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 that 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. The ``"\01"`` prefix
83 can be used on global variables to suppress mangling.
84 #. Unnamed values are represented as an unsigned numeric value with
85 their prefix. For example, ``%12``, ``@2``, ``%44``.
86 #. Constants, which are described in the section Constants_ below.
88 LLVM requires that values start with a prefix for two reasons: Compilers
89 don't need to worry about name clashes with reserved words, and the set
90 of reserved words may be expanded in the future without penalty.
91 Additionally, unnamed identifiers allow a compiler to quickly come up
92 with a temporary variable without having to avoid symbol table
95 Reserved words in LLVM are very similar to reserved words in other
96 languages. There are keywords for different opcodes ('``add``',
97 '``bitcast``', '``ret``', etc...), for primitive type names ('``void``',
98 '``i32``', etc...), and others. These reserved words cannot conflict
99 with variable names, because none of them start with a prefix character
100 (``'%'`` or ``'@'``).
102 Here is an example of LLVM code to multiply the integer variable
109 %result = mul i32 %X, 8
111 After strength reduction:
115 %result = shl i32 %X, 3
121 %0 = add i32 %X, %X ; yields i32:%0
122 %1 = add i32 %0, %0 ; yields i32:%1
123 %result = add i32 %1, %1
125 This last way of multiplying ``%X`` by 8 illustrates several important
126 lexical features of LLVM:
128 #. Comments are delimited with a '``;``' and go until the end of line.
129 #. Unnamed temporaries are created when the result of a computation is
130 not assigned to a named value.
131 #. Unnamed temporaries are numbered sequentially (using a per-function
132 incrementing counter, starting with 0). Note that basic blocks and unnamed
133 function parameters are included in this numbering. For example, if the
134 entry basic block is not given a label name and all function parameters are
135 named, then it will get number 0.
137 It also shows a convention that we follow in this document. When
138 demonstrating instructions, we will follow an instruction with a comment
139 that defines the type and name of value produced.
147 LLVM programs are composed of ``Module``'s, each of which is a
148 translation unit of the input programs. Each module consists of
149 functions, global variables, and symbol table entries. Modules may be
150 combined together with the LLVM linker, which merges function (and
151 global variable) definitions, resolves forward declarations, and merges
152 symbol table entries. Here is an example of the "hello world" module:
156 ; Declare the string constant as a global constant.
157 @.str = private unnamed_addr constant [13 x i8] c"hello world\0A\00"
159 ; External declaration of the puts function
160 declare i32 @puts(i8* nocapture) nounwind
162 ; Definition of main function
163 define i32 @main() { ; i32()*
164 ; Convert [13 x i8]* to i8 *...
165 %cast210 = getelementptr [13 x i8]* @.str, i64 0, i64 0
167 ; Call puts function to write out the string to stdout.
168 call i32 @puts(i8* %cast210)
173 !0 = metadata !{i32 42, null, metadata !"string"}
176 This example is made up of a :ref:`global variable <globalvars>` named
177 "``.str``", an external declaration of the "``puts``" function, a
178 :ref:`function definition <functionstructure>` for "``main``" and
179 :ref:`named metadata <namedmetadatastructure>` "``foo``".
181 In general, a module is made up of a list of global values (where both
182 functions and global variables are global values). Global values are
183 represented by a pointer to a memory location (in this case, a pointer
184 to an array of char, and a pointer to a function), and have one of the
185 following :ref:`linkage types <linkage>`.
192 All Global Variables and Functions have one of the following types of
196 Global values with "``private``" linkage are only directly
197 accessible by objects in the current module. In particular, linking
198 code into a module with an private global value may cause the
199 private to be renamed as necessary to avoid collisions. Because the
200 symbol is private to the module, all references can be updated. This
201 doesn't show up in any symbol table in the object file.
203 Similar to private, but the value shows as a local symbol
204 (``STB_LOCAL`` in the case of ELF) in the object file. This
205 corresponds to the notion of the '``static``' keyword in C.
206 ``available_externally``
207 Globals with "``available_externally``" linkage are never emitted
208 into the object file corresponding to the LLVM module. They exist to
209 allow inlining and other optimizations to take place given knowledge
210 of the definition of the global, which is known to be somewhere
211 outside the module. Globals with ``available_externally`` linkage
212 are allowed to be discarded at will, and are otherwise the same as
213 ``linkonce_odr``. This linkage type is only allowed on definitions,
216 Globals with "``linkonce``" linkage are merged with other globals of
217 the same name when linkage occurs. This can be used to implement
218 some forms of inline functions, templates, or other code which must
219 be generated in each translation unit that uses it, but where the
220 body may be overridden with a more definitive definition later.
221 Unreferenced ``linkonce`` globals are allowed to be discarded. Note
222 that ``linkonce`` linkage does not actually allow the optimizer to
223 inline the body of this function into callers because it doesn't
224 know if this definition of the function is the definitive definition
225 within the program or whether it will be overridden by a stronger
226 definition. To enable inlining and other optimizations, use
227 "``linkonce_odr``" linkage.
229 "``weak``" linkage has the same merging semantics as ``linkonce``
230 linkage, except that unreferenced globals with ``weak`` linkage may
231 not be discarded. This is used for globals that are declared "weak"
234 "``common``" linkage is most similar to "``weak``" linkage, but they
235 are used for tentative definitions in C, such as "``int X;``" at
236 global scope. Symbols with "``common``" linkage are merged in the
237 same way as ``weak symbols``, and they may not be deleted if
238 unreferenced. ``common`` symbols may not have an explicit section,
239 must have a zero initializer, and may not be marked
240 ':ref:`constant <globalvars>`'. Functions and aliases may not have
243 .. _linkage_appending:
246 "``appending``" linkage may only be applied to global variables of
247 pointer to array type. When two global variables with appending
248 linkage are linked together, the two global arrays are appended
249 together. This is the LLVM, typesafe, equivalent of having the
250 system linker append together "sections" with identical names when
253 The semantics of this linkage follow the ELF object file model: the
254 symbol is weak until linked, if not linked, the symbol becomes null
255 instead of being an undefined reference.
256 ``linkonce_odr``, ``weak_odr``
257 Some languages allow differing globals to be merged, such as two
258 functions with different semantics. Other languages, such as
259 ``C++``, ensure that only equivalent globals are ever merged (the
260 "one definition rule" --- "ODR"). Such languages can use the
261 ``linkonce_odr`` and ``weak_odr`` linkage types to indicate that the
262 global will only be merged with equivalent globals. These linkage
263 types are otherwise the same as their non-``odr`` versions.
265 If none of the above identifiers are used, the global is externally
266 visible, meaning that it participates in linkage and can be used to
267 resolve external symbol references.
269 It is illegal for a function *declaration* to have any linkage type
270 other than ``external`` or ``extern_weak``.
277 LLVM :ref:`functions <functionstructure>`, :ref:`calls <i_call>` and
278 :ref:`invokes <i_invoke>` can all have an optional calling convention
279 specified for the call. The calling convention of any pair of dynamic
280 caller/callee must match, or the behavior of the program is undefined.
281 The following calling conventions are supported by LLVM, and more may be
284 "``ccc``" - The C calling convention
285 This calling convention (the default if no other calling convention
286 is specified) matches the target C calling conventions. This calling
287 convention supports varargs function calls and tolerates some
288 mismatch in the declared prototype and implemented declaration of
289 the function (as does normal C).
290 "``fastcc``" - The fast calling convention
291 This calling convention attempts to make calls as fast as possible
292 (e.g. by passing things in registers). This calling convention
293 allows the target to use whatever tricks it wants to produce fast
294 code for the target, without having to conform to an externally
295 specified ABI (Application Binary Interface). `Tail calls can only
296 be optimized when this, the GHC or the HiPE convention is
297 used. <CodeGenerator.html#id80>`_ This calling convention does not
298 support varargs and requires the prototype of all callees to exactly
299 match the prototype of the function definition.
300 "``coldcc``" - The cold calling convention
301 This calling convention attempts to make code in the caller as
302 efficient as possible under the assumption that the call is not
303 commonly executed. As such, these calls often preserve all registers
304 so that the call does not break any live ranges in the caller side.
305 This calling convention does not support varargs and requires the
306 prototype of all callees to exactly match the prototype of the
307 function definition. Furthermore the inliner doesn't consider such function
309 "``cc 10``" - GHC convention
310 This calling convention has been implemented specifically for use by
311 the `Glasgow Haskell Compiler (GHC) <http://www.haskell.org/ghc>`_.
312 It passes everything in registers, going to extremes to achieve this
313 by disabling callee save registers. This calling convention should
314 not be used lightly but only for specific situations such as an
315 alternative to the *register pinning* performance technique often
316 used when implementing functional programming languages. At the
317 moment only X86 supports this convention and it has the following
320 - On *X86-32* only supports up to 4 bit type parameters. No
321 floating point types are supported.
322 - On *X86-64* only supports up to 10 bit type parameters and 6
323 floating point parameters.
325 This calling convention supports `tail call
326 optimization <CodeGenerator.html#id80>`_ but requires both the
327 caller and callee are using it.
328 "``cc 11``" - The HiPE calling convention
329 This calling convention has been implemented specifically for use by
330 the `High-Performance Erlang
331 (HiPE) <http://www.it.uu.se/research/group/hipe/>`_ compiler, *the*
332 native code compiler of the `Ericsson's Open Source Erlang/OTP
333 system <http://www.erlang.org/download.shtml>`_. It uses more
334 registers for argument passing than the ordinary C calling
335 convention and defines no callee-saved registers. The calling
336 convention properly supports `tail call
337 optimization <CodeGenerator.html#id80>`_ but requires that both the
338 caller and the callee use it. It uses a *register pinning*
339 mechanism, similar to GHC's convention, for keeping frequently
340 accessed runtime components pinned to specific hardware registers.
341 At the moment only X86 supports this convention (both 32 and 64
343 "``webkit_jscc``" - WebKit's JavaScript calling convention
344 This calling convention has been implemented for `WebKit FTL JIT
345 <https://trac.webkit.org/wiki/FTLJIT>`_. It passes arguments on the
346 stack right to left (as cdecl does), and returns a value in the
347 platform's customary return register.
348 "``anyregcc``" - Dynamic calling convention for code patching
349 This is a special convention that supports patching an arbitrary code
350 sequence in place of a call site. This convention forces the call
351 arguments into registers but allows them to be dynamcially
352 allocated. This can currently only be used with calls to
353 llvm.experimental.patchpoint because only this intrinsic records
354 the location of its arguments in a side table. See :doc:`StackMaps`.
355 "``preserve_mostcc``" - The `PreserveMost` calling convention
356 This calling convention attempts to make the code in the caller as little
357 intrusive as possible. This calling convention behaves identical to the `C`
358 calling convention on how arguments and return values are passed, but it
359 uses a different set of caller/callee-saved registers. This alleviates the
360 burden of saving and recovering a large register set before and after the
361 call in the caller. If the arguments are passed in callee-saved registers,
362 then they will be preserved by the callee across the call. This doesn't
363 apply for values returned in callee-saved registers.
365 - On X86-64 the callee preserves all general purpose registers, except for
366 R11. R11 can be used as a scratch register. Floating-point registers
367 (XMMs/YMMs) are not preserved and need to be saved by the caller.
369 The idea behind this convention is to support calls to runtime functions
370 that have a hot path and a cold path. The hot path is usually a small piece
371 of code that doesn't many registers. The cold path might need to call out to
372 another function and therefore only needs to preserve the caller-saved
373 registers, which haven't already been saved by the caller. The
374 `PreserveMost` calling convention is very similar to the `cold` calling
375 convention in terms of caller/callee-saved registers, but they are used for
376 different types of function calls. `coldcc` is for function calls that are
377 rarely executed, whereas `preserve_mostcc` function calls are intended to be
378 on the hot path and definitely executed a lot. Furthermore `preserve_mostcc`
379 doesn't prevent the inliner from inlining the function call.
381 This calling convention will be used by a future version of the ObjectiveC
382 runtime and should therefore still be considered experimental at this time.
383 Although this convention was created to optimize certain runtime calls to
384 the ObjectiveC runtime, it is not limited to this runtime and might be used
385 by other runtimes in the future too. The current implementation only
386 supports X86-64, but the intention is to support more architectures in the
388 "``preserve_allcc``" - The `PreserveAll` calling convention
389 This calling convention attempts to make the code in the caller even less
390 intrusive than the `PreserveMost` calling convention. This calling
391 convention also behaves identical to the `C` calling convention on how
392 arguments and return values are passed, but it uses a different set of
393 caller/callee-saved registers. This removes the burden of saving and
394 recovering a large register set before and after the call in the caller. If
395 the arguments are passed in callee-saved registers, then they will be
396 preserved by the callee across the call. This doesn't apply for values
397 returned in callee-saved registers.
399 - On X86-64 the callee preserves all general purpose registers, except for
400 R11. R11 can be used as a scratch register. Furthermore it also preserves
401 all floating-point registers (XMMs/YMMs).
403 The idea behind this convention is to support calls to runtime functions
404 that don't need to call out to any other functions.
406 This calling convention, like the `PreserveMost` calling convention, will be
407 used by a future version of the ObjectiveC runtime and should be considered
408 experimental at this time.
409 "``cc <n>``" - Numbered convention
410 Any calling convention may be specified by number, allowing
411 target-specific calling conventions to be used. Target specific
412 calling conventions start at 64.
414 More calling conventions can be added/defined on an as-needed basis, to
415 support Pascal conventions or any other well-known target-independent
418 .. _visibilitystyles:
423 All Global Variables and Functions have one of the following visibility
426 "``default``" - Default style
427 On targets that use the ELF object file format, default visibility
428 means that the declaration is visible to other modules and, in
429 shared libraries, means that the declared entity may be overridden.
430 On Darwin, default visibility means that the declaration is visible
431 to other modules. Default visibility corresponds to "external
432 linkage" in the language.
433 "``hidden``" - Hidden style
434 Two declarations of an object with hidden visibility refer to the
435 same object if they are in the same shared object. Usually, hidden
436 visibility indicates that the symbol will not be placed into the
437 dynamic symbol table, so no other module (executable or shared
438 library) can reference it directly.
439 "``protected``" - Protected style
440 On ELF, protected visibility indicates that the symbol will be
441 placed in the dynamic symbol table, but that references within the
442 defining module will bind to the local symbol. That is, the symbol
443 cannot be overridden by another module.
445 A symbol with ``internal`` or ``private`` linkage must have ``default``
453 All Global Variables, Functions and Aliases can have one of the following
457 "``dllimport``" causes the compiler to reference a function or variable via
458 a global pointer to a pointer that is set up by the DLL exporting the
459 symbol. On Microsoft Windows targets, the pointer name is formed by
460 combining ``__imp_`` and the function or variable name.
462 "``dllexport``" causes the compiler to provide a global pointer to a pointer
463 in a DLL, so that it can be referenced with the ``dllimport`` attribute. On
464 Microsoft Windows targets, the pointer name is formed by combining
465 ``__imp_`` and the function or variable name. Since this storage class
466 exists for defining a dll interface, the compiler, assembler and linker know
467 it is externally referenced and must refrain from deleting the symbol.
471 Thread Local Storage Models
472 ---------------------------
474 A variable may be defined as ``thread_local``, which means that it will
475 not be shared by threads (each thread will have a separated copy of the
476 variable). Not all targets support thread-local variables. Optionally, a
477 TLS model may be specified:
480 For variables that are only used within the current shared library.
482 For variables in modules that will not be loaded dynamically.
484 For variables defined in the executable and only used within it.
486 If no explicit model is given, the "general dynamic" model is used.
488 The models correspond to the ELF TLS models; see `ELF Handling For
489 Thread-Local Storage <http://people.redhat.com/drepper/tls.pdf>`_ for
490 more information on under which circumstances the different models may
491 be used. The target may choose a different TLS model if the specified
492 model is not supported, or if a better choice of model can be made.
494 A model can also be specified in a alias, but then it only governs how
495 the alias is accessed. It will not have any effect in the aliasee.
502 LLVM IR allows you to specify both "identified" and "literal" :ref:`structure
503 types <t_struct>`. Literal types are uniqued structurally, but identified types
504 are never uniqued. An :ref:`opaque structural type <t_opaque>` can also be used
505 to forward declare a type that is not yet available.
507 An example of a identified structure specification is:
511 %mytype = type { %mytype*, i32 }
513 Prior to the LLVM 3.0 release, identified types were structurally uniqued. Only
514 literal types are uniqued in recent versions of LLVM.
521 Global variables define regions of memory allocated at compilation time
524 Global variables definitions must be initialized.
526 Global variables in other translation units can also be declared, in which
527 case they don't have an initializer.
529 Either global variable definitions or declarations may have an explicit section
530 to be placed in and may have an optional explicit alignment specified.
532 A variable may be defined as a global ``constant``, which indicates that
533 the contents of the variable will **never** be modified (enabling better
534 optimization, allowing the global data to be placed in the read-only
535 section of an executable, etc). Note that variables that need runtime
536 initialization cannot be marked ``constant`` as there is a store to the
539 LLVM explicitly allows *declarations* of global variables to be marked
540 constant, even if the final definition of the global is not. This
541 capability can be used to enable slightly better optimization of the
542 program, but requires the language definition to guarantee that
543 optimizations based on the 'constantness' are valid for the translation
544 units that do not include the definition.
546 As SSA values, global variables define pointer values that are in scope
547 (i.e. they dominate) all basic blocks in the program. Global variables
548 always define a pointer to their "content" type because they describe a
549 region of memory, and all memory objects in LLVM are accessed through
552 Global variables can be marked with ``unnamed_addr`` which indicates
553 that the address is not significant, only the content. Constants marked
554 like this can be merged with other constants if they have the same
555 initializer. Note that a constant with significant address *can* be
556 merged with a ``unnamed_addr`` constant, the result being a constant
557 whose address is significant.
559 A global variable may be declared to reside in a target-specific
560 numbered address space. For targets that support them, address spaces
561 may affect how optimizations are performed and/or what target
562 instructions are used to access the variable. The default address space
563 is zero. The address space qualifier must precede any other attributes.
565 LLVM allows an explicit section to be specified for globals. If the
566 target supports it, it will emit globals to the section specified.
567 Additionally, the global can placed in a comdat if the target has the necessary
570 By default, global initializers are optimized by assuming that global
571 variables defined within the module are not modified from their
572 initial values before the start of the global initializer. This is
573 true even for variables potentially accessible from outside the
574 module, including those with external linkage or appearing in
575 ``@llvm.used`` or dllexported variables. This assumption may be suppressed
576 by marking the variable with ``externally_initialized``.
578 An explicit alignment may be specified for a global, which must be a
579 power of 2. If not present, or if the alignment is set to zero, the
580 alignment of the global is set by the target to whatever it feels
581 convenient. If an explicit alignment is specified, the global is forced
582 to have exactly that alignment. Targets and optimizers are not allowed
583 to over-align the global if the global has an assigned section. In this
584 case, the extra alignment could be observable: for example, code could
585 assume that the globals are densely packed in their section and try to
586 iterate over them as an array, alignment padding would break this
587 iteration. The maximum alignment is ``1 << 29``.
589 Globals can also have a :ref:`DLL storage class <dllstorageclass>`.
591 Variables and aliasaes can have a
592 :ref:`Thread Local Storage Model <tls_model>`.
596 [@<GlobalVarName> =] [Linkage] [Visibility] [DLLStorageClass] [ThreadLocal]
597 [unnamed_addr] [AddrSpace] [ExternallyInitialized]
598 <global | constant> <Type> [<InitializerConstant>]
599 [, section "name"] [, align <Alignment>]
601 For example, the following defines a global in a numbered address space
602 with an initializer, section, and alignment:
606 @G = addrspace(5) constant float 1.0, section "foo", align 4
608 The following example just declares a global variable
612 @G = external global i32
614 The following example defines a thread-local global with the
615 ``initialexec`` TLS model:
619 @G = thread_local(initialexec) global i32 0, align 4
621 .. _functionstructure:
626 LLVM function definitions consist of the "``define``" keyword, an
627 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
628 style <visibility>`, an optional :ref:`DLL storage class <dllstorageclass>`,
629 an optional :ref:`calling convention <callingconv>`,
630 an optional ``unnamed_addr`` attribute, a return type, an optional
631 :ref:`parameter attribute <paramattrs>` for the return type, a function
632 name, a (possibly empty) argument list (each with optional :ref:`parameter
633 attributes <paramattrs>`), optional :ref:`function attributes <fnattrs>`,
634 an optional section, an optional alignment,
635 an optional :ref:`comdat <langref_comdats>`,
636 an optional :ref:`garbage collector name <gc>`, an optional :ref:`prefix <prefixdata>`, an opening
637 curly brace, a list of basic blocks, and a closing curly brace.
639 LLVM function declarations consist of the "``declare``" keyword, an
640 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
641 style <visibility>`, an optional :ref:`DLL storage class <dllstorageclass>`,
642 an optional :ref:`calling convention <callingconv>`,
643 an optional ``unnamed_addr`` attribute, a return type, an optional
644 :ref:`parameter attribute <paramattrs>` for the return type, a function
645 name, a possibly empty list of arguments, an optional alignment, an optional
646 :ref:`garbage collector name <gc>` and an optional :ref:`prefix <prefixdata>`.
648 A function definition contains a list of basic blocks, forming the CFG (Control
649 Flow Graph) for the function. Each basic block may optionally start with a label
650 (giving the basic block a symbol table entry), contains a list of instructions,
651 and ends with a :ref:`terminator <terminators>` instruction (such as a branch or
652 function return). If an explicit label is not provided, a block is assigned an
653 implicit numbered label, using the next value from the same counter as used for
654 unnamed temporaries (:ref:`see above<identifiers>`). For example, if a function
655 entry block does not have an explicit label, it will be assigned label "%0",
656 then the first unnamed temporary in that block will be "%1", etc.
658 The first basic block in a function is special in two ways: it is
659 immediately executed on entrance to the function, and it is not allowed
660 to have predecessor basic blocks (i.e. there can not be any branches to
661 the entry block of a function). Because the block can have no
662 predecessors, it also cannot have any :ref:`PHI nodes <i_phi>`.
664 LLVM allows an explicit section to be specified for functions. If the
665 target supports it, it will emit functions to the section specified.
666 Additionally, the function can placed in a COMDAT.
668 An explicit alignment may be specified for a function. If not present,
669 or if the alignment is set to zero, the alignment of the function is set
670 by the target to whatever it feels convenient. If an explicit alignment
671 is specified, the function is forced to have at least that much
672 alignment. All alignments must be a power of 2.
674 If the ``unnamed_addr`` attribute is given, the address is know to not
675 be significant and two identical functions can be merged.
679 define [linkage] [visibility] [DLLStorageClass]
681 <ResultType> @<FunctionName> ([argument list])
682 [unnamed_addr] [fn Attrs] [section "name"] [comdat $<ComdatName>]
683 [align N] [gc] [prefix Constant] { ... }
685 The argument list is a comma seperated sequence of arguments where each
686 argument is of the following form
690 <type> [parameter Attrs] [name]
698 Aliases, unlike function or variables, don't create any new data. They
699 are just a new symbol and metadata for an existing position.
701 Aliases have a name and an aliasee that is either a global value or a
704 Aliases may have an optional :ref:`linkage type <linkage>`, an optional
705 :ref:`visibility style <visibility>`, an optional :ref:`DLL storage class
706 <dllstorageclass>` and an optional :ref:`tls model <tls_model>`.
710 @<Name> = [Linkage] [Visibility] [DLLStorageClass] [ThreadLocal] [unnamed_addr] alias <AliaseeTy> @<Aliasee>
712 The linkage must be one of ``private``, ``internal``, ``linkonce``, ``weak``,
713 ``linkonce_odr``, ``weak_odr``, ``external``. Note that some system linkers
714 might not correctly handle dropping a weak symbol that is aliased.
716 Alias that are not ``unnamed_addr`` are guaranteed to have the same address as
717 the aliasee expression. ``unnamed_addr`` ones are only guaranteed to point
720 Since aliases are only a second name, some restrictions apply, of which
721 some can only be checked when producing an object file:
723 * The expression defining the aliasee must be computable at assembly
724 time. Since it is just a name, no relocations can be used.
726 * No alias in the expression can be weak as the possibility of the
727 intermediate alias being overridden cannot be represented in an
730 * No global value in the expression can be a declaration, since that
731 would require a relocation, which is not possible.
738 Comdat IR provides access to COFF and ELF object file COMDAT functionality.
740 Comdats have a name which represents the COMDAT key. All global objects that
741 specify this key will only end up in the final object file if the linker chooses
742 that key over some other key. Aliases are placed in the same COMDAT that their
743 aliasee computes to, if any.
745 Comdats have a selection kind to provide input on how the linker should
746 choose between keys in two different object files.
750 $<Name> = comdat SelectionKind
752 The selection kind must be one of the following:
755 The linker may choose any COMDAT key, the choice is arbitrary.
757 The linker may choose any COMDAT key but the sections must contain the
760 The linker will choose the section containing the largest COMDAT key.
762 The linker requires that only section with this COMDAT key exist.
764 The linker may choose any COMDAT key but the sections must contain the
767 Note that the Mach-O platform doesn't support COMDATs and ELF only supports
768 ``any`` as a selection kind.
770 Here is an example of a COMDAT group where a function will only be selected if
771 the COMDAT key's section is the largest:
775 $foo = comdat largest
776 @foo = global i32 2, comdat $foo
778 define void @bar() comdat $foo {
782 In a COFF object file, this will create a COMDAT section with selection kind
783 ``IMAGE_COMDAT_SELECT_LARGEST`` containing the contents of the ``@foo`` symbol
784 and another COMDAT section with selection kind
785 ``IMAGE_COMDAT_SELECT_ASSOCIATIVE`` which is associated with the first COMDAT
786 section and contains the contents of the ``@bar`` symbol.
788 There are some restrictions on the properties of the global object.
789 It, or an alias to it, must have the same name as the COMDAT group when
791 The contents and size of this object may be used during link-time to determine
792 which COMDAT groups get selected depending on the selection kind.
793 Because the name of the object must match the name of the COMDAT group, the
794 linkage of the global object must not be local; local symbols can get renamed
795 if a collision occurs in the symbol table.
797 The combined use of COMDATS and section attributes may yield surprising results.
804 @g1 = global i32 42, section "sec", comdat $foo
805 @g2 = global i32 42, section "sec", comdat $bar
807 From the object file perspective, this requires the creation of two sections
808 with the same name. This is necessary because both globals belong to different
809 COMDAT groups and COMDATs, at the object file level, are represented by
812 Note that certain IR constructs like global variables and functions may create
813 COMDATs in the object file in addition to any which are specified using COMDAT
814 IR. This arises, for example, when a global variable has linkonce_odr linkage.
816 .. _namedmetadatastructure:
821 Named metadata is a collection of metadata. :ref:`Metadata
822 nodes <metadata>` (but not metadata strings) are the only valid
823 operands for a named metadata.
827 ; Some unnamed metadata nodes, which are referenced by the named metadata.
828 !0 = metadata !{metadata !"zero"}
829 !1 = metadata !{metadata !"one"}
830 !2 = metadata !{metadata !"two"}
832 !name = !{!0, !1, !2}
839 The return type and each parameter of a function type may have a set of
840 *parameter attributes* associated with them. Parameter attributes are
841 used to communicate additional information about the result or
842 parameters of a function. Parameter attributes are considered to be part
843 of the function, not of the function type, so functions with different
844 parameter attributes can have the same function type.
846 Parameter attributes are simple keywords that follow the type specified.
847 If multiple parameter attributes are needed, they are space separated.
852 declare i32 @printf(i8* noalias nocapture, ...)
853 declare i32 @atoi(i8 zeroext)
854 declare signext i8 @returns_signed_char()
856 Note that any attributes for the function result (``nounwind``,
857 ``readonly``) come immediately after the argument list.
859 Currently, only the following parameter attributes are defined:
862 This indicates to the code generator that the parameter or return
863 value should be zero-extended to the extent required by the target's
864 ABI (which is usually 32-bits, but is 8-bits for a i1 on x86-64) by
865 the caller (for a parameter) or the callee (for a return value).
867 This indicates to the code generator that the parameter or return
868 value should be sign-extended to the extent required by the target's
869 ABI (which is usually 32-bits) by the caller (for a parameter) or
870 the callee (for a return value).
872 This indicates that this parameter or return value should be treated
873 in a special target-dependent fashion during while emitting code for
874 a function call or return (usually, by putting it in a register as
875 opposed to memory, though some targets use it to distinguish between
876 two different kinds of registers). Use of this attribute is
879 This indicates that the pointer parameter should really be passed by
880 value to the function. The attribute implies that a hidden copy of
881 the pointee is made between the caller and the callee, so the callee
882 is unable to modify the value in the caller. This attribute is only
883 valid on LLVM pointer arguments. It is generally used to pass
884 structs and arrays by value, but is also valid on pointers to
885 scalars. The copy is considered to belong to the caller not the
886 callee (for example, ``readonly`` functions should not write to
887 ``byval`` parameters). This is not a valid attribute for return
890 The byval attribute also supports specifying an alignment with the
891 align attribute. It indicates the alignment of the stack slot to
892 form and the known alignment of the pointer specified to the call
893 site. If the alignment is not specified, then the code generator
894 makes a target-specific assumption.
900 The ``inalloca`` argument attribute allows the caller to take the
901 address of outgoing stack arguments. An ``inalloca`` argument must
902 be a pointer to stack memory produced by an ``alloca`` instruction.
903 The alloca, or argument allocation, must also be tagged with the
904 inalloca keyword. Only the last argument may have the ``inalloca``
905 attribute, and that argument is guaranteed to be passed in memory.
907 An argument allocation may be used by a call at most once because
908 the call may deallocate it. The ``inalloca`` attribute cannot be
909 used in conjunction with other attributes that affect argument
910 storage, like ``inreg``, ``nest``, ``sret``, or ``byval``. The
911 ``inalloca`` attribute also disables LLVM's implicit lowering of
912 large aggregate return values, which means that frontend authors
913 must lower them with ``sret`` pointers.
915 When the call site is reached, the argument allocation must have
916 been the most recent stack allocation that is still live, or the
917 results are undefined. It is possible to allocate additional stack
918 space after an argument allocation and before its call site, but it
919 must be cleared off with :ref:`llvm.stackrestore
922 See :doc:`InAlloca` for more information on how to use this
926 This indicates that the pointer parameter specifies the address of a
927 structure that is the return value of the function in the source
928 program. This pointer must be guaranteed by the caller to be valid:
929 loads and stores to the structure may be assumed by the callee
930 not to trap and to be properly aligned. This may only be applied to
931 the first parameter. This is not a valid attribute for return
935 This indicates that the pointer value may be assumed by the optimizer to
936 have the specified alignment.
938 Note that this attribute has additional semantics when combined with the
944 This indicates that pointer values :ref:`based <pointeraliasing>` on
945 the argument or return value do not alias pointer values that are
946 not *based* on it, ignoring certain "irrelevant" dependencies. For a
947 call to the parent function, dependencies between memory references
948 from before or after the call and from those during the call are
949 "irrelevant" to the ``noalias`` keyword for the arguments and return
950 value used in that call. The caller shares the responsibility with
951 the callee for ensuring that these requirements are met. For further
952 details, please see the discussion of the NoAlias response in :ref:`alias
953 analysis <Must, May, or No>`.
955 Note that this definition of ``noalias`` is intentionally similar
956 to the definition of ``restrict`` in C99 for function arguments,
957 though it is slightly weaker.
