced is the integer difference of the two operands.
3647 If the difference has unsigned overflow, the result returned is the
3648 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
3651 Because LLVM integers use a two's complement representation, this
3652 instruction is appropriate for both signed and unsigned integers.
3654 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3655 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3656 result value of the ``sub`` is a :ref:`poison value <poisonvalues>` if
3657 unsigned and/or signed overflow, respectively, occurs.
3662 .. code-block:: llvm
3664 <result> = sub i32 4, %var ; yields {i32}:result = 4 - %var
3665 <result> = sub i32 0, %val ; yields {i32}:result = -%var
3669 '``fsub``' Instruction
3670 ^^^^^^^^^^^^^^^^^^^^^^
3677 <result> = fsub [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3682 The '``fsub``' instruction returns the difference of its two operands.
3684 Note that the '``fsub``' instruction is used to represent the '``fneg``'
3685 instruction present in most other intermediate representations.
3690 The two arguments to the '``fsub``' instruction must be :ref:`floating
3691 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3692 Both arguments must have identical types.
3697 The value produced is the floating point difference of the two operands.
3698 This instruction can also take any number of :ref:`fast-math
3699 flags <fastmath>`, which are optimization hints to enable otherwise
3700 unsafe floating point optimizations:
3705 .. code-block:: llvm
3707 <result> = fsub float 4.0, %var ; yields {float}:result = 4.0 - %var
3708 <result> = fsub float -0.0, %val ; yields {float}:result = -%var
3710 '``mul``' Instruction
3711 ^^^^^^^^^^^^^^^^^^^^^
3718 <result> = mul <ty> <op1>, <op2> ; yields {ty}:result
3719 <result> = mul nuw <ty> <op1>, <op2> ; yields {ty}:result
3720 <result> = mul nsw <ty> <op1>, <op2> ; yields {ty}:result
3721 <result> = mul nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3726 The '``mul``' instruction returns the product of its two operands.
3731 The two arguments to the '``mul``' instruction must be
3732 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3733 arguments must have identical types.
3738 The value produced is the integer product of the two operands.
3740 If the result of the multiplication has unsigned overflow, the result
3741 returned is the mathematical result modulo 2\ :sup:`n`\ , where n is the
3742 bit width of the result.
3744 Because LLVM integers use a two's complement representation, and the
3745 result is the same width as the operands, this instruction returns the
3746 correct result for both signed and unsigned integers. If a full product
3747 (e.g. ``i32`` * ``i32`` -> ``i64``) is needed, the operands should be
3748 sign-extended or zero-extended as appropriate to the width of the full
3751 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3752 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3753 result value of the ``mul`` is a :ref:`poison value <poisonvalues>` if
3754 unsigned and/or signed overflow, respectively, occurs.
3759 .. code-block:: llvm
3761 <result> = mul i32 4, %var ; yields {i32}:result = 4 * %var
3765 '``fmul``' Instruction
3766 ^^^^^^^^^^^^^^^^^^^^^^
3773 <result> = fmul [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3778 The '``fmul``' instruction returns the product of its two operands.
3783 The two arguments to the '``fmul``' instruction must be :ref:`floating
3784 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3785 Both arguments must have identical types.
3790 The value produced is the floating point product of the two operands.
3791 This instruction can also take any number of :ref:`fast-math
3792 flags <fastmath>`, which are optimization hints to enable otherwise
3793 unsafe floating point optimizations:
3798 .. code-block:: llvm
3800 <result> = fmul float 4.0, %var ; yields {float}:result = 4.0 * %var
3802 '``udiv``' Instruction
3803 ^^^^^^^^^^^^^^^^^^^^^^
3810 <result> = udiv <ty> <op1>, <op2> ; yields {ty}:result
3811 <result> = udiv exact <ty> <op1>, <op2> ; yields {ty}:result
3816 The '``udiv``' instruction returns the quotient of its two operands.
3821 The two arguments to the '``udiv``' instruction must be
3822 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3823 arguments must have identical types.
3828 The value produced is the unsigned integer quotient of the two operands.
3830 Note that unsigned integer division and signed integer division are
3831 distinct operations; for signed integer division, use '``sdiv``'.
3833 Division by zero leads to undefined behavior.
3835 If the ``exact`` keyword is present, the result value of the ``udiv`` is
3836 a :ref:`poison value <poisonvalues>` if %op1 is not a multiple of %op2 (as
3837 such, "((a udiv exact b) mul b) == a").
3842 .. code-block:: llvm
3844 <result> = udiv i32 4, %var ; yields {i32}:result = 4 / %var
3846 '``sdiv``' Instruction
3847 ^^^^^^^^^^^^^^^^^^^^^^
3854 <result> = sdiv <ty> <op1>, <op2> ; yields {ty}:result
3855 <result> = sdiv exact <ty> <op1>, <op2> ; yields {ty}:result
3860 The '``sdiv``' instruction returns the quotient of its two operands.
3865 The two arguments to the '``sdiv``' instruction must be
3866 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3867 arguments must have identical types.
3872 The value produced is the signed integer quotient of the two operands
3873 rounded towards zero.
3875 Note that signed integer division and unsigned integer division are
3876 distinct operations; for unsigned integer division, use '``udiv``'.
3878 Division by zero leads to undefined behavior. Overflow also leads to
3879 undefined behavior; this is a rare case, but can occur, for example, by
3880 doing a 32-bit division of -2147483648 by -1.
3882 If the ``exact`` keyword is present, the result value of the ``sdiv`` is
3883 a :ref:`poison value <poisonvalues>` if the result would be rounded.
3888 .. code-block:: llvm
3890 <result> = sdiv i32 4, %var ; yields {i32}:result = 4 / %var
3894 '``fdiv``' Instruction
3895 ^^^^^^^^^^^^^^^^^^^^^^
3902 <result> = fdiv [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3907 The '``fdiv``' instruction returns the quotient of its two operands.
3912 The two arguments to the '``fdiv``' instruction must be :ref:`floating
3913 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3914 Both arguments must have identical types.
3919 The value produced is the floating point quotient of the two operands.
3920 This instruction can also take any number of :ref:`fast-math
3921 flags <fastmath>`, which are optimization hints to enable otherwise
3922 unsafe floating point optimizations:
3927 .. code-block:: llvm
3929 <result> = fdiv float 4.0, %var ; yields {float}:result = 4.0 / %var
3931 '``urem``' Instruction
3932 ^^^^^^^^^^^^^^^^^^^^^^
3939 <result> = urem <ty> <op1>, <op2> ; yields {ty}:result
3944 The '``urem``' instruction returns the remainder from the unsigned
3945 division of its two arguments.
3950 The two arguments to the '``urem``' instruction must be
3951 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3952 arguments must have identical types.
3957 This instruction returns the unsigned integer *remainder* of a division.
3958 This instruction always performs an unsigned division to get the
3961 Note that unsigned integer remainder and signed integer remainder are
3962 distinct operations; for signed integer remainder, use '``srem``'.
3964 Taking the remainder of a division by zero leads to undefined behavior.
3969 .. code-block:: llvm
3971 <result> = urem i32 4, %var ; yields {i32}:result = 4 % %var
3973 '``srem``' Instruction
3974 ^^^^^^^^^^^^^^^^^^^^^^
3981 <result> = srem <ty> <op1>, <op2> ; yields {ty}:result
3986 The '``srem``' instruction returns the remainder from the signed
3987 division of its two operands. This instruction can also take
3988 :ref:`vector <t_vector>` versions of the values in which case the elements
3994 The two arguments to the '``srem``' instruction must be
3995 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3996 arguments must have identical types.
4001 This instruction returns the *remainder* of a division (where the result
4002 is either zero or has the same sign as the dividend, ``op1``), not the
4003 *modulo* operator (where the result is either zero or has the same sign
4004 as the divisor, ``op2``) of a value. For more information about the
4005 difference, see `The Math
4006 Forum <http://mathforum.org/dr.math/problems/anne.4.28.99.html>`_. For a
4007 table of how this is implemented in various languages, please see
4009 operation <http://en.wikipedia.org/wiki/Modulo_operation>`_.
4011 Note that signed integer remainder and unsigned integer remainder are
4012 distinct operations; for unsigned integer remainder, use '``urem``'.
4014 Taking the remainder of a division by zero leads to undefined behavior.
4015 Overflow also leads to undefined behavior; this is a rare case, but can
4016 occur, for example, by taking the remainder of a 32-bit division of
4017 -2147483648 by -1. (The remainder doesn't actually overflow, but this
4018 rule lets srem be implemented using instructions that return both the
4019 result of the division and the remainder.)
4024 .. code-block:: llvm
4026 <result> = srem i32 4, %var ; yields {i32}:result = 4 % %var
4030 '``frem``' Instruction
4031 ^^^^^^^^^^^^^^^^^^^^^^
4038 <result> = frem [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
4043 The '``frem``' instruction returns the remainder from the division of
4049 The two arguments to the '``frem``' instruction must be :ref:`floating
4050 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
4051 Both arguments must have identical types.
4056 This instruction returns the *remainder* of a division. The remainder
4057 has the same sign as the dividend. This instruction can also take any
4058 number of :ref:`fast-math flags <fastmath>`, which are optimization hints
4059 to enable otherwise unsafe floating point optimizations:
4064 .. code-block:: llvm
4066 <result> = frem float 4.0, %var ; yields {float}:result = 4.0 % %var
4070 Bitwise Binary Operations
4071 -------------------------
4073 Bitwise binary operators are used to do various forms of bit-twiddling
4074 in a program. They are generally very efficient instructions and can
4075 commonly be strength reduced from other instructions. They require two
4076 operands of the same type, execute an operation on them, and produce a
4077 single value. The resulting value is the same type as its operands.
4079 '``shl``' Instruction
4080 ^^^^^^^^^^^^^^^^^^^^^
4087 <result> = shl <ty> <op1>, <op2> ; yields {ty}:result
4088 <result> = shl nuw <ty> <op1>, <op2> ; yields {ty}:result
4089 <result> = shl nsw <ty> <op1>, <op2> ; yields {ty}:result
4090 <result> = shl nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
4095 The '``shl``' instruction returns the first operand shifted to the left
4096 a specified number of bits.
4101 Both arguments to the '``shl``' instruction must be the same
4102 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4103 '``op2``' is treated as an unsigned value.
4108 The value produced is ``op1`` \* 2\ :sup:`op2` mod 2\ :sup:`n`,
4109 where ``n`` is the width of the result. If ``op2`` is (statically or
4110 dynamically) negative or equal to or larger than the number of bits in
4111 ``op1``, the result is undefined. If the arguments are vectors, each
4112 vector element of ``op1`` is shifted by the corresponding shift amount
4115 If the ``nuw`` keyword is present, then the shift produces a :ref:`poison
4116 value <poisonvalues>` if it shifts out any non-zero bits. If the
4117 ``nsw`` keyword is present, then the shift produces a :ref:`poison
4118 value <poisonvalues>` if it shifts out any bits that disagree with the
4119 resultant sign bit. As such, NUW/NSW have the same semantics as they
4120 would if the shift were expressed as a mul instruction with the same
4121 nsw/nuw bits in (mul %op1, (shl 1, %op2)).
4126 .. code-block:: llvm
4128 <result> = shl i32 4, %var ; yields {i32}: 4 << %var
4129 <result> = shl i32 4, 2 ; yields {i32}: 16
4130 <result> = shl i32 1, 10 ; yields {i32}: 1024
4131 <result> = shl i32 1, 32 ; undefined
4132 <result> = shl <2 x i32> < i32 1, i32 1>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 2, i32 4>
4134 '``lshr``' Instruction
4135 ^^^^^^^^^^^^^^^^^^^^^^
4142 <result> = lshr <ty> <op1>, <op2> ; yields {ty}:result
4143 <result> = lshr exact <ty> <op1>, <op2> ; yields {ty}:result
4148 The '``lshr``' instruction (logical shift right) returns the first
4149 operand shifted to the right a specified number of bits with zero fill.
4154 Both arguments to the '``lshr``' instruction must be the same
4155 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4156 '``op2``' is treated as an unsigned value.
4161 This instruction always performs a logical shift right operation. The
4162 most significant bits of the result will be filled with zero bits after
4163 the shift. If ``op2`` is (statically or dynamically) equal to or larger
4164 than the number of bits in ``op1``, the result is undefined. If the
4165 arguments are vectors, each vector element of ``op1`` is shifted by the
4166 corresponding shift amount in ``op2``.
4168 If the ``exact`` keyword is present, the result value of the ``lshr`` is
4169 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
4175 .. code-block:: llvm
4177 <result> = lshr i32 4, 1 ; yields {i32}:result = 2
4178 <result> = lshr i32 4, 2 ; yields {i32}:result = 1
4179 <result> = lshr i8 4, 3 ; yields {i8}:result = 0
4180 <result> = lshr i8 -2, 1 ; yields {i8}:result = 0x7F
4181 <result> = lshr i32 1, 32 ; undefined
4182 <result> = lshr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 0x7FFFFFFF, i32 1>
4184 '``ashr``' Instruction
4185 ^^^^^^^^^^^^^^^^^^^^^^
4192 <result> = ashr <ty> <op1>, <op2> ; yields {ty}:result
4193 <result> = ashr exact <ty> <op1>, <op2> ; yields {ty}:result
4198 The '``ashr``' instruction (arithmetic shift right) returns the first
4199 operand shifted to the right a specified number of bits with sign
4205 Both arguments to the '``ashr``' instruction must be the same
4206 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4207 '``op2``' is treated as an unsigned value.
4212 This instruction always performs an arithmetic shift right operation,
4213 The most significant bits of the result will be filled with the sign bit
4214 of ``op1``. If ``op2`` is (statically or dynamically) equal to or larger
4215 than the number of bits in ``op1``, the result is undefined. If the
4216 arguments are vectors, each vector element of ``op1`` is shifted by the
4217 corresponding shift amount in ``op2``.
4219 If the ``exact`` keyword is present, the result value of the ``ashr`` is
4220 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
4226 .. code-block:: llvm
4228 <result> = ashr i32 4, 1 ; yields {i32}:result = 2
4229 <result> = ashr i32 4, 2 ; yields {i32}:result = 1
4230 <result> = ashr i8 4, 3 ; yields {i8}:result = 0
4231 <result> = ashr i8 -2, 1 ; yields {i8}:result = -1
4232 <result> = ashr i32 1, 32 ; undefined
4233 <result> = ashr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 3> ; yields: result=<2 x i32> < i32 -1, i32 0>
4235 '``and``' Instruction
4236 ^^^^^^^^^^^^^^^^^^^^^
4243 <result> = and <ty> <op1>, <op2> ; yields {ty}:result
4248 The '``and``' instruction returns the bitwise logical and of its two
4254 The two arguments to the '``and``' instruction must be
4255 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4256 arguments must have identical types.
4261 The truth table used for the '``and``' instruction is:
4278 .. code-block:: llvm
4280 <result> = and i32 4, %var ; yields {i32}:result = 4 & %var
4281 <result> = and i32 15, 40 ; yields {i32}:result = 8
4282 <result> = and i32 4, 8 ; yields {i32}:result = 0
4284 '``or``' Instruction
4285 ^^^^^^^^^^^^^^^^^^^^
4292 <result> = or <ty> <op1>, <op2> ; yields {ty}:result
4297 The '``or``' instruction returns the bitwise logical inclusive or of its
4303 The two arguments to the '``or``' instruction must be
4304 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4305 arguments must have identical types.
4310 The truth table used for the '``or``' instruction is:
4329 <result> = or i32 4, %var ; yields {i32}:result = 4 | %var
4330 <result> = or i32 15, 40 ; yields {i32}:result = 47
4331 <result> = or i32 4, 8 ; yields {i32}:result = 12
4333 '``xor``' Instruction
4334 ^^^^^^^^^^^^^^^^^^^^^
4341 <result> = xor <ty> <op1>, <op2> ; yields {ty}:result
4346 The '``xor``' instruction returns the bitwise logical exclusive or of
4347 its two operands. The ``xor`` is used to implement the "one's
4348 complement" operation, which is the "~" operator in C.
4353 The two arguments to the '``xor``' instruction must be
4354 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4355 arguments must have identical types.
4360 The truth table used for the '``xor``' instruction is:
4377 .. code-block:: llvm
4379 <result> = xor i32 4, %var ; yields {i32}:result = 4 ^ %var
4380 <result> = xor i32 15, 40 ; yields {i32}:result = 39
4381 <result> = xor i32 4, 8 ; yields {i32}:result = 12
4382 <result> = xor i32 %V, -1 ; yields {i32}:result = ~%V
4387 LLVM supports several instructions to represent vector operations in a
4388 target-independent manner. These instructions cover the element-access
4389 and vector-specific operations needed to process vectors effectively.
4390 While LLVM does directly support these vector operations, many
4391 sophisticated algorithms will want to use target-specific intrinsics to
4392 take full advantage of a specific target.
4394 .. _i_extractelement:
4396 '``extractelement``' Instruction
4397 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4404 <result> = extractelement <n x <ty>> <val>, i32 <idx> ; yields <ty>
4409 The '``extractelement``' instruction extracts a single scalar element
4410 from a vector at a specified index.
4415 The first operand of an '``extractelement``' instruction is a value of
4416 :ref:`vector <t_vector>` type. The second operand is an index indicating
4417 the position from which to extract the element. The index may be a
4423 The result is a scalar of the same type as the element type of ``val``.
4424 Its value is the value at position ``idx`` of ``val``. If ``idx``
4425 exceeds the length of ``val``, the results are undefined.