959 For function return values, C99's ``restrict`` is not meaningful,
960 while LLVM's ``noalias`` is.
962 This indicates that the callee does not make any copies of the
963 pointer that outlive the callee itself. This is not a valid
964 attribute for return values.
969 This indicates that the pointer parameter can be excised using the
970 :ref:`trampoline intrinsics <int_trampoline>`. This is not a valid
971 attribute for return values and can only be applied to one parameter.
974 This indicates that the function always returns the argument as its return
975 value. This is an optimization hint to the code generator when generating
976 the caller, allowing tail call optimization and omission of register saves
977 and restores in some cases; it is not checked or enforced when generating
978 the callee. The parameter and the function return type must be valid
979 operands for the :ref:`bitcast instruction <i_bitcast>`. This is not a
980 valid attribute for return values and can only be applied to one parameter.
983 This indicates that the parameter or return pointer is not null. This
984 attribute may only be applied to pointer typed parameters. This is not
985 checked or enforced by LLVM, the caller must ensure that the pointer
986 passed in is non-null, or the callee must ensure that the returned pointer
989 ``dereferenceable(<n>)``
990 This indicates that the parameter or return pointer is dereferenceable. This
991 attribute may only be applied to pointer typed parameters. A pointer that
992 is dereferenceable can be loaded from speculatively without a risk of
993 trapping. The number of bytes known to be dereferenceable must be provided
994 in parentheses. It is legal for the number of bytes to be less than the
995 size of the pointee type. The ``nonnull`` attribute does not imply
996 dereferenceability (consider a pointer to one element past the end of an
997 array), however ``dereferenceable(<n>)`` does imply ``nonnull`` in
998 ``addrspace(0)`` (which is the default address space).
1002 Garbage Collector Names
1003 -----------------------
1005 Each function may specify a garbage collector name, which is simply a
1008 .. code-block:: llvm
1010 define void @f() gc "name" { ... }
1012 The compiler declares the supported values of *name*. Specifying a
1013 collector will cause the compiler to alter its output in order to
1014 support the named garbage collection algorithm.
1021 Prefix data is data associated with a function which the code generator
1022 will emit immediately before the function body. The purpose of this feature
1023 is to allow frontends to associate language-specific runtime metadata with
1024 specific functions and make it available through the function pointer while
1025 still allowing the function pointer to be called. To access the data for a
1026 given function, a program may bitcast the function pointer to a pointer to
1027 the constant's type. This implies that the IR symbol points to the start
1030 To maintain the semantics of ordinary function calls, the prefix data must
1031 have a particular format. Specifically, it must begin with a sequence of
1032 bytes which decode to a sequence of machine instructions, valid for the
1033 module's target, which transfer control to the point immediately succeeding
1034 the prefix data, without performing any other visible action. This allows
1035 the inliner and other passes to reason about the semantics of the function
1036 definition without needing to reason about the prefix data. Obviously this
1037 makes the format of the prefix data highly target dependent.
1039 Prefix data is laid out as if it were an initializer for a global variable
1040 of the prefix data's type. No padding is automatically placed between the
1041 prefix data and the function body. If padding is required, it must be part
1044 A trivial example of valid prefix data for the x86 architecture is ``i8 144``,
1045 which encodes the ``nop`` instruction:
1047 .. code-block:: llvm
1049 define void @f() prefix i8 144 { ... }
1051 Generally prefix data can be formed by encoding a relative branch instruction
1052 which skips the metadata, as in this example of valid prefix data for the
1053 x86_64 architecture, where the first two bytes encode ``jmp .+10``:
1055 .. code-block:: llvm
1057 %0 = type <{ i8, i8, i8* }>
1059 define void @f() prefix %0 <{ i8 235, i8 8, i8* @md}> { ... }
1061 A function may have prefix data but no body. This has similar semantics
1062 to the ``available_externally`` linkage in that the data may be used by the
1063 optimizers but will not be emitted in the object file.
1070 Attribute groups are groups of attributes that are referenced by objects within
1071 the IR. They are important for keeping ``.ll`` files readable, because a lot of
1072 functions will use the same set of attributes. In the degenerative case of a
1073 ``.ll`` file that corresponds to a single ``.c`` file, the single attribute
1074 group will capture the important command line flags used to build that file.
1076 An attribute group is a module-level object. To use an attribute group, an
1077 object references the attribute group's ID (e.g. ``#37``). An object may refer
1078 to more than one attribute group. In that situation, the attributes from the
1079 different groups are merged.
1081 Here is an example of attribute groups for a function that should always be
1082 inlined, has a stack alignment of 4, and which shouldn't use SSE instructions:
1084 .. code-block:: llvm
1086 ; Target-independent attributes:
1087 attributes #0 = { alwaysinline alignstack=4 }
1089 ; Target-dependent attributes:
1090 attributes #1 = { "no-sse" }
1092 ; Function @f has attributes: alwaysinline, alignstack=4, and "no-sse".
1093 define void @f() #0 #1 { ... }
1100 Function attributes are set to communicate additional information about
1101 a function. Function attributes are considered to be part of the
1102 function, not of the function type, so functions with different function
1103 attributes can have the same function type.
1105 Function attributes are simple keywords that follow the type specified.
1106 If multiple attributes are needed, they are space separated. For
1109 .. code-block:: llvm
1111 define void @f() noinline { ... }
1112 define void @f() alwaysinline { ... }
1113 define void @f() alwaysinline optsize { ... }
1114 define void @f() optsize { ... }
1117 This attribute indicates that, when emitting the prologue and
1118 epilogue, the backend should forcibly align the stack pointer.
1119 Specify the desired alignment, which must be a power of two, in
1122 This attribute indicates that the inliner should attempt to inline
1123 this function into callers whenever possible, ignoring any active
1124 inlining size threshold for this caller.
1126 This indicates that the callee function at a call site should be
1127 recognized as a built-in function, even though the function's declaration
1128 uses the ``nobuiltin`` attribute. This is only valid at call sites for
1129 direct calls to functions that are declared with the ``nobuiltin``
1132 This attribute indicates that this function is rarely called. When
1133 computing edge weights, basic blocks post-dominated by a cold
1134 function call are also considered to be cold; and, thus, given low
1137 This attribute indicates that the source code contained a hint that
1138 inlining this function is desirable (such as the "inline" keyword in
1139 C/C++). It is just a hint; it imposes no requirements on the
1142 This attribute indicates that the function should be added to a
1143 jump-instruction table at code-generation time, and that all address-taken
1144 references to this function should be replaced with a reference to the
1145 appropriate jump-instruction-table function pointer. Note that this creates
1146 a new pointer for the original function, which means that code that depends
1147 on function-pointer identity can break. So, any function annotated with
1148 ``jumptable`` must also be ``unnamed_addr``.
1150 This attribute suggests that optimization passes and code generator
1151 passes make choices that keep the code size of this function as small
1152 as possible and perform optimizations that may sacrifice runtime
1153 performance in order to minimize the size of the generated code.
1155 This attribute disables prologue / epilogue emission for the
1156 function. This can have very system-specific consequences.
1158 This indicates that the callee function at a call site is not recognized as
1159 a built-in function. LLVM will retain the original call and not replace it
1160 with equivalent code based on the semantics of the built-in function, unless
1161 the call site uses the ``builtin`` attribute. This is valid at call sites
1162 and on function declarations and definitions.
1164 This attribute indicates that calls to the function cannot be
1165 duplicated. A call to a ``noduplicate`` function may be moved
1166 within its parent function, but may not be duplicated within
1167 its parent function.
1169 A function containing a ``noduplicate`` call may still
1170 be an inlining candidate, provided that the call is not
1171 duplicated by inlining. That implies that the function has
1172 internal linkage and only has one call site, so the original
1173 call is dead after inlining.
1175 This attributes disables implicit floating point instructions.
1177 This attribute indicates that the inliner should never inline this
1178 function in any situation. This attribute may not be used together
1179 with the ``alwaysinline`` attribute.
1181 This attribute suppresses lazy symbol binding for the function. This
1182 may make calls to the function faster, at the cost of extra program
1183 startup time if the function is not called during program startup.
1185 This attribute indicates that the code generator should not use a
1186 red zone, even if the target-specific ABI normally permits it.
1188 This function attribute indicates that the function never returns
1189 normally. This produces undefined behavior at runtime if the
1190 function ever does dynamically return.
1192 This function attribute indicates that the function never returns
1193 with an unwind or exceptional control flow. If the function does
1194 unwind, its runtime behavior is undefined.
1196 This function attribute indicates that the function is not optimized
1197 by any optimization or code generator passes with the
1198 exception of interprocedural optimization passes.
1199 This attribute cannot be used together with the ``alwaysinline``
1200 attribute; this attribute is also incompatible
1201 with the ``minsize`` attribute and the ``optsize`` attribute.
1203 This attribute requires the ``noinline`` attribute to be specified on
1204 the function as well, so the function is never inlined into any caller.
1205 Only functions with the ``alwaysinline`` attribute are valid
1206 candidates for inlining into the body of this function.
1208 This attribute suggests that optimization passes and code generator
1209 passes make choices that keep the code size of this function low,
1210 and otherwise do optimizations specifically to reduce code size as
1211 long as they do not significantly impact runtime performance.
1213 On a function, this attribute indicates that the function computes its
1214 result (or decides to unwind an exception) based strictly on its arguments,
1215 without dereferencing any pointer arguments or otherwise accessing
1216 any mutable state (e.g. memory, control registers, etc) visible to
1217 caller functions. It does not write through any pointer arguments
1218 (including ``byval`` arguments) and never changes any state visible
1219 to callers. This means that it cannot unwind exceptions by calling
1220 the ``C++`` exception throwing methods.
1222 On an argument, this attribute indicates that the function does not
1223 dereference that pointer argument, even though it may read or write the
1224 memory that the pointer points to if accessed through other pointers.
1226 On a function, this attribute indicates that the function does not write
1227 through any pointer arguments (including ``byval`` arguments) or otherwise
1228 modify any state (e.g. memory, control registers, etc) visible to
1229 caller functions. It may dereference pointer arguments and read
1230 state that may be set in the caller. A readonly function always
1231 returns the same value (or unwinds an exception identically) when
1232 called with the same set of arguments and global state. It cannot
1233 unwind an exception by calling the ``C++`` exception throwing
1236 On an argument, this attribute indicates that the function does not write
1237 through this pointer argument, even though it may write to the memory that
1238 the pointer points to.
1240 This attribute indicates that this function can return twice. The C
1241 ``setjmp`` is an example of such a function. The compiler disables
1242 some optimizations (like tail calls) in the caller of these
1244 ``sanitize_address``
1245 This attribute indicates that AddressSanitizer checks
1246 (dynamic address safety analysis) are enabled for this function.
1248 This attribute indicates that MemorySanitizer checks (dynamic detection
1249 of accesses to uninitialized memory) are enabled for this function.
1251 This attribute indicates that ThreadSanitizer checks
1252 (dynamic thread safety analysis) are enabled for this function.
1254 This attribute indicates that the function should emit a stack
1255 smashing protector. It is in the form of a "canary" --- a random value
1256 placed on the stack before the local variables that's checked upon
1257 return from the function to see if it has been overwritten. A
1258 heuristic is used to determine if a function needs stack protectors
1259 or not. The heuristic used will enable protectors for functions with:
1261 - Character arrays larger than ``ssp-buffer-size`` (default 8).
1262 - Aggregates containing character arrays larger than ``ssp-buffer-size``.
1263 - Calls to alloca() with variable sizes or constant sizes greater than
1264 ``ssp-buffer-size``.
1266 Variables that are identified as requiring a protector will be arranged
1267 on the stack such that they are adjacent to the stack protector guard.
1269 If a function that has an ``ssp`` attribute is inlined into a
1270 function that doesn't have an ``ssp`` attribute, then the resulting
1271 function will have an ``ssp`` attribute.
1273 This attribute indicates that the function should *always* emit a
1274 stack smashing protector. This overrides the ``ssp`` function
1277 Variables that are identified as requiring a protector will be arranged
1278 on the stack such that they are adjacent to the stack protector guard.
1279 The specific layout rules are:
1281 #. Large arrays and structures containing large arrays
1282 (``>= ssp-buffer-size``) are closest to the stack protector.
1283 #. Small arrays and structures containing small arrays
1284 (``< ssp-buffer-size``) are 2nd closest to the protector.
1285 #. Variables that have had their address taken are 3rd closest to the
1288 If a function that has an ``sspreq`` attribute is inlined into a
1289 function that doesn't have an ``sspreq`` attribute or which has an
1290 ``ssp`` or ``sspstrong`` attribute, then the resulting function will have
1291 an ``sspreq`` attribute.
1293 This attribute indicates that the function should emit a stack smashing
1294 protector. This attribute causes a strong heuristic to be used when
1295 determining if a function needs stack protectors. The strong heuristic
1296 will enable protectors for functions with:
1298 - Arrays of any size and type
1299 - Aggregates containing an array of any size and type.
1300 - Calls to alloca().
1301 - Local variables that have had their address taken.
1303 Variables that are identified as requiring a protector will be arranged
1304 on the stack such that they are adjacent to the stack protector guard.
1305 The specific layout rules are:
1307 #. Large arrays and structures containing large arrays
1308 (``>= ssp-buffer-size``) are closest to the stack protector.
1309 #. Small arrays and structures containing small arrays
1310 (``< ssp-buffer-size``) are 2nd closest to the protector.
1311 #. Variables that have had their address taken are 3rd closest to the
1314 This overrides the ``ssp`` function attribute.
1316 If a function that has an ``sspstrong`` attribute is inlined into a
1317 function that doesn't have an ``sspstrong`` attribute, then the
1318 resulting function will have an ``sspstrong`` attribute.
1320 This attribute indicates that the ABI being targeted requires that
1321 an unwind table entry be produce for this function even if we can
1322 show that no exceptions passes by it. This is normally the case for
1323 the ELF x86-64 abi, but it can be disabled for some compilation
1328 Module-Level Inline Assembly
1329 ----------------------------
1331 Modules may contain "module-level inline asm" blocks, which corresponds
1332 to the GCC "file scope inline asm" blocks. These blocks are internally
1333 concatenated by LLVM and treated as a single unit, but may be separated
1334 in the ``.ll`` file if desired. The syntax is very simple:
1336 .. code-block:: llvm
1338 module asm "inline asm code goes here"
1339 module asm "more can go here"
1341 The strings can contain any character by escaping non-printable
1342 characters. The escape sequence used is simply "\\xx" where "xx" is the
1343 two digit hex code for the number.
1345 The inline asm code is simply printed to the machine code .s file when
1346 assembly code is generated.
1348 .. _langref_datalayout:
1353 A module may specify a target specific data layout string that specifies
1354 how data is to be laid out in memory. The syntax for the data layout is
1357 .. code-block:: llvm
1359 target datalayout = "layout specification"
1361 The *layout specification* consists of a list of specifications
1362 separated by the minus sign character ('-'). Each specification starts
1363 with a letter and may include other information after the letter to
1364 define some aspect of the data layout. The specifications accepted are
1368 Specifies that the target lays out data in big-endian form. That is,
1369 the bits with the most significance have the lowest address
1372 Specifies that the target lays out data in little-endian form. That
1373 is, the bits with the least significance have the lowest address
1376 Specifies the natural alignment of the stack in bits. Alignment
1377 promotion of stack variables is limited to the natural stack
1378 alignment to avoid dynamic stack realignment. The stack alignment
1379 must be a multiple of 8-bits. If omitted, the natural stack
1380 alignment defaults to "unspecified", which does not prevent any
1381 alignment promotions.
1382 ``p[n]:<size>:<abi>:<pref>``
1383 This specifies the *size* of a pointer and its ``<abi>`` and
1384 ``<pref>``\erred alignments for address space ``n``. All sizes are in
1385 bits. The address space, ``n`` is optional, and if not specified,
1386 denotes the default address space 0. The value of ``n`` must be
1387 in the range [1,2^23).
1388 ``i<size>:<abi>:<pref>``
1389 This specifies the alignment for an integer type of a given bit
1390 ``<size>``. The value of ``<size>`` must be in the range [1,2^23).
1391 ``v<size>:<abi>:<pref>``
1392 This specifies the alignment for a vector type of a given bit
1394 ``f<size>:<abi>:<pref>``
1395 This specifies the alignment for a floating point type of a given bit
1396 ``<size>``. Only values of ``<size>`` that are supported by the target
1397 will work. 32 (float) and 64 (double) are supported on all targets; 80
1398 or 128 (different flavors of long double) are also supported on some
1401 This specifies the alignment for an object of aggregate type.
1403 If present, specifies that llvm names are mangled in the output. The
1406 * ``e``: ELF mangling: Private symbols get a ``.L`` prefix.
1407 * ``m``: Mips mangling: Private symbols get a ``$`` prefix.
1408 * ``o``: Mach-O mangling: Private symbols get ``L`` prefix. Other
1409 symbols get a ``_`` prefix.
1410 * ``w``: Windows COFF prefix: Similar to Mach-O, but stdcall and fastcall
1411 functions also get a suffix based on the frame size.
1412 ``n<size1>:<size2>:<size3>...``
1413 This specifies a set of native integer widths for the target CPU in
1414 bits. For example, it might contain ``n32`` for 32-bit PowerPC,
1415 ``n32:64`` for PowerPC 64, or ``n8:16:32:64`` for X86-64. Elements of
1416 this set are considered to support most general arithmetic operations
1419 On every specification that takes a ``<abi>:<pref>``, specifying the
1420 ``<pref>`` alignment is optional. If omitted, the preceding ``:``
1421 should be omitted too and ``<pref>`` will be equal to ``<abi>``.
1423 When constructing the data layout for a given target, LLVM starts with a
1424 default set of specifications which are then (possibly) overridden by
1425 the specifications in the ``datalayout`` keyword. The default
1426 specifications are given in this list:
1428 - ``E`` - big endian
1429 - ``p:64:64:64`` - 64-bit pointers with 64-bit alignment.
1430 - ``p[n]:64:64:64`` - Other address spaces are assumed to be the
1431 same as the default address space.
1432 - ``S0`` - natural stack alignment is unspecified
1433 - ``i1:8:8`` - i1 is 8-bit (byte) aligned
1434 - ``i8:8:8`` - i8 is 8-bit (byte) aligned
1435 - ``i16:16:16`` - i16 is 16-bit aligned
1436 - ``i32:32:32`` - i32 is 32-bit aligned
1437 - ``i64:32:64`` - i64 has ABI alignment of 32-bits but preferred
1438 alignment of 64-bits
1439 - ``f16:16:16`` - half is 16-bit aligned
1440 - ``f32:32:32`` - float is 32-bit aligned
1441 - ``f64:64:64`` - double is 64-bit aligned
1442 - ``f128:128:128`` - quad is 128-bit aligned
1443 - ``v64:64:64`` - 64-bit vector is 64-bit aligned
1444 - ``v128:128:128`` - 128-bit vector is 128-bit aligned
1445 - ``a:0:64`` - aggregates are 64-bit aligned
1447 When LLVM is determining the alignment for a given type, it uses the
1450 #. If the type sought is an exact match for one of the specifications,
1451 that specification is used.
1452 #. If no match is found, and the type sought is an integer type, then
1453 the smallest integer type that is larger than the bitwidth of the
1454 sought type is used. If none of the specifications are larger than
1455 the bitwidth then the largest integer type is used. For example,
1456 given the default specifications above, the i7 type will use the
1457 alignment of i8 (next largest) while both i65 and i256 will use the
1458 alignment of i64 (largest specified).
1459 #. If no match is found, and the type sought is a vector type, then the
1460 largest vector type that is smaller than the sought vector type will
1461 be used as a fall back. This happens because <128 x double> can be
1462 implemented in terms of 64 <2 x double>, for example.
1464 The function of the data layout string may not be what you expect.
1465 Notably, this is not a specification from the frontend of what alignment
1466 the code generator should use.
1468 Instead, if specified, the target data layout is required to match what
1469 the ultimate *code generator* expects. This string is used by the
1470 mid-level optimizers to improve code, and this only works if it matches
1471 what the ultimate code generator uses. If you would like to generate IR
1472 that does not embed this target-specific detail into the IR, then you
1473 don't have to specify the string. This will disable some optimizations
1474 that require precise layout information, but this also prevents those
1475 optimizations from introducing target specificity into the IR.
1482 A module may specify a target triple string that describes the target
1483 host. The syntax for the target triple is simply:
1485 .. code-block:: llvm
1487 target triple = "x86_64-apple-macosx10.7.0"
1489 The *target triple* string consists of a series of identifiers delimited
1490 by the minus sign character ('-'). The canonical forms are:
1494 ARCHITECTURE-VENDOR-OPERATING_SYSTEM
1495 ARCHITECTURE-VENDOR-OPERATING_SYSTEM-ENVIRONMENT
1497 This information is passed along to the backend so that it generates
1498 code for the proper architecture. It's possible to override this on the
1499 command line with the ``-mtriple`` command line option.
1501 .. _pointeraliasing:
1503 Pointer Aliasing Rules
1504 ----------------------
1506 Any memory access must be done through a pointer value associated with
1507 an address range of the memory access, otherwise the behavior is
1508 undefined. Pointer values are associated with address ranges according
1509 to the following rules:
1511 - A pointer value is associated with the addresses associated with any
1512 value it is *based* on.
1513 - An address of a global variable is associated with the address range
1514 of the variable's storage.
1515 - The result value of an allocation instruction is associated with the
1516 address range of the allocated storage.
1517 - A null pointer in the default address-space is associated with no
1519 - An integer constant other than zero or a pointer value returned from
1520 a function not defined within LLVM may be associated with address
1521 ranges allocated through mechanisms other than those provided by
1522 LLVM. Such ranges shall not overlap with any ranges of addresses
1523 allocated by mechanisms provided by LLVM.
1525 A pointer value is *based* on another pointer value according to the
1528 - A pointer value formed from a ``getelementptr`` operation is *based*
1529 on the first operand of the ``getelementptr``.
1530 - The result value of a ``bitcast`` is *based* on the operand of the
1532 - A pointer value formed by an ``inttoptr`` is *based* on all pointer
1533 values that contribute (directly or indirectly) to the computation of
1534 the pointer's value.
1535 - The "*based* on" relationship is transitive.
1537 Note that this definition of *"based"* is intentionally similar to the
1538 definition of *"based"* in C99, though it is slightly weaker.
1540 LLVM IR does not associate types with memory. The result type of a
1541 ``load`` merely indicates the size and alignment of the memory from
1542 which to load, as well as the interpretation of the value. The first
1543 operand type of a ``store`` similarly only indicates the size and
1544 alignment of the store.
1546 Consequently, type-based alias analysis, aka TBAA, aka
1547 ``-fstrict-aliasing``, is not applicable to general unadorned LLVM IR.
1548 :ref:`Metadata <metadata>` may be used to encode additional information
1549 which specialized optimization passes may use to implement type-based
1554 Volatile Memory Accesses
1555 ------------------------
1557 Certain memory accesses, such as :ref:`load <i_load>`'s,
1558 :ref:`store <i_store>`'s, and :ref:`llvm.memcpy <int_memcpy>`'s may be
1559 marked ``volatile``. The optimizers must not change the number of
1560 volatile operations or change their order of execution relative to other
1561 volatile operations. The optimizers *may* change the order of volatile
1562 operations relative to non-volatile operations. This is not Java's
1563 "volatile" and has no cross-thread synchronization behavior.
1565 IR-level volatile loads and stores cannot safely be optimized into
1566 llvm.memcpy or llvm.memmove intrinsics even when those intrinsics are
1567 flagged volatile. Likewise, the backend should never split or merge
1568 target-legal volatile load/store instructions.
1570 .. admonition:: Rationale
1572 Platforms may rely on volatile loads and stores of natively supported
1573 data width to be executed as single instruction. For example, in C
1574 this holds for an l-value of volatile primitive type with native
1575 hardware support, but not necessarily for aggregate types. The
1576 frontend upholds these expectations, which are intentionally
1577 unspecified in the IR. The rules above ensure that IR transformation
1578 do not violate the frontend's contract with the language.
1582 Memory Model for Concurrent Operations
1583 --------------------------------------
1585 The LLVM IR does not define any way to start parallel threads of
1586 execution or to register signal handlers. Nonetheless, there are
1587 platform-specific ways to create them, and we define LLVM IR's behavior
1588 in their presence. This model is inspired by the C++0x memory model.
1590 For a more informal introduction to this model, see the :doc:`Atomics`.
1592 We define a *happens-before* partial order as the least partial order
1595 - Is a superset of single-thread program order, and
1596 - When a *synchronizes-with* ``b``, includes an edge from ``a`` to
1597 ``b``. *Synchronizes-with* pairs are introduced by platform-specific
1598 techniques, like pthread locks, thread creation, thread joining,
1599 etc., and by atomic instructions. (See also :ref:`Atomic Memory Ordering
1600 Constraints <ordering>`).
1602 Note that program order does not introduce *happens-before* edges
1603 between a thread and signals executing inside that thread.
1605 Every (defined) read operation (load instructions, memcpy, atomic
1606 loads/read-modify-writes, etc.) R reads a series of bytes written by
1607 (defined) write operations (store instructions, atomic
1608 stores/read-modify-writes, memcpy, etc.). For the purposes of this
1609 section, initialized globals are considered to have a write of the
1610 initializer which is atomic and happens before any other read or write
1611 of the memory in question. For each byte of a read R, R\ :sub:`byte`
1612 may see any write to the same byte, except:
1614 - If write\ :sub:`1` happens before write\ :sub:`2`, and
1615 write\ :sub:`2` happens before R\ :sub:`byte`, then
1616 R\ :sub:`byte` does not see write\ :sub:`1`.
1617 - If R\ :sub:`byte` happens before write\ :sub:`3`, then
1618 R\ :sub:`byte` does not see write\ :sub:`3`.
1620 Given that definition, R\ :sub:`byte` is defined as follows:
1622 - If R is volatile, the result is target-dependent. (Volatile is
1623 supposed to give guarantees which can support ``sig_atomic_t`` in
1624 C/C++, and may be used for accesses to addresses that do not behave
1625 like normal memory. It does not generally provide cross-thread
1627 - Otherwise, if there is no write to the same byte that happens before
1628 R\ :sub:`byte`, R\ :sub:`byte` returns ``undef`` for that byte.
1629 - Otherwise, if R\ :sub:`byte` may see exactly one write,
1630 R\ :sub:`byte` returns the value written by that write.
1631 - Otherwise, if R is atomic, and all the writes R\ :sub:`byte` may
1632 see are atomic, it chooses one of the values written. See the :ref:`Atomic
1633 Memory Ordering Constraints <ordering>` section for additional
1634 constraints on how the choice is made.
1635 - Otherwise R\ :sub:`byte` returns ``undef``.
1637 R returns the value composed of the series of bytes it read. This
1638 implies that some bytes within the value may be ``undef`` **without**
1639 the entire value being ``undef``. Note that this only defines the
1640 semantics of the operation; it doesn't mean that targets will emit more
1641 than one instruction to read the series of bytes.
1643 Note that in cases where none of the atomic intrinsics are used, this
1644 model places only one restriction on IR transformations on top of what
1645 is required for single-threaded execution: introducing a store to a byte
1646 which might not otherwise be stored is not allowed in general.
1647 (Specifically, in the case where another thread might write to and read
1648 from an address, introducing a store can change a load that may see
1649 exactly one write into a load that may see multiple writes.)
1653 Atomic Memory Ordering Constraints
1654 ----------------------------------
1656 Atomic instructions (:ref:`cmpxchg <i_cmpxchg>`,
1657 :ref:`atomicrmw <i_atomicrmw>`, :ref:`fence <i_fence>`,
1658 :ref:`atomic load <i_load>`, and :ref:`atomic store <i_store>`) take
1659 ordering parameters that determine which other atomic instructions on
1660 the same address they *synchronize with*. These semantics are borrowed
1661 from Java and C++0x, but are somewhat more colloquial. If these
1662 descriptions aren't precise enough, check those specs (see spec
1663 references in the :doc:`atomics guide <Atomics>`).
1664 :ref:`fence <i_fence>` instructions treat these orderings somewhat
1665 differently since they don't take an address. See that instruction's
1666 documentation for details.
1668 For a simpler introduction to the ordering constraints, see the
1672 The set of values that can be read is governed by the happens-before
1673 partial order. A value cannot be read unless some operation wrote
1674 it. This is intended to provide a guarantee strong enough to model
1675 Java's non-volatile shared variables. This ordering cannot be
1676 specified for read-modify-write operations; it is not strong enough
1677 to make them atomic in any interesting way.
1679 In addition to the guarantees of ``unordered``, there is a single
1680 total order for modifications by ``monotonic`` operations on each
1681 address. All modification orders must be compatible with the
1682 happens-before order. There is no guarantee that the modification
1683 orders can be combined to a global total order for the whole program
1684 (and this often will not be possible). The read in an atomic
1685 read-modify-write operation (:ref:`cmpxchg <i_cmpxchg>` and
1686 :ref:`atomicrmw <i_atomicrmw>`) reads the value in the modification
1687 order immediately before the value it writes. If one atomic read
1688 happens before another atomic read of the same address, the later
1689 read must see the same value or a later value in the address's
1690 modification order. This disallows reordering of ``monotonic`` (or
1691 stronger) operations on the same address. If an address is written
1692 ``monotonic``-ally by one thread, and other threads ``monotonic``-ally
1693 read that address repeatedly, the other threads must eventually see
1694 the write. This corresponds to the C++0x/C1x
1695 ``memory_order_relaxed``.
1697 In addition to the guarantees of ``monotonic``, a
1698 *synchronizes-with* edge may be formed with a ``release`` operation.
1699 This is intended to model C++'s ``memory_order_acquire``.
1701 In addition to the guarantees of ``monotonic``, if this operation
1702 writes a value which is subsequently read by an ``acquire``
1703 operation, it *synchronizes-with* that operation. (This isn't a
1704 complete description; see the C++0x definition of a release
1705 sequence.) This corresponds to the C++0x/C1x
1706 ``memory_order_release``.
1707 ``acq_rel`` (acquire+release)
1708 Acts as both an ``acquire`` and ``release`` operation on its
1709 address. This corresponds to the C++0x/C1x ``memory_order_acq_rel``.
1710 ``seq_cst`` (sequentially consistent)
1711 In addition to the guarantees of ``acq_rel`` (``acquire`` for an
1712 operation that only reads, ``release`` for an operation that only
1713 writes), there is a global total order on all
1714 sequentially-consistent operations on all addresses, which is
1715 consistent with the *happens-before* partial order and with the
1716 modification orders of all the affected addresses. Each
1717 sequentially-consistent read sees the last preceding write to the
1718 same address in this global order. This corresponds to the C++0x/C1x
1719 ``memory_order_seq_cst`` and Java volatile.
1723 If an atomic operation is marked ``singlethread``, it only *synchronizes
1724 with* or participates in modification and seq\_cst total orderings with
1725 other operations running in the same thread (for example, in signal
1733 LLVM IR floating-point binary ops (:ref:`fadd <i_fadd>`,
1734 :ref:`fsub <i_fsub>`, :ref:`fmul <i_fmul>`, :ref:`fdiv <i_fdiv>`,
1735 :ref:`frem <i_frem>`) have the following flags that can set to enable
1736 otherwise unsafe floating point operations
1739 No NaNs - Allow optimizations to assume the arguments and result are not
1740 NaN. Such optimizations are required to retain defined behavior over
1741 NaNs, but the value of the result is undefined.
1744 No Infs - Allow optimizations to assume the arguments and result are not
1745 +/-Inf. Such optimizations are required to retain defined behavior over
1746 +/-Inf, but the value of the result is undefined.