4430 .. code-block:: llvm
4432 <result> = extractelement <4 x i32> %vec, i32 0 ; yields i32
4434 .. _i_insertelement:
4436 '``insertelement``' Instruction
4437 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4444 <result> = insertelement <n x <ty>> <val>, <ty> <elt>, i32 <idx> ; yields <n x <ty>>
4449 The '``insertelement``' instruction inserts a scalar element into a
4450 vector at a specified index.
4455 The first operand of an '``insertelement``' instruction is a value of
4456 :ref:`vector <t_vector>` type. The second operand is a scalar value whose
4457 type must equal the element type of the first operand. The third operand
4458 is an index indicating the position at which to insert the value. The
4459 index may be a variable.
4464 The result is a vector of the same type as ``val``. Its element values
4465 are those of ``val`` except at position ``idx``, where it gets the value
4466 ``elt``. If ``idx`` exceeds the length of ``val``, the results are
4472 .. code-block:: llvm
4474 <result> = insertelement <4 x i32> %vec, i32 1, i32 0 ; yields <4 x i32>
4476 .. _i_shufflevector:
4478 '``shufflevector``' Instruction
4479 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4486 <result> = shufflevector <n x <ty>> <v1>, <n x <ty>> <v2>, <m x i32> <mask> ; yields <m x <ty>>
4491 The '``shufflevector``' instruction constructs a permutation of elements
4492 from two input vectors, returning a vector with the same element type as
4493 the input and length that is the same as the shuffle mask.
4498 The first two operands of a '``shufflevector``' instruction are vectors
4499 with the same type. The third argument is a shuffle mask whose element
4500 type is always 'i32'. The result of the instruction is a vector whose
4501 length is the same as the shuffle mask and whose element type is the
4502 same as the element type of the first two operands.
4504 The shuffle mask operand is required to be a constant vector with either
4505 constant integer or undef values.
4510 The elements of the two input vectors are numbered from left to right
4511 across both of the vectors. The shuffle mask operand specifies, for each
4512 element of the result vector, which element of the two input vectors the
4513 result element gets. The element selector may be undef (meaning "don't
4514 care") and the second operand may be undef if performing a shuffle from
4520 .. code-block:: llvm
4522 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
4523 <4 x i32> <i32 0, i32 4, i32 1, i32 5> ; yields <4 x i32>
4524 <result> = shufflevector <4 x i32> %v1, <4 x i32> undef,
4525 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32> - Identity shuffle.
4526 <result> = shufflevector <8 x i32> %v1, <8 x i32> undef,
4527 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32>
4528 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
4529 <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7 > ; yields <8 x i32>
4531 Aggregate Operations
4532 --------------------
4534 LLVM supports several instructions for working with
4535 :ref:`aggregate <t_aggregate>` values.
4539 '``extractvalue``' Instruction
4540 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4547 <result> = extractvalue <aggregate type> <val>, <idx>{, <idx>}*
4552 The '``extractvalue``' instruction extracts the value of a member field
4553 from an :ref:`aggregate <t_aggregate>` value.
4558 The first operand of an '``extractvalue``' instruction is a value of
4559 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The operands are
4560 constant indices to specify which value to extract in a similar manner
4561 as indices in a '``getelementptr``' instruction.
4563 The major differences to ``getelementptr`` indexing are:
4565 - Since the value being indexed is not a pointer, the first index is
4566 omitted and assumed to be zero.
4567 - At least one index must be specified.
4568 - Not only struct indices but also array indices must be in bounds.
4573 The result is the value at the position in the aggregate specified by
4579 .. code-block:: llvm
4581 <result> = extractvalue {i32, float} %agg, 0 ; yields i32
4585 '``insertvalue``' Instruction
4586 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4593 <result> = insertvalue <aggregate type> <val>, <ty> <elt>, <idx>{, <idx>}* ; yields <aggregate type>
4598 The '``insertvalue``' instruction inserts a value into a member field in
4599 an :ref:`aggregate <t_aggregate>` value.
4604 The first operand of an '``insertvalue``' instruction is a value of
4605 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The second operand is
4606 a first-class value to insert. The following operands are constant
4607 indices indicating the position at which to insert the value in a
4608 similar manner as indices in a '``extractvalue``' instruction. The value
4609 to insert must have the same type as the value identified by the
4615 The result is an aggregate of the same type as ``val``. Its value is
4616 that of ``val`` except that the value at the position specified by the
4617 indices is that of ``elt``.
4622 .. code-block:: llvm
4624 %agg1 = insertvalue {i32, float} undef, i32 1, 0 ; yields {i32 1, float undef}
4625 %agg2 = insertvalue {i32, float} %agg1, float %val, 1 ; yields {i32 1, float %val}
4626 %agg3 = insertvalue {i32, {float}} %agg1, float %val, 1, 0 ; yields {i32 1, float %val}
4630 Memory Access and Addressing Operations
4631 ---------------------------------------
4633 A key design point of an SSA-based representation is how it represents
4634 memory. In LLVM, no memory locations are in SSA form, which makes things
4635 very simple. This section describes how to read, write, and allocate
4640 '``alloca``' Instruction
4641 ^^^^^^^^^^^^^^^^^^^^^^^^
4648 <result> = alloca <type>[, <ty> <NumElements>][, align <alignment>] ; yields {type*}:result
4653 The '``alloca``' instruction allocates memory on the stack frame of the
4654 currently executing function, to be automatically released when this
4655 function returns to its caller. The object is always allocated in the
4656 generic address space (address space zero).
4661 The '``alloca``' instruction allocates ``sizeof(<type>)*NumElements``
4662 bytes of memory on the runtime stack, returning a pointer of the
4663 appropriate type to the program. If "NumElements" is specified, it is
4664 the number of elements allocated, otherwise "NumElements" is defaulted
4665 to be one. If a constant alignment is specified, the value result of the
4666 allocation is guaranteed to be aligned to at least that boundary. If not
4667 specified, or if zero, the target can choose to align the allocation on
4668 any convenient boundary compatible with the type.
4670 '``type``' may be any sized type.
4675 Memory is allocated; a pointer is returned. The operation is undefined
4676 if there is insufficient stack space for the allocation. '``alloca``'d
4677 memory is automatically released when the function returns. The
4678 '``alloca``' instruction is commonly used to represent automatic
4679 variables that must have an address available. When the function returns
4680 (either with the ``ret`` or ``resume`` instructions), the memory is
4681 reclaimed. Allocating zero bytes is legal, but the result is undefined.
4682 The order in which memory is allocated (ie., which way the stack grows)
4688 .. code-block:: llvm
4690 %ptr = alloca i32 ; yields {i32*}:ptr
4691 %ptr = alloca i32, i32 4 ; yields {i32*}:ptr
4692 %ptr = alloca i32, i32 4, align 1024 ; yields {i32*}:ptr
4693 %ptr = alloca i32, align 1024 ; yields {i32*}:ptr
4697 '``load``' Instruction
4698 ^^^^^^^^^^^^^^^^^^^^^^
4705 <result> = load [volatile] <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>][, !invariant.load !<index>]
4706 <result> = load atomic [volatile] <ty>* <pointer> [singlethread] <ordering>, align <alignment>
4707 !<index> = !{ i32 1 }
4712 The '``load``' instruction is used to read from memory.
4717 The argument to the ``load`` instruction specifies the memory address
4718 from which to load. The pointer must point to a :ref:`first
4719 class <t_firstclass>` type. If the ``load`` is marked as ``volatile``,
4720 then the optimizer is not allowed to modify the number or order of
4721 execution of this ``load`` with other :ref:`volatile
4722 operations <volatile>`.
4724 If the ``load`` is marked as ``atomic``, it takes an extra
4725 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
4726 ``release`` and ``acq_rel`` orderings are not valid on ``load``
4727 instructions. Atomic loads produce :ref:`defined <memmodel>` results
4728 when they may see multiple atomic stores. The type of the pointee must
4729 be an integer type whose bit width is a power of two greater than or
4730 equal to eight and less than or equal to a target-specific size limit.
4731 ``align`` must be explicitly specified on atomic loads, and the load has
4732 undefined behavior if the alignment is not set to a value which is at
4733 least the size in bytes of the pointee. ``!nontemporal`` does not have
4734 any defined semantics for atomic loads.
4736 The optional constant ``align`` argument specifies the alignment of the
4737 operation (that is, the alignment of the memory address). A value of 0
4738 or an omitted ``align`` argument means that the operation has the ABI
4739 alignment for the target. It is the responsibility of the code emitter
4740 to ensure that the alignment information is correct. Overestimating the
4741 alignment results in undefined behavior. Underestimating the alignment
4742 may produce less efficient code. An alignment of 1 is always safe.
4744 The optional ``!nontemporal`` metadata must reference a single
4745 metadata name ``<index>`` corresponding to a metadata node with one
4746 ``i32`` entry of value 1. The existence of the ``!nontemporal``
4747 metadata on the instruction tells the optimizer and code generator
4748 that this load is not expected to be reused in the cache. The code
4749 generator may select special instructions to save cache bandwidth, such
4750 as the ``MOVNT`` instruction on x86.
4752 The optional ``!invariant.load`` metadata must reference a single
4753 metadata name ``<index>`` corresponding to a metadata node with no
4754 entries. The existence of the ``!invariant.load`` metadata on the
4755 instruction tells the optimizer and code generator that this load
4756 address points to memory which does not change value during program
4757 execution. The optimizer may then move this load around, for example, by
4758 hoisting it out of loops using loop invariant code motion.
4763 The location of memory pointed to is loaded. If the value being loaded
4764 is of scalar type then the number of bytes read does not exceed the
4765 minimum number of bytes needed to hold all bits of the type. For
4766 example, loading an ``i24`` reads at most three bytes. When loading a
4767 value of a type like ``i20`` with a size that is not an integral number
4768 of bytes, the result is undefined if the value was not originally
4769 written using a store of the same type.
4774 .. code-block:: llvm
4776 %ptr = alloca i32 ; yields {i32*}:ptr
4777 store i32 3, i32* %ptr ; yields {void}
4778 %val = load i32* %ptr ; yields {i32}:val = i32 3
4782 '``store``' Instruction
4783 ^^^^^^^^^^^^^^^^^^^^^^^
4790 store [volatile] <ty> <value>, <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>] ; yields {void}
4791 store atomic [volatile] <ty> <value>, <ty>* <pointer> [singlethread] <ordering>, align <alignment> ; yields {void}
4796 The '``store``' instruction is used to write to memory.
4801 There are two arguments to the ``store`` instruction: a value to store
4802 and an address at which to store it. The type of the ``<pointer>``
4803 operand must be a pointer to the :ref:`first class <t_firstclass>` type of
4804 the ``<value>`` operand. If the ``store`` is marked as ``volatile``,
4805 then the optimizer is not allowed to modify the number or order of
4806 execution of this ``store`` with other :ref:`volatile
4807 operations <volatile>`.
4809 If the ``store`` is marked as ``atomic``, it takes an extra
4810 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
4811 ``acquire`` and ``acq_rel`` orderings aren't valid on ``store``
4812 instructions. Atomic loads produce :ref:`defined <memmodel>` results
4813 when they may see multiple atomic stores. The type of the pointee must
4814 be an integer type whose bit width is a power of two greater than or
4815 equal to eight and less than or equal to a target-specific size limit.
4816 ``align`` must be explicitly specified on atomic stores, and the store
4817 has undefined behavior if the alignment is not set to a value which is
4818 at least the size in bytes of the pointee. ``!nontemporal`` does not
4819 have any defined semantics for atomic stores.
4821 The optional constant ``align`` argument specifies the alignment of the
4822 operation (that is, the alignment of the memory address). A value of 0
4823 or an omitted ``align`` argument means that the operation has the ABI
4824 alignment for the target. It is the responsibility of the code emitter
4825 to ensure that the alignment information is correct. Overestimating the
4826 alignment results in undefined behavior. Underestimating the
4827 alignment may produce less efficient code. An alignment of 1 is always
4830 The optional ``!nontemporal`` metadata must reference a single metadata
4831 name ``<index>`` corresponding to a metadata node with one ``i32`` entry of
4832 value 1. The existence of the ``!nontemporal`` metadata on the instruction
4833 tells the optimizer and code generator that this load is not expected to
4834 be reused in the cache. The code generator may select special
4835 instructions to save cache bandwidth, such as the MOVNT instruction on
4841 The contents of memory are updated to contain ``<value>`` at the
4842 location specified by the ``<pointer>`` operand. If ``<value>`` is
4843 of scalar type then the number of bytes written does not exceed the
4844 minimum number of bytes needed to hold all bits of the type. For
4845 example, storing an ``i24`` writes at most three bytes. When writing a
4846 value of a type like ``i20`` with a size that is not an integral number
4847 of bytes, it is unspecified what happens to the extra bits that do not
4848 belong to the type, but they will typically be overwritten.
4853 .. code-block:: llvm
4855 %ptr = alloca i32 ; yields {i32*}:ptr
4856 store i32 3, i32* %ptr ; yields {void}
4857 %val = load i32* %ptr ; yields {i32}:val = i32 3
4861 '``fence``' Instruction
4862 ^^^^^^^^^^^^^^^^^^^^^^^
4869 fence [singlethread] <ordering> ; yields {void}
4874 The '``fence``' instruction is used to introduce happens-before edges
4880 '``fence``' instructions take an :ref:`ordering <ordering>` argument which
4881 defines what *synchronizes-with* edges they add. They can only be given
4882 ``acquire``, ``release``, ``acq_rel``, and ``seq_cst`` orderings.
4887 A fence A which has (at least) ``release`` ordering semantics
4888 *synchronizes with* a fence B with (at least) ``acquire`` ordering
4889 semantics if and only if there exist atomic operations X and Y, both
4890 operating on some atomic object M, such that A is sequenced before X, X
4891 modifies M (either directly or through some side effect of a sequence
4892 headed by X), Y is sequenced before B, and Y observes M. This provides a
4893 *happens-before* dependency between A and B. Rather than an explicit
4894 ``fence``, one (but not both) of the atomic operations X or Y might
4895 provide a ``release`` or ``acquire`` (resp.) ordering constraint and
4896 still *synchronize-with* the explicit ``fence`` and establish the
4897 *happens-before* edge.
4899 A ``fence`` which has ``seq_cst`` ordering, in addition to having both
4900 ``acquire`` and ``release`` semantics specified above, participates in
4901 the global program order of other ``seq_cst`` operations and/or fences.
4903 The optional ":ref:`singlethread <singlethread>`" argument specifies
4904 that the fence only synchronizes with other fences in the same thread.
4905 (This is useful for interacting with signal handlers.)
4910 .. code-block:: llvm
4912 fence acquire ; yields {void}
4913 fence singlethread seq_cst ; yields {void}
4917 '``cmpxchg``' Instruction
4918 ^^^^^^^^^^^^^^^^^^^^^^^^^
4925 cmpxchg [volatile] <ty>* <pointer>, <ty> <cmp>, <ty> <new> [singlethread] <ordering> ; yields {ty}
4930 The '``cmpxchg``' instruction is used to atomically modify memory. It
4931 loads a value in memory and compares it to a given value. If they are
4932 equal, it stores a new value into the memory.
4937 There are three arguments to the '``cmpxchg``' instruction: an address
4938 to operate on, a value to compare to the value currently be at that
4939 address, and a new value to place at that address if the compared values
4940 are equal. The type of '<cmp>' must be an integer type whose bit width
4941 is a power of two greater than or equal to eight and less than or equal
4942 to a target-specific size limit. '<cmp>' and '<new>' must have the same
4943 type, and the type of '<pointer>' must be a pointer to that type. If the
4944 ``cmpxchg`` is marked as ``volatile``, then the optimizer is not allowed
4945 to modify the number or order of execution of this ``cmpxchg`` with
4946 other :ref:`volatile operations <volatile>`.
4948 The :ref:`ordering <ordering>` argument specifies how this ``cmpxchg``
4949 synchronizes with other atomic operations.
4951 The optional "``singlethread``" argument declares that the ``cmpxchg``
4952 is only atomic with respect to code (usually signal handlers) running in
4953 the same thread as the ``cmpxchg``. Otherwise the cmpxchg is atomic with
4954 respect to all other code in the system.
4956 The pointer passed into cmpxchg must have alignment greater than or
4957 equal to the size in memory of the operand.
4962 The contents of memory at the location specified by the '``<pointer>``'
4963 operand is read and compared to '``<cmp>``'; if the read value is the
4964 equal, '``<new>``' is written. The original value at the location is
4967 A successful ``cmpxchg`` is a read-modify-write instruction for the purpose
4968 of identifying release sequences. A failed ``cmpxchg`` is equivalent to an
4969 atomic load with an ordering parameter determined by dropping any
4970 ``release`` part of the ``cmpxchg``'s ordering.
4975 .. code-block:: llvm
4978 %orig = atomic load i32* %ptr unordered ; yields {i32}
4982 %cmp = phi i32 [ %orig, %entry ], [%old, %loop]
4983 %squared = mul i32 %cmp, %cmp
4984 %old = cmpxchg i32* %ptr, i32 %cmp, i32 %squared ; yields {i32}
4985 %success = icmp eq i32 %cmp, %old
4986 br i1 %success, label %done, label %loop
4993 '``atomicrmw``' Instruction
4994 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
5001 atomicrmw [volatile] <operation> <ty>* <pointer>, <ty> <value> [singlethread] <ordering> ; yields {ty}
5006 The '``atomicrmw``' instruction is used to atomically modify memory.
5011 There are three arguments to the '``atomicrmw``' instruction: an
5012 operation to apply, an address whose value to modify, an argument to the
5013 operation. The operation must be one of the following keywords:
5027 The type of '<value>' must be an integer type whose bit width is a power
5028 of two greater than or equal to eight and less than or equal to a
5029 target-specific size limit. The type of the '``<pointer>``' operand must
5030 be a pointer to that type. If the ``atomicrmw`` is marked as
5031 ``volatile``, then the optimizer is not allowed to modify the number or
5032 order of execution of this ``atomicrmw`` with other :ref:`volatile
5033 operations <volatile>`.
5038 The contents of memory at the location specified by the '``<pointer>``'
5039 operand are atomically read, modified, and written back. The original
5040 value at the location is returned. The modification is specified by the
5043 - xchg: ``*ptr = val``
5044 - add: ``*ptr = *ptr + val``
5045 - sub: ``*ptr = *ptr - val``
5046 - and: ``*ptr = *ptr & val``
5047 - nand: ``*ptr = ~(*ptr & val)``
5048 - or: ``*ptr = *ptr | val``
5049 - xor: ``*ptr = *ptr ^ val``
5050 - max: ``*ptr = *ptr > val ? *ptr : val`` (using a signed comparison)
5051 - min: ``*ptr = *ptr < val ? *ptr : val`` (using a signed comparison)
5052 - umax: ``*ptr = *ptr > val ? *ptr : val`` (using an unsigned
5054 - umin: ``*ptr = *ptr < val ? *ptr : val`` (using an unsigned
5060 .. code-block:: llvm
5062 %old = atomicrmw add i32* %ptr, i32 1 acquire ; yields {i32}
5064 .. _i_getelementptr:
5066 '``getelementptr``' Instruction
5067 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5074 <result> = getelementptr <pty>* <ptrval>{, <ty> <idx>}*
5075 <result> = getelementptr inbounds <pty>* <ptrval>{, <ty> <idx>}*
5076 <result> = getelementptr <ptr vector> ptrval, <vector index type> idx
5081 The '``getelementptr``' instruction is used to get the address of a
5082 subelement of an :ref:`aggregate <t_aggregate>` data structure. It performs
5083 address calculation only and does not access memory.
5088 The first argument is always a pointer or a vector of pointers, and
5089 forms the basis of the calculation. The remaining arguments are indices
5090 that indicate which of the elements of the aggregate object are indexed.
5091 The interpretation of each index is dependent on the type being indexed
5092 into. The first index always indexes the pointer value given as the
5093 first argument, the second index indexes a value of the type pointed to
5094 (not necessarily the value directly pointed to, since the first index
5095 can be non-zero), etc. The first type indexed into must be a pointer
5096 value, subsequent types can be arrays, vectors, and structs. Note that
5097 subsequent types being indexed into can never be pointers, since that
5098 would require loading the pointer before continuing calculation.
5100 The type of each index argument depends on the type it is indexing into.
5101 When indexing into a (optionally packed) structure, only ``i32`` integer
5102 **constants** are allowed (when using a vector of indices they must all
5103 be the **same** ``i32`` integer constant). When indexing into an array,
5104 pointer or vector, integers of any width are allowed, and they are not
5105 required to be constant. These integers are treated as signed values
5108 For example, let's consider a C code fragment and how it gets compiled
5124 int *foo(struct ST *s) {
5125 return &s[1].Z.B[5][13];
5128 The LLVM code generated by Clang is:
5130 .. code-block:: llvm
5132 %struct.RT = type { i8, [10 x [20 x i32]], i8 }
5133 %struct.ST = type { i32, double, %struct.RT }
5135 define i32* @foo(%struct.ST* %s) nounwind uwtable readnone optsize ssp {
5137 %arrayidx = getelementptr inbounds %struct.ST* %s, i64 1, i32 2, i32 1, i64 5, i64 13
5144 In the example above, the first index is indexing into the
5145 '``%struct.ST*``' type, which is a pointer, yielding a '``%struct.ST``'
5146 = '``{ i32, double, %struct.RT }``' type, a structure. The second index
5147 indexes into the third element of the structure, yielding a
5148 '``%struct.RT``' = '``{ i8 , [10 x [20 x i32]], i8 }``' type, another
5149 structure. The third index indexes into the second element of the
5150 structure, yielding a '``[10 x [20 x i32]]``' type, an array. The two
5151 dimensions of the array are subscripted into, yielding an '``i32``'
5152 type. The '``getelementptr``' instruction returns a pointer to this
5153 element, thus computing a value of '``i32*``' type.
5155 Note that it is perfectly legal to index partially through a structure,
5156 returning a pointer to an inner element. Because of this, the LLVM code
5157 for the given testcase is equivalent to:
5159 .. code-block:: llvm
5161 define i32* @foo(%struct.ST* %s) {
5162 %t1 = getelementptr %struct.ST* %s, i32 1 ; yields %struct.ST*:%t1
5163 %t2 = getelementptr %struct.ST* %t1, i32 0, i32 2 ; yields %struct.RT*:%t2
5164 %t3 = getelementptr %struct.RT* %t2, i32 0, i32 1 ; yields [10 x [20 x i32]]*:%t3
5165 %t4 = getelementptr [10 x [20 x i32]]* %t3, i32 0, i32 5 ; yields [20 x i32]*:%t4
5166 %t5 = getelementptr [20 x i32]* %t4, i32 0, i32 13 ; yields i32*:%t5
5170 If the ``inbounds`` keyword is present, the result value of the
5171 ``getelementptr`` is a :ref:`poison value <poisonvalues>` if the base
5172 pointer is not an *in bounds* address of an allocated object, or if any
5173 of the addresses that would be formed by successive addition of the
5174 offsets implied by the indices to the base address with infinitely
5175 precise signed arithmetic are not an *in bounds* address of that
5176 allocated object. The *in bounds* addresses for an allocated object are
5177 all the addresses that point into the object, plus the address one byte
5178 past the end. In cases where the base is a vector of pointers the
5179 ``inbounds`` keyword applies to each of the computations element-wise.
5181 If the ``inbounds`` keyword is not present, the offsets are added to the
5182 base address with silently-wrapping two's complement arithmetic. If the
5183 offsets have a different width from the pointer, they are sign-extended
5184 or truncated to the width of the pointer. The result value of the
5185 ``getelementptr`` may be outside the object pointed to by the base
5186 pointer. The result value may not necessarily be used to access memory
5187 though, even if it happens to point into allocated storage. See the
5188 :ref:`Pointer Aliasing Rules <pointeraliasing>` section for more
5191 The getelementptr instruction is often confusing. For some more insight
5192 into how it works, see :doc:`the getelementptr FAQ <GetElementPtr>`.
5197 .. code-block:: llvm
5199 ; yields [12 x i8]*:aptr
5200 %aptr = getelementptr {i32, [12 x i8]}* %saptr, i64 0, i32 1
5202 %vptr = getelementptr {i32, <2 x i8>}* %svptr, i64 0, i32 1, i32 1
5204 %eptr = getelementptr [12 x i8]* %aptr, i64 0, i32 1
5206 %iptr = getelementptr [10 x i32]* @arr, i16 0, i16 0
5208 In cases where the pointer argument is a vector of pointers, each index
5209 must be a vector with the same number of elements. For example:
5211 .. code-block:: llvm
5213 %A = getelementptr <4 x i8*> %ptrs, <4 x i64> %offsets,
5215 Conversion Operations
5216 ---------------------
5218 The instructions in this category are the conversion instructions
5219 (casting) which all take a single operand and a type. They perform
5220 various bit conversions on the operand.
5222 '``trunc .. to``' Instruction
5223 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5230 <result> = trunc <ty> <value> to <ty2> ; yields ty2
5235 The '``trunc``' instruction truncates its operand to the type ``ty2``.
5240 The '``trunc``' instruction takes a value to trunc, and a type to trunc
5241 it to. Both types must be of :ref:`integer <t_integer>` types, or vectors
5242 of the same number of integers. The bit size of the ``value`` must be
5243 larger than the bit size of the destination type, ``ty2``. Equal sized
5244 types are not allowed.
5249 The '``trunc``' instruction truncates the high order bits in ``value``
5250 and converts the remaining bits to ``ty2``. Since the source size must
5251 be larger than the destination size, ``trunc`` cannot be a *no-op cast*.
5252 It will always truncate bits.
5257 .. code-block:: llvm
5259 %X = trunc i32 257 to i8 ; yields i8:1
5260 %Y = trunc i32 123 to i1 ; yields i1:true
5261 %Z = trunc i32 122 to i1 ; yields i1:false
5262 %W = trunc <2 x i16> <i16 8, i16 7> to <2 x i8> ; yields <i8 8, i8 7>
5264 '``zext .. to``' Instruction
5265 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5272 <result> = zext <ty> <value> to <ty2> ; yields ty2
5277 The '``zext``' instruction zero extends its operand to type ``ty2``.
5282 The '``zext``' instruction takes a value to cast, and a type to cast it
5283 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
5284 the same number of integers. The bit size of the ``value`` must be
5285 smaller than the bit size of the destination type, ``ty2``.
5290 The ``zext`` fills the high order bits of the ``value`` with zero bits
5291 until it reaches the size of the destination type, ``ty2``.
5293 When zero extending from i1, the result will always be either 0 or 1.
5298 .. code-block:: llvm
5300 %X = zext i32 257 to i64 ; yields i64:257
5301 %Y = zext i1 true to i32 ; yields i32:1
5302 %Z = zext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
5304 '``sext .. to``' Instruction
5305 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5312 <result> = sext <ty> <value> to <ty2> ; yields ty2
5317 The '``sext``' sign extends ``value`` to the type ``ty2``.
5322 The '``sext``' instruction takes a value to cast, and a type to cast it
5323 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
5324 the same number of integers. The bit size of the ``value`` must be
5325 smaller than the bit size of the destination type, ``ty2``.
5330 The '``sext``' instruction performs a sign extension by copying the sign
5331 bit (highest order bit) of the ``value`` until it reaches the bit size
5332 of the type ``ty2``.
5334 When sign extending from i1, the extension always results in -1 or 0.
5339 .. code-block:: llvm
5341 %X = sext i8 -1 to i16 ; yields i16 :65535
5342 %Y = sext i1 true to i32 ; yields i32:-1
5343 %Z = sext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
5345 '``fptrunc .. to``' Instruction
5346 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5353 <result> = fptrunc <ty> <value> to <ty2> ; yields ty2
5358 The '``fptrunc``' instruction truncates ``value`` to type ``ty2``.
5363 The '``fptrunc``' instruction takes a :ref:`floating point <t_floating>`
5364 value to cast and a :ref:`floating point <t_floating>` type to cast it to.
5365 The size of ``value`` must be larger than the size of ``ty2``. This
5366 implies that ``fptrunc`` cannot be used to make a *no-op cast*.
5371 The '``fptrunc``' instruction truncates a ``value`` from a larger
5372 :ref:`floating point <t_floating>` type to a smaller :ref:`floating
5373 point <t_floating>` type. If the value cannot fit within the
5374 destination type, ``ty2``, then the results are undefined.
5379 .. code-block:: llvm
5381 %X = fptrunc double 123.0 to float ; yields float:123.0
5382 %Y = fptrunc double 1.0E+300 to float ; yields undefined
5384 '``fpext .. to``' Instruction
5385 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5392 <result> = fpext <ty> <value> to <ty2> ; yields ty2
5397 The '``fpext``' extends a floating point ``value`` to a larger floating
5403 The '``fpext``' instruction takes a :ref:`floating point <t_floating>`
5404 ``value`` to cast, and a :ref:`floating point <t_floating>` type to cast it
5405 to. The source type must be smaller than the destination type.
5410 The '``fpext``' instruction extends the ``value`` from a smaller
5411 :ref:`floating point <t_floating>` type to a larger :ref:`floating
5412 point <t_floating>` type. The ``fpext`` cannot be used to make a
5413 *no-op cast* because it always changes bits. Use ``bitcast`` to make a
5414 *no-op cast* for a floating point cast.
5419 .. code-block:: llvm
5421 %X = fpext float 3.125 to double ; yields double:3.125000e+00
5422 %Y = fpext double %X to fp128 ; yields fp128:0xL00000000000000004000900000000000
5424 '``fptoui .. to``' Instruction
5425 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5432 <result> = fptoui <ty> <value> to <ty2> ; yields ty2
5437 The '``fptoui``' converts a floating point ``value`` to its unsigned
5438 integer equivalent of type ``ty2``.
5443 The '``fptoui``' instruction takes a value to cast, which must be a
5444 scalar or vector :ref:`floating point <t_floating>` value, and a type to
5445 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
5446 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
5447 type with the same number of elements as ``ty``
5452 The '``fptoui``' instruction converts its :ref:`floating
5453 point <t_floating>` operand into the nearest (rounding towards zero)
5454 unsigned integer value. If the value cannot fit in ``ty2``, the results
5460 .. code-block:: llvm
5462 %X = fptoui double 123.0 to i32 ; yields i32:123
5463 %Y = fptoui float 1.0E+300 to i1 ; yields undefined:1
5464 %Z = fptoui float 1.04E+17 to i8 ; yields undefined:1
5466 '``fptosi .. to``' Instruction
5467 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5474 <result> = fptosi <ty> <value> to <ty2> ; yields ty2
5479 The '``fptosi``' instruction converts :ref:`floating point <t_floating>`
5480 ``value`` to type ``ty2``.
5485 The '``fptosi``' instruction takes a value to cast, which must be a
5486 scalar or vector :ref:`floating point <t_floating>` value, and a type to
5487 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
5488 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
5489 type with the same number of elements as ``ty``
5494 The '``fptosi``' instruction converts its :ref:`floating
5495 point <t_floating>` operand into the nearest (rounding towards zero)
5496 signed integer value. If the value cannot fit in ``ty2``, the results
5502 .. code-block:: llvm
5504 %X = fptosi double -123.0 to i32 ; yields i32:-123
5505 %Y = fptosi float 1.0E-247 to i1 ; yields undefined:1
5506 %Z = fptosi float 1.04E+17 to i8 ; yields undefined:1
5508 '``uitofp .. to``' Instruction
5509 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5516 <result> = uitofp <ty> <value> to <ty2> ; yields ty2
5521 The '``uitofp``' instruction regards ``value`` as an unsigned integer
5522 and converts that value to the ``ty2`` type.
5527 The '``uitofp``' instruction takes a value to cast, which must be a
5528 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
5529 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
5530 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
5531 type with the same number of elements as ``ty``
5536 The '``uitofp``' instruction interprets its operand as an unsigned
5537 integer quantity and converts it to the corresponding floating point
5538 value. If the value cannot fit in the floating point value, the results
5544 .. code-block:: llvm
5546 %X = uitofp i32 257 to float ; yields float:257.0
5547 %Y = uitofp i8 -1 to double ; yields double:255.0
5549 '``sitofp .. to``' Instruction
5550 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5557 <result> = sitofp <ty> <value> to <ty2> ; yields ty2
5562 The '``sitofp``' instruction regards ``value`` as a signed integer and
5563 converts that value to the ``ty2`` type.
5568 The '``sitofp``' instruction takes a value to cast, which must be a
5569 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
5570 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
5571 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
5572 type with the same number of elements as ``ty``
5577 The '``sitofp``' instruction interprets its operand as a signed integer
5578 quantity and converts it to the corresponding floating point value. If
5579 the value cannot fit in the floating point value, the results are
5585 .. code-block:: llvm
5587 %X = sitofp i32 257 to float ; yields float:257.0
5588 %Y = sitofp i8 -1 to double ; yields double:-1.0
5592 '``ptrtoint .. to``' Instruction
5593 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5600 <result> = ptrtoint <ty> <value> to <ty2> ; yields ty2
5605 The '``ptrtoint``' instruction converts the pointer or a vector of
5606 pointers ``value`` to the integer (or vector of integers) type ``ty2``.
5611 The '``ptrtoint``' instruction takes a ``value`` to cast, which must be
5612 a a value of type :ref:`pointer <t_pointer>` or a vector of pointers, and a
5613 type to cast it to ``ty2``, which must be an :ref:`integer <t_integer>` or
5614 a vector of integers type.
5619 The '``ptrtoint``' instruction converts ``value`` to integer type
5620 ``ty2`` by interpreting the pointer value as an integer and either
5621 truncating or zero extending that value to the size of the integer type.
5622 If ``value`` is smaller than ``ty2`` then a zero extension is done. If
5623 ``value`` is larger than ``ty2`` then a truncation is done. If they are
5624 the same size, then nothing is done (*no-op cast*) other than a type
5630 .. code-block:: llvm
5632 %X = ptrtoint i32* %P to i8 ; yields truncation on 32-bit architecture
5633 %Y = ptrtoint i32* %P to i64 ; yields zero extension on 32-bit architecture
5634 %Z = ptrtoint <4 x i32*> %P to <4 x i64>; yields vector zero extension for a vector of addresses on 32-bit architecture
5638 '``inttoptr .. to``' Instruction
5639 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5646 <result> = inttoptr <ty> <value> to <ty2> ; yields ty2
5651 The '``inttoptr``' instruction converts an integer ``value`` to a
5652 pointer type, ``ty2``.
5657 The '``inttoptr``' instruction takes an :ref:`integer <t_integer>` value to
5658 cast, and a type to cast it to, which must be a :ref:`pointer <t_pointer>`
5664 The '``inttoptr``' instruction converts ``value`` to type ``ty2`` by
5665 applying either a zero extension or a truncation depending on the size
5666 of the integer ``value``. If ``value`` is larger than the size of a
5667 pointer then a truncation is done. If ``value`` is smaller than the size
5668 of a pointer then a zero extension is done. If they are the same size,
5669 nothing is done (*no-op cast*).