1749 No Signed Zeros - Allow optimizations to treat the sign of a zero
1750 argument or result as insignificant.
1753 Allow Reciprocal - Allow optimizations to use the reciprocal of an
1754 argument rather than perform division.
1757 Fast - Allow algebraically equivalent transformations that may
1758 dramatically change results in floating point (e.g. reassociate). This
1759 flag implies all the others.
1763 Use-list Order Directives
1764 -------------------------
1766 Use-list directives encode the in-memory order of each use-list, allowing the
1767 order to be recreated. ``<order-indexes>`` is a comma-separated list of
1768 indexes that are assigned to the referenced value's uses. The referenced
1769 value's use-list is immediately sorted by these indexes.
1771 Use-list directives may appear at function scope or global scope. They are not
1772 instructions, and have no effect on the semantics of the IR. When they're at
1773 function scope, they must appear after the terminator of the final basic block.
1775 If basic blocks have their address taken via ``blockaddress()`` expressions,
1776 ``uselistorder_bb`` can be used to reorder their use-lists from outside their
1783 uselistorder <ty> <value>, { <order-indexes> }
1784 uselistorder_bb @function, %block { <order-indexes> }
1790 define void @foo(i32 %arg1, i32 %arg2) {
1792 ; ... instructions ...
1794 ; ... instructions ...
1796 ; At function scope.
1797 uselistorder i32 %arg1, { 1, 0, 2 }
1798 uselistorder label %bb, { 1, 0 }
1802 uselistorder i32* @global, { 1, 2, 0 }
1803 uselistorder i32 7, { 1, 0 }
1804 uselistorder i32 (i32) @bar, { 1, 0 }
1805 uselistorder_bb @foo, %bb, { 5, 1, 3, 2, 0, 4 }
1812 The LLVM type system is one of the most important features of the
1813 intermediate representation. Being typed enables a number of
1814 optimizations to be performed on the intermediate representation
1815 directly, without having to do extra analyses on the side before the
1816 transformation. A strong type system makes it easier to read the
1817 generated code and enables novel analyses and transformations that are
1818 not feasible to perform on normal three address code representations.
1828 The void type does not represent any value and has no size.
1846 The function type can be thought of as a function signature. It consists of a
1847 return type and a list of formal parameter types. The return type of a function
1848 type is a void type or first class type --- except for :ref:`label <t_label>`
1849 and :ref:`metadata <t_metadata>` types.
1855 <returntype> (<parameter list>)
1857 ...where '``<parameter list>``' is a comma-separated list of type
1858 specifiers. Optionally, the parameter list may include a type ``...``, which
1859 indicates that the function takes a variable number of arguments. Variable
1860 argument functions can access their arguments with the :ref:`variable argument
1861 handling intrinsic <int_varargs>` functions. '``<returntype>``' is any type
1862 except :ref:`label <t_label>` and :ref:`metadata <t_metadata>`.
1866 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1867 | ``i32 (i32)`` | function taking an ``i32``, returning an ``i32`` |
1868 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1869 | ``float (i16, i32 *) *`` | :ref:`Pointer <t_pointer>` to a function that takes an ``i16`` and a :ref:`pointer <t_pointer>` to ``i32``, returning ``float``. |
1870 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1871 | ``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. |
1872 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1873 | ``{i32, i32} (i32)`` | A function taking an ``i32``, returning a :ref:`structure <t_struct>` containing two ``i32`` values |
1874 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1881 The :ref:`first class <t_firstclass>` types are perhaps the most important.
1882 Values of these types are the only ones which can be produced by
1890 These are the types that are valid in registers from CodeGen's perspective.
1899 The integer type is a very simple type that simply specifies an
1900 arbitrary bit width for the integer type desired. Any bit width from 1
1901 bit to 2\ :sup:`23`\ -1 (about 8 million) can be specified.
1909 The number of bits the integer will occupy is specified by the ``N``
1915 +----------------+------------------------------------------------+
1916 | ``i1`` | a single-bit integer. |
1917 +----------------+------------------------------------------------+
1918 | ``i32`` | a 32-bit integer. |
1919 +----------------+------------------------------------------------+
1920 | ``i1942652`` | a really big integer of over 1 million bits. |
1921 +----------------+------------------------------------------------+
1925 Floating Point Types
1926 """"""""""""""""""""
1935 - 16-bit floating point value
1938 - 32-bit floating point value
1941 - 64-bit floating point value
1944 - 128-bit floating point value (112-bit mantissa)
1947 - 80-bit floating point value (X87)
1950 - 128-bit floating point value (two 64-bits)
1957 The x86_mmx type represents a value held in an MMX register on an x86
1958 machine. The operations allowed on it are quite limited: parameters and
1959 return values, load and store, and bitcast. User-specified MMX
1960 instructions are represented as intrinsic or asm calls with arguments
1961 and/or results of this type. There are no arrays, vectors or constants
1978 The pointer type is used to specify memory locations. Pointers are
1979 commonly used to reference objects in memory.
1981 Pointer types may have an optional address space attribute defining the
1982 numbered address space where the pointed-to object resides. The default
1983 address space is number zero. The semantics of non-zero address spaces
1984 are target-specific.
1986 Note that LLVM does not permit pointers to void (``void*``) nor does it
1987 permit pointers to labels (``label*``). Use ``i8*`` instead.
1997 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1998 | ``[4 x i32]*`` | A :ref:`pointer <t_pointer>` to :ref:`array <t_array>` of four ``i32`` values. |
1999 +-------------------------+--------------------------------------------------------------------------------------------------------------+
2000 | ``i32 (i32*) *`` | A :ref:`pointer <t_pointer>` to a :ref:`function <t_function>` that takes an ``i32*``, returning an ``i32``. |
2001 +-------------------------+--------------------------------------------------------------------------------------------------------------+
2002 | ``i32 addrspace(5)*`` | A :ref:`pointer <t_pointer>` to an ``i32`` value that resides in address space #5. |
2003 +-------------------------+--------------------------------------------------------------------------------------------------------------+
2012 A vector type is a simple derived type that represents a vector of
2013 elements. Vector types are used when multiple primitive data are
2014 operated in parallel using a single instruction (SIMD). A vector type
2015 requires a size (number of elements) and an underlying primitive data
2016 type. Vector types are considered :ref:`first class <t_firstclass>`.
2022 < <# elements> x <elementtype> >
2024 The number of elements is a constant integer value larger than 0;
2025 elementtype may be any integer, floating point or pointer type. Vectors
2026 of size zero are not allowed.
2030 +-------------------+--------------------------------------------------+
2031 | ``<4 x i32>`` | Vector of 4 32-bit integer values. |
2032 +-------------------+--------------------------------------------------+
2033 | ``<8 x float>`` | Vector of 8 32-bit floating-point values. |
2034 +-------------------+--------------------------------------------------+
2035 | ``<2 x i64>`` | Vector of 2 64-bit integer values. |
2036 +-------------------+--------------------------------------------------+
2037 | ``<4 x i64*>`` | Vector of 4 pointers to 64-bit integer values. |
2038 +-------------------+--------------------------------------------------+
2047 The label type represents code labels.
2062 The metadata type represents embedded metadata. No derived types may be
2063 created from metadata except for :ref:`function <t_function>` arguments.
2076 Aggregate Types are a subset of derived types that can contain multiple
2077 member types. :ref:`Arrays <t_array>` and :ref:`structs <t_struct>` are
2078 aggregate types. :ref:`Vectors <t_vector>` are not considered to be
2088 The array type is a very simple derived type that arranges elements
2089 sequentially in memory. The array type requires a size (number of
2090 elements) and an underlying data type.
2096 [<# elements> x <elementtype>]
2098 The number of elements is a constant integer value; ``elementtype`` may
2099 be any type with a size.
2103 +------------------+--------------------------------------+
2104 | ``[40 x i32]`` | Array of 40 32-bit integer values. |
2105 +------------------+--------------------------------------+
2106 | ``[41 x i32]`` | Array of 41 32-bit integer values. |
2107 +------------------+--------------------------------------+
2108 | ``[4 x i8]`` | Array of 4 8-bit integer values. |
2109 +------------------+--------------------------------------+
2111 Here are some examples of multidimensional arrays:
2113 +-----------------------------+----------------------------------------------------------+
2114 | ``[3 x [4 x i32]]`` | 3x4 array of 32-bit integer values. |
2115 +-----------------------------+----------------------------------------------------------+
2116 | ``[12 x [10 x float]]`` | 12x10 array of single precision floating point values. |
2117 +-----------------------------+----------------------------------------------------------+
2118 | ``[2 x [3 x [4 x i16]]]`` | 2x3x4 array of 16-bit integer values. |
2119 +-----------------------------+----------------------------------------------------------+
2121 There is no restriction on indexing beyond the end of the array implied
2122 by a static type (though there are restrictions on indexing beyond the
2123 bounds of an allocated object in some cases). This means that
2124 single-dimension 'variable sized array' addressing can be implemented in
2125 LLVM with a zero length array type. An implementation of 'pascal style
2126 arrays' in LLVM could use the type "``{ i32, [0 x float]}``", for
2136 The structure type is used to represent a collection of data members
2137 together in memory. The elements of a structure may be any type that has
2140 Structures in memory are accessed using '``load``' and '``store``' by
2141 getting a pointer to a field with the '``getelementptr``' instruction.
2142 Structures in registers are accessed using the '``extractvalue``' and
2143 '``insertvalue``' instructions.
2145 Structures may optionally be "packed" structures, which indicate that
2146 the alignment of the struct is one byte, and that there is no padding
2147 between the elements. In non-packed structs, padding between field types
2148 is inserted as defined by the DataLayout string in the module, which is
2149 required to match what the underlying code generator expects.
2151 Structures can either be "literal" or "identified". A literal structure
2152 is defined inline with other types (e.g. ``{i32, i32}*``) whereas
2153 identified types are always defined at the top level with a name.
2154 Literal types are uniqued by their contents and can never be recursive
2155 or opaque since there is no way to write one. Identified types can be
2156 recursive, can be opaqued, and are never uniqued.
2162 %T1 = type { <type list> } ; Identified normal struct type
2163 %T2 = type <{ <type list> }> ; Identified packed struct type
2167 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2168 | ``{ i32, i32, i32 }`` | A triple of three ``i32`` values |
2169 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2170 | ``{ 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``. |
2171 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2172 | ``<{ i8, i32 }>`` | A packed struct known to be 5 bytes in size. |
2173 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2177 Opaque Structure Types
2178 """"""""""""""""""""""
2182 Opaque structure types are used to represent named structure types that
2183 do not have a body specified. This corresponds (for example) to the C
2184 notion of a forward declared structure.
2195 +--------------+-------------------+
2196 | ``opaque`` | An opaque type. |
2197 +--------------+-------------------+
2204 LLVM has several different basic types of constants. This section
2205 describes them all and their syntax.
2210 **Boolean constants**
2211 The two strings '``true``' and '``false``' are both valid constants
2213 **Integer constants**
2214 Standard integers (such as '4') are constants of the
2215 :ref:`integer <t_integer>` type. Negative numbers may be used with
2217 **Floating point constants**
2218 Floating point constants use standard decimal notation (e.g.
2219 123.421), exponential notation (e.g. 1.23421e+2), or a more precise
2220 hexadecimal notation (see below). The assembler requires the exact
2221 decimal value of a floating-point constant. For example, the
2222 assembler accepts 1.25 but rejects 1.3 because 1.3 is a repeating
2223 decimal in binary. Floating point constants must have a :ref:`floating
2224 point <t_floating>` type.
2225 **Null pointer constants**
2226 The identifier '``null``' is recognized as a null pointer constant
2227 and must be of :ref:`pointer type <t_pointer>`.
2229 The one non-intuitive notation for constants is the hexadecimal form of
2230 floating point constants. For example, the form
2231 '``double 0x432ff973cafa8000``' is equivalent to (but harder to read
2232 than) '``double 4.5e+15``'. The only time hexadecimal floating point
2233 constants are required (and the only time that they are generated by the
2234 disassembler) is when a floating point constant must be emitted but it
2235 cannot be represented as a decimal floating point number in a reasonable
2236 number of digits. For example, NaN's, infinities, and other special
2237 values are represented in their IEEE hexadecimal format so that assembly
2238 and disassembly do not cause any bits to change in the constants.
2240 When using the hexadecimal form, constants of types half, float, and
2241 double are represented using the 16-digit form shown above (which
2242 matches the IEEE754 representation for double); half and float values
2243 must, however, be exactly representable as IEEE 754 half and single
2244 precision, respectively. Hexadecimal format is always used for long
2245 double, and there are three forms of long double. The 80-bit format used
2246 by x86 is represented as ``0xK`` followed by 20 hexadecimal digits. The
2247 128-bit format used by PowerPC (two adjacent doubles) is represented by
2248 ``0xM`` followed by 32 hexadecimal digits. The IEEE 128-bit format is
2249 represented by ``0xL`` followed by 32 hexadecimal digits. Long doubles
2250 will only work if they match the long double format on your target.
2251 The IEEE 16-bit format (half precision) is represented by ``0xH``
2252 followed by 4 hexadecimal digits. All hexadecimal formats are big-endian
2253 (sign bit at the left).
2255 There are no constants of type x86_mmx.
2257 .. _complexconstants:
2262 Complex constants are a (potentially recursive) combination of simple
2263 constants and smaller complex constants.
2265 **Structure constants**
2266 Structure constants are represented with notation similar to
2267 structure type definitions (a comma separated list of elements,
2268 surrounded by braces (``{}``)). For example:
2269 "``{ i32 4, float 17.0, i32* @G }``", where "``@G``" is declared as
2270 "``@G = external global i32``". Structure constants must have
2271 :ref:`structure type <t_struct>`, and the number and types of elements
2272 must match those specified by the type.
2274 Array constants are represented with notation similar to array type
2275 definitions (a comma separated list of elements, surrounded by
2276 square brackets (``[]``)). For example:
2277 "``[ i32 42, i32 11, i32 74 ]``". Array constants must have
2278 :ref:`array type <t_array>`, and the number and types of elements must
2279 match those specified by the type. As a special case, character array
2280 constants may also be represented as a double-quoted string using the ``c``
2281 prefix. For example: "``c"Hello World\0A\00"``".
2282 **Vector constants**
2283 Vector constants are represented with notation similar to vector
2284 type definitions (a comma separated list of elements, surrounded by
2285 less-than/greater-than's (``<>``)). For example:
2286 "``< i32 42, i32 11, i32 74, i32 100 >``". Vector constants
2287 must have :ref:`vector type <t_vector>`, and the number and types of
2288 elements must match those specified by the type.
2289 **Zero initialization**
2290 The string '``zeroinitializer``' can be used to zero initialize a
2291 value to zero of *any* type, including scalar and
2292 :ref:`aggregate <t_aggregate>` types. This is often used to avoid
2293 having to print large zero initializers (e.g. for large arrays) and
2294 is always exactly equivalent to using explicit zero initializers.
2296 A metadata node is a structure-like constant with :ref:`metadata
2297 type <t_metadata>`. For example:
2298 "``metadata !{ i32 0, metadata !"test" }``". Unlike other
2299 constants that are meant to be interpreted as part of the
2300 instruction stream, metadata is a place to attach additional
2301 information such as debug info.
2303 Global Variable and Function Addresses
2304 --------------------------------------
2306 The addresses of :ref:`global variables <globalvars>` and
2307 :ref:`functions <functionstructure>` are always implicitly valid
2308 (link-time) constants. These constants are explicitly referenced when
2309 the :ref:`identifier for the global <identifiers>` is used and always have
2310 :ref:`pointer <t_pointer>` type. For example, the following is a legal LLVM
2313 .. code-block:: llvm
2317 @Z = global [2 x i32*] [ i32* @X, i32* @Y ]
2324 The string '``undef``' can be used anywhere a constant is expected, and
2325 indicates that the user of the value may receive an unspecified
2326 bit-pattern. Undefined values may be of any type (other than '``label``'
2327 or '``void``') and be used anywhere a constant is permitted.
2329 Undefined values are useful because they indicate to the compiler that
2330 the program is well defined no matter what value is used. This gives the
2331 compiler more freedom to optimize. Here are some examples of
2332 (potentially surprising) transformations that are valid (in pseudo IR):
2334 .. code-block:: llvm
2344 This is safe because all of the output bits are affected by the undef
2345 bits. Any output bit can have a zero or one depending on the input bits.
2347 .. code-block:: llvm
2358 These logical operations have bits that are not always affected by the
2359 input. For example, if ``%X`` has a zero bit, then the output of the
2360 '``and``' operation will always be a zero for that bit, no matter what
2361 the corresponding bit from the '``undef``' is. As such, it is unsafe to
2362 optimize or assume that the result of the '``and``' is '``undef``'.
2363 However, it is safe to assume that all bits of the '``undef``' could be
2364 0, and optimize the '``and``' to 0. Likewise, it is safe to assume that
2365 all the bits of the '``undef``' operand to the '``or``' could be set,
2366 allowing the '``or``' to be folded to -1.
2368 .. code-block:: llvm
2370 %A = select undef, %X, %Y
2371 %B = select undef, 42, %Y
2372 %C = select %X, %Y, undef
2382 This set of examples shows that undefined '``select``' (and conditional
2383 branch) conditions can go *either way*, but they have to come from one
2384 of the two operands. In the ``%A`` example, if ``%X`` and ``%Y`` were
2385 both known to have a clear low bit, then ``%A`` would have to have a
2386 cleared low bit. However, in the ``%C`` example, the optimizer is
2387 allowed to assume that the '``undef``' operand could be the same as
2388 ``%Y``, allowing the whole '``select``' to be eliminated.
2390 .. code-block:: llvm
2392 %A = xor undef, undef
2409 This example points out that two '``undef``' operands are not
2410 necessarily the same. This can be surprising to people (and also matches
2411 C semantics) where they assume that "``X^X``" is always zero, even if
2412 ``X`` is undefined. This isn't true for a number of reasons, but the
2413 short answer is that an '``undef``' "variable" can arbitrarily change
2414 its value over its "live range". This is true because the variable
2415 doesn't actually *have a live range*. Instead, the value is logically
2416 read from arbitrary registers that happen to be around when needed, so
2417 the value is not necessarily consistent over time. In fact, ``%A`` and
2418 ``%C`` need to have the same semantics or the core LLVM "replace all
2419 uses with" concept would not hold.
2421 .. code-block:: llvm
2429 These examples show the crucial difference between an *undefined value*
2430 and *undefined behavior*. An undefined value (like '``undef``') is
2431 allowed to have an arbitrary bit-pattern. This means that the ``%A``
2432 operation can be constant folded to '``undef``', because the '``undef``'
2433 could be an SNaN, and ``fdiv`` is not (currently) defined on SNaN's.
2434 However, in the second example, we can make a more aggressive
2435 assumption: because the ``undef`` is allowed to be an arbitrary value,
2436 we are allowed to assume that it could be zero. Since a divide by zero
2437 has *undefined behavior*, we are allowed to assume that the operation
2438 does not execute at all. This allows us to delete the divide and all
2439 code after it. Because the undefined operation "can't happen", the
2440 optimizer can assume that it occurs in dead code.
2442 .. code-block:: llvm
2444 a: store undef -> %X
2445 b: store %X -> undef
2450 These examples reiterate the ``fdiv`` example: a store *of* an undefined
2451 value can be assumed to not have any effect; we can assume that the
2452 value is overwritten with bits that happen to match what was already
2453 there. However, a store *to* an undefined location could clobber
2454 arbitrary memory, therefore, it has undefined behavior.
2461 Poison values are similar to :ref:`undef values <undefvalues>`, however
2462 they also represent the fact that an instruction or constant expression
2463 that cannot evoke side effects has nevertheless detected a condition
2464 that results in undefined behavior.
2466 There is currently no way of representing a poison value in the IR; they
2467 only exist when produced by operations such as :ref:`add <i_add>` with
2470 Poison value behavior is defined in terms of value *dependence*:
2472 - Values other than :ref:`phi <i_phi>` nodes depend on their operands.
2473 - :ref:`Phi <i_phi>` nodes depend on the operand corresponding to
2474 their dynamic predecessor basic block.
2475 - Function arguments depend on the corresponding actual argument values
2476 in the dynamic callers of their functions.
2477 - :ref:`Call <i_call>` instructions depend on the :ref:`ret <i_ret>`
2478 instructions that dynamically transfer control back to them.
2479 - :ref:`Invoke <i_invoke>` instructions depend on the
2480 :ref:`ret <i_ret>`, :ref:`resume <i_resume>`, or exception-throwing
2481 call instructions that dynamically transfer control back to them.
2482 - Non-volatile loads and stores depend on the most recent stores to all
2483 of the referenced memory addresses, following the order in the IR
2484 (including loads and stores implied by intrinsics such as
2485 :ref:`@llvm.memcpy <int_memcpy>`.)
2486 - An instruction with externally visible side effects depends on the
2487 most recent preceding instruction with externally visible side
2488 effects, following the order in the IR. (This includes :ref:`volatile
2489 operations <volatile>`.)
2490 - An instruction *control-depends* on a :ref:`terminator
2491 instruction <terminators>` if the terminator instruction has
2492 multiple successors and the instruction is always executed when
2493 control transfers to one of the successors, and may not be executed
2494 when control is transferred to another.
2495 - Additionally, an instruction also *control-depends* on a terminator
2496 instruction if the set of instructions it otherwise depends on would
2497 be different if the terminator had transferred control to a different
2499 - Dependence is transitive.
2501 Poison values have the same behavior as :ref:`undef values <undefvalues>`,
2502 with the additional effect that any instruction that has a *dependence*
2503 on a poison value has undefined behavior.
2505 Here are some examples:
2507 .. code-block:: llvm
2510 %poison = sub nuw i32 0, 1 ; Results in a poison value.
2511 %still_poison = and i32 %poison, 0 ; 0, but also poison.
2512 %poison_yet_again = getelementptr i32* @h, i32 %still_poison
2513 store i32 0, i32* %poison_yet_again ; memory at @h[0] is poisoned
2515 store i32 %poison, i32* @g ; Poison value stored to memory.
2516 %poison2 = load i32* @g ; Poison value loaded back from memory.
2518 store volatile i32 %poison, i32* @g ; External observation; undefined behavior.
2520 %narrowaddr = bitcast i32* @g to i16*
2521 %wideaddr = bitcast i32* @g to i64*
2522 %poison3 = load i16* %narrowaddr ; Returns a poison value.
2523 %poison4 = load i64* %wideaddr ; Returns a poison value.
2525 %cmp = icmp slt i32 %poison, 0 ; Returns a poison value.
2526 br i1 %cmp, label %true, label %end ; Branch to either destination.
2529 store volatile i32 0, i32* @g ; This is control-dependent on %cmp, so
2530 ; it has undefined behavior.
2534 %p = phi i32 [ 0, %entry ], [ 1, %true ]
2535 ; Both edges into this PHI are
2536 ; control-dependent on %cmp, so this
2537 ; always results in a poison value.
2539 store volatile i32 0, i32* @g ; This would depend on the store in %true
2540 ; if %cmp is true, or the store in %entry
2541 ; otherwise, so this is undefined behavior.
2543 br i1 %cmp, label %second_true, label %second_end
2544 ; The same branch again, but this time the
2545 ; true block doesn't have side effects.
2552 store volatile i32 0, i32* @g ; This time, the instruction always depends
2553 ; on the store in %end. Also, it is
2554 ; control-equivalent to %end, so this is
2555 ; well-defined (ignoring earlier undefined
2556 ; behavior in this example).
2560 Addresses of Basic Blocks
2561 -------------------------
2563 ``blockaddress(@function, %block)``
2565 The '``blockaddress``' constant computes the address of the specified
2566 basic block in the specified function, and always has an ``i8*`` type.
2567 Taking the address of the entry block is illegal.
2569 This value only has defined behavior when used as an operand to the
2570 ':ref:`indirectbr <i_indirectbr>`' instruction, or for comparisons
2571 against null. Pointer equality tests between labels addresses results in
2572 undefined behavior --- though, again, comparison against null is ok, and
2573 no label is equal to the null pointer. This may be passed around as an
2574 opaque pointer sized value as long as the bits are not inspected. This
2575 allows ``ptrtoint`` and arithmetic to be performed on these values so
2576 long as the original value is reconstituted before the ``indirectbr``
2579 Finally, some targets may provide defined semantics when using the value
2580 as the operand to an inline assembly, but that is target specific.
2584 Constant Expressions
2585 --------------------
2587 Constant expressions are used to allow expressions involving other
2588 constants to be used as constants. Constant expressions may be of any
2589 :ref:`first class <t_firstclass>` type and may involve any LLVM operation
2590 that does not have side effects (e.g. load and call are not supported).
2591 The following is the syntax for constant expressions:
2593 ``trunc (CST to TYPE)``
2594 Truncate a constant to another type. The bit size of CST must be
2595 larger than the bit size of TYPE. Both types must be integers.
2596 ``zext (CST to TYPE)``
2597 Zero extend a constant to another type. The bit size of CST must be
2598 smaller than the bit size of TYPE. Both types must be integers.
2599 ``sext (CST to TYPE)``
2600 Sign extend a constant to another type. The bit size of CST must be
2601 smaller than the bit size of TYPE. Both types must be integers.
2602 ``fptrunc (CST to TYPE)``
2603 Truncate a floating point constant to another floating point type.
2604 The size of CST must be larger than the size of TYPE. Both types
2605 must be floating point.
2606 ``fpext (CST to TYPE)``
2607 Floating point extend a constant to another type. The size of CST
2608 must be smaller or equal to the size of TYPE. Both types must be
2610 ``fptoui (CST to TYPE)``
2611 Convert a floating point constant to the corresponding unsigned
2612 integer constant. TYPE must be a scalar or vector integer type. CST
2613 must be of scalar or vector floating point type. Both CST and TYPE
2614 must be scalars, or vectors of the same number of elements. If the
2615 value won't fit in the integer type, the results are undefined.
2616 ``fptosi (CST to TYPE)``
2617 Convert a floating point constant to the corresponding signed
2618 integer constant. TYPE must be a scalar or vector integer type. CST
2619 must be of scalar or vector floating point type. Both CST and TYPE
2620 must be scalars, or vectors of the same number of elements. If the
2621 value won't fit in the integer type, the results are undefined.
2622 ``uitofp (CST to TYPE)``
2623 Convert an unsigned integer constant to the corresponding floating
2624 point constant. TYPE must be a scalar or vector floating point type.
2625 CST must be of scalar or vector integer type. Both CST and TYPE must
2626 be scalars, or vectors of the same number of elements. If the value
2627 won't fit in the floating point type, the results are undefined.
2628 ``sitofp (CST to TYPE)``
2629 Convert a signed integer constant to the corresponding floating
2630 point constant. TYPE must be a scalar or vector floating point type.
2631 CST must be of scalar or vector integer type. Both CST and TYPE must
2632 be scalars, or vectors of the same number of elements. If the value
2633 won't fit in the floating point type, the results are undefined.
2634 ``ptrtoint (CST to TYPE)``
2635 Convert a pointer typed constant to the corresponding integer
2636 constant. ``TYPE`` must be an integer type. ``CST`` must be of
2637 pointer type. The ``CST`` value is zero extended, truncated, or
2638 unchanged to make it fit in ``TYPE``.
2639 ``inttoptr (CST to TYPE)``
2640 Convert an integer constant to a pointer constant. TYPE must be a
2641 pointer type. CST must be of integer type. The CST value is zero
2642 extended, truncated, or unchanged to make it fit in a pointer size.
2643 This one is *really* dangerous!
2644 ``bitcast (CST to TYPE)``
2645 Convert a constant, CST, to another TYPE. The constraints of the
2646 operands are the same as those for the :ref:`bitcast
2647 instruction <i_bitcast>`.
2648 ``addrspacecast (CST to TYPE)``
2649 Convert a constant pointer or constant vector of pointer, CST, to another
2650 TYPE in a different address space. The constraints of the operands are the
2651 same as those for the :ref:`addrspacecast instruction <i_addrspacecast>`.
2652 ``getelementptr (CSTPTR, IDX0, IDX1, ...)``, ``getelementptr inbounds (CSTPTR, IDX0, IDX1, ...)``
2653 Perform the :ref:`getelementptr operation <i_getelementptr>` on
2654 constants. As with the :ref:`getelementptr <i_getelementptr>`
2655 instruction, the index list may have zero or more indexes, which are
2656 required to make sense for the type of "CSTPTR".
2657 ``select (COND, VAL1, VAL2)``
2658 Perform the :ref:`select operation <i_select>` on constants.
2659 ``icmp COND (VAL1, VAL2)``
2660 Performs the :ref:`icmp operation <i_icmp>` on constants.
2661 ``fcmp COND (VAL1, VAL2)``
2662 Performs the :ref:`fcmp operation <i_fcmp>` on constants.
2663 ``extractelement (VAL, IDX)``
2664 Perform the :ref:`extractelement operation <i_extractelement>` on
2666 ``insertelement (VAL, ELT, IDX)``
2667 Perform the :ref:`insertelement operation <i_insertelement>` on
2669 ``shufflevector (VEC1, VEC2, IDXMASK)``
2670 Perform the :ref:`shufflevector operation <i_shufflevector>` on
2672 ``extractvalue (VAL, IDX0, IDX1, ...)``
2673 Perform the :ref:`extractvalue operation <i_extractvalue>` on
2674 constants. The index list is interpreted in a similar manner as
2675 indices in a ':ref:`getelementptr <i_getelementptr>`' operation. At
2676 least one index value must be specified.
2677 ``insertvalue (VAL, ELT, IDX0, IDX1, ...)``
2678 Perform the :ref:`insertvalue operation <i_insertvalue>` on constants.
2679 The index list is interpreted in a similar manner as indices in a
2680 ':ref:`getelementptr <i_getelementptr>`' operation. At least one index
2681 value must be specified.
2682 ``OPCODE (LHS, RHS)``
2683 Perform the specified operation of the LHS and RHS constants. OPCODE
2684 may be any of the :ref:`binary <binaryops>` or :ref:`bitwise
2685 binary <bitwiseops>` operations. The constraints on operands are
2686 the same as those for the corresponding instruction (e.g. no bitwise
2687 operations on floating point values are allowed).
2694 Inline Assembler Expressions
2695 ----------------------------
2697 LLVM supports inline assembler expressions (as opposed to :ref:`Module-Level
2698 Inline Assembly <moduleasm>`) through the use of a special value. This
2699 value represents the inline assembler as a string (containing the
2700 instructions to emit), a list of operand constraints (stored as a
2701 string), a flag that indicates whether or not the inline asm expression
2702 has side effects, and a flag indicating whether the function containing
2703 the asm needs to align its stack conservatively. An example inline
2704 assembler expression is:
2706 .. code-block:: llvm
2708 i32 (i32) asm "bswap $0", "=r,r"
2710 Inline assembler expressions may **only** be used as the callee operand
2711 of a :ref:`call <i_call>` or an :ref:`invoke <i_invoke>` instruction.