5674 .. code-block:: llvm
5676 %X = inttoptr i32 255 to i32* ; yields zero extension on 64-bit architecture
5677 %Y = inttoptr i32 255 to i32* ; yields no-op on 32-bit architecture
5678 %Z = inttoptr i64 0 to i32* ; yields truncation on 32-bit architecture
5679 %Z = inttoptr <4 x i32> %G to <4 x i8*>; yields truncation of vector G to four pointers
5683 '``bitcast .. to``' Instruction
5684 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5691 <result> = bitcast <ty> <value> to <ty2> ; yields ty2
5696 The '``bitcast``' instruction converts ``value`` to type ``ty2`` without
5702 The '``bitcast``' instruction takes a value to cast, which must be a
5703 non-aggregate first class value, and a type to cast it to, which must
5704 also be a non-aggregate :ref:`first class <t_firstclass>` type. The
5705 bit sizes of ``value`` and the destination type, ``ty2``, must be
5706 identical. If the source type is a pointer, the destination type must
5707 also be a pointer of the same size. This instruction supports bitwise
5708 conversion of vectors to integers and to vectors of other types (as
5709 long as they have the same size).
5714 The '``bitcast``' instruction converts ``value`` to type ``ty2``. It
5715 is always a *no-op cast* because no bits change with this
5716 conversion. The conversion is done as if the ``value`` had been stored
5717 to memory and read back as type ``ty2``. Pointer (or vector of
5718 pointers) types may only be converted to other pointer (or vector of
5719 pointers) types with the same address space through this instruction.
5720 To convert pointers to other types, use the :ref:`inttoptr <i_inttoptr>`
5721 or :ref:`ptrtoint <i_ptrtoint>` instructions first.
5726 .. code-block:: llvm
5728 %X = bitcast i8 255 to i8 ; yields i8 :-1
5729 %Y = bitcast i32* %x to sint* ; yields sint*:%x
5730 %Z = bitcast <2 x int> %V to i64; ; yields i64: %V
5731 %Z = bitcast <2 x i32*> %V to <2 x i64*> ; yields <2 x i64*>
5733 .. _i_addrspacecast:
5735 '``addrspacecast .. to``' Instruction
5736 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5743 <result> = addrspacecast <pty> <ptrval> to <pty2> ; yields pty2
5748 The '``addrspacecast``' instruction converts ``ptrval`` from ``pty`` in
5749 address space ``n`` to type ``pty2`` in address space ``m``.
5754 The '``addrspacecast``' instruction takes a pointer or vector of pointer value
5755 to cast and a pointer type to cast it to, which must have a different
5761 The '``addrspacecast``' instruction converts the pointer value
5762 ``ptrval`` to type ``pty2``. It can be a *no-op cast* or a complex
5763 value modification, depending on the target and the address space
5764 pair. Pointer conversions within the same address space must be
5765 performed with the ``bitcast`` instruction. Note that if the address space
5766 conversion is legal then both result and operand refer to the same memory
5772 .. code-block:: llvm
5774 %X = addrspacecast i32* %x to i32 addrspace(1)* ; yields i32 addrspace(1)*:%x
5775 %Y = addrspacecast i32 addrspace(1)* %y to i64 addrspace(2)* ; yields i64 addrspace(2)*:%y
5776 %Z = addrspacecast <4 x i32*> %z to <4 x float addrspace(3)*> ; yields <4 x float addrspace(3)*>:%z
5783 The instructions in this category are the "miscellaneous" instructions,
5784 which defy better classification.
5788 '``icmp``' Instruction
5789 ^^^^^^^^^^^^^^^^^^^^^^
5796 <result> = icmp <cond> <ty> <op1>, <op2> ; yields {i1} or {<N x i1>}:result
5801 The '``icmp``' instruction returns a boolean value or a vector of
5802 boolean values based on comparison of its two integer, integer vector,
5803 pointer, or pointer vector operands.
5808 The '``icmp``' instruction takes three operands. The first operand is
5809 the condition code indicating the kind of comparison to perform. It is
5810 not a value, just a keyword. The possible condition code are:
5813 #. ``ne``: not equal
5814 #. ``ugt``: unsigned greater than
5815 #. ``uge``: unsigned greater or equal
5816 #. ``ult``: unsigned less than
5817 #. ``ule``: unsigned less or equal
5818 #. ``sgt``: signed greater than
5819 #. ``sge``: signed greater or equal
5820 #. ``slt``: signed less than
5821 #. ``sle``: signed less or equal
5823 The remaining two arguments must be :ref:`integer <t_integer>` or
5824 :ref:`pointer <t_pointer>` or integer :ref:`vector <t_vector>` typed. They
5825 must also be identical types.
5830 The '``icmp``' compares ``op1`` and ``op2`` according to the condition
5831 code given as ``cond``. The comparison performed always yields either an
5832 :ref:`i1 <t_integer>` or vector of ``i1`` result, as follows:
5834 #. ``eq``: yields ``true`` if the operands are equal, ``false``
5835 otherwise. No sign interpretation is necessary or performed.
5836 #. ``ne``: yields ``true`` if the operands are unequal, ``false``
5837 otherwise. No sign interpretation is necessary or performed.
5838 #. ``ugt``: interprets the operands as unsigned values and yields
5839 ``true`` if ``op1`` is greater than ``op2``.
5840 #. ``uge``: interprets the operands as unsigned values and yields
5841 ``true`` if ``op1`` is greater than or equal to ``op2``.
5842 #. ``ult``: interprets the operands as unsigned values and yields
5843 ``true`` if ``op1`` is less than ``op2``.
5844 #. ``ule``: interprets the operands as unsigned values and yields
5845 ``true`` if ``op1`` is less than or equal to ``op2``.
5846 #. ``sgt``: interprets the operands as signed values and yields ``true``
5847 if ``op1`` is greater than ``op2``.
5848 #. ``sge``: interprets the operands as signed values and yields ``true``
5849 if ``op1`` is greater than or equal to ``op2``.
5850 #. ``slt``: interprets the operands as signed values and yields ``true``
5851 if ``op1`` is less than ``op2``.
5852 #. ``sle``: interprets the operands as signed values and yields ``true``
5853 if ``op1`` is less than or equal to ``op2``.
5855 If the operands are :ref:`pointer <t_pointer>` typed, the pointer values
5856 are compared as if they were integers.
5858 If the operands are integer vectors, then they are compared element by
5859 element. The result is an ``i1`` vector with the same number of elements
5860 as the values being compared. Otherwise, the result is an ``i1``.
5865 .. code-block:: llvm
5867 <result> = icmp eq i32 4, 5 ; yields: result=false
5868 <result> = icmp ne float* %X, %X ; yields: result=false
5869 <result> = icmp ult i16 4, 5 ; yields: result=true
5870 <result> = icmp sgt i16 4, 5 ; yields: result=false
5871 <result> = icmp ule i16 -4, 5 ; yields: result=false
5872 <result> = icmp sge i16 4, 5 ; yields: result=false
5874 Note that the code generator does not yet support vector types with the
5875 ``icmp`` instruction.
5879 '``fcmp``' Instruction
5880 ^^^^^^^^^^^^^^^^^^^^^^
5887 <result> = fcmp <cond> <ty> <op1>, <op2> ; yields {i1} or {<N x i1>}:result
5892 The '``fcmp``' instruction returns a boolean value or vector of boolean
5893 values based on comparison of its operands.
5895 If the operands are floating point scalars, then the result type is a
5896 boolean (:ref:`i1 <t_integer>`).
5898 If the operands are floating point vectors, then the result type is a
5899 vector of boolean with the same number of elements as the operands being
5905 The '``fcmp``' instruction takes three operands. The first operand is
5906 the condition code indicating the kind of comparison to perform. It is
5907 not a value, just a keyword. The possible condition code are:
5909 #. ``false``: no comparison, always returns false
5910 #. ``oeq``: ordered and equal
5911 #. ``ogt``: ordered and greater than
5912 #. ``oge``: ordered and greater than or equal
5913 #. ``olt``: ordered and less than
5914 #. ``ole``: ordered and less than or equal
5915 #. ``one``: ordered and not equal
5916 #. ``ord``: ordered (no nans)
5917 #. ``ueq``: unordered or equal
5918 #. ``ugt``: unordered or greater than
5919 #. ``uge``: unordered or greater than or equal
5920 #. ``ult``: unordered or less than
5921 #. ``ule``: unordered or less than or equal
5922 #. ``une``: unordered or not equal
5923 #. ``uno``: unordered (either nans)
5924 #. ``true``: no comparison, always returns true
5926 *Ordered* means that neither operand is a QNAN while *unordered* means
5927 that either operand may be a QNAN.
5929 Each of ``val1`` and ``val2`` arguments must be either a :ref:`floating
5930 point <t_floating>` type or a :ref:`vector <t_vector>` of floating point
5931 type. They must have identical types.
5936 The '``fcmp``' instruction compares ``op1`` and ``op2`` according to the
5937 condition code given as ``cond``. If the operands are vectors, then the
5938 vectors are compared element by element. Each comparison performed
5939 always yields an :ref:`i1 <t_integer>` result, as follows:
5941 #. ``false``: always yields ``false``, regardless of operands.
5942 #. ``oeq``: yields ``true`` if both operands are not a QNAN and ``op1``
5943 is equal to ``op2``.
5944 #. ``ogt``: yields ``true`` if both operands are not a QNAN and ``op1``
5945 is greater than ``op2``.
5946 #. ``oge``: yields ``true`` if both operands are not a QNAN and ``op1``
5947 is greater than or equal to ``op2``.
5948 #. ``olt``: yields ``true`` if both operands are not a QNAN and ``op1``
5949 is less than ``op2``.
5950 #. ``ole``: yields ``true`` if both operands are not a QNAN and ``op1``
5951 is less than or equal to ``op2``.
5952 #. ``one``: yields ``true`` if both operands are not a QNAN and ``op1``
5953 is not equal to ``op2``.
5954 #. ``ord``: yields ``true`` if both operands are not a QNAN.
5955 #. ``ueq``: yields ``true`` if either operand is a QNAN or ``op1`` is
5957 #. ``ugt``: yields ``true`` if either operand is a QNAN or ``op1`` is
5958 greater than ``op2``.
5959 #. ``uge``: yields ``true`` if either operand is a QNAN or ``op1`` is
5960 greater than or equal to ``op2``.
5961 #. ``ult``: yields ``true`` if either operand is a QNAN or ``op1`` is
5963 #. ``ule``: yields ``true`` if either operand is a QNAN or ``op1`` is
5964 less than or equal to ``op2``.
5965 #. ``une``: yields ``true`` if either operand is a QNAN or ``op1`` is
5966 not equal to ``op2``.
5967 #. ``uno``: yields ``true`` if either operand is a QNAN.
5968 #. ``true``: always yields ``true``, regardless of operands.
5973 .. code-block:: llvm
5975 <result> = fcmp oeq float 4.0, 5.0 ; yields: result=false
5976 <result> = fcmp one float 4.0, 5.0 ; yields: result=true
5977 <result> = fcmp olt float 4.0, 5.0 ; yields: result=true
5978 <result> = fcmp ueq double 1.0, 2.0 ; yields: result=false
5980 Note that the code generator does not yet support vector types with the
5981 ``fcmp`` instruction.
5985 '``phi``' Instruction
5986 ^^^^^^^^^^^^^^^^^^^^^
5993 <result> = phi <ty> [ <val0>, <label0>], ...
5998 The '``phi``' instruction is used to implement the φ node in the SSA
5999 graph representing the function.
6004 The type of the incoming values is specified with the first type field.
6005 After this, the '``phi``' instruction takes a list of pairs as
6006 arguments, with one pair for each predecessor basic block of the current
6007 block. Only values of :ref:`first class <t_firstclass>` type may be used as
6008 the value arguments to the PHI node. Only labels may be used as the
6011 There must be no non-phi instructions between the start of a basic block
6012 and the PHI instructions: i.e. PHI instructions must be first in a basic
6015 For the purposes of the SSA form, the use of each incoming value is
6016 deemed to occur on the edge from the corresponding predecessor block to
6017 the current block (but after any definition of an '``invoke``'
6018 instruction's return value on the same edge).
6023 At runtime, the '``phi``' instruction logically takes on the value
6024 specified by the pair corresponding to the predecessor basic block that
6025 executed just prior to the current block.
6030 .. code-block:: llvm
6032 Loop: ; Infinite loop that counts from 0 on up...
6033 %indvar = phi i32 [ 0, %LoopHeader ], [ %nextindvar, %Loop ]
6034 %nextindvar = add i32 %indvar, 1
6039 '``select``' Instruction
6040 ^^^^^^^^^^^^^^^^^^^^^^^^
6047 <result> = select selty <cond>, <ty> <val1>, <ty> <val2> ; yields ty
6049 selty is either i1 or {<N x i1>}
6054 The '``select``' instruction is used to choose one value based on a
6055 condition, without branching.
6060 The '``select``' instruction requires an 'i1' value or a vector of 'i1'
6061 values indicating the condition, and two values of the same :ref:`first
6062 class <t_firstclass>` type. If the val1/val2 are vectors and the
6063 condition is a scalar, then entire vectors are selected, not individual
6069 If the condition is an i1 and it evaluates to 1, the instruction returns
6070 the first value argument; otherwise, it returns the second value
6073 If the condition is a vector of i1, then the value arguments must be
6074 vectors of the same size, and the selection is done element by element.
6079 .. code-block:: llvm
6081 %X = select i1 true, i8 17, i8 42 ; yields i8:17
6085 '``call``' Instruction
6086 ^^^^^^^^^^^^^^^^^^^^^^
6093 <result> = [tail] call [cconv] [ret attrs] <ty> [<fnty>*] <fnptrval>(<function args>) [fn attrs]
6098 The '``call``' instruction represents a simple function call.
6103 This instruction requires several arguments:
6105 #. The optional "tail" marker indicates that the callee function does
6106 not access any allocas or varargs in the caller. Note that calls may
6107 be marked "tail" even if they do not occur before a
6108 :ref:`ret <i_ret>` instruction. If the "tail" marker is present, the
6109 function call is eligible for tail call optimization, but `might not
6110 in fact be optimized into a jump <CodeGenerator.html#tailcallopt>`_.
6111 The code generator may optimize calls marked "tail" with either 1)
6112 automatic `sibling call
6113 optimization <CodeGenerator.html#sibcallopt>`_ when the caller and
6114 callee have matching signatures, or 2) forced tail call optimization
6115 when the following extra requirements are met:
6117 - Caller and callee both have the calling convention ``fastcc``.
6118 - The call is in tail position (ret immediately follows call and ret
6119 uses value of call or is void).
6120 - Option ``-tailcallopt`` is enabled, or
6121 ``llvm::GuaranteedTailCallOpt`` is ``true``.
6122 - `Platform specific constraints are
6123 met. <CodeGenerator.html#tailcallopt>`_
6125 #. The optional "cconv" marker indicates which :ref:`calling
6126 convention <callingconv>` the call should use. If none is
6127 specified, the call defaults to using C calling conventions. The
6128 calling convention of the call must match the calling convention of
6129 the target function, or else the behavior is undefined.
6130 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
6131 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
6133 #. '``ty``': the type of the call instruction itself which is also the
6134 type of the return value. Functions that return no value are marked
6136 #. '``fnty``': shall be the signature of the pointer to function value
6137 being invoked. The argument types must match the types implied by
6138 this signature. This type can be omitted if the function is not
6139 varargs and if the function type does not return a pointer to a
6141 #. '``fnptrval``': An LLVM value containing a pointer to a function to
6142 be invoked. In most cases, this is a direct function invocation, but
6143 indirect ``call``'s are just as possible, calling an arbitrary pointer
6145 #. '``function args``': argument list whose types match the function
6146 signature argument types and parameter attributes. All arguments must
6147 be of :ref:`first class <t_firstclass>` type. If the function signature
6148 indicates the function accepts a variable number of arguments, the
6149 extra arguments can be specified.
6150 #. The optional :ref:`function attributes <fnattrs>` list. Only
6151 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
6152 attributes are valid here.
6157 The '``call``' instruction is used to cause control flow to transfer to
6158 a specified function, with its incoming arguments bound to the specified
6159 values. Upon a '``ret``' instruction in the called function, control
6160 flow continues with the instruction after the function call, and the
6161 return value of the function is bound to the result argument.
6166 .. code-block:: llvm
6168 %retval = call i32 @test(i32 %argc)
6169 call i32 (i8*, ...)* @printf(i8* %msg, i32 12, i8 42) ; yields i32
6170 %X = tail call i32 @foo() ; yields i32
6171 %Y = tail call fastcc i32 @foo() ; yields i32
6172 call void %foo(i8 97 signext)
6174 %struct.A = type { i32, i8 }
6175 %r = call %struct.A @foo() ; yields { 32, i8 }
6176 %gr = extractvalue %struct.A %r, 0 ; yields i32
6177 %gr1 = extractvalue %struct.A %r, 1 ; yields i8
6178 %Z = call void @foo() noreturn ; indicates that %foo never returns normally
6179 %ZZ = call zeroext i32 @bar() ; Return value is %zero extended
6181 llvm treats calls to some functions with names and arguments that match
6182 the standard C99 library as being the C99 library functions, and may
6183 perform optimizations or generate code for them under that assumption.
6184 This is something we'd like to change in the future to provide better
6185 support for freestanding environments and non-C-based languages.
6189 '``va_arg``' Instruction
6190 ^^^^^^^^^^^^^^^^^^^^^^^^
6197 <resultval> = va_arg <va_list*> <arglist>, <argty>
6202 The '``va_arg``' instruction is used to access arguments passed through
6203 the "variable argument" area of a function call. It is used to implement
6204 the ``va_arg`` macro in C.
6209 This instruction takes a ``va_list*`` value and the type of the
6210 argument. It returns a value of the specified argument type and
6211 increments the ``va_list`` to point to the next argument. The actual
6212 type of ``va_list`` is target specific.
6217 The '``va_arg``' instruction loads an argument of the specified type
6218 from the specified ``va_list`` and causes the ``va_list`` to point to
6219 the next argument. For more information, see the variable argument
6220 handling :ref:`Intrinsic Functions <int_varargs>`.