2712 Thus, typically we have:
2714 .. code-block:: llvm
2716 %X = call i32 asm "bswap $0", "=r,r"(i32 %Y)
2718 Inline asms with side effects not visible in the constraint list must be
2719 marked as having side effects. This is done through the use of the
2720 '``sideeffect``' keyword, like so:
2722 .. code-block:: llvm
2724 call void asm sideeffect "eieio", ""()
2726 In some cases inline asms will contain code that will not work unless
2727 the stack is aligned in some way, such as calls or SSE instructions on
2728 x86, yet will not contain code that does that alignment within the asm.
2729 The compiler should make conservative assumptions about what the asm
2730 might contain and should generate its usual stack alignment code in the
2731 prologue if the '``alignstack``' keyword is present:
2733 .. code-block:: llvm
2735 call void asm alignstack "eieio", ""()
2737 Inline asms also support using non-standard assembly dialects. The
2738 assumed dialect is ATT. When the '``inteldialect``' keyword is present,
2739 the inline asm is using the Intel dialect. Currently, ATT and Intel are
2740 the only supported dialects. An example is:
2742 .. code-block:: llvm
2744 call void asm inteldialect "eieio", ""()
2746 If multiple keywords appear the '``sideeffect``' keyword must come
2747 first, the '``alignstack``' keyword second and the '``inteldialect``'
2753 The call instructions that wrap inline asm nodes may have a
2754 "``!srcloc``" MDNode attached to it that contains a list of constant
2755 integers. If present, the code generator will use the integer as the
2756 location cookie value when report errors through the ``LLVMContext``
2757 error reporting mechanisms. This allows a front-end to correlate backend
2758 errors that occur with inline asm back to the source code that produced
2761 .. code-block:: llvm
2763 call void asm sideeffect "something bad", ""(), !srcloc !42
2765 !42 = !{ i32 1234567 }
2767 It is up to the front-end to make sense of the magic numbers it places
2768 in the IR. If the MDNode contains multiple constants, the code generator
2769 will use the one that corresponds to the line of the asm that the error
2774 Metadata Nodes and Metadata Strings
2775 -----------------------------------
2777 LLVM IR allows metadata to be attached to instructions in the program
2778 that can convey extra information about the code to the optimizers and
2779 code generator. One example application of metadata is source-level
2780 debug information. There are two metadata primitives: strings and nodes.
2781 All metadata has the ``metadata`` type and is identified in syntax by a
2782 preceding exclamation point ('``!``').
2784 A metadata string is a string surrounded by double quotes. It can
2785 contain any character by escaping non-printable characters with
2786 "``\xx``" where "``xx``" is the two digit hex code. For example:
2789 Metadata nodes are represented with notation similar to structure
2790 constants (a comma separated list of elements, surrounded by braces and
2791 preceded by an exclamation point). Metadata nodes can have any values as
2792 their operand. For example:
2794 .. code-block:: llvm
2796 !{ metadata !"test\00", i32 10}
2798 A :ref:`named metadata <namedmetadatastructure>` is a collection of
2799 metadata nodes, which can be looked up in the module symbol table. For
2802 .. code-block:: llvm
2804 !foo = metadata !{!4, !3}
2806 Metadata can be used as function arguments. Here ``llvm.dbg.value``
2807 function is using two metadata arguments:
2809 .. code-block:: llvm
2811 call void @llvm.dbg.value(metadata !24, i64 0, metadata !25)
2813 Metadata can be attached with an instruction. Here metadata ``!21`` is
2814 attached to the ``add`` instruction using the ``!dbg`` identifier:
2816 .. code-block:: llvm
2818 %indvar.next = add i64 %indvar, 1, !dbg !21
2820 More information about specific metadata nodes recognized by the
2821 optimizers and code generator is found below.
2826 In LLVM IR, memory does not have types, so LLVM's own type system is not
2827 suitable for doing TBAA. Instead, metadata is added to the IR to
2828 describe a type system of a higher level language. This can be used to
2829 implement typical C/C++ TBAA, but it can also be used to implement
2830 custom alias analysis behavior for other languages.
2832 The current metadata format is very simple. TBAA metadata nodes have up
2833 to three fields, e.g.:
2835 .. code-block:: llvm
2837 !0 = metadata !{ metadata !"an example type tree" }
2838 !1 = metadata !{ metadata !"int", metadata !0 }
2839 !2 = metadata !{ metadata !"float", metadata !0 }
2840 !3 = metadata !{ metadata !"const float", metadata !2, i64 1 }
2842 The first field is an identity field. It can be any value, usually a
2843 metadata string, which uniquely identifies the type. The most important
2844 name in the tree is the name of the root node. Two trees with different
2845 root node names are entirely disjoint, even if they have leaves with
2848 The second field identifies the type's parent node in the tree, or is
2849 null or omitted for a root node. A type is considered to alias all of
2850 its descendants and all of its ancestors in the tree. Also, a type is
2851 considered to alias all types in other trees, so that bitcode produced
2852 from multiple front-ends is handled conservatively.
2854 If the third field is present, it's an integer which if equal to 1
2855 indicates that the type is "constant" (meaning
2856 ``pointsToConstantMemory`` should return true; see `other useful
2857 AliasAnalysis methods <AliasAnalysis.html#OtherItfs>`_).
2859 '``tbaa.struct``' Metadata
2860 ^^^^^^^^^^^^^^^^^^^^^^^^^^
2862 The :ref:`llvm.memcpy <int_memcpy>` is often used to implement
2863 aggregate assignment operations in C and similar languages, however it
2864 is defined to copy a contiguous region of memory, which is more than
2865 strictly necessary for aggregate types which contain holes due to
2866 padding. Also, it doesn't contain any TBAA information about the fields
2869 ``!tbaa.struct`` metadata can describe which memory subregions in a
2870 memcpy are padding and what the TBAA tags of the struct are.
2872 The current metadata format is very simple. ``!tbaa.struct`` metadata
2873 nodes are a list of operands which are in conceptual groups of three.
2874 For each group of three, the first operand gives the byte offset of a
2875 field in bytes, the second gives its size in bytes, and the third gives
2878 .. code-block:: llvm
2880 !4 = metadata !{ i64 0, i64 4, metadata !1, i64 8, i64 4, metadata !2 }
2882 This describes a struct with two fields. The first is at offset 0 bytes
2883 with size 4 bytes, and has tbaa tag !1. The second is at offset 8 bytes
2884 and has size 4 bytes and has tbaa tag !2.
2886 Note that the fields need not be contiguous. In this example, there is a
2887 4 byte gap between the two fields. This gap represents padding which
2888 does not carry useful data and need not be preserved.
2890 '``noalias``' and '``alias.scope``' Metadata
2891 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2893 ``noalias`` and ``alias.scope`` metadata provide the ability to specify generic
2894 noalias memory-access sets. This means that some collection of memory access
2895 instructions (loads, stores, memory-accessing calls, etc.) that carry
2896 ``noalias`` metadata can specifically be specified not to alias with some other
2897 collection of memory access instructions that carry ``alias.scope`` metadata.
2898 Each type of metadata specifies a list of scopes where each scope has an id and
2899 a domain. When evaluating an aliasing query, if for some some domain, the set
2900 of scopes with that domain in one instruction's ``alias.scope`` list is a
2901 subset of (or qual to) the set of scopes for that domain in another
2902 instruction's ``noalias`` list, then the two memory accesses are assumed not to
2905 The metadata identifying each domain is itself a list containing one or two
2906 entries. The first entry is the name of the domain. Note that if the name is a
2907 string then it can be combined accross functions and translation units. A
2908 self-reference can be used to create globally unique domain names. A
2909 descriptive string may optionally be provided as a second list entry.
2911 The metadata identifying each scope is also itself a list containing two or
2912 three entries. The first entry is the name of the scope. Note that if the name
2913 is a string then it can be combined accross functions and translation units. A
2914 self-reference can be used to create globally unique scope names. A metadata
2915 reference to the scope's domain is the second entry. A descriptive string may
2916 optionally be provided as a third list entry.
2920 .. code-block:: llvm
2922 ; Two scope domains:
2923 !0 = metadata !{metadata !0}
2924 !1 = metadata !{metadata !1}
2926 ; Some scopes in these domains:
2927 !2 = metadata !{metadata !2, metadata !0}
2928 !3 = metadata !{metadata !3, metadata !0}
2929 !4 = metadata !{metadata !4, metadata !1}
2932 !5 = metadata !{metadata !4} ; A list containing only scope !4
2933 !6 = metadata !{metadata !4, metadata !3, metadata !2}
2934 !7 = metadata !{metadata !3}
2936 ; These two instructions don't alias:
2937 %0 = load float* %c, align 4, !alias.scope !5
2938 store float %0, float* %arrayidx.i, align 4, !noalias !5
2940 ; These two instructions also don't alias (for domain !1, the set of scopes
2941 ; in the !alias.scope equals that in the !noalias list):
2942 %2 = load float* %c, align 4, !alias.scope !5
2943 store float %2, float* %arrayidx.i2, align 4, !noalias !6
2945 ; These two instructions don't alias (for domain !0, the set of scopes in
2946 ; the !noalias list is not a superset of, or equal to, the scopes in the
2947 ; !alias.scope list):
2948 %2 = load float* %c, align 4, !alias.scope !6
2949 store float %0, float* %arrayidx.i, align 4, !noalias !7
2951 '``fpmath``' Metadata
2952 ^^^^^^^^^^^^^^^^^^^^^
2954 ``fpmath`` metadata may be attached to any instruction of floating point
2955 type. It can be used to express the maximum acceptable error in the
2956 result of that instruction, in ULPs, thus potentially allowing the
2957 compiler to use a more efficient but less accurate method of computing
2958 it. ULP is defined as follows:
2960 If ``x`` is a real number that lies between two finite consecutive
2961 floating-point numbers ``a`` and ``b``, without being equal to one
2962 of them, then ``ulp(x) = |b - a|``, otherwise ``ulp(x)`` is the
2963 distance between the two non-equal finite floating-point numbers
2964 nearest ``x``. Moreover, ``ulp(NaN)`` is ``NaN``.
2966 The metadata node shall consist of a single positive floating point
2967 number representing the maximum relative error, for example:
2969 .. code-block:: llvm
2971 !0 = metadata !{ float 2.5 } ; maximum acceptable inaccuracy is 2.5 ULPs
2973 '``range``' Metadata
2974 ^^^^^^^^^^^^^^^^^^^^
2976 ``range`` metadata may be attached only to ``load``, ``call`` and ``invoke`` of
2977 integer types. It expresses the possible ranges the loaded value or the value
2978 returned by the called function at this call site is in. The ranges are
2979 represented with a flattened list of integers. The loaded value or the value
2980 returned is known to be in the union of the ranges defined by each consecutive
2981 pair. Each pair has the following properties:
2983 - The type must match the type loaded by the instruction.
2984 - The pair ``a,b`` represents the range ``[a,b)``.
2985 - Both ``a`` and ``b`` are constants.
2986 - The range is allowed to wrap.
2987 - The range should not represent the full or empty set. That is,
2990 In addition, the pairs must be in signed order of the lower bound and
2991 they must be non-contiguous.
2995 .. code-block:: llvm
2997 %a = load i8* %x, align 1, !range !0 ; Can only be 0 or 1
2998 %b = load i8* %y, align 1, !range !1 ; Can only be 255 (-1), 0 or 1
2999 %c = call i8 @foo(), !range !2 ; Can only be 0, 1, 3, 4 or 5
3000 %d = invoke i8 @bar() to label %cont
3001 unwind label %lpad, !range !3 ; Can only be -2, -1, 3, 4 or 5
3003 !0 = metadata !{ i8 0, i8 2 }
3004 !1 = metadata !{ i8 255, i8 2 }
3005 !2 = metadata !{ i8 0, i8 2, i8 3, i8 6 }
3006 !3 = metadata !{ i8 -2, i8 0, i8 3, i8 6 }
3011 It is sometimes useful to attach information to loop constructs. Currently,
3012 loop metadata is implemented as metadata attached to the branch instruction
3013 in the loop latch block. This type of metadata refer to a metadata node that is
3014 guaranteed to be separate for each loop. The loop identifier metadata is
3015 specified with the name ``llvm.loop``.
3017 The loop identifier metadata is implemented using a metadata that refers to
3018 itself to avoid merging it with any other identifier metadata, e.g.,
3019 during module linkage or function inlining. That is, each loop should refer
3020 to their own identification metadata even if they reside in separate functions.
3021 The following example contains loop identifier metadata for two separate loop
3024 .. code-block:: llvm
3026 !0 = metadata !{ metadata !0 }
3027 !1 = metadata !{ metadata !1 }
3029 The loop identifier metadata can be used to specify additional
3030 per-loop metadata. Any operands after the first operand can be treated
3031 as user-defined metadata. For example the ``llvm.loop.unroll.count``
3032 suggests an unroll factor to the loop unroller:
3034 .. code-block:: llvm
3036 br i1 %exitcond, label %._crit_edge, label %.lr.ph, !llvm.loop !0
3038 !0 = metadata !{ metadata !0, metadata !1 }
3039 !1 = metadata !{ metadata !"llvm.loop.unroll.count", i32 4 }
3041 '``llvm.loop.vectorize``' and '``llvm.loop.interleave``'
3042 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3044 Metadata prefixed with ``llvm.loop.vectorize`` or ``llvm.loop.interleave`` are
3045 used to control per-loop vectorization and interleaving parameters such as
3046 vectorization width and interleave count. These metadata should be used in
3047 conjunction with ``llvm.loop`` loop identification metadata. The
3048 ``llvm.loop.vectorize`` and ``llvm.loop.interleave`` metadata are only
3049 optimization hints and the optimizer will only interleave and vectorize loops if
3050 it believes it is safe to do so. The ``llvm.mem.parallel_loop_access`` metadata
3051 which contains information about loop-carried memory dependencies can be helpful
3052 in determining the safety of these transformations.
3054 '``llvm.loop.interleave.count``' Metadata
3055 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3057 This metadata suggests an interleave count to the loop interleaver.
3058 The first operand is the string ``llvm.loop.interleave.count`` and the
3059 second operand is an integer specifying the interleave count. For
3062 .. code-block:: llvm
3064 !0 = metadata !{ metadata !"llvm.loop.interleave.count", i32 4 }
3066 Note that setting ``llvm.loop.interleave.count`` to 1 disables interleaving
3067 multiple iterations of the loop. If ``llvm.loop.interleave.count`` is set to 0
3068 then the interleave count will be determined automatically.
3070 '``llvm.loop.vectorize.enable``' Metadata
3071 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3073 This metadata selectively enables or disables vectorization for the loop. The
3074 first operand is the string ``llvm.loop.vectorize.enable`` and the second operand
3075 is a bit. If the bit operand value is 1 vectorization is enabled. A value of
3076 0 disables vectorization:
3078 .. code-block:: llvm
3080 !0 = metadata !{ metadata !"llvm.loop.vectorize.enable", i1 0 }
3081 !1 = metadata !{ metadata !"llvm.loop.vectorize.enable", i1 1 }
3083 '``llvm.loop.vectorize.width``' Metadata
3084 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3086 This metadata sets the target width of the vectorizer. The first
3087 operand is the string ``llvm.loop.vectorize.width`` and the second
3088 operand is an integer specifying the width. For example:
3090 .. code-block:: llvm
3092 !0 = metadata !{ metadata !"llvm.loop.vectorize.width", i32 4 }
3094 Note that setting ``llvm.loop.vectorize.width`` to 1 disables
3095 vectorization of the loop. If ``llvm.loop.vectorize.width`` is set to
3096 0 or if the loop does not have this metadata the width will be
3097 determined automatically.
3099 '``llvm.loop.unroll``'
3100 ^^^^^^^^^^^^^^^^^^^^^^
3102 Metadata prefixed with ``llvm.loop.unroll`` are loop unrolling
3103 optimization hints such as the unroll factor. ``llvm.loop.unroll``
3104 metadata should be used in conjunction with ``llvm.loop`` loop
3105 identification metadata. The ``llvm.loop.unroll`` metadata are only
3106 optimization hints and the unrolling will only be performed if the
3107 optimizer believes it is safe to do so.
3109 '``llvm.loop.unroll.count``' Metadata
3110 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3112 This metadata suggests an unroll factor to the loop unroller. The
3113 first operand is the string ``llvm.loop.unroll.count`` and the second
3114 operand is a positive integer specifying the unroll factor. For
3117 .. code-block:: llvm
3119 !0 = metadata !{ metadata !"llvm.loop.unroll.count", i32 4 }
3121 If the trip count of the loop is less than the unroll count the loop
3122 will be partially unrolled.
3124 '``llvm.loop.unroll.disable``' Metadata
3125 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3127 This metadata either disables loop unrolling. The metadata has a single operand
3128 which is the string ``llvm.loop.unroll.disable``. For example:
3130 .. code-block:: llvm
3132 !0 = metadata !{ metadata !"llvm.loop.unroll.disable" }
3134 '``llvm.loop.unroll.full``' Metadata
3135 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3137 This metadata either suggests that the loop should be unrolled fully. The
3138 metadata has a single operand which is the string ``llvm.loop.unroll.disable``.
3141 .. code-block:: llvm
3143 !0 = metadata !{ metadata !"llvm.loop.unroll.full" }
3148 Metadata types used to annotate memory accesses with information helpful
3149 for optimizations are prefixed with ``llvm.mem``.
3151 '``llvm.mem.parallel_loop_access``' Metadata
3152 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3154 The ``llvm.mem.parallel_loop_access`` metadata refers to a loop identifier,
3155 or metadata containing a list of loop identifiers for nested loops.
3156 The metadata is attached to memory accessing instructions and denotes that
3157 no loop carried memory dependence exist between it and other instructions denoted
3158 with the same loop identifier.
3160 Precisely, given two instructions ``m1`` and ``m2`` that both have the
3161 ``llvm.mem.parallel_loop_access`` metadata, with ``L1`` and ``L2`` being the
3162 set of loops associated with that metadata, respectively, then there is no loop
3163 carried dependence between ``m1`` and ``m2`` for loops in both ``L1`` and
3166 As a special case, if all memory accessing instructions in a loop have
3167 ``llvm.mem.parallel_loop_access`` metadata that refers to that loop, then the
3168 loop has no loop carried memory dependences and is considered to be a parallel
3171 Note that if not all memory access instructions have such metadata referring to
3172 the loop, then the loop is considered not being trivially parallel. Additional
3173 memory dependence analysis is required to make that determination. As a fail
3174 safe mechanism, this causes loops that were originally parallel to be considered
3175 sequential (if optimization passes that are unaware of the parallel semantics
3176 insert new memory instructions into the loop body).
3178 Example of a loop that is considered parallel due to its correct use of
3179 both ``llvm.loop`` and ``llvm.mem.parallel_loop_access``
3180 metadata types that refer to the same loop identifier metadata.
3182 .. code-block:: llvm
3186 %val0 = load i32* %arrayidx, !llvm.mem.parallel_loop_access !0
3188 store i32 %val0, i32* %arrayidx1, !llvm.mem.parallel_loop_access !0
3190 br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !0
3194 !0 = metadata !{ metadata !0 }
3196 It is also possible to have nested parallel loops. In that case the
3197 memory accesses refer to a list of loop identifier metadata nodes instead of
3198 the loop identifier metadata node directly:
3200 .. code-block:: llvm
3204 %val1 = load i32* %arrayidx3, !llvm.mem.parallel_loop_access !2
3206 br label %inner.for.body
3210 %val0 = load i32* %arrayidx1, !llvm.mem.parallel_loop_access !0
3212 store i32 %val0, i32* %arrayidx2, !llvm.mem.parallel_loop_access !0
3214 br i1 %exitcond, label %inner.for.end, label %inner.for.body, !llvm.loop !1
3218 store i32 %val1, i32* %arrayidx4, !llvm.mem.parallel_loop_access !2
3220 br i1 %exitcond, label %outer.for.end, label %outer.for.body, !llvm.loop !2
3222 outer.for.end: ; preds = %for.body
3224 !0 = metadata !{ metadata !1, metadata !2 } ; a list of loop identifiers
3225 !1 = metadata !{ metadata !1 } ; an identifier for the inner loop
3226 !2 = metadata !{ metadata !2 } ; an identifier for the outer loop
3228 Module Flags Metadata
3229 =====================
3231 Information about the module as a whole is difficult to convey to LLVM's
3232 subsystems. The LLVM IR isn't sufficient to transmit this information.
3233 The ``llvm.module.flags`` named metadata exists in order to facilitate
3234 this. These flags are in the form of key / value pairs --- much like a
3235 dictionary --- making it easy for any subsystem who cares about a flag to
3238 The ``llvm.module.flags`` metadata contains a list of metadata triplets.
3239 Each triplet has the following form:
3241 - The first element is a *behavior* flag, which specifies the behavior
3242 when two (or more) modules are merged together, and it encounters two
3243 (or more) metadata with the same ID. The supported behaviors are
3245 - The second element is a metadata string that is a unique ID for the
3246 metadata. Each module may only have one flag entry for each unique ID (not
3247 including entries with the **Require** behavior).
3248 - The third element is the value of the flag.
3250 When two (or more) modules are merged together, the resulting
3251 ``llvm.module.flags`` metadata is the union of the modules' flags. That is, for
3252 each unique metadata ID string, there will be exactly one entry in the merged
3253 modules ``llvm.module.flags`` metadata table, and the value for that entry will
3254 be determined by the merge behavior flag, as described below. The only exception
3255 is that entries with the *Require* behavior are always preserved.
3257 The following behaviors are supported:
3268 Emits an error if two values disagree, otherwise the resulting value
3269 is that of the operands.
3273 Emits a warning if two values disagree. The result value will be the
3274 operand for the flag from the first module being linked.
3278 Adds a requirement that another module flag be present and have a
3279 specified value after linking is performed. The value must be a
3280 metadata pair, where the first element of the pair is the ID of the
3281 module flag to be restricted, and the second element of the pair is
3282 the value the module flag should be restricted to. This behavior can
3283 be used to restrict the allowable results (via triggering of an
3284 error) of linking IDs with the **Override** behavior.
3288 Uses the specified value, regardless of the behavior or value of the
3289 other module. If both modules specify **Override**, but the values
3290 differ, an error will be emitted.
3294 Appends the two values, which are required to be metadata nodes.
3298 Appends the two values, which are required to be metadata
3299 nodes. However, duplicate entries in the second list are dropped
3300 during the append operation.
3302 It is an error for a particular unique flag ID to have multiple behaviors,
3303 except in the case of **Require** (which adds restrictions on another metadata
3304 value) or **Override**.
3306 An example of module flags:
3308 .. code-block:: llvm
3310 !0 = metadata !{ i32 1, metadata !"foo", i32 1 }
3311 !1 = metadata !{ i32 4, metadata !"bar", i32 37 }
3312 !2 = metadata !{ i32 2, metadata !"qux", i32 42 }
3313 !3 = metadata !{ i32 3, metadata !"qux",
3315 metadata !"foo", i32 1
3318 !llvm.module.flags = !{ !0, !1, !2, !3 }
3320 - Metadata ``!0`` has the ID ``!"foo"`` and the value '1'. The behavior
3321 if two or more ``!"foo"`` flags are seen is to emit an error if their
3322 values are not equal.
3324 - Metadata ``!1`` has the ID ``!"bar"`` and the value '37'. The
3325 behavior if two or more ``!"bar"`` flags are seen is to use the value
3328 - Metadata ``!2`` has the ID ``!"qux"`` and the value '42'. The
3329 behavior if two or more ``!"qux"`` flags are seen is to emit a
3330 warning if their values are not equal.
3332 - Metadata ``!3`` has the ID ``!"qux"`` and the value:
3336 metadata !{ metadata !"foo", i32 1 }
3338 The behavior is to emit an error if the ``llvm.module.flags`` does not
3339 contain a flag with the ID ``!"foo"`` that has the value '1' after linking is
3342 Objective-C Garbage Collection Module Flags Metadata
3343 ----------------------------------------------------
3345 On the Mach-O platform, Objective-C stores metadata about garbage
3346 collection in a special section called "image info". The metadata
3347 consists of a version number and a bitmask specifying what types of
3348 garbage collection are supported (if any) by the file. If two or more
3349 modules are linked together their garbage collection metadata needs to
3350 be merged rather than appended together.
3352 The Objective-C garbage collection module flags metadata consists of the
3353 following key-value pairs:
3362 * - ``Objective-C Version``
3363 - **[Required]** --- The Objective-C ABI version. Valid values are 1 and 2.
3365 * - ``Objective-C Image Info Version``
3366 - **[Required]** --- The version of the image info section. Currently
3369 * - ``Objective-C Image Info Section``
3370 - **[Required]** --- The section to place the metadata. Valid values are
3371 ``"__OBJC, __image_info, regular"`` for Objective-C ABI version 1, and
3372 ``"__DATA,__objc_imageinfo, regular, no_dead_strip"`` for
3373 Objective-C ABI version 2.
3375 * - ``Objective-C Garbage Collection``
3376 - **[Required]** --- Specifies whether garbage collection is supported or
3377 not. Valid values are 0, for no garbage collection, and 2, for garbage
3378 collection supported.
3380 * - ``Objective-C GC Only``
3381 - **[Optional]** --- Specifies that only garbage collection is supported.
3382 If present, its value must be 6. This flag requires that the
3383 ``Objective-C Garbage Collection`` flag have the value 2.
3385 Some important flag interactions:
3387 - If a module with ``Objective-C Garbage Collection`` set to 0 is
3388 merged with a module with ``Objective-C Garbage Collection`` set to
3389 2, then the resulting module has the
3390 ``Objective-C Garbage Collection`` flag set to 0.
3391 - A module with ``Objective-C Garbage Collection`` set to 0 cannot be
3392 merged with a module with ``Objective-C GC Only`` set to 6.
3394 Automatic Linker Flags Module Flags Metadata
3395 --------------------------------------------
3397 Some targets support embedding flags to the linker inside individual object
3398 files. Typically this is used in conjunction with language extensions which
3399 allow source files to explicitly declare the libraries they depend on, and have
3400 these automatically be transmitted to the linker via object files.
3402 These flags are encoded in the IR using metadata in the module flags section,
3403 using the ``Linker Options`` key. The merge behavior for this flag is required
3404 to be ``AppendUnique``, and the value for the key is expected to be a metadata
3405 node which should be a list of other metadata nodes, each of which should be a
3406 list of metadata strings defining linker options.
3408 For example, the following metadata section specifies two separate sets of
3409 linker options, presumably to link against ``libz`` and the ``Cocoa``
3412 !0 = metadata !{ i32 6, metadata !"Linker Options",
3414 metadata !{ metadata !"-lz" },
3415 metadata !{ metadata !"-framework", metadata !"Cocoa" } } }
3416 !llvm.module.flags = !{ !0 }
3418 The metadata encoding as lists of lists of options, as opposed to a collapsed
3419 list of options, is chosen so that the IR encoding can use multiple option
3420 strings to specify e.g., a single library, while still having that specifier be
3421 preserved as an atomic element that can be recognized by a target specific
3422 assembly writer or object file emitter.
3424 Each individual option is required to be either a valid option for the target's
3425 linker, or an option that is reserved by the target specific assembly writer or
3426 object file emitter. No other aspect of these options is defined by the IR.
3428 C type width Module Flags Metadata
3429 ----------------------------------
3431 The ARM backend emits a section into each generated object file describing the
3432 options that it was compiled with (in a compiler-independent way) to prevent
3433 linking incompatible objects, and to allow automatic library selection. Some
3434 of these options are not visible at the IR level, namely wchar_t width and enum
3437 To pass this information to the backend, these options are encoded in module
3438 flags metadata, using the following key-value pairs:
3448 - * 0 --- sizeof(wchar_t) == 4
3449 * 1 --- sizeof(wchar_t) == 2
3452 - * 0 --- Enums are at least as large as an ``int``.
3453 * 1 --- Enums are stored in the smallest integer type which can
3454 represent all of its values.
3456 For example, the following metadata section specifies that the module was
3457 compiled with a ``wchar_t`` width of 4 bytes, and the underlying type of an
3458 enum is the smallest type which can represent all of its values::
3460 !llvm.module.flags = !{!0, !1}
3461 !0 = metadata !{i32 1, metadata !"short_wchar", i32 1}
3462 !1 = metadata !{i32 1, metadata !"short_enum", i32 0}
3464 .. _intrinsicglobalvariables:
3466 Intrinsic Global Variables
3467 ==========================
3469 LLVM has a number of "magic" global variables that contain data that
3470 affect code generation or other IR semantics. These are documented here.
3471 All globals of this sort should have a section specified as
3472 "``llvm.metadata``". This section and all globals that start with
3473 "``llvm.``" are reserved for use by LLVM.
3477 The '``llvm.used``' Global Variable
3478 -----------------------------------
3480 The ``@llvm.used`` global is an array which has
3481 :ref:`appending linkage <linkage_appending>`. This array contains a list of
3482 pointers to named global variables, functions and aliases which may optionally
3483 have a pointer cast formed of bitcast or getelementptr. For example, a legal
3486 .. code-block:: llvm
3491 @llvm.used = appending global [2 x i8*] [
3493 i8* bitcast (i32* @Y to i8*)
3494 ], section "llvm.metadata"
3496 If a symbol appears in the ``@llvm.used`` list, then the compiler, assembler,
3497 and linker are required to treat the symbol as if there is a reference to the
3498 symbol that it cannot see (which is why they have to be named). For example, if
3499 a variable has internal linkage and no references other than that from the
3500 ``@llvm.used`` list, it cannot be deleted. This is commonly used to represent
3501 references from inline asms and other things the compiler cannot "see", and
3502 corresponds to "``attribute((used))``" in GNU C.
3504 On some targets, the code generator must emit a directive to the
3505 assembler or object file to prevent the assembler and linker from
3506 molesting the symbol.
3508 .. _gv_llvmcompilerused:
3510 The '``llvm.compiler.used``' Global Variable
3511 --------------------------------------------
3513 The ``@llvm.compiler.used`` directive is the same as the ``@llvm.used``
3514 directive, except that it only prevents the compiler from touching the
3515 symbol. On targets that support it, this allows an intelligent linker to
3516 optimize references to the symbol without being impeded as it would be
3519 This is a rare construct that should only be used in rare circumstances,
3520 and should not be exposed to source languages.
3522 .. _gv_llvmglobalctors:
3524 The '``llvm.global_ctors``' Global Variable
3525 -------------------------------------------
3527 .. code-block:: llvm
3529 %0 = type { i32, void ()*, i8* }
3530 @llvm.global_ctors = appending global [1 x %0] [%0 { i32 65535, void ()* @ctor, i8* @data }]
3532 The ``@llvm.global_ctors`` array contains a list of constructor
3533 functions, priorities, and an optional associated global or function.
3534 The functions referenced by this array will be called in ascending order
3535 of priority (i.e. lowest first) when the module is loaded. The order of
3536 functions with the same priority is not defined.
3538 If the third field is present, non-null, and points to a global variable
3539 or function, the initializer function will only run if the associated
3540 data from the current module is not discarded.
3542 .. _llvmglobaldtors:
3544 The '``llvm.global_dtors``' Global Variable
3545 -------------------------------------------
3547 .. code-block:: llvm
3549 %0 = type { i32, void ()*, i8* }
3550 @llvm.global_dtors = appending global [1 x %0] [%0 { i32 65535, void ()* @dtor, i8* @data }]
3552 The ``@llvm.global_dtors`` array contains a list of destructor
3553 functions, priorities, and an optional associated global or function.
3554 The functions referenced by this array will be called in descending
3555 order of priority (i.e. highest first) when the module is unloaded. The
3556 order of functions with the same priority is not defined.