6222 It is legal for this instruction to be called in a function which does
6223 not take a variable number of arguments, for example, the ``vfprintf``
6226 ``va_arg`` is an LLVM instruction instead of an :ref:`intrinsic
6227 function <intrinsics>` because it takes a type as an argument.
6232 See the :ref:`variable argument processing <int_varargs>` section.
6234 Note that the code generator does not yet fully support va\_arg on many
6235 targets. Also, it does not currently support va\_arg with aggregate
6236 types on any target.
6240 '``landingpad``' Instruction
6241 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6248 <resultval> = landingpad <resultty> personality <type> <pers_fn> <clause>+
6249 <resultval> = landingpad <resultty> personality <type> <pers_fn> cleanup <clause>*
6251 <clause> := catch <type> <value>
6252 <clause> := filter <array constant type> <array constant>
6257 The '``landingpad``' instruction is used by `LLVM's exception handling
6258 system <ExceptionHandling.html#overview>`_ to specify that a basic block
6259 is a landing pad --- one where the exception lands, and corresponds to the
6260 code found in the ``catch`` portion of a ``try``/``catch`` sequence. It
6261 defines values supplied by the personality function (``pers_fn``) upon
6262 re-entry to the function. The ``resultval`` has the type ``resultty``.
6267 This instruction takes a ``pers_fn`` value. This is the personality
6268 function associated with the unwinding mechanism. The optional
6269 ``cleanup`` flag indicates that the landing pad block is a cleanup.
6271 A ``clause`` begins with the clause type --- ``catch`` or ``filter`` --- and
6272 contains the global variable representing the "type" that may be caught
6273 or filtered respectively. Unlike the ``catch`` clause, the ``filter``
6274 clause takes an array constant as its argument. Use
6275 "``[0 x i8**] undef``" for a filter which cannot throw. The
6276 '``landingpad``' instruction must contain *at least* one ``clause`` or
6277 the ``cleanup`` flag.
6282 The '``landingpad``' instruction defines the values which are set by the
6283 personality function (``pers_fn``) upon re-entry to the function, and
6284 therefore the "result type" of the ``landingpad`` instruction. As with
6285 calling conventions, how the personality function results are
6286 represented in LLVM IR is target specific.
6288 The clauses are applied in order from top to bottom. If two
6289 ``landingpad`` instructions are merged together through inlining, the
6290 clauses from the calling function are appended to the list of clauses.
6291 When the call stack is being unwound due to an exception being thrown,
6292 the exception is compared against each ``clause`` in turn. If it doesn't
6293 match any of the clauses, and the ``cleanup`` flag is not set, then
6294 unwinding continues further up the call stack.
6296 The ``landingpad`` instruction has several restrictions:
6298 - A landing pad block is a basic block which is the unwind destination
6299 of an '``invoke``' instruction.
6300 - A landing pad block must have a '``landingpad``' instruction as its
6301 first non-PHI instruction.
6302 - There can be only one '``landingpad``' instruction within the landing
6304 - A basic block that is not a landing pad block may not include a
6305 '``landingpad``' instruction.
6306 - All '``landingpad``' instructions in a function must have the same
6307 personality function.
6312 .. code-block:: llvm
6314 ;; A landing pad which can catch an integer.
6315 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6317 ;; A landing pad that is a cleanup.
6318 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6320 ;; A landing pad which can catch an integer and can only throw a double.
6321 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6323 filter [1 x i8**] [@_ZTId]
6330 LLVM supports the notion of an "intrinsic function". These functions
6331 have well known names and semantics and are required to follow certain
6332 restrictions. Overall, these intrinsics represent an extension mechanism
6333 for the LLVM language that does not require changing all of the
6334 transformations in LLVM when adding to the language (or the bitcode
6335 reader/writer, the parser, etc...).
6337 Intrinsic function names must all start with an "``llvm.``" prefix. This
6338 prefix is reserved in LLVM for intrinsic names; thus, function names may
6339 not begin with this prefix. Intrinsic functions must always be external
6340 functions: you cannot define the body of intrinsic functions. Intrinsic
6341 functions may only be used in call or invoke instructions: it is illegal
6342 to take the address of an intrinsic function. Additionally, because
6343 intrinsic functions are part of the LLVM language, it is required if any
6344 are added that they be documented here.
6346 Some intrinsic functions can be overloaded, i.e., the intrinsic
6347 represents a family of functions that perform the same operation but on
6348 different data types. Because LLVM can represent over 8 million
6349 different integer types, overloading is used commonly to allow an
6350 intrinsic function to operate on any integer type. One or more of the
6351 argument types or the result type can be overloaded to accept any
6352 integer type. Argument types may also be defined as exactly matching a
6353 previous argument's type or the result type. This allows an intrinsic
6354 function which accepts multiple arguments, but needs all of them to be
6355 of the same type, to only be overloaded with respect to a single
6356 argument or the result.
6358 Overloaded intrinsics will have the names of its overloaded argument
6359 types encoded into its function name, each preceded by a period. Only
6360 those types which are overloaded result in a name suffix. Arguments
6361 whose type is matched against another type do not. For example, the
6362 ``llvm.ctpop`` function can take an integer of any width and returns an
6363 integer of exactly the same integer width. This leads to a family of
6364 functions such as ``i8 @llvm.ctpop.i8(i8 %val)`` and
6365 ``i29 @llvm.ctpop.i29(i29 %val)``. Only one type, the return type, is
6366 overloaded, and only one type suffix is required. Because the argument's
6367 type is matched against the return type, it does not require its own
6370 To learn how to add an intrinsic function, please see the `Extending
6371 LLVM Guide <ExtendingLLVM.html>`_.
6375 Variable Argument Handling Intrinsics
6376 -------------------------------------
6378 Variable argument support is defined in LLVM with the
6379 :ref:`va_arg <i_va_arg>` instruction and these three intrinsic
6380 functions. These functions are related to the similarly named macros
6381 defined in the ``<stdarg.h>`` header file.
6383 All of these functions operate on arguments that use a target-specific
6384 value type "``va_list``". The LLVM assembly language reference manual
6385 does not define what this type is, so all transformations should be
6386 prepared to handle these functions regardless of the type used.
6388 This example shows how the :ref:`va_arg <i_va_arg>` instruction and the
6389 variable argument handling intrinsic functions are used.
6391 .. code-block:: llvm
6393 define i32 @test(i32 %X, ...) {
6394 ; Initialize variable argument processing
6396 %ap2 = bitcast i8** %ap to i8*
6397 call void @llvm.va_start(i8* %ap2)
6399 ; Read a single integer argument
6400 %tmp = va_arg i8** %ap, i32
6402 ; Demonstrate usage of llvm.va_copy and llvm.va_end
6404 %aq2 = bitcast i8** %aq to i8*
6405 call void @llvm.va_copy(i8* %aq2, i8* %ap2)
6406 call void @llvm.va_end(i8* %aq2)
6408 ; Stop processing of arguments.
6409 call void @llvm.va_end(i8* %ap2)
6413 declare void @llvm.va_start(i8*)
6414 declare void @llvm.va_copy(i8*, i8*)
6415 declare void @llvm.va_end(i8*)
6419 '``llvm.va_start``' Intrinsic
6420 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6427 declare void @llvm.va_start(i8* <arglist>)
6432 The '``llvm.va_start``' intrinsic initializes ``*<arglist>`` for
6433 subsequent use by ``va_arg``.
6438 The argument is a pointer to a ``va_list`` element to initialize.
6443 The '``llvm.va_start``' intrinsic works just like the ``va_start`` macro
6444 available in C. In a target-dependent way, it initializes the
6445 ``va_list`` element to which the argument points, so that the next call
6446 to ``va_arg`` will produce the first variable argument passed to the
6447 function. Unlike the C ``va_start`` macro, this intrinsic does not need
6448 to know the last argument of the function as the compiler can figure
6451 '``llvm.va_end``' Intrinsic
6452 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6459 declare void @llvm.va_end(i8* <arglist>)
6464 The '``llvm.va_end``' intrinsic destroys ``*<arglist>``, which has been
6465 initialized previously with ``llvm.va_start`` or ``llvm.va_copy``.
6470 The argument is a pointer to a ``va_list`` to destroy.
6475 The '``llvm.va_end``' intrinsic works just like the ``va_end`` macro
6476 available in C. In a target-dependent way, it destroys the ``va_list``
6477 element to which the argument points. Calls to
6478 :ref:`llvm.va_start <int_va_start>` and
6479 :ref:`llvm.va_copy <int_va_copy>` must be matched exactly with calls to
6484 '``llvm.va_copy``' Intrinsic
6485 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6492 declare void @llvm.va_copy(i8* <destarglist>, i8* <srcarglist>)
6497 The '``llvm.va_copy``' intrinsic copies the current argument position
6498 from the source argument list to the destination argument list.
6503 The first argument is a pointer to a ``va_list`` element to initialize.
6504 The second argument is a pointer to a ``va_list`` element to copy from.
6509 The '``llvm.va_copy``' intrinsic works just like the ``va_copy`` macro
6510 available in C. In a target-dependent way, it copies the source
6511 ``va_list`` element into the destination ``va_list`` element. This
6512 intrinsic is necessary because the `` llvm.va_start`` intrinsic may be
6513 arbitrarily complex and require, for example, memory allocation.
6515 Accurate Garbage Collection Intrinsics
6516 --------------------------------------
6518 LLVM support for `Accurate Garbage Collection <GarbageCollection.html>`_
6519 (GC) requires the implementation and generation of these intrinsics.
6520 These intrinsics allow identification of :ref:`GC roots on the
6521 stack <int_gcroot>`, as well as garbage collector implementations that
6522 require :ref:`read <int_gcread>` and :ref:`write <int_gcwrite>` barriers.
6523 Front-ends for type-safe garbage collected languages should generate
6524 these intrinsics to make use of the LLVM garbage collectors. For more
6525 details, see `Accurate Garbage Collection with
6526 LLVM <GarbageCollection.html>`_.
6528 The garbage collection intrinsics only operate on objects in the generic
6529 address space (address space zero).
6533 '``llvm.gcroot``' Intrinsic
6534 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6541 declare void @llvm.gcroot(i8** %ptrloc, i8* %metadata)
6546 The '``llvm.gcroot``' intrinsic declares the existence of a GC root to
6547 the code generator, and allows some metadata to be associated with it.
6552 The first argument specifies the address of a stack object that contains
6553 the root pointer. The second pointer (which must be either a constant or
6554 a global value address) contains the meta-data to be associated with the
6560 At runtime, a call to this intrinsic stores a null pointer into the
6561 "ptrloc" location. At compile-time, the code generator generates
6562 information to allow the runtime to find the pointer at GC safe points.
6563 The '``llvm.gcroot``' intrinsic may only be used in a function which
6564 :ref:`specifies a GC algorithm <gc>`.
6568 '``llvm.gcread``' Intrinsic
6569 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6576 declare i8* @llvm.gcread(i8* %ObjPtr, i8** %Ptr)
6581 The '``llvm.gcread``' intrinsic identifies reads of references from heap
6582 locations, allowing garbage collector implementations that require read
6588 The second argument is the address to read from, which should be an
6589 address allocated from the garbage collector. The first object is a
6590 pointer to the start of the referenced object, if needed by the language
6591 runtime (otherwise null).
6596 The '``llvm.gcread``' intrinsic has the same semantics as a load
6597 instruction, but may be replaced with substantially more complex code by
6598 the garbage collector runtime, as needed. The '``llvm.gcread``'
6599 intrinsic may only be used in a function which :ref:`specifies a GC
6604 '``llvm.gcwrite``' Intrinsic
6605 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6612 declare void @llvm.gcwrite(i8* %P1, i8* %Obj, i8** %P2)
6617 The '``llvm.gcwrite``' intrinsic identifies writes of references to heap
6618 locations, allowing garbage collector implementations that require write
6619 barriers (such as generational or reference counting collectors).
6624 The first argument is the reference to store, the second is the start of
6625 the object to store it to, and the third is the address of the field of
6626 Obj to store to. If the runtime does not require a pointer to the
6627 object, Obj may be null.
6632 The '``llvm.gcwrite``' intrinsic has the same semantics as a store
6633 instruction, but may be replaced with substantially more complex code by
6634 the garbage collector runtime, as needed. The '``llvm.gcwrite``'
6635 intrinsic may only be used in a function which :ref:`specifies a GC
6638 Code Generator Intrinsics
6639 -------------------------
6641 These intrinsics are provided by LLVM to expose special features that
6642 may only be implemented with code generator support.
6644 '``llvm.returnaddress``' Intrinsic
6645 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6652 declare i8 *@llvm.returnaddress(i32 <level>)
6657 The '``llvm.returnaddress``' intrinsic attempts to compute a
6658 target-specific value indicating the return address of the current
6659 function or one of its callers.
6664 The argument to this intrinsic indicates which function to return the
6665 address for. Zero indicates the calling function, one indicates its
6666 caller, etc. The argument is **required** to be a constant integer
6672 The '``llvm.returnaddress``' intrinsic either returns a pointer
6673 indicating the return address of the specified call frame, or zero if it
6674 cannot be identified. The value returned by this intrinsic is likely to
6675 be incorrect or 0 for arguments other than zero, so it should only be
6676 used for debugging purposes.
6678 Note that calling this intrinsic does not prevent function inlining or
6679 other aggressive transformations, so the value returned may not be that
6680 of the obvious source-language caller.
6682 '``llvm.frameaddress``' Intrinsic
6683 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6690 declare i8* @llvm.frameaddress(i32 <level>)
6695 The '``llvm.frameaddress``' intrinsic attempts to return the
6696 target-specific frame pointer value for the specified stack frame.
6701 The argument to this intrinsic indicates which function to return the
6702 frame pointer for. Zero indicates the calling function, one indicates
6703 its caller, etc. The argument is **required** to be a constant integer
6709 The '``llvm.frameaddress``' intrinsic either returns a pointer
6710 indicating the frame address of the specified call frame, or zero if it
6711 cannot be identified. The value returned by this intrinsic is likely to
6712 be incorrect or 0 for arguments other than zero, so it should only be
6713 used for debugging purposes.
6715 Note that calling this intrinsic does not prevent function inlining or
6716 other aggressive transformations, so the value returned may not be that
6717 of the obvious source-language caller.
6721 '``llvm.stacksave``' Intrinsic
6722 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6729 declare i8* @llvm.stacksave()
6734 The '``llvm.stacksave``' intrinsic is used to remember the current state
6735 of the function stack, for use with
6736 :ref:`llvm.stackrestore <int_stackrestore>`. This is useful for
6737 implementing language features like scoped automatic variable sized
6743 This intrinsic returns a opaque pointer value that can be passed to
6744 :ref:`llvm.stackrestore <int_stackrestore>`. When an
6745 ``llvm.stackrestore`` intrinsic is executed with a value saved from
6746 ``llvm.stacksave``, it effectively restores the state of the stack to
6747 the state it was in when the ``llvm.stacksave`` intrinsic executed. In
6748 practice, this pops any :ref:`alloca <i_alloca>` blocks from the stack that
6749 were allocated after the ``llvm.stacksave`` was executed.
6751 .. _int_stackrestore:
6753 '``llvm.stackrestore``' Intrinsic
6754 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6761 declare void @llvm.stackrestore(i8* %ptr)
6766 The '``llvm.stackrestore``' intrinsic is used to restore the state of
6767 the function stack to the state it was in when the corresponding
6768 :ref:`llvm.stacksave <int_stacksave>` intrinsic executed. This is
6769 useful for implementing language features like scoped automatic variable
6770 sized arrays in C99.
6775 See the description for :ref:`llvm.stacksave <int_stacksave>`.
6777 '``llvm.prefetch``' Intrinsic
6778 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6785 declare void @llvm.prefetch(i8* <address>, i32 <rw>, i32 <locality>, i32 <cache type>)
6790 The '``llvm.prefetch``' intrinsic is a hint to the code generator to
6791 insert a prefetch instruction if supported; otherwise, it is a noop.
6792 Prefetches have no effect on the behavior of the program but can change
6793 its performance characteristics.
6798 ``address`` is the address to be prefetched, ``rw`` is the specifier
6799 determining if the fetch should be for a read (0) or write (1), and
6800 ``locality`` is a temporal locality specifier ranging from (0) - no
6801 locality, to (3) - extremely local keep in cache. The ``cache type``
6802 specifies whether the prefetch is performed on the data (1) or
6803 instruction (0) cache. The ``rw``, ``locality`` and ``cache type``
6804 arguments must be constant integers.
6809 This intrinsic does not modify the behavior of the program. In
6810 particular, prefetches cannot trap and do not produce a value. On
6811 targets that support this intrinsic, the prefetch can provide hints to
6812 the processor cache for better performance.
6814 '``llvm.pcmarker``' Intrinsic
6815 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6822 declare void @llvm.pcmarker(i32 <id>)
6827 The '``llvm.pcmarker``' intrinsic is a method to export a Program
6828 Counter (PC) in a region of code to simulators and other tools. The
6829 method is target specific, but it is expected that the marker will use
6830 exported symbols to transmit the PC of the marker. The marker makes no
6831 guarantees that it will remain with any specific instruction after
6832 optimizations. It is possible that the presence of a marker will inhibit
6833 optimizations. The intended use is to be inserted after optimizations to
6834 allow correlations of simulation runs.
6839 ``id`` is a numerical id identifying the marker.
6844 This intrinsic does not modify the behavior of the program. Backends
6845 that do not support this intrinsic may ignore it.