3558 If the third field is present, non-null, and points to a global variable
3559 or function, the destructor function will only run if the associated
3560 data from the current module is not discarded.
3562 Instruction Reference
3563 =====================
3565 The LLVM instruction set consists of several different classifications
3566 of instructions: :ref:`terminator instructions <terminators>`, :ref:`binary
3567 instructions <binaryops>`, :ref:`bitwise binary
3568 instructions <bitwiseops>`, :ref:`memory instructions <memoryops>`, and
3569 :ref:`other instructions <otherops>`.
3573 Terminator Instructions
3574 -----------------------
3576 As mentioned :ref:`previously <functionstructure>`, every basic block in a
3577 program ends with a "Terminator" instruction, which indicates which
3578 block should be executed after the current block is finished. These
3579 terminator instructions typically yield a '``void``' value: they produce
3580 control flow, not values (the one exception being the
3581 ':ref:`invoke <i_invoke>`' instruction).
3583 The terminator instructions are: ':ref:`ret <i_ret>`',
3584 ':ref:`br <i_br>`', ':ref:`switch <i_switch>`',
3585 ':ref:`indirectbr <i_indirectbr>`', ':ref:`invoke <i_invoke>`',
3586 ':ref:`resume <i_resume>`', and ':ref:`unreachable <i_unreachable>`'.
3590 '``ret``' Instruction
3591 ^^^^^^^^^^^^^^^^^^^^^
3598 ret <type> <value> ; Return a value from a non-void function
3599 ret void ; Return from void function
3604 The '``ret``' instruction is used to return control flow (and optionally
3605 a value) from a function back to the caller.
3607 There are two forms of the '``ret``' instruction: one that returns a
3608 value and then causes control flow, and one that just causes control
3614 The '``ret``' instruction optionally accepts a single argument, the
3615 return value. The type of the return value must be a ':ref:`first
3616 class <t_firstclass>`' type.
3618 A function is not :ref:`well formed <wellformed>` if it it has a non-void
3619 return type and contains a '``ret``' instruction with no return value or
3620 a return value with a type that does not match its type, or if it has a
3621 void return type and contains a '``ret``' instruction with a return
3627 When the '``ret``' instruction is executed, control flow returns back to
3628 the calling function's context. If the caller is a
3629 ":ref:`call <i_call>`" instruction, execution continues at the
3630 instruction after the call. If the caller was an
3631 ":ref:`invoke <i_invoke>`" instruction, execution continues at the
3632 beginning of the "normal" destination block. If the instruction returns
3633 a value, that value shall set the call or invoke instruction's return
3639 .. code-block:: llvm
3641 ret i32 5 ; Return an integer value of 5
3642 ret void ; Return from a void function
3643 ret { i32, i8 } { i32 4, i8 2 } ; Return a struct of values 4 and 2
3647 '``br``' Instruction
3648 ^^^^^^^^^^^^^^^^^^^^
3655 br i1 <cond>, label <iftrue>, label <iffalse>
3656 br label <dest> ; Unconditional branch
3661 The '``br``' instruction is used to cause control flow to transfer to a
3662 different basic block in the current function. There are two forms of
3663 this instruction, corresponding to a conditional branch and an
3664 unconditional branch.
3669 The conditional branch form of the '``br``' instruction takes a single
3670 '``i1``' value and two '``label``' values. The unconditional form of the
3671 '``br``' instruction takes a single '``label``' value as a target.
3676 Upon execution of a conditional '``br``' instruction, the '``i1``'
3677 argument is evaluated. If the value is ``true``, control flows to the
3678 '``iftrue``' ``label`` argument. If "cond" is ``false``, control flows
3679 to the '``iffalse``' ``label`` argument.
3684 .. code-block:: llvm
3687 %cond = icmp eq i32 %a, %b
3688 br i1 %cond, label %IfEqual, label %IfUnequal
3696 '``switch``' Instruction
3697 ^^^^^^^^^^^^^^^^^^^^^^^^
3704 switch <intty> <value>, label <defaultdest> [ <intty> <val>, label <dest> ... ]
3709 The '``switch``' instruction is used to transfer control flow to one of
3710 several different places. It is a generalization of the '``br``'
3711 instruction, allowing a branch to occur to one of many possible
3717 The '``switch``' instruction uses three parameters: an integer
3718 comparison value '``value``', a default '``label``' destination, and an
3719 array of pairs of comparison value constants and '``label``'s. The table
3720 is not allowed to contain duplicate constant entries.
3725 The ``switch`` instruction specifies a table of values and destinations.
3726 When the '``switch``' instruction is executed, this table is searched
3727 for the given value. If the value is found, control flow is transferred
3728 to the corresponding destination; otherwise, control flow is transferred
3729 to the default destination.
3734 Depending on properties of the target machine and the particular
3735 ``switch`` instruction, this instruction may be code generated in
3736 different ways. For example, it could be generated as a series of
3737 chained conditional branches or with a lookup table.
3742 .. code-block:: llvm
3744 ; Emulate a conditional br instruction
3745 %Val = zext i1 %value to i32
3746 switch i32 %Val, label %truedest [ i32 0, label %falsedest ]
3748 ; Emulate an unconditional br instruction
3749 switch i32 0, label %dest [ ]
3751 ; Implement a jump table:
3752 switch i32 %val, label %otherwise [ i32 0, label %onzero
3754 i32 2, label %ontwo ]
3758 '``indirectbr``' Instruction
3759 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3766 indirectbr <somety>* <address>, [ label <dest1>, label <dest2>, ... ]
3771 The '``indirectbr``' instruction implements an indirect branch to a
3772 label within the current function, whose address is specified by
3773 "``address``". Address must be derived from a
3774 :ref:`blockaddress <blockaddress>` constant.
3779 The '``address``' argument is the address of the label to jump to. The
3780 rest of the arguments indicate the full set of possible destinations
3781 that the address may point to. Blocks are allowed to occur multiple
3782 times in the destination list, though this isn't particularly useful.
3784 This destination list is required so that dataflow analysis has an
3785 accurate understanding of the CFG.
3790 Control transfers to the block specified in the address argument. All
3791 possible destination blocks must be listed in the label list, otherwise
3792 this instruction has undefined behavior. This implies that jumps to
3793 labels defined in other functions have undefined behavior as well.
3798 This is typically implemented with a jump through a register.
3803 .. code-block:: llvm
3805 indirectbr i8* %Addr, [ label %bb1, label %bb2, label %bb3 ]
3809 '``invoke``' Instruction
3810 ^^^^^^^^^^^^^^^^^^^^^^^^
3817 <result> = invoke [cconv] [ret attrs] <ptr to function ty> <function ptr val>(<function args>) [fn attrs]
3818 to label <normal label> unwind label <exception label>
3823 The '``invoke``' instruction causes control to transfer to a specified
3824 function, with the possibility of control flow transfer to either the
3825 '``normal``' label or the '``exception``' label. If the callee function
3826 returns with the "``ret``" instruction, control flow will return to the
3827 "normal" label. If the callee (or any indirect callees) returns via the
3828 ":ref:`resume <i_resume>`" instruction or other exception handling
3829 mechanism, control is interrupted and continued at the dynamically
3830 nearest "exception" label.
3832 The '``exception``' label is a `landing
3833 pad <ExceptionHandling.html#overview>`_ for the exception. As such,
3834 '``exception``' label is required to have the
3835 ":ref:`landingpad <i_landingpad>`" instruction, which contains the
3836 information about the behavior of the program after unwinding happens,
3837 as its first non-PHI instruction. The restrictions on the
3838 "``landingpad``" instruction's tightly couples it to the "``invoke``"
3839 instruction, so that the important information contained within the
3840 "``landingpad``" instruction can't be lost through normal code motion.
3845 This instruction requires several arguments:
3847 #. The optional "cconv" marker indicates which :ref:`calling
3848 convention <callingconv>` the call should use. If none is
3849 specified, the call defaults to using C calling conventions.
3850 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
3851 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
3853 #. '``ptr to function ty``': shall be the signature of the pointer to
3854 function value being invoked. In most cases, this is a direct
3855 function invocation, but indirect ``invoke``'s are just as possible,
3856 branching off an arbitrary pointer to function value.
3857 #. '``function ptr val``': An LLVM value containing a pointer to a
3858 function to be invoked.
3859 #. '``function args``': argument list whose types match the function
3860 signature argument types and parameter attributes. All arguments must
3861 be of :ref:`first class <t_firstclass>` type. If the function signature
3862 indicates the function accepts a variable number of arguments, the
3863 extra arguments can be specified.
3864 #. '``normal label``': the label reached when the called function
3865 executes a '``ret``' instruction.
3866 #. '``exception label``': the label reached when a callee returns via
3867 the :ref:`resume <i_resume>` instruction or other exception handling
3869 #. The optional :ref:`function attributes <fnattrs>` list. Only
3870 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
3871 attributes are valid here.
3876 This instruction is designed to operate as a standard '``call``'
3877 instruction in most regards. The primary difference is that it
3878 establishes an association with a label, which is used by the runtime
3879 library to unwind the stack.
3881 This instruction is used in languages with destructors to ensure that
3882 proper cleanup is performed in the case of either a ``longjmp`` or a
3883 thrown exception. Additionally, this is important for implementation of
3884 '``catch``' clauses in high-level languages that support them.
3886 For the purposes of the SSA form, the definition of the value returned
3887 by the '``invoke``' instruction is deemed to occur on the edge from the
3888 current block to the "normal" label. If the callee unwinds then no
3889 return value is available.
3894 .. code-block:: llvm
3896 %retval = invoke i32 @Test(i32 15) to label %Continue
3897 unwind label %TestCleanup ; i32:retval set
3898 %retval = invoke coldcc i32 %Testfnptr(i32 15) to label %Continue
3899 unwind label %TestCleanup ; i32:retval set
3903 '``resume``' Instruction
3904 ^^^^^^^^^^^^^^^^^^^^^^^^
3911 resume <type> <value>
3916 The '``resume``' instruction is a terminator instruction that has no
3922 The '``resume``' instruction requires one argument, which must have the
3923 same type as the result of any '``landingpad``' instruction in the same
3929 The '``resume``' instruction resumes propagation of an existing
3930 (in-flight) exception whose unwinding was interrupted with a
3931 :ref:`landingpad <i_landingpad>` instruction.
3936 .. code-block:: llvm
3938 resume { i8*, i32 } %exn
3942 '``unreachable``' Instruction
3943 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3955 The '``unreachable``' instruction has no defined semantics. This
3956 instruction is used to inform the optimizer that a particular portion of
3957 the code is not reachable. This can be used to indicate that the code
3958 after a no-return function cannot be reached, and other facts.
3963 The '``unreachable``' instruction has no defined semantics.
3970 Binary operators are used to do most of the computation in a program.
3971 They require two operands of the same type, execute an operation on
3972 them, and produce a single value. The operands might represent multiple
3973 data, as is the case with the :ref:`vector <t_vector>` data type. The
3974 result value has the same type as its operands.
3976 There are several different binary operators:
3980 '``add``' Instruction
3981 ^^^^^^^^^^^^^^^^^^^^^
3988 <result> = add <ty> <op1>, <op2> ; yields ty:result
3989 <result> = add nuw <ty> <op1>, <op2> ; yields ty:result
3990 <result> = add nsw <ty> <op1>, <op2> ; yields ty:result
3991 <result> = add nuw nsw <ty> <op1>, <op2> ; yields ty:result
3996 The '``add``' instruction returns the sum of its two operands.
4001 The two arguments to the '``add``' instruction must be
4002 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4003 arguments must have identical types.
4008 The value produced is the integer sum of the two operands.
4010 If the sum has unsigned overflow, the result returned is the
4011 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
4014 Because LLVM integers use a two's complement representation, this
4015 instruction is appropriate for both signed and unsigned integers.
4017 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
4018 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
4019 result value of the ``add`` is a :ref:`poison value <poisonvalues>` if
4020 unsigned and/or signed overflow, respectively, occurs.
4025 .. code-block:: llvm
4027 <result> = add i32 4, %var ; yields i32:result = 4 + %var
4031 '``fadd``' Instruction
4032 ^^^^^^^^^^^^^^^^^^^^^^
4039 <result> = fadd [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
4044 The '``fadd``' instruction returns the sum of its two operands.
4049 The two arguments to the '``fadd``' 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 The value produced is the floating point sum of the two operands. This
4057 instruction can also take any number of :ref:`fast-math flags <fastmath>`,
4058 which are optimization hints to enable otherwise unsafe floating point
4064 .. code-block:: llvm
4066 <result> = fadd float 4.0, %var ; yields float:result = 4.0 + %var
4068 '``sub``' Instruction
4069 ^^^^^^^^^^^^^^^^^^^^^
4076 <result> = sub <ty> <op1>, <op2> ; yields ty:result
4077 <result> = sub nuw <ty> <op1>, <op2> ; yields ty:result
4078 <result> = sub nsw <ty> <op1>, <op2> ; yields ty:result
4079 <result> = sub nuw nsw <ty> <op1>, <op2> ; yields ty:result
4084 The '``sub``' instruction returns the difference of its two operands.
4086 Note that the '``sub``' instruction is used to represent the '``neg``'
4087 instruction present in most other intermediate representations.
4092 The two arguments to the '``sub``' instruction must be
4093 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4094 arguments must have identical types.
4099 The value produced is the integer difference of the two operands.
4101 If the difference has unsigned overflow, the result returned is the
4102 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
4105 Because LLVM integers use a two's complement representation, this
4106 instruction is appropriate for both signed and unsigned integers.
4108 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
4109 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
4110 result value of the ``sub`` is a :ref:`poison value <poisonvalues>` if
4111 unsigned and/or signed overflow, respectively, occurs.
4116 .. code-block:: llvm
4118 <result> = sub i32 4, %var ; yields i32:result = 4 - %var
4119 <result> = sub i32 0, %val ; yields i32:result = -%var
4123 '``fsub``' Instruction
4124 ^^^^^^^^^^^^^^^^^^^^^^
4131 <result> = fsub [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
4136 The '``fsub``' instruction returns the difference of its two operands.
4138 Note that the '``fsub``' instruction is used to represent the '``fneg``'
4139 instruction present in most other intermediate representations.
4144 The two arguments to the '``fsub``' instruction must be :ref:`floating
4145 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
4146 Both arguments must have identical types.
4151 The value produced is the floating point difference of the two operands.
4152 This instruction can also take any number of :ref:`fast-math
4153 flags <fastmath>`, which are optimization hints to enable otherwise
4154 unsafe floating point optimizations:
4159 .. code-block:: llvm
4161 <result> = fsub float 4.0, %var ; yields float:result = 4.0 - %var
4162 <result> = fsub float -0.0, %val ; yields float:result = -%var
4164 '``mul``' Instruction
4165 ^^^^^^^^^^^^^^^^^^^^^
4172 <result> = mul <ty> <op1>, <op2> ; yields ty:result
4173 <result> = mul nuw <ty> <op1>, <op2> ; yields ty:result
4174 <result> = mul nsw <ty> <op1>, <op2> ; yields ty:result
4175 <result> = mul nuw nsw <ty> <op1>, <op2> ; yields ty:result
4180 The '``mul``' instruction returns the product of its two operands.
4185 The two arguments to the '``mul``' instruction must be
4186 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4187 arguments must have identical types.
4192 The value produced is the integer product of the two operands.
4194 If the result of the multiplication has unsigned overflow, the result
4195 returned is the mathematical result modulo 2\ :sup:`n`\ , where n is the
4196 bit width of the result.
4198 Because LLVM integers use a two's complement representation, and the
4199 result is the same width as the operands, this instruction returns the
4200 correct result for both signed and unsigned integers. If a full product
4201 (e.g. ``i32`` * ``i32`` -> ``i64``) is needed, the operands should be
4202 sign-extended or zero-extended as appropriate to the width of the full
4205 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
4206 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
4207 result value of the ``mul`` is a :ref:`poison value <poisonvalues>` if
4208 unsigned and/or signed overflow, respectively, occurs.
4213 .. code-block:: llvm
4215 <result> = mul i32 4, %var ; yields i32:result = 4 * %var
4219 '``fmul``' Instruction
4220 ^^^^^^^^^^^^^^^^^^^^^^
4227 <result> = fmul [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
4232 The '``fmul``' instruction returns the product of its two operands.
4237 The two arguments to the '``fmul``' instruction must be :ref:`floating
4238 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
4239 Both arguments must have identical types.
4244 The value produced is the floating point product of the two operands.
4245 This instruction can also take any number of :ref:`fast-math
4246 flags <fastmath>`, which are optimization hints to enable otherwise
4247 unsafe floating point optimizations:
4252 .. code-block:: llvm
4254 <result> = fmul float 4.0, %var ; yields float:result = 4.0 * %var
4256 '``udiv``' Instruction
4257 ^^^^^^^^^^^^^^^^^^^^^^
4264 <result> = udiv <ty> <op1>, <op2> ; yields ty:result
4265 <result> = udiv exact <ty> <op1>, <op2> ; yields ty:result
4270 The '``udiv``' instruction returns the quotient of its two operands.
4275 The two arguments to the '``udiv``' instruction must be
4276 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4277 arguments must have identical types.
4282 The value produced is the unsigned integer quotient of the two operands.
4284 Note that unsigned integer division and signed integer division are
4285 distinct operations; for signed integer division, use '``sdiv``'.
4287 Division by zero leads to undefined behavior.
4289 If the ``exact`` keyword is present, the result value of the ``udiv`` is
4290 a :ref:`poison value <poisonvalues>` if %op1 is not a multiple of %op2 (as
4291 such, "((a udiv exact b) mul b) == a").
4296 .. code-block:: llvm
4298 <result> = udiv i32 4, %var ; yields i32:result = 4 / %var
4300 '``sdiv``' Instruction
4301 ^^^^^^^^^^^^^^^^^^^^^^
4308 <result> = sdiv <ty> <op1>, <op2> ; yields ty:result
4309 <result> = sdiv exact <ty> <op1>, <op2> ; yields ty:result
4314 The '``sdiv``' instruction returns the quotient of its two operands.
4319 The two arguments to the '``sdiv``' instruction must be
4320 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4321 arguments must have identical types.
4326 The value produced is the signed integer quotient of the two operands
4327 rounded towards zero.
4329 Note that signed integer division and unsigned integer division are
4330 distinct operations; for unsigned integer division, use '``udiv``'.
4332 Division by zero leads to undefined behavior. Overflow also leads to
4333 undefined behavior; this is a rare case, but can occur, for example, by
4334 doing a 32-bit division of -2147483648 by -1.
4336 If the ``exact`` keyword is present, the result value of the ``sdiv`` is
4337 a :ref:`poison value <poisonvalues>` if the result would be rounded.
4342 .. code-block:: llvm
4344 <result> = sdiv i32 4, %var ; yields i32:result = 4 / %var
4348 '``fdiv``' Instruction
4349 ^^^^^^^^^^^^^^^^^^^^^^
4356 <result> = fdiv [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
4361 The '``fdiv``' instruction returns the quotient of its two operands.
4366 The two arguments to the '``fdiv``' instruction must be :ref:`floating
4367 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
4368 Both arguments must have identical types.
4373 The value produced is the floating point quotient of the two operands.
4374 This instruction can also take any number of :ref:`fast-math
4375 flags <fastmath>`, which are optimization hints to enable otherwise
4376 unsafe floating point optimizations:
4381 .. code-block:: llvm
4383 <result> = fdiv float 4.0, %var ; yields float:result = 4.0 / %var
4385 '``urem``' Instruction
4386 ^^^^^^^^^^^^^^^^^^^^^^
4393 <result> = urem <ty> <op1>, <op2> ; yields ty:result
4398 The '``urem``' instruction returns the remainder from the unsigned
4399 division of its two arguments.
4404 The two arguments to the '``urem``' instruction must be
4405 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4406 arguments must have identical types.
4411 This instruction returns the unsigned integer *remainder* of a division.
4412 This instruction always performs an unsigned division to get the
4415 Note that unsigned integer remainder and signed integer remainder are
4416 distinct operations; for signed integer remainder, use '``srem``'.
4418 Taking the remainder of a division by zero leads to undefined behavior.
4423 .. code-block:: llvm
4425 <result> = urem i32 4, %var ; yields i32:result = 4 % %var
4427 '``srem``' Instruction
4428 ^^^^^^^^^^^^^^^^^^^^^^
4435 <result> = srem <ty> <op1>, <op2> ; yields ty:result
4440 The '``srem``' instruction returns the remainder from the signed
4441 division of its two operands. This instruction can also take
4442 :ref:`vector <t_vector>` versions of the values in which case the elements
4448 The two arguments to the '``srem``' instruction must be
4449 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4450 arguments must have identical types.
4455 This instruction returns the *remainder* of a division (where the result
4456 is either zero or has the same sign as the dividend, ``op1``), not the
4457 *modulo* operator (where the result is either zero or has the same sign
4458 as the divisor, ``op2``) of a value. For more information about the
4459 difference, see `The Math
4460 Forum <http://mathforum.org/dr.math/problems/anne.4.28.99.html>`_. For a
4461 table of how this is implemented in various languages, please see
4463 operation <http://en.wikipedia.org/wiki/Modulo_operation>`_.
4465 Note that signed integer remainder and unsigned integer remainder are
4466 distinct operations; for unsigned integer remainder, use '``urem``'.
4468 Taking the remainder of a division by zero leads to undefined behavior.
4469 Overflow also leads to undefined behavior; this is a rare case, but can
4470 occur, for example, by taking the remainder of a 32-bit division of
4471 -2147483648 by -1. (The remainder doesn't actually overflow, but this
4472 rule lets srem be implemented using instructions that return both the
4473 result of the division and the remainder.)
4478 .. code-block:: llvm
4480 <result> = srem i32 4, %var ; yields i32:result = 4 % %var
4484 '``frem``' Instruction
4485 ^^^^^^^^^^^^^^^^^^^^^^
4492 <result> = frem [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
4497 The '``frem``' instruction returns the remainder from the division of
4503 The two arguments to the '``frem``' instruction must be :ref:`floating
4504 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
4505 Both arguments must have identical types.
4510 This instruction returns the *remainder* of a division. The remainder
4511 has the same sign as the dividend. This instruction can also take any
4512 number of :ref:`fast-math flags <fastmath>`, which are optimization hints
4513 to enable otherwise unsafe floating point optimizations:
4518 .. code-block:: llvm
4520 <result> = frem float 4.0, %var ; yields float:result = 4.0 % %var
4524 Bitwise Binary Operations
4525 -------------------------
4527 Bitwise binary operators are used to do various forms of bit-twiddling
4528 in a program. They are generally very efficient instructions and can
4529 commonly be strength reduced from other instructions. They require two
4530 operands of the same type, execute an operation on them, and produce a
4531 single value. The resulting value is the same type as its operands.
4533 '``shl``' Instruction
4534 ^^^^^^^^^^^^^^^^^^^^^
4541 <result> = shl <ty> <op1>, <op2> ; yields ty:result
4542 <result> = shl nuw <ty> <op1>, <op2> ; yields ty:result
4543 <result> = shl nsw <ty> <op1>, <op2> ; yields ty:result
4544 <result> = shl nuw nsw <ty> <op1>, <op2> ; yields ty:result
4549 The '``shl``' instruction returns the first operand shifted to the left
4550 a specified number of bits.
4555 Both arguments to the '``shl``' instruction must be the same
4556 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4557 '``op2``' is treated as an unsigned value.
4562 The value produced is ``op1`` \* 2\ :sup:`op2` mod 2\ :sup:`n`,
4563 where ``n`` is the width of the result. If ``op2`` is (statically or
4564 dynamically) negative or equal to or larger than the number of bits in
4565 ``op1``, the result is undefined. If the arguments are vectors, each
4566 vector element of ``op1`` is shifted by the corresponding shift amount
4569 If the ``nuw`` keyword is present, then the shift produces a :ref:`poison
4570 value <poisonvalues>` if it shifts out any non-zero bits. If the
4571 ``nsw`` keyword is present, then the shift produces a :ref:`poison
4572 value <poisonvalues>` if it shifts out any bits that disagree with the
4573 resultant sign bit. As such, NUW/NSW have the same semantics as they
4574 would if the shift were expressed as a mul instruction with the same
4575 nsw/nuw bits in (mul %op1, (shl 1, %op2)).
4580 .. code-block:: llvm
4582 <result> = shl i32 4, %var ; yields i32: 4 << %var
4583 <result> = shl i32 4, 2 ; yields i32: 16
4584 <result> = shl i32 1, 10 ; yields i32: 1024
4585 <result> = shl i32 1, 32 ; undefined
4586 <result> = shl <2 x i32> < i32 1, i32 1>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 2, i32 4>
4588 '``lshr``' Instruction
4589 ^^^^^^^^^^^^^^^^^^^^^^
4596 <result> = lshr <ty> <op1>, <op2> ; yields ty:result
4597 <result> = lshr exact <ty> <op1>, <op2> ; yields ty:result
4602 The '``lshr``' instruction (logical shift right) returns the first
4603 operand shifted to the right a specified number of bits with zero fill.
4608 Both arguments to the '``lshr``' instruction must be the same
4609 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4610 '``op2``' is treated as an unsigned value.
4615 This instruction always performs a logical shift right operation. The
4616 most significant bits of the result will be filled with zero bits after
4617 the shift. If ``op2`` is (statically or dynamically) equal to or larger
4618 than the number of bits in ``op1``, the result is undefined. If the
4619 arguments are vectors, each vector element of ``op1`` is shifted by the
4620 corresponding shift amount in ``op2``.
4622 If the ``exact`` keyword is present, the result value of the ``lshr`` is
4623 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
4629 .. code-block:: llvm
4631 <result> = lshr i32 4, 1 ; yields i32:result = 2
4632 <result> = lshr i32 4, 2 ; yields i32:result = 1
4633 <result> = lshr i8 4, 3 ; yields i8:result = 0
4634 <result> = lshr i8 -2, 1 ; yields i8:result = 0x7F
4635 <result> = lshr i32 1, 32 ; undefined
4636 <result> = lshr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 0x7FFFFFFF, i32 1>
4638 '``ashr``' Instruction
4639 ^^^^^^^^^^^^^^^^^^^^^^
4646 <result> = ashr <ty> <op1>, <op2> ; yields ty:result
4647 <result> = ashr exact <ty> <op1>, <op2> ; yields ty:result
4652 The '``ashr``' instruction (arithmetic shift right) returns the first
4653 operand shifted to the right a specified number of bits with sign
4659 Both arguments to the '``ashr``' instruction must be the same
4660 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4661 '``op2``' is treated as an unsigned value.
4666 This instruction always performs an arithmetic shift right operation,
4667 The most significant bits of the result will be filled with the sign bit
4668 of ``op1``. If ``op2`` is (statically or dynamically) equal to or larger
4669 than the number of bits in ``op1``, the result is undefined. If the
4670 arguments are vectors, each vector element of ``op1`` is shifted by the
4671 corresponding shift amount in ``op2``.
4673 If the ``exact`` keyword is present, the result value of the ``ashr`` is
4674 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
4680 .. code-block:: llvm
4682 <result> = ashr i32 4, 1 ; yields i32:result = 2
4683 <result> = ashr i32 4, 2 ; yields i32:result = 1
4684 <result> = ashr i8 4, 3 ; yields i8:result = 0
4685 <result> = ashr i8 -2, 1 ; yields i8:result = -1
4686 <result> = ashr i32 1, 32 ; undefined
4687 <result> = ashr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 3> ; yields: result=<2 x i32> < i32 -1, i32 0>
4689 '``and``' Instruction
4690 ^^^^^^^^^^^^^^^^^^^^^
4697 <result> = and <ty> <op1>, <op2> ; yields ty:result
4702 The '``and``' instruction returns the bitwise logical and of its two
4708 The two arguments to the '``and``' instruction must be
4709 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4710 arguments must have identical types.
4715 The truth table used for the '``and``' instruction is:
4732 .. code-block:: llvm
4734 <result> = and i32 4, %var ; yields i32:result = 4 & %var
4735 <result> = and i32 15, 40 ; yields i32:result = 8
4736 <result> = and i32 4, 8 ; yields i32:result = 0
4738 '``or``' Instruction
4739 ^^^^^^^^^^^^^^^^^^^^
4746 <result> = or <ty> <op1>, <op2> ; yields ty:result
4751 The '``or``' instruction returns the bitwise logical inclusive or of its
4757 The two arguments to the '``or``' instruction must be
4758 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4759 arguments must have identical types.
4764 The truth table used for the '``or``' instruction is:
4783 <result> = or i32 4, %var ; yields i32:result = 4 | %var
4784 <result> = or i32 15, 40 ; yields i32:result = 47
4785 <result> = or i32 4, 8 ; yields i32:result = 12
4787 '``xor``' Instruction
4788 ^^^^^^^^^^^^^^^^^^^^^
4795 <result> = xor <ty> <op1>, <op2> ; yields ty:result
4800 The '``xor``' instruction returns the bitwise logical exclusive or of
4801 its two operands. The ``xor`` is used to implement the "one's
4802 complement" operation, which is the "~" operator in C.
4807 The two arguments to the '``xor``' instruction must be
4808 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4809 arguments must have identical types.
4814 The truth table used for the '``xor``' instruction is:
4831 .. code-block:: llvm
4833 <result> = xor i32 4, %var ; yields i32:result = 4 ^ %var
4834 <result> = xor i32 15, 40 ; yields i32:result = 39
4835 <result> = xor i32 4, 8 ; yields i32:result = 12
4836 <result> = xor i32 %V, -1 ; yields i32:result = ~%V
4841 LLVM supports several instructions to represent vector operations in a
4842 target-independent manner. These instructions cover the element-access
4843 and vector-specific operations needed to process vectors effectively.
4844 While LLVM does directly support these vector operations, many
4845 sophisticated algorithms will want to use target-specific intrinsics to
4846 take full advantage of a specific target.
4848 .. _i_extractelement:
4850 '``extractelement``' Instruction
4851 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4858 <result> = extractelement <n x <ty>> <val>, <ty2> <idx> ; yields <ty>
4863 The '``extractelement``' instruction extracts a single scalar element
4864 from a vector at a specified index.
4869 The first operand of an '``extractelement``' instruction is a value of
4870 :ref:`vector <t_vector>` type. The second operand is an index indicating
4871 the position from which to extract the element. The index may be a
4872 variable of any integer type.
4877 The result is a scalar of the same type as the element type of ``val``.
4878 Its value is the value at position ``idx`` of ``val``. If ``idx``
4879 exceeds the length of ``val``, the results are undefined.
4884 .. code-block:: llvm
4886 <result> = extractelement <4 x i32> %vec, i32 0 ; yields i32
4888 .. _i_insertelement:
4890 '``insertelement``' Instruction
4891 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4898 <result> = insertelement <n x <ty>> <val>, <ty> <elt>, <ty2> <idx> ; yields <n x <ty>>
4903 The '``insertelement``' instruction inserts a scalar element into a
4904 vector at a specified index.
4909 The first operand of an '``insertelement``' instruction is a value of
4910 :ref:`vector <t_vector>` type. The second operand is a scalar value whose
4911 type must equal the element type of the first operand. The third operand
4912 is an index indicating the position at which to insert the value. The
4913 index may be a variable of any integer type.
4918 The result is a vector of the same type as ``val``. Its element values
4919 are those of ``val`` except at position ``idx``, where it gets the value
4920 ``elt``. If ``idx`` exceeds the length of ``val``, the results are
4926 .. code-block:: llvm
4928 <result> = insertelement <4 x i32> %vec, i32 1, i32 0 ; yields <4 x i32>
4930 .. _i_shufflevector:
4932 '``shufflevector``' Instruction
4933 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4940 <result> = shufflevector <n x <ty>> <v1>, <n x <ty>> <v2>, <m x i32> <mask> ; yields <m x <ty>>
4945 The '``shufflevector``' instruction constructs a permutation of elements
4946 from two input vectors, returning a vector with the same element type as
4947 the input and length that is the same as the shuffle mask.
4952 The first two operands of a '``shufflevector``' instruction are vectors
4953 with the same type. The third argument is a shuffle mask whose element
4954 type is always 'i32'. The result of the instruction is a vector whose
4955 length is the same as the shuffle mask and whose element type is the
4956 same as the element type of the first two operands.
4958 The shuffle mask operand is required to be a constant vector with either
4959 constant integer or undef values.
4964 The elements of the two input vectors are numbered from left to right
4965 across both of the vectors. The shuffle mask operand specifies, for each
4966 element of the result vector, which element of the two input vectors the
4967 result element gets. The element selector may be undef (meaning "don't
4968 care") and the second operand may be undef if performing a shuffle from
4974 .. code-block:: llvm
4976 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
4977 <4 x i32> <i32 0, i32 4, i32 1, i32 5> ; yields <4 x i32>
4978 <result> = shufflevector <4 x i32> %v1, <4 x i32> undef,
4979 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32> - Identity shuffle.
4980 <result> = shufflevector <8 x i32> %v1, <8 x i32> undef,
4981 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32>
4982 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
4983 <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7 > ; yields <8 x i32>
4985 Aggregate Operations
4986 --------------------
4988 LLVM supports several instructions for working with
4989 :ref:`aggregate <t_aggregate>` values.
4993 '``extractvalue``' Instruction
4994 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5001 <result> = extractvalue <aggregate type> <val>, <idx>{, <idx>}*
5006 The '``extractvalue``' instruction extracts the value of a member field
5007 from an :ref:`aggregate <t_aggregate>` value.
5012 The first operand of an '``extractvalue``' instruction is a value of
5013 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The operands are
5014 constant indices to specify which value to extract in a similar manner
5015 as indices in a '``getelementptr``' instruction.
5017 The major differences to ``getelementptr`` indexing are:
5019 - Since the value being indexed is not a pointer, the first index is
5020 omitted and assumed to be zero.
5021 - At least one index must be specified.
5022 - Not only struct indices but also array indices must be in bounds.
5027 The result is the value at the position in the aggregate specified by
5033 .. code-block:: llvm
5035 <result> = extractvalue {i32, float} %agg, 0 ; yields i32
5039 '``insertvalue``' Instruction
5040 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5047 <result> = insertvalue <aggregate type> <val>, <ty> <elt>, <idx>{, <idx>}* ; yields <aggregate type>
5052 The '``insertvalue``' instruction inserts a value into a member field in
5053 an :ref:`aggregate <t_aggregate>` value.
5058 The first operand of an '``insertvalue``' instruction is a value of
5059 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The second operand is
5060 a first-class value to insert. The following operands are constant
5061 indices indicating the position at which to insert the value in a
5062 similar manner as indices in a '``extractvalue``' instruction. The value
5063 to insert must have the same type as the value identified by the
5069 The result is an aggregate of the same type as ``val``. Its value is
5070 that of ``val`` except that the value at the position specified by the
5071 indices is that of ``elt``.
5076 .. code-block:: llvm
5078 %agg1 = insertvalue {i32, float} undef, i32 1, 0 ; yields {i32 1, float undef}
5079 %agg2 = insertvalue {i32, float} %agg1, float %val, 1 ; yields {i32 1, float %val}
5080 %agg3 = insertvalue {i32, {float}} undef, float %val, 1, 0 ; yields {i32 undef, {float %val}}
5084 Memory Access and Addressing Operations
5085 ---------------------------------------
5087 A key design point of an SSA-based representation is how it represents
5088 memory. In LLVM, no memory locations are in SSA form, which makes things
5089 very simple. This section describes how to read, write, and allocate
5094 '``alloca``' Instruction
5095 ^^^^^^^^^^^^^^^^^^^^^^^^
5102 <result> = alloca [inalloca] <type> [, <ty> <NumElements>] [, align <alignment>] ; yields type*:result
5107 The '``alloca``' instruction allocates memory on the stack frame of the
5108 currently executing function, to be automatically released when this
5109 function returns to its caller. The object is always allocated in the
5110 generic address space (address space zero).
5115 The '``alloca``' instruction allocates ``sizeof(<type>)*NumElements``
5116 bytes of memory on the runtime stack, returning a pointer of the
5117 appropriate type to the program. If "NumElements" is specified, it is
5118 the number of elements allocated, otherwise "NumElements" is defaulted
5119 to be one. If a constant alignment is specified, the value result of the
5120 allocation is guaranteed to be aligned to at least that boundary. The
5121 alignment may not be greater than ``1 << 29``. If not specified, or if
5122 zero, the target can choose to align the allocation on any convenient
5123 boundary compatible with the type.
5125 '``type``' may be any sized type.
5130 Memory is allocated; a pointer is returned. The operation is undefined
5131 if there is insufficient stack space for the allocation. '``alloca``'d
5132 memory is automatically released when the function returns. The
5133 '``alloca``' instruction is commonly used to represent automatic
5134 variables that must have an address available. When the function returns
5135 (either with the ``ret`` or ``resume`` instructions), the memory is
5136 reclaimed. Allocating zero bytes is legal, but the result is undefined.
5137 The order in which memory is allocated (ie., which way the stack grows)
5143 .. code-block:: llvm
5145 %ptr = alloca i32 ; yields i32*:ptr
5146 %ptr = alloca i32, i32 4 ; yields i32*:ptr
5147 %ptr = alloca i32, i32 4, align 1024 ; yields i32*:ptr
5148 %ptr = alloca i32, align 1024 ; yields i32*:ptr
5152 '``load``' Instruction
5153 ^^^^^^^^^^^^^^^^^^^^^^
5160 <result> = load [volatile] <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>][, !invariant.load !<index>][, !nonnull !<index>]
5161 <result> = load atomic [volatile] <ty>* <pointer> [singlethread] <ordering>, align <alignment>
5162 !<index> = !{ i32 1 }
5167 The '``load``' instruction is used to read from memory.
5172 The argument to the ``load`` instruction specifies the memory address
5173 from which to load. The pointer must point to a :ref:`first
5174 class <t_firstclass>` type. If the ``load`` is marked as ``volatile``,
5175 then the optimizer is not allowed to modify the number or order of
5176 execution of this ``load`` with other :ref:`volatile
5177 operations <volatile>`.
5179 If the ``load`` is marked as ``atomic``, it takes an extra
5180 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
5181 ``release`` and ``acq_rel`` orderings are not valid on ``load``
5182 instructions. Atomic loads produce :ref:`defined <memmodel>` results
5183 when they may see multiple atomic stores. The type of the pointee must
5184 be an integer type whose bit width is a power of two greater than or
5185 equal to eight and less than or equal to a target-specific size limit.
5186 ``align`` must be explicitly specified on atomic loads, and the load has
5187 undefined behavior if the alignment is not set to a value which is at
5188 least the size in bytes of the pointee. ``!nontemporal`` does not have
5189 any defined semantics for atomic loads.
5191 The optional constant ``align`` argument specifies the alignment of the
5192 operation (that is, the alignment of the memory address). A value of 0
5193 or an omitted ``align`` argument means that the operation has the ABI
5194 alignment for the target. It is the responsibility of the code emitter
5195 to ensure that the alignment information is correct. Overestimating the
5196 alignment results in undefined behavior. Underestimating the alignment
5197 may produce less efficient code. An alignment of 1 is always safe. The
5198 maximum possible alignment is ``1 << 29``.
5200 The optional ``!nontemporal`` metadata must reference a single
5201 metadata name ``<index>`` corresponding to a metadata node with one
5202 ``i32`` entry of value 1. The existence of the ``!nontemporal``
5203 metadata on the instruction tells the optimizer and code generator
5204 that this load is not expected to be reused in the cache. The code
5205 generator may select special instructions to save cache bandwidth, such
5206 as the ``MOVNT`` instruction on x86.
5208 The optional ``!invariant.load`` metadata must reference a single
5209 metadata name ``<index>`` corresponding to a metadata node with no
5210 entries. The existence of the ``!invariant.load`` metadata on the
5211 instruction tells the optimizer and code generator that this load
5212 address points to memory which does not change value during program
5213 execution. The optimizer may then move this load around, for example, by
5214 hoisting it out of loops using loop invariant code motion.
5216 The optional ``!nonnull`` metadata must reference a single
5217 metadata name ``<index>`` corresponding to a metadata node with no
5218 entries. The existence of the ``!nonnull`` metadata on the
5219 instruction tells the optimizer that the value loaded is known to
5220 never be null. This is analogous to the ''nonnull'' attribute
5221 on parameters and return values. This metadata can only be applied
5222 to loads of a pointer type.
5227 The location of memory pointed to is loaded. If the value being loaded
5228 is of scalar type then the number of bytes read does not exceed the
5229 minimum number of bytes needed to hold all bits of the type. For
5230 example, loading an ``i24`` reads at most three bytes. When loading a
5231 value of a type like ``i20`` with a size that is not an integral number
5232 of bytes, the result is undefined if the value was not originally
5233 written using a store of the same type.
5238 .. code-block:: llvm
5240 %ptr = alloca i32 ; yields i32*:ptr
5241 store i32 3, i32* %ptr ; yields void
5242 %val = load i32* %ptr ; yields i32:val = i32 3
5246 '``store``' Instruction
5247 ^^^^^^^^^^^^^^^^^^^^^^^
5254 store [volatile] <ty> <value>, <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>] ; yields void
5255 store atomic [volatile] <ty> <value>, <ty>* <pointer> [singlethread] <ordering>, align <alignment> ; yields void
5260 The '``store``' instruction is used to write to memory.
5265 There are two arguments to the ``store`` instruction: a value to store
5266 and an address at which to store it. The type of the ``<pointer>``
5267 operand must be a pointer to the :ref:`first class <t_firstclass>` type of
5268 the ``<value>`` operand. If the ``store`` is marked as ``volatile``,
5269 then the optimizer is not allowed to modify the number or order of
5270 execution of this ``store`` with other :ref:`volatile
5271 operations <volatile>`.
5273 If the ``store`` is marked as ``atomic``, it takes an extra
5274 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
5275 ``acquire`` and ``acq_rel`` orderings aren't valid on ``store``
5276 instructions. Atomic loads produce :ref:`defined <memmodel>` results
5277 when they may see multiple atomic stores. The type of the pointee must
5278 be an integer type whose bit width is a power of two greater than or
5279 equal to eight and less than or equal to a target-specific size limit.
5280 ``align`` must be explicitly specified on atomic stores, and the store
5281 has undefined behavior if the alignment is not set to a value which is
5282 at least the size in bytes of the pointee. ``!nontemporal`` does not
5283 have any defined semantics for atomic stores.
5285 The optional constant ``align`` argument specifies the alignment of the
5286 operation (that is, the alignment of the memory address). A value of 0
5287 or an omitted ``align`` argument means that the operation has the ABI
5288 alignment for the target. It is the responsibility of the code emitter
5289 to ensure that the alignment information is correct. Overestimating the
5290 alignment results in undefined behavior. Underestimating the
5291 alignment may produce less efficient code. An alignment of 1 is always
5292 safe. The maximum possible alignment is ``1 << 29``.
5294 The optional ``!nontemporal`` metadata must reference a single metadata
5295 name ``<index>`` corresponding to a metadata node with one ``i32`` entry of
5296 value 1. The existence of the ``!nontemporal`` metadata on the instruction
5297 tells the optimizer and code generator that this load is not expected to
5298 be reused in the cache. The code generator may select special
5299 instructions to save cache bandwidth, such as the MOVNT instruction on
5305 The contents of memory are updated to contain ``<value>`` at the
5306 location specified by the ``<pointer>`` operand. If ``<value>`` is
5307 of scalar type then the number of bytes written does not exceed the
5308 minimum number of bytes needed to hold all bits of the type. For
5309 example, storing an ``i24`` writes at most three bytes. When writing a
5310 value of a type like ``i20`` with a size that is not an integral number
5311 of bytes, it is unspecified what happens to the extra bits that do not
5312 belong to the type, but they will typically be overwritten.
5317 .. code-block:: llvm
5319 %ptr = alloca i32 ; yields i32*:ptr
5320 store i32 3, i32* %ptr ; yields void
5321 %val = load i32* %ptr ; yields i32:val = i32 3
5325 '``fence``' Instruction
5326 ^^^^^^^^^^^^^^^^^^^^^^^
5333 fence [singlethread] <ordering> ; yields void
5338 The '``fence``' instruction is used to introduce happens-before edges
5344 '``fence``' instructions take an :ref:`ordering <ordering>` argument which
5345 defines what *synchronizes-with* edges they add. They can only be given
5346 ``acquire``, ``release``, ``acq_rel``, and ``seq_cst`` orderings.
5351 A fence A which has (at least) ``release`` ordering semantics
5352 *synchronizes with* a fence B with (at least) ``acquire`` ordering
5353 semantics if and only if there exist atomic operations X and Y, both
5354 operating on some atomic object M, such that A is sequenced before X, X
5355 modifies M (either directly or through some side effect of a sequence
5356 headed by X), Y is sequenced before B, and Y observes M. This provides a
5357 *happens-before* dependency between A and B. Rather than an explicit
5358 ``fence``, one (but not both) of the atomic operations X or Y might
5359 provide a ``release`` or ``acquire`` (resp.) ordering constraint and
5360 still *synchronize-with* the explicit ``fence`` and establish the
5361 *happens-before* edge.
5363 A ``fence`` which has ``seq_cst`` ordering, in addition to having both
5364 ``acquire`` and ``release`` semantics specified above, participates in
5365 the global program order of other ``seq_cst`` operations and/or fences.
5367 The optional ":ref:`singlethread <singlethread>`" argument specifies
5368 that the fence only synchronizes with other fences in the same thread.
5369 (This is useful for interacting with signal handlers.)
5374 .. code-block:: llvm
5376 fence acquire ; yields void
5377 fence singlethread seq_cst ; yields void
5381 '``cmpxchg``' Instruction
5382 ^^^^^^^^^^^^^^^^^^^^^^^^^
5389 cmpxchg [weak] [volatile] <ty>* <pointer>, <ty> <cmp>, <ty> <new> [singlethread] <success ordering> <failure ordering> ; yields { ty, i1 }
5394 The '``cmpxchg``' instruction is used to atomically modify memory. It
5395 loads a value in memory and compares it to a given value. If they are
5396 equal, it tries to store a new value into the memory.
5401 There are three arguments to the '``cmpxchg``' instruction: an address
5402 to operate on, a value to compare to the value currently be at that
5403 address, and a new value to place at that address if the compared values
5404 are equal. The type of '<cmp>' must be an integer type whose bit width
5405 is a power of two greater than or equal to eight and less than or equal
5406 to a target-specific size limit. '<cmp>' and '<new>' must have the same
5407 type, and the type of '<pointer>' must be a pointer to that type. If the
5408 ``cmpxchg`` is marked as ``volatile``, then the optimizer is not allowed
5409 to modify the number or order of execution of this ``cmpxchg`` with
5410 other :ref:`volatile operations <volatile>`.
5412 The success and failure :ref:`ordering <ordering>` arguments specify how this
5413 ``cmpxchg`` synchronizes with other atomic operations. Both ordering parameters
5414 must be at least ``monotonic``, the ordering constraint on failure must be no
5415 stronger than that on success, and the failure ordering cannot be either
5416 ``release`` or ``acq_rel``.
5418 The optional "``singlethread``" argument declares that the ``cmpxchg``
5419 is only atomic with respect to code (usually signal handlers) running in
5420 the same thread as the ``cmpxchg``. Otherwise the cmpxchg is atomic with
5421 respect to all other code in the system.
5423 The pointer passed into cmpxchg must have alignment greater than or
5424 equal to the size in memory of the operand.
5429 The contents of memory at the location specified by the '``<pointer>``' operand
5430 is read and compared to '``<cmp>``'; if the read value is the equal, the
5431 '``<new>``' is written. The original value at the location is returned, together
5432 with a flag indicating success (true) or failure (false).
5434 If the cmpxchg operation is marked as ``weak`` then a spurious failure is
5435 permitted: the operation may not write ``<new>`` even if the comparison
5438 If the cmpxchg operation is strong (the default), the i1 value is 1 if and only
5439 if the value loaded equals ``cmp``.
5441 A successful ``cmpxchg`` is a read-modify-write instruction for the purpose of
5442 identifying release sequences. A failed ``cmpxchg`` is equivalent to an atomic
5443 load with an ordering parameter determined the second ordering parameter.
5448 .. code-block:: llvm
5451 %orig = atomic load i32* %ptr unordered ; yields i32
5455 %cmp = phi i32 [ %orig, %entry ], [%old, %loop]
5456 %squared = mul i32 %cmp, %cmp
5457 %val_success = cmpxchg i32* %ptr, i32 %cmp, i32 %squared acq_rel monotonic ; yields { i32, i1 }
5458 %value_loaded = extractvalue { i32, i1 } %val_success, 0
5459 %success = extractvalue { i32, i1 } %val_success, 1
5460 br i1 %success, label %done, label %loop
5467 '``atomicrmw``' Instruction
5468 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
5475 atomicrmw [volatile] <operation> <ty>* <pointer>, <ty> <value> [singlethread] <ordering> ; yields ty
5480 The '``atomicrmw``' instruction is used to atomically modify memory.
5485 There are three arguments to the '``atomicrmw``' instruction: an
5486 operation to apply, an address whose value to modify, an argument to the
5487 operation. The operation must be one of the following keywords:
5501 The type of '<value>' must be an integer type whose bit width is a power
5502 of two greater than or equal to eight and less than or equal to a
5503 target-specific size limit. The type of the '``<pointer>``' operand must
5504 be a pointer to that type. If the ``atomicrmw`` is marked as
5505 ``volatile``, then the optimizer is not allowed to modify the number or
5506 order of execution of this ``atomicrmw`` with other :ref:`volatile
5507 operations <volatile>`.
5512 The contents of memory at the location specified by the '``<pointer>``'
5513 operand are atomically read, modified, and written back. The original
5514 value at the location is returned. The modification is specified by the
5517 - xchg: ``*ptr = val``
5518 - add: ``*ptr = *ptr + val``
5519 - sub: ``*ptr = *ptr - val``
5520 - and: ``*ptr = *ptr & val``
5521 - nand: ``*ptr = ~(*ptr & val)``
5522 - or: ``*ptr = *ptr | val``
5523 - xor: ``*ptr = *ptr ^ val``
5524 - max: ``*ptr = *ptr > val ? *ptr : val`` (using a signed comparison)
5525 - min: ``*ptr = *ptr < val ? *ptr : val`` (using a signed comparison)
5526 - umax: ``*ptr = *ptr > val ? *ptr : val`` (using an unsigned
5528 - umin: ``*ptr = *ptr < val ? *ptr : val`` (using an unsigned
5534 .. code-block:: llvm
5536 %old = atomicrmw add i32* %ptr, i32 1 acquire ; yields i32
5538 .. _i_getelementptr:
5540 '``getelementptr``' Instruction
5541 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5548 <result> = getelementptr <pty>* <ptrval>{, <ty> <idx>}*
5549 <result> = getelementptr inbounds <pty>* <ptrval>{, <ty> <idx>}*
5550 <result> = getelementptr <ptr vector> ptrval, <vector index type> idx
5555 The '``getelementptr``' instruction is used to get the address of a
5556 subelement of an :ref:`aggregate <t_aggregate>` data structure. It performs
5557 address calculation only and does not access memory.
5562 The first argument is always a pointer or a vector of pointers, and
5563 forms the basis of the calculation. The remaining arguments are indices
5564 that indicate which of the elements of the aggregate object are indexed.
5565 The interpretation of each index is dependent on the type being indexed
5566 into. The first index always indexes the pointer value given as the
5567 first argument, the second index indexes a value of the type pointed to
5568 (not necessarily the value directly pointed to, since the first index
5569 can be non-zero), etc. The first type indexed into must be a pointer
5570 value, subsequent types can be arrays, vectors, and structs. Note that
5571 subsequent types being indexed into can never be pointers, since that
5572 would require loading the pointer before continuing calculation.
5574 The type of each index argument depends on the type it is indexing into.
5575 When indexing into a (optionally packed) structure, only ``i32`` integer
5576 **constants** are allowed (when using a vector of indices they must all
5577 be the **same** ``i32`` integer constant). When indexing into an array,
5578 pointer or vector, integers of any width are allowed, and they are not
5579 required to be constant. These integers are treated as signed values
5582 For example, let's consider a C code fragment and how it gets compiled
5598 int *foo(struct ST *s) {
5599 return &s[1].Z.B[5][13];
5602 The LLVM code generated by Clang is:
5604 .. code-block:: llvm
5606 %struct.RT = type { i8, [10 x [20 x i32]], i8 }
5607 %struct.ST = type { i32, double, %struct.RT }
5609 define i32* @foo(%struct.ST* %s) nounwind uwtable readnone optsize ssp {
5611 %arrayidx = getelementptr inbounds %struct.ST* %s, i64 1, i32 2, i32 1, i64 5, i64 13
5618 In the example above, the first index is indexing into the
5619 '``%struct.ST*``' type, which is a pointer, yielding a '``%struct.ST``'
5620 = '``{ i32, double, %struct.RT }``' type, a structure. The second index
5621 indexes into the third element of the structure, yielding a
5622 '``%struct.RT``' = '``{ i8 , [10 x [20 x i32]], i8 }``' type, another
5623 structure. The third index indexes into the second element of the
5624 structure, yielding a '``[10 x [20 x i32]]``' type, an array. The two
5625 dimensions of the array are subscripted into, yielding an '``i32``'
5626 type. The '``getelementptr``' instruction returns a pointer to this
5627 element, thus computing a value of '``i32*``' type.
5629 Note that it is perfectly legal to index partially through a structure,
5630 returning a pointer to an inner element. Because of this, the LLVM code
5631 for the given testcase is equivalent to:
5633 .. code-block:: llvm
5635 define i32* @foo(%struct.ST* %s) {
5636 %t1 = getelementptr %struct.ST* %s, i32 1 ; yields %struct.ST*:%t1
5637 %t2 = getelementptr %struct.ST* %t1, i32 0, i32 2 ; yields %struct.RT*:%t2
5638 %t3 = getelementptr %struct.RT* %t2, i32 0, i32 1 ; yields [10 x [20 x i32]]*:%t3
5639 %t4 = getelementptr [10 x [20 x i32]]* %t3, i32 0, i32 5 ; yields [20 x i32]*:%t4
5640 %t5 = getelementptr [20 x i32]* %t4, i32 0, i32 13 ; yields i32*:%t5
5644 If the ``inbounds`` keyword is present, the result value of the
5645 ``getelementptr`` is a :ref:`poison value <poisonvalues>` if the base
5646 pointer is not an *in bounds* address of an allocated object, or if any
5647 of the addresses that would be formed by successive addition of the
5648 offsets implied by the indices to the base address with infinitely
5649 precise signed arithmetic are not an *in bounds* address of that
5650 allocated object. The *in bounds* addresses for an allocated object are
5651 all the addresses that point into the object, plus the address one byte
5652 past the end. In cases where the base is a vector of pointers the
5653 ``inbounds`` keyword applies to each of the computations element-wise.
5655 If the ``inbounds`` keyword is not present, the offsets are added to the
5656 base address with silently-wrapping two's complement arithmetic. If the
5657 offsets have a different width from the pointer, they are sign-extended
5658 or truncated to the width of the pointer. The result value of the
5659 ``getelementptr`` may be outside the object pointed to by the base
5660 pointer. The result value may not necessarily be used to access memory
5661 though, even if it happens to point into allocated storage. See the
5662 :ref:`Pointer Aliasing Rules <pointeraliasing>` section for more
5665 The getelementptr instruction is often confusing. For some more insight
5666 into how it works, see :doc:`the getelementptr FAQ <GetElementPtr>`.
5671 .. code-block:: llvm
5673 ; yields [12 x i8]*:aptr
5674 %aptr = getelementptr {i32, [12 x i8]}* %saptr, i64 0, i32 1
5676 %vptr = getelementptr {i32, <2 x i8>}* %svptr, i64 0, i32 1, i32 1
5678 %eptr = getelementptr [12 x i8]* %aptr, i64 0, i32 1
5680 %iptr = getelementptr [10 x i32]* @arr, i16 0, i16 0
5682 In cases where the pointer argument is a vector of pointers, each index
5683 must be a vector with the same number of elements. For example:
5685 .. code-block:: llvm
5687 %A = getelementptr <4 x i8*> %ptrs, <4 x i64> %offsets,
5689 Conversion Operations
5690 ---------------------
5692 The instructions in this category are the conversion instructions
5693 (casting) which all take a single operand and a type. They perform
5694 various bit conversions on the operand.
5696 '``trunc .. to``' Instruction
5697 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5704 <result> = trunc <ty> <value> to <ty2> ; yields ty2
5709 The '``trunc``' instruction truncates its operand to the type ``ty2``.
5714 The '``trunc``' instruction takes a value to trunc, and a type to trunc
5715 it to. Both types must be of :ref:`integer <t_integer>` types, or vectors
5716 of the same number of integers. The bit size of the ``value`` must be
5717 larger than the bit size of the destination type, ``ty2``. Equal sized
5718 types are not allowed.
5723 The '``trunc``' instruction truncates the high order bits in ``value``
5724 and converts the remaining bits to ``ty2``. Since the source size must
5725 be larger than the destination size, ``trunc`` cannot be a *no-op cast*.
5726 It will always truncate bits.
5731 .. code-block:: llvm
5733 %X = trunc i32 257 to i8 ; yields i8:1
5734 %Y = trunc i32 123 to i1 ; yields i1:true
5735 %Z = trunc i32 122 to i1 ; yields i1:false
5736 %W = trunc <2 x i16> <i16 8, i16 7> to <2 x i8> ; yields <i8 8, i8 7>
5738 '``zext .. to``' Instruction
5739 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5746 <result> = zext <ty> <value> to <ty2> ; yields ty2
5751 The '``zext``' instruction zero extends its operand to type ``ty2``.
5756 The '``zext``' instruction takes a value to cast, and a type to cast it
5757 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
5758 the same number of integers. The bit size of the ``value`` must be
5759 smaller than the bit size of the destination type, ``ty2``.
5764 The ``zext`` fills the high order bits of the ``value`` with zero bits
5765 until it reaches the size of the destination type, ``ty2``.
5767 When zero extending from i1, the result will always be either 0 or 1.
5772 .. code-block:: llvm
5774 %X = zext i32 257 to i64 ; yields i64:257
5775 %Y = zext i1 true to i32 ; yields i32:1
5776 %Z = zext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
5778 '``sext .. to``' Instruction
5779 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5786 <result> = sext <ty> <value> to <ty2> ; yields ty2
5791 The '``sext``' sign extends ``value`` to the type ``ty2``.
5796 The '``sext``' instruction takes a value to cast, and a type to cast it
5797 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
5798 the same number of integers. The bit size of the ``value`` must be
5799 smaller than the bit size of the destination type, ``ty2``.
5804 The '``sext``' instruction performs a sign extension by copying the sign
5805 bit (highest order bit) of the ``value`` until it reaches the bit size
5806 of the type ``ty2``.
5808 When sign extending from i1, the extension always results in -1 or 0.
5813 .. code-block:: llvm
5815 %X = sext i8 -1 to i16 ; yields i16 :65535
5816 %Y = sext i1 true to i32 ; yields i32:-1
5817 %Z = sext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
5819 '``fptrunc .. to``' Instruction
5820 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5827 <result> = fptrunc <ty> <value> to <ty2> ; yields ty2
5832 The '``fptrunc``' instruction truncates ``value`` to type ``ty2``.
5837 The '``fptrunc``' instruction takes a :ref:`floating point <t_floating>`
5838 value to cast and a :ref:`floating point <t_floating>` type to cast it to.
5839 The size of ``value`` must be larger than the size of ``ty2``. This
5840 implies that ``fptrunc`` cannot be used to make a *no-op cast*.
5845 The '``fptrunc``' instruction truncates a ``value`` from a larger
5846 :ref:`floating point <t_floating>` type to a smaller :ref:`floating
5847 point <t_floating>` type. If the value cannot fit within the
5848 destination type, ``ty2``, then the results are undefined.
5853 .. code-block:: llvm
5855 %X = fptrunc double 123.0 to float ; yields float:123.0
5856 %Y = fptrunc double 1.0E+300 to float ; yields undefined
5858 '``fpext .. to``' Instruction
5859 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5866 <result> = fpext <ty> <value> to <ty2> ; yields ty2
5871 The '``fpext``' extends a floating point ``value`` to a larger floating
5877 The '``fpext``' instruction takes a :ref:`floating point <t_floating>`
5878 ``value`` to cast, and a :ref:`floating point <t_floating>` type to cast it
5879 to. The source type must be smaller than the destination type.
5884 The '``fpext``' instruction extends the ``value`` from a smaller
5885 :ref:`floating point <t_floating>` type to a larger :ref:`floating
5886 point <t_floating>` type. The ``fpext`` cannot be used to make a
5887 *no-op cast* because it always changes bits. Use ``bitcast`` to make a
5888 *no-op cast* for a floating point cast.
5893 .. code-block:: llvm
5895 %X = fpext float 3.125 to double ; yields double:3.125000e+00
5896 %Y = fpext double %X to fp128 ; yields fp128:0xL00000000000000004000900000000000
5898 '``fptoui .. to``' Instruction
5899 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5906 <result> = fptoui <ty> <value> to <ty2> ; yields ty2
5911 The '``fptoui``' converts a floating point ``value`` to its unsigned
5912 integer equivalent of type ``ty2``.
5917 The '``fptoui``' instruction takes a value to cast, which must be a
5918 scalar or vector :ref:`floating point <t_floating>` value, and a type to
5919 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
5920 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
5921 type with the same number of elements as ``ty``
5926 The '``fptoui``' instruction converts its :ref:`floating
5927 point <t_floating>` operand into the nearest (rounding towards zero)
5928 unsigned integer value. If the value cannot fit in ``ty2``, the results
5934 .. code-block:: llvm
5936 %X = fptoui double 123.0 to i32 ; yields i32:123
5937 %Y = fptoui float 1.0E+300 to i1 ; yields undefined:1
5938 %Z = fptoui float 1.04E+17 to i8 ; yields undefined:1
5940 '``fptosi .. to``' Instruction
5941 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5948 <result> = fptosi <ty> <value> to <ty2> ; yields ty2
5953 The '``fptosi``' instruction converts :ref:`floating point <t_floating>`
5954 ``value`` to type ``ty2``.
5959 The '``fptosi``' instruction takes a value to cast, which must be a
5960 scalar or vector :ref:`floating point <t_floating>` value, and a type to
5961 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
5962 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
5963 type with the same number of elements as ``ty``
5968 The '``fptosi``' instruction converts its :ref:`floating
5969 point <t_floating>` operand into the nearest (rounding towards zero)
5970 signed integer value. If the value cannot fit in ``ty2``, the results
5976 .. code-block:: llvm
5978 %X = fptosi double -123.0 to i32 ; yields i32:-123
5979 %Y = fptosi float 1.0E-247 to i1 ; yields undefined:1
5980 %Z = fptosi float 1.04E+17 to i8 ; yields undefined:1
5982 '``uitofp .. to``' Instruction
5983 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5990 <result> = uitofp <ty> <value> to <ty2> ; yields ty2
5995 The '``uitofp``' instruction regards ``value`` as an unsigned integer
5996 and converts that value to the ``ty2`` type.