6847 '``llvm.readcyclecounter``' Intrinsic
6848 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6855 declare i64 @llvm.readcyclecounter()
6860 The '``llvm.readcyclecounter``' intrinsic provides access to the cycle
6861 counter register (or similar low latency, high accuracy clocks) on those
6862 targets that support it. On X86, it should map to RDTSC. On Alpha, it
6863 should map to RPCC. As the backing counters overflow quickly (on the
6864 order of 9 seconds on alpha), this should only be used for small
6870 When directly supported, reading the cycle counter should not modify any
6871 memory. Implementations are allowed to either return a application
6872 specific value or a system wide value. On backends without support, this
6873 is lowered to a constant 0.
6875 Note that runtime support may be conditional on the privilege-level code is
6876 running at and the host platform.
6878 Standard C Library Intrinsics
6879 -----------------------------
6881 LLVM provides intrinsics for a few important standard C library
6882 functions. These intrinsics allow source-language front-ends to pass
6883 information about the alignment of the pointer arguments to the code
6884 generator, providing opportunity for more efficient code generation.
6888 '``llvm.memcpy``' Intrinsic
6889 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6894 This is an overloaded intrinsic. You can use ``llvm.memcpy`` on any
6895 integer bit width and for different address spaces. Not all targets
6896 support all bit widths however.
6900 declare void @llvm.memcpy.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
6901 i32 <len>, i32 <align>, i1 <isvolatile>)
6902 declare void @llvm.memcpy.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
6903 i64 <len>, i32 <align>, i1 <isvolatile>)
6908 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
6909 source location to the destination location.
6911 Note that, unlike the standard libc function, the ``llvm.memcpy.*``
6912 intrinsics do not return a value, takes extra alignment/isvolatile
6913 arguments and the pointers can be in specified address spaces.
6918 The first argument is a pointer to the destination, the second is a
6919 pointer to the source. The third argument is an integer argument
6920 specifying the number of bytes to copy, the fourth argument is the
6921 alignment of the source and destination locations, and the fifth is a
6922 boolean indicating a volatile access.
6924 If the call to this intrinsic has an alignment value that is not 0 or 1,
6925 then the caller guarantees that both the source and destination pointers
6926 are aligned to that boundary.
6928 If the ``isvolatile`` parameter is ``true``, the ``llvm.memcpy`` call is
6929 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
6930 very cleanly specified and it is unwise to depend on it.
6935 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
6936 source location to the destination location, which are not allowed to
6937 overlap. It copies "len" bytes of memory over. If the argument is known
6938 to be aligned to some boundary, this can be specified as the fourth
6939 argument, otherwise it should be set to 0 or 1 (both meaning no alignment).
6941 '``llvm.memmove``' Intrinsic
6942 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6947 This is an overloaded intrinsic. You can use llvm.memmove on any integer
6948 bit width and for different address space. Not all targets support all
6953 declare void @llvm.memmove.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
6954 i32 <len>, i32 <align>, i1 <isvolatile>)
6955 declare void @llvm.memmove.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
6956 i64 <len>, i32 <align>, i1 <isvolatile>)
6961 The '``llvm.memmove.*``' intrinsics move a block of memory from the
6962 source location to the destination location. It is similar to the
6963 '``llvm.memcpy``' intrinsic but allows the two memory locations to
6966 Note that, unlike the standard libc function, the ``llvm.memmove.*``
6967 intrinsics do not return a value, takes extra alignment/isvolatile
6968 arguments and the pointers can be in specified address spaces.
6973 The first argument is a pointer to the destination, the second is a
6974 pointer to the source. The third argument is an integer argument
6975 specifying the number of bytes to copy, the fourth argument is the
6976 alignment of the source and destination locations, and the fifth is a
6977 boolean indicating a volatile access.
6979 If the call to this intrinsic has an alignment value that is not 0 or 1,
6980 then the caller guarantees that the source and destination pointers are
6981 aligned to that boundary.
6983 If the ``isvolatile`` parameter is ``true``, the ``llvm.memmove`` call
6984 is a :ref:`volatile operation <volatile>`. The detailed access behavior is
6985 not very cleanly specified and it is unwise to depend on it.
6990 The '``llvm.memmove.*``' intrinsics copy a block of memory from the
6991 source location to the destination location, which may overlap. It
6992 copies "len" bytes of memory over. If the argument is known to be
6993 aligned to some boundary, this can be specified as the fourth argument,
6994 otherwise it should be set to 0 or 1 (both meaning no alignment).
6996 '``llvm.memset.*``' Intrinsics
6997 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7002 This is an overloaded intrinsic. You can use llvm.memset on any integer
7003 bit width and for different address spaces. However, not all targets
7004 support all bit widths.
7008 declare void @llvm.memset.p0i8.i32(i8* <dest>, i8 <val>,
7009 i32 <len>, i32 <align>, i1 <isvolatile>)
7010 declare void @llvm.memset.p0i8.i64(i8* <dest>, i8 <val>,
7011 i64 <len>, i32 <align>, i1 <isvolatile>)
7016 The '``llvm.memset.*``' intrinsics fill a block of memory with a
7017 particular byte value.
7019 Note that, unlike the standard libc function, the ``llvm.memset``
7020 intrinsic does not return a value and takes extra alignment/volatile
7021 arguments. Also, the destination can be in an arbitrary address space.
7026 The first argument is a pointer to the destination to fill, the second
7027 is the byte value with which to fill it, the third argument is an
7028 integer argument specifying the number of bytes to fill, and the fourth
7029 argument is the known alignment of the destination location.
7031 If the call to this intrinsic has an alignment value that is not 0 or 1,
7032 then the caller guarantees that the destination pointer is aligned to
7035 If the ``isvolatile`` parameter is ``true``, the ``llvm.memset`` call is
7036 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
7037 very cleanly specified and it is unwise to depend on it.
7042 The '``llvm.memset.*``' intrinsics fill "len" bytes of memory starting
7043 at the destination location. If the argument is known to be aligned to
7044 some boundary, this can be specified as the fourth argument, otherwise
7045 it should be set to 0 or 1 (both meaning no alignment).
7047 '``llvm.sqrt.*``' Intrinsic
7048 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7053 This is an overloaded intrinsic. You can use ``llvm.sqrt`` on any
7054 floating point or vector of floating point type. Not all targets support
7059 declare float @llvm.sqrt.f32(float %Val)
7060 declare double @llvm.sqrt.f64(double %Val)
7061 declare x86_fp80 @llvm.sqrt.f80(x86_fp80 %Val)
7062 declare fp128 @llvm.sqrt.f128(fp128 %Val)
7063 declare ppc_fp128 @llvm.sqrt.ppcf128(ppc_fp128 %Val)
7068 The '``llvm.sqrt``' intrinsics return the sqrt of the specified operand,
7069 returning the same value as the libm '``sqrt``' functions would. Unlike
7070 ``sqrt`` in libm, however, ``llvm.sqrt`` has undefined behavior for
7071 negative numbers other than -0.0 (which allows for better optimization,
7072 because there is no need to worry about errno being set).
7073 ``llvm.sqrt(-0.0)`` is defined to return -0.0 like IEEE sqrt.
7078 The argument and return value are floating point numbers of the same
7084 This function returns the sqrt of the specified operand if it is a
7085 nonnegative floating point number.
7087 '``llvm.powi.*``' Intrinsic
7088 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7093 This is an overloaded intrinsic. You can use ``llvm.powi`` on any
7094 floating point or vector of floating point type. Not all targets support
7099 declare float @llvm.powi.f32(float %Val, i32 %power)
7100 declare double @llvm.powi.f64(double %Val, i32 %power)
7101 declare x86_fp80 @llvm.powi.f80(x86_fp80 %Val, i32 %power)
7102 declare fp128 @llvm.powi.f128(fp128 %Val, i32 %power)
7103 declare ppc_fp128 @llvm.powi.ppcf128(ppc_fp128 %Val, i32 %power)
7108 The '``llvm.powi.*``' intrinsics return the first operand raised to the
7109 specified (positive or negative) power. The order of evaluation of
7110 multiplications is not defined. When a vector of floating point type is
7111 used, the second argument remains a scalar integer value.
7116 The second argument is an integer power, and the first is a value to
7117 raise to that power.
7122 This function returns the first value raised to the second power with an
7123 unspecified sequence of rounding operations.
7125 '``llvm.sin.*``' Intrinsic
7126 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7131 This is an overloaded intrinsic. You can use ``llvm.sin`` on any
7132 floating point or vector of floating point type. Not all targets support
7137 declare float @llvm.sin.f32(float %Val)
7138 declare double @llvm.sin.f64(double %Val)
7139 declare x86_fp80 @llvm.sin.f80(x86_fp80 %Val)
7140 declare fp128 @llvm.sin.f128(fp128 %Val)
7141 declare ppc_fp128 @llvm.sin.ppcf128(ppc_fp128 %Val)
7146 The '``llvm.sin.*``' intrinsics return the sine of the operand.
7151 The argument and return value are floating point numbers of the same
7157 This function returns the sine of the specified operand, returning the
7158 same values as the libm ``sin`` functions would, and handles error
7159 conditions in the same way.
7161 '``llvm.cos.*``' Intrinsic
7162 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7167 This is an overloaded intrinsic. You can use ``llvm.cos`` on any
7168 floating point or vector of floating point type. Not all targets support
7173 declare float @llvm.cos.f32(float %Val)
7174 declare double @llvm.cos.f64(double %Val)
7175 declare x86_fp80 @llvm.cos.f80(x86_fp80 %Val)
7176 declare fp128 @llvm.cos.f128(fp128 %Val)
7177 declare ppc_fp128 @llvm.cos.ppcf128(ppc_fp128 %Val)
7182 The '``llvm.cos.*``' intrinsics return the cosine of the operand.
7187 The argument and return value are floating point numbers of the same
7193 This function returns the cosine of the specified operand, returning the
7194 same values as the libm ``cos`` functions would, and handles error
7195 conditions in the same way.
7197 '``llvm.pow.*``' Intrinsic
7198 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7203 This is an overloaded intrinsic. You can use ``llvm.pow`` on any
7204 floating point or vector of floating point type. Not all targets support
7209 declare float @llvm.pow.f32(float %Val, float %Power)
7210 declare double @llvm.pow.f64(double %Val, double %Power)
7211 declare x86_fp80 @llvm.pow.f80(x86_fp80 %Val, x86_fp80 %Power)
7212 declare fp128 @llvm.pow.f128(fp128 %Val, fp128 %Power)
7213 declare ppc_fp128 @llvm.pow.ppcf128(ppc_fp128 %Val, ppc_fp128 Power)
7218 The '``llvm.pow.*``' intrinsics return the first operand raised to the
7219 specified (positive or negative) power.
7224 The second argument is a floating point power, and the first is a value
7225 to raise to that power.
7230 This function returns the first value raised to the second power,
7231 returning the same values as the libm ``pow`` functions would, and
7232 handles error conditions in the same way.
7234 '``llvm.exp.*``' Intrinsic
7235 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7240 This is an overloaded intrinsic. You can use ``llvm.exp`` on any
7241 floating point or vector of floating point type. Not all targets support
7246 declare float @llvm.exp.f32(float %Val)
7247 declare double @llvm.exp.f64(double %Val)
7248 declare x86_fp80 @llvm.exp.f80(x86_fp80 %Val)
7249 declare fp128 @llvm.exp.f128(fp128 %Val)
7250 declare ppc_fp128 @llvm.exp.ppcf128(ppc_fp128 %Val)
7255 The '``llvm.exp.*``' intrinsics perform the exp function.
7260 The argument and return value are floating point numbers of the same
7266 This function returns the same values as the libm ``exp`` functions
7267 would, and handles error conditions in the same way.
7269 '``llvm.exp2.*``' Intrinsic
7270 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7275 This is an overloaded intrinsic. You can use ``llvm.exp2`` on any
7276 floating point or vector of floating point type. Not all targets support
7281 declare float @llvm.exp2.f32(float %Val)
7282 declare double @llvm.exp2.f64(double %Val)
7283 declare x86_fp80 @llvm.exp2.f80(x86_fp80 %Val)
7284 declare fp128 @llvm.exp2.f128(fp128 %Val)
7285 declare ppc_fp128 @llvm.exp2.ppcf128(ppc_fp128 %Val)
7290 The '``llvm.exp2.*``' intrinsics perform the exp2 function.
7295 The argument and return value are floating point numbers of the same
7301 This function returns the same values as the libm ``exp2`` functions
7302 would, and handles error conditions in the same way.
7304 '``llvm.log.*``' Intrinsic
7305 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7310 This is an overloaded intrinsic. You can use ``llvm.log`` on any
7311 floating point or vector of floating point type. Not all targets support
7316 declare float @llvm.log.f32(float %Val)
7317 declare double @llvm.log.f64(double %Val)
7318 declare x86_fp80 @llvm.log.f80(x86_fp80 %Val)
7319 declare fp128 @llvm.log.f128(fp128 %Val)
7320 declare ppc_fp128 @llvm.log.ppcf128(ppc_fp128 %Val)
7325 The '``llvm.log.*``' intrinsics perform the log function.
7330 The argument and return value are floating point numbers of the same
7336 This function returns the same values as the libm ``log`` functions
7337 would, and handles error conditions in the same way.
7339 '``llvm.log10.*``' Intrinsic
7340 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7345 This is an overloaded intrinsic. You can use ``llvm.log10`` on any
7346 floating point or vector of floating point type. Not all targets support
7351 declare float @llvm.log10.f32(float %Val)
7352 declare double @llvm.log10.f64(double %Val)
7353 declare x86_fp80 @llvm.log10.f80(x86_fp80 %Val)
7354 declare fp128 @llvm.log10.f128(fp128 %Val)
7355 declare ppc_fp128 @llvm.log10.ppcf128(ppc_fp128 %Val)
7360 The '``llvm.log10.*``' intrinsics perform the log10 function.
7365 The argument and return value are floating point numbers of the same
7371 This function returns the same values as the libm ``log10`` functions
7372 would, and handles error conditions in the same way.
7374 '``llvm.log2.*``' Intrinsic
7375 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7380 This is an overloaded intrinsic. You can use ``llvm.log2`` on any
7381 floating point or vector of floating point type. Not all targets support
7386 declare float @llvm.log2.f32(float %Val)
7387 declare double @llvm.log2.f64(double %Val)
7388 declare x86_fp80 @llvm.log2.f80(x86_fp80 %Val)
7389 declare fp128 @llvm.log2.f128(fp128 %Val)
7390 declare ppc_fp128 @llvm.log2.ppcf128(ppc_fp128 %Val)
7395 The '``llvm.log2.*``' intrinsics perform the log2 function.
7400 The argument and return value are floating point numbers of the same
7406 This function returns the same values as the libm ``log2`` functions
7407 would, and handles error conditions in the same way.
7409 '``llvm.fma.*``' Intrinsic
7410 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7415 This is an overloaded intrinsic. You can use ``llvm.fma`` on any
7416 floating point or vector of floating point type. Not all targets support
7421 declare float @llvm.fma.f32(float %a, float %b, float %c)
7422 declare double @llvm.fma.f64(double %a, double %b, double %c)
7423 declare x86_fp80 @llvm.fma.f80(x86_fp80 %a, x86_fp80 %b, x86_fp80 %c)
7424 declare fp128 @llvm.fma.f128(fp128 %a, fp128 %b, fp128 %c)
7425 declare ppc_fp128 @llvm.fma.ppcf128(ppc_fp128 %a, ppc_fp128 %b, ppc_fp128 %c)
7430 The '``llvm.fma.*``' intrinsics perform the fused multiply-add
7436 The argument and return value are floating point numbers of the same
7442 This function returns the same values as the libm ``fma`` functions
7445 '``llvm.fabs.*``' Intrinsic
7446 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7451 This is an overloaded intrinsic. You can use ``llvm.fabs`` on any
7452 floating point or vector of floating point type. Not all targets support
7457 declare float @llvm.fabs.f32(float %Val)
7458 declare double @llvm.fabs.f64(double %Val)
7459 declare x86_fp80 @llvm.fabs.f80(x86_fp80 %Val)
7460 declare fp128 @llvm.fabs.f128(fp128 %Val)
7461 declare ppc_fp128 @llvm.fabs.ppcf128(ppc_fp128 %Val)
7466 The '``llvm.fabs.*``' intrinsics return the absolute value of the
7472 The argument and return value are floating point numbers of the same
7478 This function returns the same values as the libm ``fabs`` functions
7479 would, and handles error conditions in the same way.
7481 '``llvm.copysign.*``' Intrinsic
7482 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7487 This is an overloaded intrinsic. You can use ``llvm.copysign`` on any
7488 floating point or vector of floating point type. Not all targets support
7493 declare float @llvm.copysign.f32(float %Mag, float %Sgn)
7494 declare double @llvm.copysign.f64(double %Mag, double %Sgn)
7495 declare x86_fp80 @llvm.copysign.f80(x86_fp80 %Mag, x86_fp80 %Sgn)
7496 declare fp128 @llvm.copysign.f128(fp128 %Mag, fp128 %Sgn)
7497 declare ppc_fp128 @llvm.copysign.ppcf128(ppc_fp128 %Mag, ppc_fp128 %Sgn)
7502 The '``llvm.copysign.*``' intrinsics return a value with the magnitude of the
7503 first operand and the sign of the second operand.
7508 The arguments and return value are floating point numbers of the same
7514 This function returns the same values as the libm ``copysign``
7515 functions would, and handles error conditions in the same way.
7517 '``llvm.floor.*``' Intrinsic
7518 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7523 This is an overloaded intrinsic. You can use ``llvm.floor`` on any
7524 floating point or vector of floating point type. Not all targets support
7529 declare float @llvm.floor.f32(float %Val)
7530 declare double @llvm.floor.f64(double %Val)
7531 declare x86_fp80 @llvm.floor.f80(x86_fp80 %Val)
7532 declare fp128 @llvm.floor.f128(fp128 %Val)
7533 declare ppc_fp128 @llvm.floor.ppcf128(ppc_fp128 %Val)
7538 The '``llvm.floor.*``' intrinsics return the floor of the operand.
7543 The argument and return value are floating point numbers of the same
7549 This function returns the same values as the libm ``floor`` functions
7550 would, and handles error conditions in the same way.