6001 The '``uitofp``' instruction takes a value to cast, which must be a
6002 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
6003 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
6004 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
6005 type with the same number of elements as ``ty``
6010 The '``uitofp``' instruction interprets its operand as an unsigned
6011 integer quantity and converts it to the corresponding floating point
6012 value. If the value cannot fit in the floating point value, the results
6018 .. code-block:: llvm
6020 %X = uitofp i32 257 to float ; yields float:257.0
6021 %Y = uitofp i8 -1 to double ; yields double:255.0
6023 '``sitofp .. to``' Instruction
6024 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6031 <result> = sitofp <ty> <value> to <ty2> ; yields ty2
6036 The '``sitofp``' instruction regards ``value`` as a signed integer and
6037 converts that value to the ``ty2`` type.
6042 The '``sitofp``' instruction takes a value to cast, which must be a
6043 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
6044 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
6045 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
6046 type with the same number of elements as ``ty``
6051 The '``sitofp``' instruction interprets its operand as a signed integer
6052 quantity and converts it to the corresponding floating point value. If
6053 the value cannot fit in the floating point value, the results are
6059 .. code-block:: llvm
6061 %X = sitofp i32 257 to float ; yields float:257.0
6062 %Y = sitofp i8 -1 to double ; yields double:-1.0
6066 '``ptrtoint .. to``' Instruction
6067 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6074 <result> = ptrtoint <ty> <value> to <ty2> ; yields ty2
6079 The '``ptrtoint``' instruction converts the pointer or a vector of
6080 pointers ``value`` to the integer (or vector of integers) type ``ty2``.
6085 The '``ptrtoint``' instruction takes a ``value`` to cast, which must be
6086 a a value of type :ref:`pointer <t_pointer>` or a vector of pointers, and a
6087 type to cast it to ``ty2``, which must be an :ref:`integer <t_integer>` or
6088 a vector of integers type.
6093 The '``ptrtoint``' instruction converts ``value`` to integer type
6094 ``ty2`` by interpreting the pointer value as an integer and either
6095 truncating or zero extending that value to the size of the integer type.
6096 If ``value`` is smaller than ``ty2`` then a zero extension is done. If
6097 ``value`` is larger than ``ty2`` then a truncation is done. If they are
6098 the same size, then nothing is done (*no-op cast*) other than a type
6104 .. code-block:: llvm
6106 %X = ptrtoint i32* %P to i8 ; yields truncation on 32-bit architecture
6107 %Y = ptrtoint i32* %P to i64 ; yields zero extension on 32-bit architecture
6108 %Z = ptrtoint <4 x i32*> %P to <4 x i64>; yields vector zero extension for a vector of addresses on 32-bit architecture
6112 '``inttoptr .. to``' Instruction
6113 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6120 <result> = inttoptr <ty> <value> to <ty2> ; yields ty2
6125 The '``inttoptr``' instruction converts an integer ``value`` to a
6126 pointer type, ``ty2``.
6131 The '``inttoptr``' instruction takes an :ref:`integer <t_integer>` value to
6132 cast, and a type to cast it to, which must be a :ref:`pointer <t_pointer>`
6138 The '``inttoptr``' instruction converts ``value`` to type ``ty2`` by
6139 applying either a zero extension or a truncation depending on the size
6140 of the integer ``value``. If ``value`` is larger than the size of a
6141 pointer then a truncation is done. If ``value`` is smaller than the size
6142 of a pointer then a zero extension is done. If they are the same size,
6143 nothing is done (*no-op cast*).
6148 .. code-block:: llvm
6150 %X = inttoptr i32 255 to i32* ; yields zero extension on 64-bit architecture
6151 %Y = inttoptr i32 255 to i32* ; yields no-op on 32-bit architecture
6152 %Z = inttoptr i64 0 to i32* ; yields truncation on 32-bit architecture
6153 %Z = inttoptr <4 x i32> %G to <4 x i8*>; yields truncation of vector G to four pointers
6157 '``bitcast .. to``' Instruction
6158 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6165 <result> = bitcast <ty> <value> to <ty2> ; yields ty2
6170 The '``bitcast``' instruction converts ``value`` to type ``ty2`` without
6176 The '``bitcast``' instruction takes a value to cast, which must be a
6177 non-aggregate first class value, and a type to cast it to, which must
6178 also be a non-aggregate :ref:`first class <t_firstclass>` type. The
6179 bit sizes of ``value`` and the destination type, ``ty2``, must be
6180 identical. If the source type is a pointer, the destination type must
6181 also be a pointer of the same size. This instruction supports bitwise
6182 conversion of vectors to integers and to vectors of other types (as
6183 long as they have the same size).
6188 The '``bitcast``' instruction converts ``value`` to type ``ty2``. It
6189 is always a *no-op cast* because no bits change with this
6190 conversion. The conversion is done as if the ``value`` had been stored
6191 to memory and read back as type ``ty2``. Pointer (or vector of
6192 pointers) types may only be converted to other pointer (or vector of
6193 pointers) types with the same address space through this instruction.
6194 To convert pointers to other types, use the :ref:`inttoptr <i_inttoptr>`
6195 or :ref:`ptrtoint <i_ptrtoint>` instructions first.
6200 .. code-block:: llvm
6202 %X = bitcast i8 255 to i8 ; yields i8 :-1
6203 %Y = bitcast i32* %x to sint* ; yields sint*:%x
6204 %Z = bitcast <2 x int> %V to i64; ; yields i64: %V
6205 %Z = bitcast <2 x i32*> %V to <2 x i64*> ; yields <2 x i64*>
6207 .. _i_addrspacecast:
6209 '``addrspacecast .. to``' Instruction
6210 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6217 <result> = addrspacecast <pty> <ptrval> to <pty2> ; yields pty2
6222 The '``addrspacecast``' instruction converts ``ptrval`` from ``pty`` in
6223 address space ``n`` to type ``pty2`` in address space ``m``.
6228 The '``addrspacecast``' instruction takes a pointer or vector of pointer value
6229 to cast and a pointer type to cast it to, which must have a different
6235 The '``addrspacecast``' instruction converts the pointer value
6236 ``ptrval`` to type ``pty2``. It can be a *no-op cast* or a complex
6237 value modification, depending on the target and the address space
6238 pair. Pointer conversions within the same address space must be
6239 performed with the ``bitcast`` instruction. Note that if the address space
6240 conversion is legal then both result and operand refer to the same memory
6246 .. code-block:: llvm
6248 %X = addrspacecast i32* %x to i32 addrspace(1)* ; yields i32 addrspace(1)*:%x
6249 %Y = addrspacecast i32 addrspace(1)* %y to i64 addrspace(2)* ; yields i64 addrspace(2)*:%y
6250 %Z = addrspacecast <4 x i32*> %z to <4 x float addrspace(3)*> ; yields <4 x float addrspace(3)*>:%z
6257 The instructions in this category are the "miscellaneous" instructions,
6258 which defy better classification.
6262 '``icmp``' Instruction
6263 ^^^^^^^^^^^^^^^^^^^^^^
6270 <result> = icmp <cond> <ty> <op1>, <op2> ; yields i1 or <N x i1>:result
6275 The '``icmp``' instruction returns a boolean value or a vector of
6276 boolean values based on comparison of its two integer, integer vector,
6277 pointer, or pointer vector operands.
6282 The '``icmp``' instruction takes three operands. The first operand is
6283 the condition code indicating the kind of comparison to perform. It is
6284 not a value, just a keyword. The possible condition code are:
6287 #. ``ne``: not equal
6288 #. ``ugt``: unsigned greater than
6289 #. ``uge``: unsigned greater or equal
6290 #. ``ult``: unsigned less than
6291 #. ``ule``: unsigned less or equal
6292 #. ``sgt``: signed greater than
6293 #. ``sge``: signed greater or equal
6294 #. ``slt``: signed less than
6295 #. ``sle``: signed less or equal
6297 The remaining two arguments must be :ref:`integer <t_integer>` or
6298 :ref:`pointer <t_pointer>` or integer :ref:`vector <t_vector>` typed. They
6299 must also be identical types.
6304 The '``icmp``' compares ``op1`` and ``op2`` according to the condition
6305 code given as ``cond``. The comparison performed always yields either an
6306 :ref:`i1 <t_integer>` or vector of ``i1`` result, as follows:
6308 #. ``eq``: yields ``true`` if the operands are equal, ``false``
6309 otherwise. No sign interpretation is necessary or performed.
6310 #. ``ne``: yields ``true`` if the operands are unequal, ``false``
6311 otherwise. No sign interpretation is necessary or performed.
6312 #. ``ugt``: interprets the operands as unsigned values and yields
6313 ``true`` if ``op1`` is greater than ``op2``.
6314 #. ``uge``: interprets the operands as unsigned values and yields
6315 ``true`` if ``op1`` is greater than or equal to ``op2``.
6316 #. ``ult``: interprets the operands as unsigned values and yields
6317 ``true`` if ``op1`` is less than ``op2``.
6318 #. ``ule``: interprets the operands as unsigned values and yields
6319 ``true`` if ``op1`` is less than or equal to ``op2``.
6320 #. ``sgt``: interprets the operands as signed values and yields ``true``
6321 if ``op1`` is greater than ``op2``.
6322 #. ``sge``: interprets the operands as signed values and yields ``true``
6323 if ``op1`` is greater than or equal to ``op2``.
6324 #. ``slt``: interprets the operands as signed values and yields ``true``
6325 if ``op1`` is less than ``op2``.
6326 #. ``sle``: interprets the operands as signed values and yields ``true``
6327 if ``op1`` is less than or equal to ``op2``.
6329 If the operands are :ref:`pointer <t_pointer>` typed, the pointer values
6330 are compared as if they were integers.
6332 If the operands are integer vectors, then they are compared element by
6333 element. The result is an ``i1`` vector with the same number of elements
6334 as the values being compared. Otherwise, the result is an ``i1``.
6339 .. code-block:: llvm
6341 <result> = icmp eq i32 4, 5 ; yields: result=false
6342 <result> = icmp ne float* %X, %X ; yields: result=false
6343 <result> = icmp ult i16 4, 5 ; yields: result=true
6344 <result> = icmp sgt i16 4, 5 ; yields: result=false
6345 <result> = icmp ule i16 -4, 5 ; yields: result=false
6346 <result> = icmp sge i16 4, 5 ; yields: result=false
6348 Note that the code generator does not yet support vector types with the
6349 ``icmp`` instruction.
6353 '``fcmp``' Instruction
6354 ^^^^^^^^^^^^^^^^^^^^^^
6361 <result> = fcmp <cond> <ty> <op1>, <op2> ; yields i1 or <N x i1>:result
6366 The '``fcmp``' instruction returns a boolean value or vector of boolean
6367 values based on comparison of its operands.
6369 If the operands are floating point scalars, then the result type is a
6370 boolean (:ref:`i1 <t_integer>`).
6372 If the operands are floating point vectors, then the result type is a
6373 vector of boolean with the same number of elements as the operands being
6379 The '``fcmp``' instruction takes three operands. The first operand is
6380 the condition code indicating the kind of comparison to perform. It is
6381 not a value, just a keyword. The possible condition code are:
6383 #. ``false``: no comparison, always returns false
6384 #. ``oeq``: ordered and equal
6385 #. ``ogt``: ordered and greater than
6386 #. ``oge``: ordered and greater than or equal
6387 #. ``olt``: ordered and less than
6388 #. ``ole``: ordered and less than or equal
6389 #. ``one``: ordered and not equal
6390 #. ``ord``: ordered (no nans)
6391 #. ``ueq``: unordered or equal
6392 #. ``ugt``: unordered or greater than
6393 #. ``uge``: unordered or greater than or equal
6394 #. ``ult``: unordered or less than
6395 #. ``ule``: unordered or less than or equal
6396 #. ``une``: unordered or not equal
6397 #. ``uno``: unordered (either nans)
6398 #. ``true``: no comparison, always returns true
6400 *Ordered* means that neither operand is a QNAN while *unordered* means
6401 that either operand may be a QNAN.
6403 Each of ``val1`` and ``val2`` arguments must be either a :ref:`floating
6404 point <t_floating>` type or a :ref:`vector <t_vector>` of floating point
6405 type. They must have identical types.
6410 The '``fcmp``' instruction compares ``op1`` and ``op2`` according to the
6411 condition code given as ``cond``. If the operands are vectors, then the
6412 vectors are compared element by element. Each comparison performed
6413 always yields an :ref:`i1 <t_integer>` result, as follows:
6415 #. ``false``: always yields ``false``, regardless of operands.
6416 #. ``oeq``: yields ``true`` if both operands are not a QNAN and ``op1``
6417 is equal to ``op2``.
6418 #. ``ogt``: yields ``true`` if both operands are not a QNAN and ``op1``
6419 is greater than ``op2``.
6420 #. ``oge``: yields ``true`` if both operands are not a QNAN and ``op1``
6421 is greater than or equal to ``op2``.
6422 #. ``olt``: yields ``true`` if both operands are not a QNAN and ``op1``
6423 is less than ``op2``.
6424 #. ``ole``: yields ``true`` if both operands are not a QNAN and ``op1``
6425 is less than or equal to ``op2``.
6426 #. ``one``: yields ``true`` if both operands are not a QNAN and ``op1``
6427 is not equal to ``op2``.
6428 #. ``ord``: yields ``true`` if both operands are not a QNAN.
6429 #. ``ueq``: yields ``true`` if either operand is a QNAN or ``op1`` is
6431 #. ``ugt``: yields ``true`` if either operand is a QNAN or ``op1`` is
6432 greater than ``op2``.
6433 #. ``uge``: yields ``true`` if either operand is a QNAN or ``op1`` is
6434 greater than or equal to ``op2``.
6435 #. ``ult``: yields ``true`` if either operand is a QNAN or ``op1`` is
6437 #. ``ule``: yields ``true`` if either operand is a QNAN or ``op1`` is
6438 less than or equal to ``op2``.
6439 #. ``une``: yields ``true`` if either operand is a QNAN or ``op1`` is
6440 not equal to ``op2``.
6441 #. ``uno``: yields ``true`` if either operand is a QNAN.
6442 #. ``true``: always yields ``true``, regardless of operands.
6447 .. code-block:: llvm
6449 <result> = fcmp oeq float 4.0, 5.0 ; yields: result=false
6450 <result> = fcmp one float 4.0, 5.0 ; yields: result=true
6451 <result> = fcmp olt float 4.0, 5.0 ; yields: result=true
6452 <result> = fcmp ueq double 1.0, 2.0 ; yields: result=false
6454 Note that the code generator does not yet support vector types with the
6455 ``fcmp`` instruction.
6459 '``phi``' Instruction
6460 ^^^^^^^^^^^^^^^^^^^^^
6467 <result> = phi <ty> [ <val0>, <label0>], ...
6472 The '``phi``' instruction is used to implement the φ node in the SSA
6473 graph representing the function.
6478 The type of the incoming values is specified with the first type field.
6479 After this, the '``phi``' instruction takes a list of pairs as
6480 arguments, with one pair for each predecessor basic block of the current
6481 block. Only values of :ref:`first class <t_firstclass>` type may be used as
6482 the value arguments to the PHI node. Only labels may be used as the
6485 There must be no non-phi instructions between the start of a basic block
6486 and the PHI instructions: i.e. PHI instructions must be first in a basic
6489 For the purposes of the SSA form, the use of each incoming value is
6490 deemed to occur on the edge from the corresponding predecessor block to
6491 the current block (but after any definition of an '``invoke``'
6492 instruction's return value on the same edge).
6497 At runtime, the '``phi``' instruction logically takes on the value
6498 specified by the pair corresponding to the predecessor basic block that
6499 executed just prior to the current block.
6504 .. code-block:: llvm
6506 Loop: ; Infinite loop that counts from 0 on up...
6507 %indvar = phi i32 [ 0, %LoopHeader ], [ %nextindvar, %Loop ]
6508 %nextindvar = add i32 %indvar, 1
6513 '``select``' Instruction
6514 ^^^^^^^^^^^^^^^^^^^^^^^^
6521 <result> = select selty <cond>, <ty> <val1>, <ty> <val2> ; yields ty
6523 selty is either i1 or {<N x i1>}
6528 The '``select``' instruction is used to choose one value based on a
6529 condition, without IR-level branching.
6534 The '``select``' instruction requires an 'i1' value or a vector of 'i1'
6535 values indicating the condition, and two values of the same :ref:`first
6536 class <t_firstclass>` type. If the val1/val2 are vectors and the
6537 condition is a scalar, then entire vectors are selected, not individual
6543 If the condition is an i1 and it evaluates to 1, the instruction returns
6544 the first value argument; otherwise, it returns the second value
6547 If the condition is a vector of i1, then the value arguments must be
6548 vectors of the same size, and the selection is done element by element.
6553 .. code-block:: llvm
6555 %X = select i1 true, i8 17, i8 42 ; yields i8:17
6559 '``call``' Instruction
6560 ^^^^^^^^^^^^^^^^^^^^^^
6567 <result> = [tail | musttail] call [cconv] [ret attrs] <ty> [<fnty>*] <fnptrval>(<function args>) [fn attrs]
6572 The '``call``' instruction represents a simple function call.
6577 This instruction requires several arguments:
6579 #. The optional ``tail`` and ``musttail`` markers indicate that the optimizers
6580 should perform tail call optimization. The ``tail`` marker is a hint that
6581 `can be ignored <CodeGenerator.html#sibcallopt>`_. The ``musttail`` marker
6582 means that the call must be tail call optimized in order for the program to
6583 be correct. The ``musttail`` marker provides these guarantees:
6585 #. The call will not cause unbounded stack growth if it is part of a
6586 recursive cycle in the call graph.
6587 #. Arguments with the :ref:`inalloca <attr_inalloca>` attribute are
6590 Both markers imply that the callee does not access allocas or varargs from
6591 the caller. Calls marked ``musttail`` must obey the following additional
6594 - The call must immediately precede a :ref:`ret <i_ret>` instruction,
6595 or a pointer bitcast followed by a ret instruction.
6596 - The ret instruction must return the (possibly bitcasted) value
6597 produced by the call or void.
6598 - The caller and callee prototypes must match. Pointer types of
6599 parameters or return types may differ in pointee type, but not
6601 - The calling conventions of the caller and callee must match.
6602 - All ABI-impacting function attributes, such as sret, byval, inreg,
6603 returned, and inalloca, must match.
6604 - The callee must be varargs iff the caller is varargs. Bitcasting a
6605 non-varargs function to the appropriate varargs type is legal so
6606 long as the non-varargs prefixes obey the other rules.
6608 Tail call optimization for calls marked ``tail`` is guaranteed to occur if
6609 the following conditions are met:
6611 - Caller and callee both have the calling convention ``fastcc``.
6612 - The call is in tail position (ret immediately follows call and ret
6613 uses value of call or is void).
6614 - Option ``-tailcallopt`` is enabled, or
6615 ``llvm::GuaranteedTailCallOpt`` is ``true``.
6616 - `Platform-specific constraints are
6617 met. <CodeGenerator.html#tailcallopt>`_
6619 #. The optional "cconv" marker indicates which :ref:`calling
6620 convention <callingconv>` the call should use. If none is
6621 specified, the call defaults to using C calling conventions. The
6622 calling convention of the call must match the calling convention of
6623 the target function, or else the behavior is undefined.
6624 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
6625 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
6627 #. '``ty``': the type of the call instruction itself which is also the
6628 type of the return value. Functions that return no value are marked
6630 #. '``fnty``': shall be the signature of the pointer to function value
6631 being invoked. The argument types must match the types implied by
6632 this signature. This type can be omitted if the function is not
6633 varargs and if the function type does not return a pointer to a
6635 #. '``fnptrval``': An LLVM value containing a pointer to a function to
6636 be invoked. In most cases, this is a direct function invocation, but
6637 indirect ``call``'s are just as possible, calling an arbitrary pointer
6639 #. '``function args``': argument list whose types match the function
6640 signature argument types and parameter attributes. All arguments must
6641 be of :ref:`first class <t_firstclass>` type. If the function signature
6642 indicates the function accepts a variable number of arguments, the
6643 extra arguments can be specified.
6644 #. The optional :ref:`function attributes <fnattrs>` list. Only
6645 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
6646 attributes are valid here.
6651 The '``call``' instruction is used to cause control flow to transfer to
6652 a specified function, with its incoming arguments bound to the specified
6653 values. Upon a '``ret``' instruction in the called function, control
6654 flow continues with the instruction after the function call, and the
6655 return value of the function is bound to the result argument.
6660 .. code-block:: llvm
6662 %retval = call i32 @test(i32 %argc)
6663 call i32 (i8*, ...)* @printf(i8* %msg, i32 12, i8 42) ; yields i32
6664 %X = tail call i32 @foo() ; yields i32
6665 %Y = tail call fastcc i32 @foo() ; yields i32
6666 call void %foo(i8 97 signext)
6668 %struct.A = type { i32, i8 }
6669 %r = call %struct.A @foo() ; yields { i32, i8 }
6670 %gr = extractvalue %struct.A %r, 0 ; yields i32
6671 %gr1 = extractvalue %struct.A %r, 1 ; yields i8
6672 %Z = call void @foo() noreturn ; indicates that %foo never returns normally
6673 %ZZ = call zeroext i32 @bar() ; Return value is %zero extended
6675 llvm treats calls to some functions with names and arguments that match
6676 the standard C99 library as being the C99 library functions, and may
6677 perform optimizations or generate code for them under that assumption.
6678 This is something we'd like to change in the future to provide better
6679 support for freestanding environments and non-C-based languages.
6683 '``va_arg``' Instruction
6684 ^^^^^^^^^^^^^^^^^^^^^^^^
6691 <resultval> = va_arg <va_list*> <arglist>, <argty>
6696 The '``va_arg``' instruction is used to access arguments passed through
6697 the "variable argument" area of a function call. It is used to implement
6698 the ``va_arg`` macro in C.
6703 This instruction takes a ``va_list*`` value and the type of the
6704 argument. It returns a value of the specified argument type and
6705 increments the ``va_list`` to point to the next argument. The actual
6706 type of ``va_list`` is target specific.
6711 The '``va_arg``' instruction loads an argument of the specified type
6712 from the specified ``va_list`` and causes the ``va_list`` to point to
6713 the next argument. For more information, see the variable argument
6714 handling :ref:`Intrinsic Functions <int_varargs>`.
6716 It is legal for this instruction to be called in a function which does
6717 not take a variable number of arguments, for example, the ``vfprintf``
6720 ``va_arg`` is an LLVM instruction instead of an :ref:`intrinsic
6721 function <intrinsics>` because it takes a type as an argument.
6726 See the :ref:`variable argument processing <int_varargs>` section.
6728 Note that the code generator does not yet fully support va\_arg on many
6729 targets. Also, it does not currently support va\_arg with aggregate
6730 types on any target.
6734 '``landingpad``' Instruction
6735 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6742 <resultval> = landingpad <resultty> personality <type> <pers_fn> <clause>+
6743 <resultval> = landingpad <resultty> personality <type> <pers_fn> cleanup <clause>*
6745 <clause> := catch <type> <value>
6746 <clause> := filter <array constant type> <array constant>
6751 The '``landingpad``' instruction is used by `LLVM's exception handling
6752 system <ExceptionHandling.html#overview>`_ to specify that a basic block
6753 is a landing pad --- one where the exception lands, and corresponds to the
6754 code found in the ``catch`` portion of a ``try``/``catch`` sequence. It
6755 defines values supplied by the personality function (``pers_fn``) upon
6756 re-entry to the function. The ``resultval`` has the type ``resultty``.
6761 This instruction takes a ``pers_fn`` value. This is the personality
6762 function associated with the unwinding mechanism. The optional
6763 ``cleanup`` flag indicates that the landing pad block is a cleanup.
6765 A ``clause`` begins with the clause type --- ``catch`` or ``filter`` --- and
6766 contains the global variable representing the "type" that may be caught
6767 or filtered respectively. Unlike the ``catch`` clause, the ``filter``
6768 clause takes an array constant as its argument. Use
6769 "``[0 x i8**] undef``" for a filter which cannot throw. The
6770 '``landingpad``' instruction must contain *at least* one ``clause`` or
6771 the ``cleanup`` flag.
6776 The '``landingpad``' instruction defines the values which are set by the
6777 personality function (``pers_fn``) upon re-entry to the function, and
6778 therefore the "result type" of the ``landingpad`` instruction. As with
6779 calling conventions, how the personality function results are
6780 represented in LLVM IR is target specific.
6782 The clauses are applied in order from top to bottom. If two
6783 ``landingpad`` instructions are merged together through inlining, the
6784 clauses from the calling function are appended to the list of clauses.
6785 When the call stack is being unwound due to an exception being thrown,
6786 the exception is compared against each ``clause`` in turn. If it doesn't
6787 match any of the clauses, and the ``cleanup`` flag is not set, then
6788 unwinding continues further up the call stack.
6790 The ``landingpad`` instruction has several restrictions:
6792 - A landing pad block is a basic block which is the unwind destination
6793 of an '``invoke``' instruction.
6794 - A landing pad block must have a '``landingpad``' instruction as its
6795 first non-PHI instruction.
6796 - There can be only one '``landingpad``' instruction within the landing
6798 - A basic block that is not a landing pad block may not include a
6799 '``landingpad``' instruction.
6800 - All '``landingpad``' instructions in a function must have the same
6801 personality function.
6806 .. code-block:: llvm
6808 ;; A landing pad which can catch an integer.
6809 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6811 ;; A landing pad that is a cleanup.
6812 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6814 ;; A landing pad which can catch an integer and can only throw a double.
6815 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6817 filter [1 x i8**] [@_ZTId]
6824 LLVM supports the notion of an "intrinsic function". These functions
6825 have well known names and semantics and are required to follow certain
6826 restrictions. Overall, these intrinsics represent an extension mechanism
6827 for the LLVM language that does not require changing all of the
6828 transformations in LLVM when adding to the language (or the bitcode
6829 reader/writer, the parser, etc...).
6831 Intrinsic function names must all start with an "``llvm.``" prefix. This
6832 prefix is reserved in LLVM for intrinsic names; thus, function names may
6833 not begin with this prefix. Intrinsic functions must always be external
6834 functions: you cannot define the body of intrinsic functions. Intrinsic
6835 functions may only be used in call or invoke instructions: it is illegal
6836 to take the address of an intrinsic function. Additionally, because
6837 intrinsic functions are part of the LLVM language, it is required if any
6838 are added that they be documented here.
6840 Some intrinsic functions can be overloaded, i.e., the intrinsic
6841 represents a family of functions that perform the same operation but on
6842 different data types. Because LLVM can represent over 8 million
6843 different integer types, overloading is used commonly to allow an
6844 intrinsic function to operate on any integer type. One or more of the
6845 argument types or the result type can be overloaded to accept any
6846 integer type. Argument types may also be defined as exactly matching a
6847 previous argument's type or the result type. This allows an intrinsic
6848 function which accepts multiple arguments, but needs all of them to be
6849 of the same type, to only be overloaded with respect to a single
6850 argument or the result.
6852 Overloaded intrinsics will have the names of its overloaded argument
6853 types encoded into its function name, each preceded by a period. Only
6854 those types which are overloaded result in a name suffix. Arguments
6855 whose type is matched against another type do not. For example, the
6856 ``llvm.ctpop`` function can take an integer of any width and returns an
6857 integer of exactly the same integer width. This leads to a family of
6858 functions such as ``i8 @llvm.ctpop.i8(i8 %val)`` and
6859 ``i29 @llvm.ctpop.i29(i29 %val)``. Only one type, the return type, is
6860 overloaded, and only one type suffix is required. Because the argument's
6861 type is matched against the return type, it does not require its own
6864 To learn how to add an intrinsic function, please see the `Extending
6865 LLVM Guide <ExtendingLLVM.html>`_.
6869 Variable Argument Handling Intrinsics
6870 -------------------------------------
6872 Variable argument support is defined in LLVM with the
6873 :ref:`va_arg <i_va_arg>` instruction and these three intrinsic
6874 functions. These functions are related to the similarly named macros
6875 defined in the ``<stdarg.h>`` header file.
6877 All of these functions operate on arguments that use a target-specific
6878 value type "``va_list``". The LLVM assembly language reference manual
6879 does not define what this type is, so all transformations should be
6880 prepared to handle these functions regardless of the type used.
6882 This example shows how the :ref:`va_arg <i_va_arg>` instruction and the
6883 variable argument handling intrinsic functions are used.
6885 .. code-block:: llvm
6887 define i32 @test(i32 %X, ...) {
6888 ; Initialize variable argument processing
6890 %ap2 = bitcast i8** %ap to i8*
6891 call void @llvm.va_start(i8* %ap2)
6893 ; Read a single integer argument
6894 %tmp = va_arg i8** %ap, i32
6896 ; Demonstrate usage of llvm.va_copy and llvm.va_end
6898 %aq2 = bitcast i8** %aq to i8*
6899 call void @llvm.va_copy(i8* %aq2, i8* %ap2)
6900 call void @llvm.va_end(i8* %aq2)
6902 ; Stop processing of arguments.
6903 call void @llvm.va_end(i8* %ap2)
6907 declare void @llvm.va_start(i8*)
6908 declare void @llvm.va_copy(i8*, i8*)
6909 declare void @llvm.va_end(i8*)
6913 '``llvm.va_start``' Intrinsic
6914 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6921 declare void @llvm.va_start(i8* <arglist>)
6926 The '``llvm.va_start``' intrinsic initializes ``*<arglist>`` for
6927 subsequent use by ``va_arg``.
6932 The argument is a pointer to a ``va_list`` element to initialize.
6937 The '``llvm.va_start``' intrinsic works just like the ``va_start`` macro
6938 available in C. In a target-dependent way, it initializes the
6939 ``va_list`` element to which the argument points, so that the next call
6940 to ``va_arg`` will produce the first variable argument passed to the
6941 function. Unlike the C ``va_start`` macro, this intrinsic does not need
6942 to know the last argument of the function as the compiler can figure
6945 '``llvm.va_end``' Intrinsic
6946 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6953 declare void @llvm.va_end(i8* <arglist>)
6958 The '``llvm.va_end``' intrinsic destroys ``*<arglist>``, which has been
6959 initialized previously with ``llvm.va_start`` or ``llvm.va_copy``.
6964 The argument is a pointer to a ``va_list`` to destroy.
6969 The '``llvm.va_end``' intrinsic works just like the ``va_end`` macro
6970 available in C. In a target-dependent way, it destroys the ``va_list``
6971 element to which the argument points. Calls to
6972 :ref:`llvm.va_start <int_va_start>` and
6973 :ref:`llvm.va_copy <int_va_copy>` must be matched exactly with calls to
6978 '``llvm.va_copy``' Intrinsic
6979 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6986 declare void @llvm.va_copy(i8* <destarglist>, i8* <srcarglist>)
6991 The '``llvm.va_copy``' intrinsic copies the current argument position
6992 from the source argument list to the destination argument list.