7552 '``llvm.ceil.*``' Intrinsic
7553 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7558 This is an overloaded intrinsic. You can use ``llvm.ceil`` on any
7559 floating point or vector of floating point type. Not all targets support
7564 declare float @llvm.ceil.f32(float %Val)
7565 declare double @llvm.ceil.f64(double %Val)
7566 declare x86_fp80 @llvm.ceil.f80(x86_fp80 %Val)
7567 declare fp128 @llvm.ceil.f128(fp128 %Val)
7568 declare ppc_fp128 @llvm.ceil.ppcf128(ppc_fp128 %Val)
7573 The '``llvm.ceil.*``' intrinsics return the ceiling of the operand.
7578 The argument and return value are floating point numbers of the same
7584 This function returns the same values as the libm ``ceil`` functions
7585 would, and handles error conditions in the same way.
7587 '``llvm.trunc.*``' Intrinsic
7588 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7593 This is an overloaded intrinsic. You can use ``llvm.trunc`` on any
7594 floating point or vector of floating point type. Not all targets support
7599 declare float @llvm.trunc.f32(float %Val)
7600 declare double @llvm.trunc.f64(double %Val)
7601 declare x86_fp80 @llvm.trunc.f80(x86_fp80 %Val)
7602 declare fp128 @llvm.trunc.f128(fp128 %Val)
7603 declare ppc_fp128 @llvm.trunc.ppcf128(ppc_fp128 %Val)
7608 The '``llvm.trunc.*``' intrinsics returns the operand rounded to the
7609 nearest integer not larger in magnitude than the operand.
7614 The argument and return value are floating point numbers of the same
7620 This function returns the same values as the libm ``trunc`` functions
7621 would, and handles error conditions in the same way.
7623 '``llvm.rint.*``' Intrinsic
7624 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7629 This is an overloaded intrinsic. You can use ``llvm.rint`` on any
7630 floating point or vector of floating point type. Not all targets support
7635 declare float @llvm.rint.f32(float %Val)
7636 declare double @llvm.rint.f64(double %Val)
7637 declare x86_fp80 @llvm.rint.f80(x86_fp80 %Val)
7638 declare fp128 @llvm.rint.f128(fp128 %Val)
7639 declare ppc_fp128 @llvm.rint.ppcf128(ppc_fp128 %Val)
7644 The '``llvm.rint.*``' intrinsics returns the operand rounded to the
7645 nearest integer. It may raise an inexact floating-point exception if the
7646 operand isn't an integer.
7651 The argument and return value are floating point numbers of the same
7657 This function returns the same values as the libm ``rint`` functions
7658 would, and handles error conditions in the same way.
7660 '``llvm.nearbyint.*``' Intrinsic
7661 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7666 This is an overloaded intrinsic. You can use ``llvm.nearbyint`` on any
7667 floating point or vector of floating point type. Not all targets support
7672 declare float @llvm.nearbyint.f32(float %Val)
7673 declare double @llvm.nearbyint.f64(double %Val)
7674 declare x86_fp80 @llvm.nearbyint.f80(x86_fp80 %Val)
7675 declare fp128 @llvm.nearbyint.f128(fp128 %Val)
7676 declare ppc_fp128 @llvm.nearbyint.ppcf128(ppc_fp128 %Val)
7681 The '``llvm.nearbyint.*``' intrinsics returns the operand rounded to the
7687 The argument and return value are floating point numbers of the same
7693 This function returns the same values as the libm ``nearbyint``
7694 functions would, and handles error conditions in the same way.
7696 '``llvm.round.*``' Intrinsic
7697 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7702 This is an overloaded intrinsic. You can use ``llvm.round`` on any
7703 floating point or vector of floating point type. Not all targets support
7708 declare float @llvm.round.f32(float %Val)
7709 declare double @llvm.round.f64(double %Val)
7710 declare x86_fp80 @llvm.round.f80(x86_fp80 %Val)
7711 declare fp128 @llvm.round.f128(fp128 %Val)
7712 declare ppc_fp128 @llvm.round.ppcf128(ppc_fp128 %Val)
7717 The '``llvm.round.*``' intrinsics returns the operand rounded to the
7723 The argument and return value are floating point numbers of the same
7729 This function returns the same values as the libm ``round``
7730 functions would, and handles error conditions in the same way.
7732 Bit Manipulation Intrinsics
7733 ---------------------------
7735 LLVM provides intrinsics for a few important bit manipulation
7736 operations. These allow efficient code generation for some algorithms.
7738 '``llvm.bswap.*``' Intrinsics
7739 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7744 This is an overloaded intrinsic function. You can use bswap on any
7745 integer type that is an even number of bytes (i.e. BitWidth % 16 == 0).
7749 declare i16 @llvm.bswap.i16(i16 <id>)
7750 declare i32 @llvm.bswap.i32(i32 <id>)
7751 declare i64 @llvm.bswap.i64(i64 <id>)
7756 The '``llvm.bswap``' family of intrinsics is used to byte swap integer
7757 values with an even number of bytes (positive multiple of 16 bits).
7758 These are useful for performing operations on data that is not in the
7759 target's native byte order.
7764 The ``llvm.bswap.i16`` intrinsic returns an i16 value that has the high
7765 and low byte of the input i16 swapped. Similarly, the ``llvm.bswap.i32``
7766 intrinsic returns an i32 value that has the four bytes of the input i32
7767 swapped, so that if the input bytes are numbered 0, 1, 2, 3 then the
7768 returned i32 will have its bytes in 3, 2, 1, 0 order. The
7769 ``llvm.bswap.i48``, ``llvm.bswap.i64`` and other intrinsics extend this
7770 concept to additional even-byte lengths (6 bytes, 8 bytes and more,
7773 '``llvm.ctpop.*``' Intrinsic
7774 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7779 This is an overloaded intrinsic. You can use llvm.ctpop on any integer
7780 bit width, or on any vector with integer elements. Not all targets
7781 support all bit widths or vector types, however.
7785 declare i8 @llvm.ctpop.i8(i8 <src>)
7786 declare i16 @llvm.ctpop.i16(i16 <src>)
7787 declare i32 @llvm.ctpop.i32(i32 <src>)
7788 declare i64 @llvm.ctpop.i64(i64 <src>)
7789 declare i256 @llvm.ctpop.i256(i256 <src>)
7790 declare <2 x i32> @llvm.ctpop.v2i32(<2 x i32> <src>)
7795 The '``llvm.ctpop``' family of intrinsics counts the number of bits set
7801 The only argument is the value to be counted. The argument may be of any
7802 integer type, or a vector with integer elements. The return type must
7803 match the argument type.
7808 The '``llvm.ctpop``' intrinsic counts the 1's in a variable, or within
7809 each element of a vector.
7811 '``llvm.ctlz.*``' Intrinsic
7812 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7817 This is an overloaded intrinsic. You can use ``llvm.ctlz`` on any
7818 integer bit width, or any vector whose elements are integers. Not all
7819 targets support all bit widths or vector types, however.
7823 declare i8 @llvm.ctlz.i8 (i8 <src>, i1 <is_zero_undef>)
7824 declare i16 @llvm.ctlz.i16 (i16 <src>, i1 <is_zero_undef>)
7825 declare i32 @llvm.ctlz.i32 (i32 <src>, i1 <is_zero_undef>)
7826 declare i64 @llvm.ctlz.i64 (i64 <src>, i1 <is_zero_undef>)
7827 declare i256 @llvm.ctlz.i256(i256 <src>, i1 <is_zero_undef>)
7828 declase <2 x i32> @llvm.ctlz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
7833 The '``llvm.ctlz``' family of intrinsic functions counts the number of
7834 leading zeros in a variable.
7839 The first argument is the value to be counted. This argument may be of
7840 any integer type, or a vectory with integer element type. The return
7841 type must match the first argument type.
7843 The second argument must be a constant and is a flag to indicate whether
7844 the intrinsic should ensure that a zero as the first argument produces a
7845 defined result. Historically some architectures did not provide a
7846 defined result for zero values as efficiently, and many algorithms are
7847 now predicated on avoiding zero-value inputs.
7852 The '``llvm.ctlz``' intrinsic counts the leading (most significant)
7853 zeros in a variable, or within each element of the vector. If
7854 ``src == 0`` then the result is the size in bits of the type of ``src``
7855 if ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
7856 ``llvm.ctlz(i32 2) = 30``.
7858 '``llvm.cttz.*``' Intrinsic
7859 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7864 This is an overloaded intrinsic. You can use ``llvm.cttz`` on any
7865 integer bit width, or any vector of integer elements. Not all targets
7866 support all bit widths or vector types, however.
7870 declare i8 @llvm.cttz.i8 (i8 <src>, i1 <is_zero_undef>)
7871 declare i16 @llvm.cttz.i16 (i16 <src>, i1 <is_zero_undef>)
7872 declare i32 @llvm.cttz.i32 (i32 <src>, i1 <is_zero_undef>)
7873 declare i64 @llvm.cttz.i64 (i64 <src>, i1 <is_zero_undef>)
7874 declare i256 @llvm.cttz.i256(i256 <src>, i1 <is_zero_undef>)
7875 declase <2 x i32> @llvm.cttz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
7880 The '``llvm.cttz``' family of intrinsic functions counts the number of
7886 The first argument is the value to be counted. This argument may be of
7887 any integer type, or a vectory with integer element type. The return
7888 type must match the first argument type.
7890 The second argument must be a constant and is a flag to indicate whether
7891 the intrinsic should ensure that a zero as the first argument produces a
7892 defined result. Historically some architectures did not provide a
7893 defined result for zero values as efficiently, and many algorithms are
7894 now predicated on avoiding zero-value inputs.
7899 The '``llvm.cttz``' intrinsic counts the trailing (least significant)
7900 zeros in a variable, or within each element of a vector. If ``src == 0``
7901 then the result is the size in bits of the type of ``src`` if
7902 ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
7903 ``llvm.cttz(2) = 1``.
7905 Arithmetic with Overflow Intrinsics
7906 -----------------------------------
7908 LLVM provides intrinsics for some arithmetic with overflow operations.
7910 '``llvm.sadd.with.overflow.*``' Intrinsics
7911 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7916 This is an overloaded intrinsic. You can use ``llvm.sadd.with.overflow``
7917 on any integer bit width.
7921 declare {i16, i1} @llvm.sadd.with.overflow.i16(i16 %a, i16 %b)
7922 declare {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
7923 declare {i64, i1} @llvm.sadd.with.overflow.i64(i64 %a, i64 %b)
7928 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
7929 a signed addition of the two arguments, and indicate whether an overflow
7930 occurred during the signed summation.
7935 The arguments (%a and %b) and the first element of the result structure
7936 may be of integer types of any bit width, but they must have the same
7937 bit width. The second element of the result structure must be of type
7938 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
7944 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
7945 a signed addition of the two variables. They return a structure --- the
7946 first element of which is the signed summation, and the second element
7947 of which is a bit specifying if the signed summation resulted in an
7953 .. code-block:: llvm
7955 %res = call {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
7956 %sum = extractvalue {i32, i1} %res, 0
7957 %obit = extractvalue {i32, i1} %res, 1
7958 br i1 %obit, label %overflow, label %normal
7960 '``llvm.uadd.with.overflow.*``' Intrinsics
7961 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7966 This is an overloaded intrinsic. You can use ``llvm.uadd.with.overflow``
7967 on any integer bit width.
7971 declare {i16, i1} @llvm.uadd.with.overflow.i16(i16 %a, i16 %b)
7972 declare {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
7973 declare {i64, i1} @llvm.uadd.with.overflow.i64(i64 %a, i64 %b)
7978 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
7979 an unsigned addition of the two arguments, and indicate whether a carry
7980 occurred during the unsigned summation.
7985 The arguments (%a and %b) and the first element of the result structure
7986 may be of integer types of any bit width, but they must have the same
7987 bit width. The second element of the result structure must be of type
7988 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
7994 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
7995 an unsigned addition of the two arguments. They return a structure --- the
7996 first element of which is the sum, and the second element of which is a
7997 bit specifying if the unsigned summation resulted in a carry.
8002 .. code-block:: llvm
8004 %res = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
8005 %sum = extractvalue {i32, i1} %res, 0
8006 %obit = extractvalue {i32, i1} %res, 1
8007 br i1 %obit, label %carry, label %normal
8009 '``llvm.ssub.with.overflow.*``' Intrinsics
8010 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8015 This is an overloaded intrinsic. You can use ``llvm.ssub.with.overflow``
8016 on any integer bit width.
8020 declare {i16, i1} @llvm.ssub.with.overflow.i16(i16 %a, i16 %b)
8021 declare {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
8022 declare {i64, i1} @llvm.ssub.with.overflow.i64(i64 %a, i64 %b)
8027 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
8028 a signed subtraction of the two arguments, and indicate whether an
8029 overflow occurred during the signed subtraction.
8034 The arguments (%a and %b) and the first element of the result structure
8035 may be of integer types of any bit width, but they must have the same
8036 bit width. The second element of the result structure must be of type
8037 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
8043 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
8044 a signed subtraction of the two arguments. They return a structure --- the
8045 first element of which is the subtraction, and the second element of
8046 which is a bit specifying if the signed subtraction resulted in an
8052 .. code-block:: llvm
8054 %res = call {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
8055 %sum = extractvalue {i32, i1} %res, 0
8056 %obit = extractvalue {i32, i1} %res, 1
8057 br i1 %obit, label %overflow, label %normal
8059 '``llvm.usub.with.overflow.*``' Intrinsics
8060 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8065 This is an overloaded intrinsic. You can use ``llvm.usub.with.overflow``
8066 on any integer bit width.
8070 declare {i16, i1} @llvm.usub.with.overflow.i16(i16 %a, i16 %b)
8071 declare {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
8072 declare {i64, i1} @llvm.usub.with.overflow.i64(i64 %a, i64 %b)
8077 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
8078 an unsigned subtraction of the two arguments, and indicate whether an
8079 overflow occurred during the unsigned subtraction.
8084 The arguments (%a and %b) and the first element of the result structure
8085 may be of integer types of any bit width, but they must have the same
8086 bit width. The second element of the result structure must be of type
8087 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
8093 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
8094 an unsigned subtraction of the two arguments. They return a structure ---
8095 the first element of which is the subtraction, and the second element of
8096 which is a bit specifying if the unsigned subtraction resulted in an
8102 .. code-block:: llvm
8104 %res = call {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
8105 %sum = extractvalue {i32, i1} %res, 0
8106 %obit = extractvalue {i32, i1} %res, 1
8107 br i1 %obit, label %overflow, label %normal
8109 '``llvm.smul.with.overflow.*``' Intrinsics
8110 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8115 This is an overloaded intrinsic. You can use ``llvm.smul.with.overflow``
8116 on any integer bit width.
8120 declare {i16, i1} @llvm.smul.with.overflow.i16(i16 %a, i16 %b)
8121 declare {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
8122 declare {i64, i1} @llvm.smul.with.overflow.i64(i64 %a, i64 %b)
8127 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
8128 a signed multiplication of the two arguments, and indicate whether an
8129 overflow occurred during the signed multiplication.
8134 The arguments (%a and %b) and the first element of the result structure
8135 may be of integer types of any bit width, but they must have the same
8136 bit width. The second element of the result structure must be of type
8137 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
8143 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
8144 a signed multiplication of the two arguments. They return a structure ---
8145 the first element of which is the multiplication, and the second element
8146 of which is a bit specifying if the signed multiplication resulted in an
8152 .. code-block:: llvm
8154 %res = call {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
8155 %sum = extractvalue {i32, i1} %res, 0
8156 %obit = extractvalue {i32, i1} %res, 1
8157 br i1 %obit, label %overflow, label %normal
8159 '``llvm.umul.with.overflow.*``' Intrinsics
8160 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8165 This is an overloaded intrinsic. You can use ``llvm.umul.with.overflow``
8166 on any integer bit width.
8170 declare {i16, i1} @llvm.umul.with.overflow.i16(i16 %a, i16 %b)
8171 declare {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
8172 declare {i64, i1} @llvm.umul.with.overflow.i64(i64 %a, i64 %b)
8177 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
8178 a unsigned multiplication of the two arguments, and indicate whether an
8179 overflow occurred during the unsigned multiplication.
8184 The arguments (%a and %b) and the first element of the result structure
8185 may be of integer types of any bit width, but they must have the same
8186 bit width. The second element of the result structure must be of type
8187 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
8193 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
8194 an unsigned multiplication of the two arguments. They return a structure ---
8195 the first element of which is the multiplication, and the second
8196 element of which is a bit specifying if the unsigned multiplication
8197 resulted in an overflow.
8202 .. code-block:: llvm
8204 %res = call {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
8205 %sum = extractvalue {i32, i1} %res, 0
8206 %obit = extractvalue {i32, i1} %res, 1
8207 br i1 %obit, label %overflow, label %normal
8209 Specialised Arithmetic Intrinsics
8210 ---------------------------------
8212 '``llvm.fmuladd.*``' Intrinsic
8213 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8220 declare float @llvm.fmuladd.f32(float %a, float %b, float %c)
8221 declare double @llvm.fmuladd.f64(double %a, double %b, double %c)
8226 The '``llvm.fmuladd.*``' intrinsic functions represent multiply-add
8227 expressions that can be fused if the code generator determines that (a) the
8228 target instruction set has support for a fused operation, and (b) that the
8229 fused operation is more efficient than the equivalent, separate pair of mul
8230 and add instructions.
8235 The '``llvm.fmuladd.*``' intrinsics each take three arguments: two
8236 multiplicands, a and b, and an addend c.
8245 %0 = call float @llvm.fmuladd.f32(%a, %b, %c)
8247 is equivalent to the expression a \* b + c, except that rounding will
8248 not be performed between the multiplication and addition steps if the
8249 code generator fuses the operations. Fusion is not guaranteed, even if
8250 the target platform supports it. If a fused multiply-add is required the
8251 corresponding llvm.fma.\* intrinsic function should be used instead.
8256 .. code-block:: llvm
8258 %r2 = call float @llvm.fmuladd.f32(float %a, float %b, float %c) ; yields {float}:r2 = (a * b) + c
8260 Half Precision Floating Point Intrinsics
8261 ----------------------------------------
8263 For most target platforms, half precision floating point is a
8264 storage-only format. This means that it is a dense encoding (in memory)
8265 but does not support computation in the format.
8267 This means that code must first load the half-precision floating point
8268 value as an i16, then convert it to float with
8269 :ref:`llvm.convert.from.fp16 <int_convert_from_fp16>`. Computation can
8270 then be performed on the float value (including extending to double
8271 etc). To store the value back to memory, it is first converted to float
8272 if needed, then converted to i16 with
8273 :ref:`llvm.convert.to.fp16 <int_convert_to_fp16>`, then storing as an
8276 .. _int_convert_to_fp16:
8278 '``llvm.convert.to.fp16``' Intrinsic
8279 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8286 declare i16 @llvm.convert.to.fp16(f32 %a)
8291 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion
8292 from single precision floating point format to half precision floating
8298 The intrinsic function contains single argument - the value to be
8304 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion
8305 from single precision floating point format to half precision floating
8306 point format. The return value is an ``i16`` which contains the
8312 .. code-block:: llvm
8314 %res = call i16 @llvm.convert.to.fp16(f32 %a)
8315 store i16 %res, i16* @x, align 2
8317 .. _int_convert_from_fp16:
8319 '``llvm.convert.from.fp16``' Intrinsic
8320 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8327 declare f32 @llvm.convert.from.fp16(i16 %a)
8332 The '``llvm.convert.from.fp16``' intrinsic function performs a
8333 conversion from half precision floating point format to single precision
8334 floating point format.
8339 The intrinsic function contains single argument - the value to be
8345 The '``llvm.convert.from.fp16``' intrinsic function performs a
8346 conversion from half single precision floating point format to single
8347 precision floating point format. The input half-float value is
8348 represented by an ``i16`` value.
8353 .. code-block:: llvm
8355 %a = load i16* @x, align 2
8356 %res = call f32 @llvm.convert.from.fp16(i16 %a)
8361 The LLVM debugger intrinsics (which all start with ``llvm.dbg.``
8362 prefix), are described in the `LLVM Source Level
8363 Debugging <SourceLevelDebugging.html#format_common_intrinsics>`_
8366 Exception Handling Intrinsics
8367 -----------------------------
8369 The LLVM exception handling intrinsics (which all start with
8370 ``llvm.eh.`` prefix), are described in the `LLVM Exception
8371 Handling <ExceptionHandling.html#format_common_intrinsics>`_ document.
8375 Trampoline Intrinsics
8376 ---------------------
8378 These intrinsics make it possible to excise one parameter, marked with
8379 the :ref:`nest <nest>` attribute, from a function. The result is a
8380 callable function pointer lacking the nest parameter - the caller does
8381 not need to provide a value for it. Instead, the value to use is stored
8382 in advance in a "trampoline", a block of memory usually allocated on the
8383 stack, which also contains code to splice the nest value into the
8384 argument list. This is used to implement the GCC nested function address
8387 For example, if the function is ``i32 f(i8* nest %c, i32 %x, i32 %y)``
8388 then the resulting function pointer has signature ``i32 (i32, i32)*``.
8389 It can be created as follows:
8391 .. code-block:: llvm
8393 %tramp = alloca [10 x i8], align 4 ; size and alignment only correct for X86
8394 %tramp1 = getelementptr [10 x i8]* %tramp, i32 0, i32 0
8395 call i8* @llvm.init.trampoline(i8* %tramp1, i8* bitcast (i32 (i8*, i32, i32)* @f to i8*), i8* %nval)
8396 %p = call i8* @llvm.adjust.trampoline(i8* %tramp1)
8397 %fp = bitcast i8* %p to i32 (i32, i32)*
8399 The call ``%val = call i32 %fp(i32 %x, i32 %y)`` is then equivalent to
8400 ``%val = call i32 %f(i8* %nval, i32 %x, i32 %y)``.
8404 '``llvm.init.trampoline``' Intrinsic
8405 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8412 declare void @llvm.init.trampoline(i8* <tramp>, i8* <func>, i8* <nval>)
8417 This fills the memory pointed to by ``tramp`` with executable code,
8418 turning it into a trampoline.
8423 The ``llvm.init.trampoline`` intrinsic takes three arguments, all
8424 pointers. The ``tramp`` argument must point to a sufficiently large and
8425 sufficiently aligned block of memory; this memory is written to by the
8426 intrinsic. Note that the size and the alignment are target-specific -
8427 LLVM currently provides no portable way of determining them, so a
8428 front-end that generates this intrinsic needs to have some
8429 target-specific knowledge. The ``func`` argument must hold a function
8430 bitcast to an ``i8*``.
8435 The block of memory pointed to by ``tramp`` is filled with target
8436 dependent code, turning it into a function. Then ``tramp`` needs to be
8437 passed to :ref:`llvm.adjust.trampoline <int_at>` to get a pointer which can
8438 be :ref:`bitcast (to a new function) and called <int_trampoline>`. The new
8439 function's signature is the same as that of ``func`` with any arguments
8440 marked with the ``nest`` attribute removed. At most one such ``nest``
8441 argument is allowed, and it must be of pointer type. Calling the new
8442 function is equivalent to calling ``func`` with the same argument list,
8443 but with ``nval`` used for the missing ``nest`` argument. If, after
8444 calling ``llvm.init.trampoline``, the memory pointed to by ``tramp`` is
8445 modified, then the effect of any later call to the returned function
8446 pointer is undefined.
8450 '``llvm.adjust.trampoline``' Intrinsic
8451 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8458 declare i8* @llvm.adjust.trampoline(i8* <tramp>)
8463 This performs any required machine-specific adjustment to the address of
8464 a trampoline (passed as ``tramp``).
8469 ``tramp`` must point to a block of memory which already has trampoline
8470 code filled in by a previous call to
8471 :ref:`llvm.init.trampoline <int_it>`.
8476 On some architectures the address of the code to be executed needs to be
8477 different to the address where the trampoline is actually stored. This
8478 intrinsic returns the executable address corresponding to ``tramp``
8479 after performing the required machine specific adjustments. The pointer
8480 returned can then be :ref:`bitcast and executed <int_trampoline>`.
8485 This class of intrinsics exists to information about the lifetime of
8486 memory objects and ranges where variables are immutable.
8490 '``llvm.lifetime.start``' Intrinsic
8491 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8498 declare void @llvm.lifetime.start(i64 <size>, i8* nocapture <ptr>)
8503 The '``llvm.lifetime.start``' intrinsic specifies the start of a memory
8509 The first argument is a constant integer representing the size of the
8510 object, or -1 if it is variable sized. The second argument is a pointer
8516 This intrinsic indicates that before this point in the code, the value
8517 of the memory pointed to by ``ptr`` is dead. This means that it is known
8518 to never be used and has an undefined value. A load from the pointer
8519 that precedes this intrinsic can be replaced with ``'undef'``.
8523 '``llvm.lifetime.end``' Intrinsic
8524 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8531 declare void @llvm.lifetime.end(i64 <size>, i8* nocapture <ptr>)
8536 The '``llvm.lifetime.end``' intrinsic specifies the end of a memory
8542 The first argument is a constant integer representing the size of the
8543 object, or -1 if it is variable sized. The second argument is a pointer
8549 This intrinsic indicates that after this point in the code, the value of
8550 the memory pointed to by ``ptr`` is dead. This means that it is known to
8551 never be used and has an undefined value. Any stores into the memory
8552 object following this intrinsic may be removed as dead.
8554 '``llvm.invariant.start``' Intrinsic
8555 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8562 declare {}* @llvm.invariant.start(i64 <size>, i8* nocapture <ptr>)
8567 The '``llvm.invariant.start``' intrinsic specifies that the contents of
8568 a memory object will not change.
8573 The first argument is a constant integer representing the size of the
8574 object, or -1 if it is variable sized. The second argument is a pointer
8580 This intrinsic indicates that until an ``llvm.invariant.end`` that uses
8581 the return value, the referenced memory location is constant and
8584 '``llvm.invariant.end``' Intrinsic
8585 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8592 declare void @llvm.invariant.end({}* <start>, i64 <size>, i8* nocapture <ptr>)
8597 The '``llvm.invariant.end``' intrinsic specifies that the contents of a
8598 memory object are mutable.
8603 The first argument is the matching ``llvm.invariant.start`` intrinsic.
8604 The second argument is a constant integer representing the size of the
8605 object, or -1 if it is variable sized and the third argument is a
8606 pointer to the object.
8611 This intrinsic indicates that the memory is mutable again.
8616 This class of intrinsics is designed to be generic and has no specific
8619 '``llvm.var.annotation``' Intrinsic
8620 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8627 declare void @llvm.var.annotation(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
8632 The '``llvm.var.annotation``' intrinsic.
8637 The first argument is a pointer to a value, the second is a pointer to a
8638 global string, the third is a pointer to a global string which is the
8639 source file name, and the last argument is the line number.
8644 This intrinsic allows annotation of local variables with arbitrary
8645 strings. This can be useful for special purpose optimizations that want
8646 to look for these annotations. These have no other defined use; they are
8647 ignored by code generation and optimization.
8649 '``llvm.ptr.annotation.*``' Intrinsic
8650 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8655 This is an overloaded intrinsic. You can use '``llvm.ptr.annotation``' on a
8656 pointer to an integer of any width. *NOTE* you must specify an address space for
8657 the pointer. The identifier for the default address space is the integer
8662 declare i8* @llvm.ptr.annotation.p<address space>i8(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
8663 declare i16* @llvm.ptr.annotation.p<address space>i16(i16* <val>, i8* <str>, i8* <str>, i32 <int>)
8664 declare i32* @llvm.ptr.annotation.p<address space>i32(i32* <val>, i8* <str>, i8* <str>, i32 <int>)
8665 declare i64* @llvm.ptr.annotation.p<address space>i64(i64* <val>, i8* <str>, i8* <str>, i32 <int>)
8666 declare i256* @llvm.ptr.annotation.p<address space>i256(i256* <val>, i8* <str>, i8* <str>, i32 <int>)
8671 The '``llvm.ptr.annotation``' intrinsic.
8676 The first argument is a pointer to an integer value of arbitrary bitwidth
8677 (result of some expression), the second is a pointer to a global string, the
8678 third is a pointer to a global string which is the source file name, and the
8679 last argument is the line number. It returns the value of the first argument.
8684 This intrinsic allows annotation of a pointer to an integer with arbitrary
8685 strings. This can be useful for special purpose optimizations that want to look
8686 for these annotations. These have no other defined use; they are ignored by code
8687 generation and optimization.
8689 '``llvm.annotation.*``' Intrinsic
8690 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8695 This is an overloaded intrinsic. You can use '``llvm.annotation``' on
8696 any integer bit width.
8700 declare i8 @llvm.annotation.i8(i8 <val>, i8* <str>, i8* <str>, i32 <int>)
8701 declare i16 @llvm.annotation.i16(i16 <val>, i8* <str>, i8* <str>, i32 <int>)
8702 declare i32 @llvm.annotation.i32(i32 <val>, i8* <str>, i8* <str>, i32 <int>)
8703 declare i64 @llvm.annotation.i64(i64 <val>, i8* <str>, i8* <str>, i32 <int>)
8704 declare i256 @llvm.annotation.i256(i256 <val>, i8* <str>, i8* <str>, i32 <int>)
8709 The '``llvm.annotation``' intrinsic.
8714 The first argument is an integer value (result of some expression), the
8715 second is a pointer to a global string, the third is a pointer to a
8716 global string which is the source file name, and the last argument is
8717 the line number. It returns the value of the first argument.
8722 This intrinsic allows annotations to be put on arbitrary expressions
8723 with arbitrary strings. This can be useful for special purpose
8724 optimizations that want to look for these annotations. These have no
8725 other defined use; they are ignored by code generation and optimization.
8727 '``llvm.trap``' Intrinsic
8728 ^^^^^^^^^^^^^^^^^^^^^^^^^
8735 declare void @llvm.trap() noreturn nounwind
8740 The '``llvm.trap``' intrinsic.
8750 This intrinsic is lowered to the target dependent trap instruction. If
8751 the target does not have a trap instruction, this intrinsic will be
8752 lowered to a call of the ``abort()`` function.
8754 '``llvm.debugtrap``' Intrinsic
8755 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8762 declare void @llvm.debugtrap() nounwind
8767 The '``llvm.debugtrap``' intrinsic.
8777 This intrinsic is lowered to code which is intended to cause an
8778 execution trap with the intention of requesting the attention of a
8781 '``llvm.stackprotector``' Intrinsic
8782 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8789 declare void @llvm.stackprotector(i8* <guard>, i8** <slot>)
8794 The ``llvm.stackprotector`` intrinsic takes the ``guard`` and stores it
8795 onto the stack at ``slot``. The stack slot is adjusted to ensure that it
8796 is placed on the stack before local variables.
8801 The ``llvm.stackprotector`` intrinsic requires two pointer arguments.
8802 The first argument is the value loaded from the stack guard
8803 ``@__stack_chk_guard``. The second variable is an ``alloca`` that has
8804 enough space to hold the value of the guard.
8809 This intrinsic causes the prologue/epilogue inserter to force the position of
8810 the ``AllocaInst`` stack slot to be before local variables on the stack. This is
8811 to ensure that if a local variable on the stack is overwritten, it will destroy
8812 the value of the guard. When the function exits, the guard on the stack is
8813 checked against the original guard by ``llvm.stackprotectorcheck``. If they are
8814 different, then ``llvm.stackprotectorcheck`` causes the program to abort by
8815 calling the ``__stack_chk_fail()`` function.
8817 '``llvm.stackprotectorcheck``' Intrinsic
8818 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8825 declare void @llvm.stackprotectorcheck(i8** <guard>)
8830 The ``llvm.stackprotectorcheck`` intrinsic compares ``guard`` against an already
8831 created stack protector and if they are not equal calls the
8832 ``__stack_chk_fail()`` function.
8837 The ``llvm.stackprotectorcheck`` intrinsic requires one pointer argument, the
8838 the variable ``@__stack_chk_guard``.
8843 This intrinsic is provided to perform the stack protector check by comparing
8844 ``guard`` with the stack slot created by ``llvm.stackprotector`` and if the
8845 values do not match call the ``__stack_chk_fail()`` function.
8847 The reason to provide this as an IR level intrinsic instead of implementing it
8848 via other IR operations is that in order to perform this operation at the IR
8849 level without an intrinsic, one would need to create additional basic blocks to
8850 handle the success/failure cases. This makes it difficult to stop the stack
8851 protector check from disrupting sibling tail calls in Codegen. With this
8852 intrinsic, we are able to generate the stack protector basic blocks late in
8853 codegen after the tail call decision has occurred.
8855 '``llvm.objectsize``' Intrinsic
8856 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8863 declare i32 @llvm.objectsize.i32(i8* <object>, i1 <min>)
8864 declare i64 @llvm.objectsize.i64(i8* <object>, i1 <min>)
8869 The ``llvm.objectsize`` intrinsic is designed to provide information to
8870 the optimizers to determine at compile time whether a) an operation
8871 (like memcpy) will overflow a buffer that corresponds to an object, or
8872 b) that a runtime check for overflow isn't necessary. An object in this
8873 context means an allocation of a specific class, structure, array, or
8879 The ``llvm.objectsize`` intrinsic takes two arguments. The first
8880 argument is a pointer to or into the ``object``. The second argument is
8881 a boolean and determines whether ``llvm.objectsize`` returns 0 (if true)
8882 or -1 (if false) when the object size is unknown. The second argument
8883 only accepts constants.
8888 The ``llvm.objectsize`` intrinsic is lowered to a constant representing
8889 the size of the object concerned. If the size cannot be determined at
8890 compile time, ``llvm.objectsize`` returns ``i32/i64 -1 or 0`` (depending
8891 on the ``min`` argument).
8893 '``llvm.expect``' Intrinsic
8894 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8901 declare i32 @llvm.expect.i32(i32 <val>, i32 <expected_val>)
8902 declare i64 @llvm.expect.i64(i64 <val>, i64 <expected_val>)
8907 The ``llvm.expect`` intrinsic provides information about expected (the
8908 most probable) value of ``val``, which can be used by optimizers.
8913 The ``llvm.expect`` intrinsic takes two arguments. The first argument is
8914 a value. The second argument is an expected value, this needs to be a
8915 constant value, variables are not allowed.
8920 This intrinsic is lowered to the ``val``.
8922 '``llvm.donothing``' Intrinsic
8923 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8930 declare void @llvm.donothing() nounwind readnone
8935 The ``llvm.donothing`` intrinsic doesn't perform any operation. It's the
8936 only intrinsic that can be called with an invoke instruction.
8946 This intrinsic does nothing, and it's removed by optimizers and ignored
8949 Stack Map Intrinsics
8950 --------------------
8952 LLVM provides experimental intrinsics to support runtime patching
8953 mechanisms commonly desired in dynamic language JITs. These intrinsics
8954 are described in :doc:`StackMaps`.