6997 The first argument is a pointer to a ``va_list`` element to initialize.
6998 The second argument is a pointer to a ``va_list`` element to copy from.
7003 The '``llvm.va_copy``' intrinsic works just like the ``va_copy`` macro
7004 available in C. In a target-dependent way, it copies the source
7005 ``va_list`` element into the destination ``va_list`` element. This
7006 intrinsic is necessary because the `` llvm.va_start`` intrinsic may be
7007 arbitrarily complex and require, for example, memory allocation.
7009 Accurate Garbage Collection Intrinsics
7010 --------------------------------------
7012 LLVM support for `Accurate Garbage Collection <GarbageCollection.html>`_
7013 (GC) requires the implementation and generation of these intrinsics.
7014 These intrinsics allow identification of :ref:`GC roots on the
7015 stack <int_gcroot>`, as well as garbage collector implementations that
7016 require :ref:`read <int_gcread>` and :ref:`write <int_gcwrite>` barriers.
7017 Front-ends for type-safe garbage collected languages should generate
7018 these intrinsics to make use of the LLVM garbage collectors. For more
7019 details, see `Accurate Garbage Collection with
7020 LLVM <GarbageCollection.html>`_.
7022 The garbage collection intrinsics only operate on objects in the generic
7023 address space (address space zero).
7027 '``llvm.gcroot``' Intrinsic
7028 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7035 declare void @llvm.gcroot(i8** %ptrloc, i8* %metadata)
7040 The '``llvm.gcroot``' intrinsic declares the existence of a GC root to
7041 the code generator, and allows some metadata to be associated with it.
7046 The first argument specifies the address of a stack object that contains
7047 the root pointer. The second pointer (which must be either a constant or
7048 a global value address) contains the meta-data to be associated with the
7054 At runtime, a call to this intrinsic stores a null pointer into the
7055 "ptrloc" location. At compile-time, the code generator generates
7056 information to allow the runtime to find the pointer at GC safe points.
7057 The '``llvm.gcroot``' intrinsic may only be used in a function which
7058 :ref:`specifies a GC algorithm <gc>`.
7062 '``llvm.gcread``' Intrinsic
7063 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7070 declare i8* @llvm.gcread(i8* %ObjPtr, i8** %Ptr)
7075 The '``llvm.gcread``' intrinsic identifies reads of references from heap
7076 locations, allowing garbage collector implementations that require read
7082 The second argument is the address to read from, which should be an
7083 address allocated from the garbage collector. The first object is a
7084 pointer to the start of the referenced object, if needed by the language
7085 runtime (otherwise null).
7090 The '``llvm.gcread``' intrinsic has the same semantics as a load
7091 instruction, but may be replaced with substantially more complex code by
7092 the garbage collector runtime, as needed. The '``llvm.gcread``'
7093 intrinsic may only be used in a function which :ref:`specifies a GC
7098 '``llvm.gcwrite``' Intrinsic
7099 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7106 declare void @llvm.gcwrite(i8* %P1, i8* %Obj, i8** %P2)
7111 The '``llvm.gcwrite``' intrinsic identifies writes of references to heap
7112 locations, allowing garbage collector implementations that require write
7113 barriers (such as generational or reference counting collectors).
7118 The first argument is the reference to store, the second is the start of
7119 the object to store it to, and the third is the address of the field of
7120 Obj to store to. If the runtime does not require a pointer to the
7121 object, Obj may be null.
7126 The '``llvm.gcwrite``' intrinsic has the same semantics as a store
7127 instruction, but may be replaced with substantially more complex code by
7128 the garbage collector runtime, as needed. The '``llvm.gcwrite``'
7129 intrinsic may only be used in a function which :ref:`specifies a GC
7132 Code Generator Intrinsics
7133 -------------------------
7135 These intrinsics are provided by LLVM to expose special features that
7136 may only be implemented with code generator support.
7138 '``llvm.returnaddress``' Intrinsic
7139 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7146 declare i8 *@llvm.returnaddress(i32 <level>)
7151 The '``llvm.returnaddress``' intrinsic attempts to compute a
7152 target-specific value indicating the return address of the current
7153 function or one of its callers.
7158 The argument to this intrinsic indicates which function to return the
7159 address for. Zero indicates the calling function, one indicates its
7160 caller, etc. The argument is **required** to be a constant integer
7166 The '``llvm.returnaddress``' intrinsic either returns a pointer
7167 indicating the return address of the specified call frame, or zero if it
7168 cannot be identified. The value returned by this intrinsic is likely to
7169 be incorrect or 0 for arguments other than zero, so it should only be
7170 used for debugging purposes.
7172 Note that calling this intrinsic does not prevent function inlining or
7173 other aggressive transformations, so the value returned may not be that
7174 of the obvious source-language caller.
7176 '``llvm.frameaddress``' Intrinsic
7177 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7184 declare i8* @llvm.frameaddress(i32 <level>)
7189 The '``llvm.frameaddress``' intrinsic attempts to return the
7190 target-specific frame pointer value for the specified stack frame.
7195 The argument to this intrinsic indicates which function to return the
7196 frame pointer for. Zero indicates the calling function, one indicates
7197 its caller, etc. The argument is **required** to be a constant integer
7203 The '``llvm.frameaddress``' intrinsic either returns a pointer
7204 indicating the frame address of the specified call frame, or zero if it
7205 cannot be identified. The value returned by this intrinsic is likely to
7206 be incorrect or 0 for arguments other than zero, so it should only be
7207 used for debugging purposes.
7209 Note that calling this intrinsic does not prevent function inlining or
7210 other aggressive transformations, so the value returned may not be that
7211 of the obvious source-language caller.
7213 .. _int_read_register:
7214 .. _int_write_register:
7216 '``llvm.read_register``' and '``llvm.write_register``' Intrinsics
7217 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7224 declare i32 @llvm.read_register.i32(metadata)
7225 declare i64 @llvm.read_register.i64(metadata)
7226 declare void @llvm.write_register.i32(metadata, i32 @value)
7227 declare void @llvm.write_register.i64(metadata, i64 @value)
7228 !0 = metadata !{metadata !"sp\00"}
7233 The '``llvm.read_register``' and '``llvm.write_register``' intrinsics
7234 provides access to the named register. The register must be valid on
7235 the architecture being compiled to. The type needs to be compatible
7236 with the register being read.
7241 The '``llvm.read_register``' intrinsic returns the current value of the
7242 register, where possible. The '``llvm.write_register``' intrinsic sets
7243 the current value of the register, where possible.
7245 This is useful to implement named register global variables that need
7246 to always be mapped to a specific register, as is common practice on
7247 bare-metal programs including OS kernels.
7249 The compiler doesn't check for register availability or use of the used
7250 register in surrounding code, including inline assembly. Because of that,
7251 allocatable registers are not supported.
7253 Warning: So far it only works with the stack pointer on selected
7254 architectures (ARM, AArch64, PowerPC and x86_64). Significant amount of
7255 work is needed to support other registers and even more so, allocatable
7260 '``llvm.stacksave``' Intrinsic
7261 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7268 declare i8* @llvm.stacksave()
7273 The '``llvm.stacksave``' intrinsic is used to remember the current state
7274 of the function stack, for use with
7275 :ref:`llvm.stackrestore <int_stackrestore>`. This is useful for
7276 implementing language features like scoped automatic variable sized
7282 This intrinsic returns a opaque pointer value that can be passed to
7283 :ref:`llvm.stackrestore <int_stackrestore>`. When an
7284 ``llvm.stackrestore`` intrinsic is executed with a value saved from
7285 ``llvm.stacksave``, it effectively restores the state of the stack to
7286 the state it was in when the ``llvm.stacksave`` intrinsic executed. In
7287 practice, this pops any :ref:`alloca <i_alloca>` blocks from the stack that
7288 were allocated after the ``llvm.stacksave`` was executed.
7290 .. _int_stackrestore:
7292 '``llvm.stackrestore``' Intrinsic
7293 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7300 declare void @llvm.stackrestore(i8* %ptr)
7305 The '``llvm.stackrestore``' intrinsic is used to restore the state of
7306 the function stack to the state it was in when the corresponding
7307 :ref:`llvm.stacksave <int_stacksave>` intrinsic executed. This is
7308 useful for implementing language features like scoped automatic variable
7309 sized arrays in C99.
7314 See the description for :ref:`llvm.stacksave <int_stacksave>`.
7316 '``llvm.prefetch``' Intrinsic
7317 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7324 declare void @llvm.prefetch(i8* <address>, i32 <rw>, i32 <locality>, i32 <cache type>)
7329 The '``llvm.prefetch``' intrinsic is a hint to the code generator to
7330 insert a prefetch instruction if supported; otherwise, it is a noop.
7331 Prefetches have no effect on the behavior of the program but can change
7332 its performance characteristics.
7337 ``address`` is the address to be prefetched, ``rw`` is the specifier
7338 determining if the fetch should be for a read (0) or write (1), and
7339 ``locality`` is a temporal locality specifier ranging from (0) - no
7340 locality, to (3) - extremely local keep in cache. The ``cache type``
7341 specifies whether the prefetch is performed on the data (1) or
7342 instruction (0) cache. The ``rw``, ``locality`` and ``cache type``
7343 arguments must be constant integers.
7348 This intrinsic does not modify the behavior of the program. In
7349 particular, prefetches cannot trap and do not produce a value. On
7350 targets that support this intrinsic, the prefetch can provide hints to
7351 the processor cache for better performance.
7353 '``llvm.pcmarker``' Intrinsic
7354 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7361 declare void @llvm.pcmarker(i32 <id>)
7366 The '``llvm.pcmarker``' intrinsic is a method to export a Program
7367 Counter (PC) in a region of code to simulators and other tools. The
7368 method is target specific, but it is expected that the marker will use
7369 exported symbols to transmit the PC of the marker. The marker makes no
7370 guarantees that it will remain with any specific instruction after
7371 optimizations. It is possible that the presence of a marker will inhibit
7372 optimizations. The intended use is to be inserted after optimizations to
7373 allow correlations of simulation runs.
7378 ``id`` is a numerical id identifying the marker.
7383 This intrinsic does not modify the behavior of the program. Backends
7384 that do not support this intrinsic may ignore it.
7386 '``llvm.readcyclecounter``' Intrinsic
7387 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7394 declare i64 @llvm.readcyclecounter()
7399 The '``llvm.readcyclecounter``' intrinsic provides access to the cycle
7400 counter register (or similar low latency, high accuracy clocks) on those
7401 targets that support it. On X86, it should map to RDTSC. On Alpha, it
7402 should map to RPCC. As the backing counters overflow quickly (on the
7403 order of 9 seconds on alpha), this should only be used for small
7409 When directly supported, reading the cycle counter should not modify any
7410 memory. Implementations are allowed to either return a application
7411 specific value or a system wide value. On backends without support, this
7412 is lowered to a constant 0.
7414 Note that runtime support may be conditional on the privilege-level code is
7415 running at and the host platform.
7417 '``llvm.clear_cache``' Intrinsic
7418 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7425 declare void @llvm.clear_cache(i8*, i8*)
7430 The '``llvm.clear_cache``' intrinsic ensures visibility of modifications
7431 in the specified range to the execution unit of the processor. On
7432 targets with non-unified instruction and data cache, the implementation
7433 flushes the instruction cache.
7438 On platforms with coherent instruction and data caches (e.g. x86), this
7439 intrinsic is a nop. On platforms with non-coherent instruction and data
7440 cache (e.g. ARM, MIPS), the intrinsic is lowered either to appropriate
7441 instructions or a system call, if cache flushing requires special
7444 The default behavior is to emit a call to ``__clear_cache`` from the run
7447 This instrinsic does *not* empty the instruction pipeline. Modifications
7448 of the current function are outside the scope of the intrinsic.
7450 Standard C Library Intrinsics
7451 -----------------------------
7453 LLVM provides intrinsics for a few important standard C library
7454 functions. These intrinsics allow source-language front-ends to pass
7455 information about the alignment of the pointer arguments to the code
7456 generator, providing opportunity for more efficient code generation.
7460 '``llvm.memcpy``' Intrinsic
7461 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7466 This is an overloaded intrinsic. You can use ``llvm.memcpy`` on any
7467 integer bit width and for different address spaces. Not all targets
7468 support all bit widths however.
7472 declare void @llvm.memcpy.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
7473 i32 <len>, i32 <align>, i1 <isvolatile>)
7474 declare void @llvm.memcpy.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
7475 i64 <len>, i32 <align>, i1 <isvolatile>)
7480 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
7481 source location to the destination location.
7483 Note that, unlike the standard libc function, the ``llvm.memcpy.*``
7484 intrinsics do not return a value, takes extra alignment/isvolatile
7485 arguments and the pointers can be in specified address spaces.
7490 The first argument is a pointer to the destination, the second is a
7491 pointer to the source. The third argument is an integer argument
7492 specifying the number of bytes to copy, the fourth argument is the
7493 alignment of the source and destination locations, and the fifth is a
7494 boolean indicating a volatile access.
7496 If the call to this intrinsic has an alignment value that is not 0 or 1,
7497 then the caller guarantees that both the source and destination pointers
7498 are aligned to that boundary.
7500 If the ``isvolatile`` parameter is ``true``, the ``llvm.memcpy`` call is
7501 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
7502 very cleanly specified and it is unwise to depend on it.
7507 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
7508 source location to the destination location, which are not allowed to
7509 overlap. It copies "len" bytes of memory over. If the argument is known
7510 to be aligned to some boundary, this can be specified as the fourth
7511 argument, otherwise it should be set to 0 or 1 (both meaning no alignment).
7513 '``llvm.memmove``' Intrinsic
7514 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7519 This is an overloaded intrinsic. You can use llvm.memmove on any integer
7520 bit width and for different address space. Not all targets support all
7525 declare void @llvm.memmove.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
7526 i32 <len>, i32 <align>, i1 <isvolatile>)
7527 declare void @llvm.memmove.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
7528 i64 <len>, i32 <align>, i1 <isvolatile>)
7533 The '``llvm.memmove.*``' intrinsics move a block of memory from the
7534 source location to the destination location. It is similar to the
7535 '``llvm.memcpy``' intrinsic but allows the two memory locations to
7538 Note that, unlike the standard libc function, the ``llvm.memmove.*``
7539 intrinsics do not return a value, takes extra alignment/isvolatile
7540 arguments and the pointers can be in specified address spaces.
7545 The first argument is a pointer to the destination, the second is a
7546 pointer to the source. The third argument is an integer argument
7547 specifying the number of bytes to copy, the fourth argument is the
7548 alignment of the source and destination locations, and the fifth is a
7549 boolean indicating a volatile access.
7551 If the call to this intrinsic has an alignment value that is not 0 or 1,
7552 then the caller guarantees that the source and destination pointers are
7553 aligned to that boundary.
7555 If the ``isvolatile`` parameter is ``true``, the ``llvm.memmove`` call
7556 is a :ref:`volatile operation <volatile>`. The detailed access behavior is
7557 not very cleanly specified and it is unwise to depend on it.
7562 The '``llvm.memmove.*``' intrinsics copy a block of memory from the
7563 source location to the destination location, which may overlap. It
7564 copies "len" bytes of memory over. If the argument is known to be
7565 aligned to some boundary, this can be specified as the fourth argument,
7566 otherwise it should be set to 0 or 1 (both meaning no alignment).
7568 '``llvm.memset.*``' Intrinsics
7569 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7574 This is an overloaded intrinsic. You can use llvm.memset on any integer
7575 bit width and for different address spaces. However, not all targets
7576 support all bit widths.
7580 declare void @llvm.memset.p0i8.i32(i8* <dest>, i8 <val>,
7581 i32 <len>, i32 <align>, i1 <isvolatile>)
7582 declare void @llvm.memset.p0i8.i64(i8* <dest>, i8 <val>,
7583 i64 <len>, i32 <align>, i1 <isvolatile>)
7588 The '``llvm.memset.*``' intrinsics fill a block of memory with a
7589 particular byte value.
7591 Note that, unlike the standard libc function, the ``llvm.memset``
7592 intrinsic does not return a value and takes extra alignment/volatile
7593 arguments. Also, the destination can be in an arbitrary address space.
7598 The first argument is a pointer to the destination to fill, the second
7599 is the byte value with which to fill it, the third argument is an
7600 integer argument specifying the number of bytes to fill, and the fourth
7601 argument is the known alignment of the destination location.
7603 If the call to this intrinsic has an alignment value that is not 0 or 1,
7604 then the caller guarantees that the destination pointer is aligned to
7607 If the ``isvolatile`` parameter is ``true``, the ``llvm.memset`` call is
7608 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
7609 very cleanly specified and it is unwise to depend on it.
7614 The '``llvm.memset.*``' intrinsics fill "len" bytes of memory starting
7615 at the destination location. If the argument is known to be aligned to
7616 some boundary, this can be specified as the fourth argument, otherwise
7617 it should be set to 0 or 1 (both meaning no alignment).
7619 '``llvm.sqrt.*``' Intrinsic
7620 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7625 This is an overloaded intrinsic. You can use ``llvm.sqrt`` on any
7626 floating point or vector of floating point type. Not all targets support
7631 declare float @llvm.sqrt.f32(float %Val)
7632 declare double @llvm.sqrt.f64(double %Val)
7633 declare x86_fp80 @llvm.sqrt.f80(x86_fp80 %Val)
7634 declare fp128 @llvm.sqrt.f128(fp128 %Val)
7635 declare ppc_fp128 @llvm.sqrt.ppcf128(ppc_fp128 %Val)
7640 The '``llvm.sqrt``' intrinsics return the sqrt of the specified operand,
7641 returning the same value as the libm '``sqrt``' functions would. Unlike
7642 ``sqrt`` in libm, however, ``llvm.sqrt`` has undefined behavior for
7643 negative numbers other than -0.0 (which allows for better optimization,
7644 because there is no need to worry about errno being set).
7645 ``llvm.sqrt(-0.0)`` is defined to return -0.0 like IEEE sqrt.
7650 The argument and return value are floating point numbers of the same
7656 This function returns the sqrt of the specified operand if it is a
7657 nonnegative floating point number.
7659 '``llvm.powi.*``' Intrinsic
7660 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7665 This is an overloaded intrinsic. You can use ``llvm.powi`` on any
7666 floating point or vector of floating point type. Not all targets support
7671 declare float @llvm.powi.f32(float %Val, i32 %power)
7672 declare double @llvm.powi.f64(double %Val, i32 %power)
7673 declare x86_fp80 @llvm.powi.f80(x86_fp80 %Val, i32 %power)
7674 declare fp128 @llvm.powi.f128(fp128 %Val, i32 %power)
7675 declare ppc_fp128 @llvm.powi.ppcf128(ppc_fp128 %Val, i32 %power)
7680 The '``llvm.powi.*``' intrinsics return the first operand raised to the
7681 specified (positive or negative) power. The order of evaluation of
7682 multiplications is not defined. When a vector of floating point type is
7683 used, the second argument remains a scalar integer value.
7688 The second argument is an integer power, and the first is a value to
7689 raise to that power.
7694 This function returns the first value raised to the second power with an
7695 unspecified sequence of rounding operations.
7697 '``llvm.sin.*``' Intrinsic
7698 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7703 This is an overloaded intrinsic. You can use ``llvm.sin`` on any
7704 floating point or vector of floating point type. Not all targets support
7709 declare float @llvm.sin.f32(float %Val)
7710 declare double @llvm.sin.f64(double %Val)
7711 declare x86_fp80 @llvm.sin.f80(x86_fp80 %Val)
7712 declare fp128 @llvm.sin.f128(fp128 %Val)
7713 declare ppc_fp128 @llvm.sin.ppcf128(ppc_fp128 %Val)
7718 The '``llvm.sin.*``' intrinsics return the sine of the operand.
7723 The argument and return value are floating point numbers of the same
7729 This function returns the sine of the specified operand, returning the
7730 same values as the libm ``sin`` functions would, and handles error
7731 conditions in the same way.
7733 '``llvm.cos.*``' Intrinsic
7734 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7739 This is an overloaded intrinsic. You can use ``llvm.cos`` on any
7740 floating point or vector of floating point type. Not all targets support
7745 declare float @llvm.cos.f32(float %Val)
7746 declare double @llvm.cos.f64(double %Val)
7747 declare x86_fp80 @llvm.cos.f80(x86_fp80 %Val)
7748 declare fp128 @llvm.cos.f128(fp128 %Val)
7749 declare ppc_fp128 @llvm.cos.ppcf128(ppc_fp128 %Val)
7754 The '``llvm.cos.*``' intrinsics return the cosine of the operand.
7759 The argument and return value are floating point numbers of the same
7765 This function returns the cosine of the specified operand, returning the
7766 same values as the libm ``cos`` functions would, and handles error
7767 conditions in the same way.
7769 '``llvm.pow.*``' Intrinsic
7770 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7775 This is an overloaded intrinsic. You can use ``llvm.pow`` on any
7776 floating point or vector of floating point type. Not all targets support
7781 declare float @llvm.pow.f32(float %Val, float %Power)
7782 declare double @llvm.pow.f64(double %Val, double %Power)
7783 declare x86_fp80 @llvm.pow.f80(x86_fp80 %Val, x86_fp80 %Power)
7784 declare fp128 @llvm.pow.f128(fp128 %Val, fp128 %Power)
7785 declare ppc_fp128 @llvm.pow.ppcf128(ppc_fp128 %Val, ppc_fp128 Power)
7790 The '``llvm.pow.*``' intrinsics return the first operand raised to the
7791 specified (positive or negative) power.
7796 The second argument is a floating point power, and the first is a value
7797 to raise to that power.
7802 This function returns the first value raised to the second power,
7803 returning the same values as the libm ``pow`` functions would, and
7804 handles error conditions in the same way.
7806 '``llvm.exp.*``' Intrinsic
7807 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7812 This is an overloaded intrinsic. You can use ``llvm.exp`` on any
7813 floating point or vector of floating point type. Not all targets support
7818 declare float @llvm.exp.f32(float %Val)
7819 declare double @llvm.exp.f64(double %Val)
7820 declare x86_fp80 @llvm.exp.f80(x86_fp80 %Val)
7821 declare fp128 @llvm.exp.f128(fp128 %Val)
7822 declare ppc_fp128 @llvm.exp.ppcf128(ppc_fp128 %Val)
7827 The '``llvm.exp.*``' intrinsics perform the exp function.
7832 The argument and return value are floating point numbers of the same
7838 This function returns the same values as the libm ``exp`` functions
7839 would, and handles error conditions in the same way.
7841 '``llvm.exp2.*``' Intrinsic
7842 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7847 This is an overloaded intrinsic. You can use ``llvm.exp2`` on any
7848 floating point or vector of floating point type. Not all targets support
7853 declare float @llvm.exp2.f32(float %Val)
7854 declare double @llvm.exp2.f64(double %Val)
7855 declare x86_fp80 @llvm.exp2.f80(x86_fp80 %Val)
7856 declare fp128 @llvm.exp2.f128(fp128 %Val)
7857 declare ppc_fp128 @llvm.exp2.ppcf128(ppc_fp128 %Val)
7862 The '``llvm.exp2.*``' intrinsics perform the exp2 function.
7867 The argument and return value are floating point numbers of the same
7873 This function returns the same values as the libm ``exp2`` functions
7874 would, and handles error conditions in the same way.
7876 '``llvm.log.*``' Intrinsic
7877 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7882 This is an overloaded intrinsic. You can use ``llvm.log`` on any
7883 floating point or vector of floating point type. Not all targets support
7888 declare float @llvm.log.f32(float %Val)
7889 declare double @llvm.log.f64(double %Val)
7890 declare x86_fp80 @llvm.log.f80(x86_fp80 %Val)
7891 declare fp128 @llvm.log.f128(fp128 %Val)
7892 declare ppc_fp128 @llvm.log.ppcf128(ppc_fp128 %Val)
7897 The '``llvm.log.*``' intrinsics perform the log function.
7902 The argument and return value are floating point numbers of the same
7908 This function returns the same values as the libm ``log`` functions
7909 would, and handles error conditions in the same way.
7911 '``llvm.log10.*``' Intrinsic
7912 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7917 This is an overloaded intrinsic. You can use ``llvm.log10`` on any
7918 floating point or vector of floating point type. Not all targets support
7923 declare float @llvm.log10.f32(float %Val)
7924 declare double @llvm.log10.f64(double %Val)
7925 declare x86_fp80 @llvm.log10.f80(x86_fp80 %Val)
7926 declare fp128 @llvm.log10.f128(fp128 %Val)
7927 declare ppc_fp128 @llvm.log10.ppcf128(ppc_fp128 %Val)
7932 The '``llvm.log10.*``' intrinsics perform the log10 function.
7937 The argument and return value are floating point numbers of the same
7943 This function returns the same values as the libm ``log10`` functions
7944 would, and handles error conditions in the same way.
7946 '``llvm.log2.*``' Intrinsic
7947 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7952 This is an overloaded intrinsic. You can use ``llvm.log2`` on any
7953 floating point or vector of floating point type. Not all targets support
7958 declare float @llvm.log2.f32(float %Val)
7959 declare double @llvm.log2.f64(double %Val)
7960 declare x86_fp80 @llvm.log2.f80(x86_fp80 %Val)
7961 declare fp128 @llvm.log2.f128(fp128 %Val)
7962 declare ppc_fp128 @llvm.log2.ppcf128(ppc_fp128 %Val)
7967 The '``llvm.log2.*``' intrinsics perform the log2 function.
7972 The argument and return value are floating point numbers of the same
7978 This function returns the same values as the libm ``log2`` functions
7979 would, and handles error conditions in the same way.
7981 '``llvm.fma.*``' Intrinsic
7982 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7987 This is an overloaded intrinsic. You can use ``llvm.fma`` on any
7988 floating point or vector of floating point type. Not all targets support
7993 declare float @llvm.fma.f32(float %a, float %b, float %c)
7994 declare double @llvm.fma.f64(double %a, double %b, double %c)
7995 declare x86_fp80 @llvm.fma.f80(x86_fp80 %a, x86_fp80 %b, x86_fp80 %c)
7996 declare fp128 @llvm.fma.f128(fp128 %a, fp128 %b, fp128 %c)
7997 declare ppc_fp128 @llvm.fma.ppcf128(ppc_fp128 %a, ppc_fp128 %b, ppc_fp128 %c)
8002 The '``llvm.fma.*``' intrinsics perform the fused multiply-add
8008 The argument and return value are floating point numbers of the same
8014 This function returns the same values as the libm ``fma`` functions
8015 would, and does not set errno.
8017 '``llvm.fabs.*``' Intrinsic
8018 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8023 This is an overloaded intrinsic. You can use ``llvm.fabs`` on any
8024 floating point or vector of floating point type. Not all targets support
8029 declare float @llvm.fabs.f32(float %Val)
8030 declare double @llvm.fabs.f64(double %Val)
8031 declare x86_fp80 @llvm.fabs.f80(x86_fp80 %Val)
8032 declare fp128 @llvm.fabs.f128(fp128 %Val)
8033 declare ppc_fp128 @llvm.fabs.ppcf128(ppc_fp128 %Val)
8038 The '``llvm.fabs.*``' intrinsics return the absolute value of the
8044 The argument and return value are floating point numbers of the same
8050 This function returns the same values as the libm ``fabs`` functions
8051 would, and handles error conditions in the same way.
8053 '``llvm.minnum.*``' Intrinsic
8054 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8059 This is an overloaded intrinsic. You can use ``llvm.minnum`` on any
8060 floating point or vector of floating point type. Not all targets support
8065 declare float @llvm.minnum.f32(float %Val)
8066 declare double @llvm.minnum.f64(double %Val)
8067 declare x86_fp80 @llvm.minnum.f80(x86_fp80 %Val)
8068 declare fp128 @llvm.minnum.f128(fp128 %Val)
8069 declare ppc_fp128 @llvm.minnum.ppcf128(ppc_fp128 %Val)
8074 The '``llvm.minnum.*``' intrinsics return the minimum of the two
8081 The arguments and return value are floating point numbers of the same
8087 Follows the IEEE-754 semantics for minNum, which also match for libm's
8090 If either operand is a NaN, returns the other non-NaN operand. Returns
8091 NaN only if both operands are NaN. If the operands compare equal,
8092 returns a value that compares equal to both operands. This means that
8093 fmin(+/-0.0, +/-0.0) could return either -0.0 or 0.0.
8095 '``llvm.maxnum.*``' Intrinsic
8096 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8101 This is an overloaded intrinsic. You can use ``llvm.maxnum`` on any
8102 floating point or vector of floating point type. Not all targets support
8107 declare float @llvm.maxnum.f32(float %Val)
8108 declare double @llvm.maxnum.f64(double %Val)
8109 declare x86_fp80 @llvm.maxnum.f80(x86_fp80 %Val)
8110 declare fp128 @llvm.maxnum.f128(fp128 %Val)
8111 declare ppc_fp128 @llvm.maxnum.ppcf128(ppc_fp128 %Val)
8116 The '``llvm.maxnum.*``' intrinsics return the maximum of the two
8123 The arguments and return value are floating point numbers of the same
8128 Follows the IEEE-754 semantics for maxNum, which also match for libm's
8131 If either operand is a NaN, returns the other non-NaN operand. Returns
8132 NaN only if both operands are NaN. If the operands compare equal,
8133 returns a value that compares equal to both operands. This means that
8134 fmax(+/-0.0, +/-0.0) could return either -0.0 or 0.0.
8136 '``llvm.copysign.*``' Intrinsic
8137 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8142 This is an overloaded intrinsic. You can use ``llvm.copysign`` on any
8143 floating point or vector of floating point type. Not all targets support
8148 declare float @llvm.copysign.f32(float %Mag, float %Sgn)
8149 declare double @llvm.copysign.f64(double %Mag, double %Sgn)
8150 declare x86_fp80 @llvm.copysign.f80(x86_fp80 %Mag, x86_fp80 %Sgn)
8151 declare fp128 @llvm.copysign.f128(fp128 %Mag, fp128 %Sgn)
8152 declare ppc_fp128 @llvm.copysign.ppcf128(ppc_fp128 %Mag, ppc_fp128 %Sgn)
8157 The '``llvm.copysign.*``' intrinsics return a value with the magnitude of the
8158 first operand and the sign of the second operand.
8163 The arguments and return value are floating point numbers of the same
8169 This function returns the same values as the libm ``copysign``
8170 functions would, and handles error conditions in the same way.
8172 '``llvm.floor.*``' Intrinsic
8173 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8178 This is an overloaded intrinsic. You can use ``llvm.floor`` on any
8179 floating point or vector of floating point type. Not all targets support
8184 declare float @llvm.floor.f32(float %Val)
8185 declare double @llvm.floor.f64(double %Val)
8186 declare x86_fp80 @llvm.floor.f80(x86_fp80 %Val)
8187 declare fp128 @llvm.floor.f128(fp128 %Val)
8188 declare ppc_fp128 @llvm.floor.ppcf128(ppc_fp128 %Val)
8193 The '``llvm.floor.*``' intrinsics return the floor of the operand.
8198 The argument and return value are floating point numbers of the same
8204 This function returns the same values as the libm ``floor`` functions
8205 would, and handles error conditions in the same way.
8207 '``llvm.ceil.*``' Intrinsic
8208 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8213 This is an overloaded intrinsic. You can use ``llvm.ceil`` on any
8214 floating point or vector of floating point type. Not all targets support
8219 declare float @llvm.ceil.f32(float %Val)
8220 declare double @llvm.ceil.f64(double %Val)
8221 declare x86_fp80 @llvm.ceil.f80(x86_fp80 %Val)
8222 declare fp128 @llvm.ceil.f128(fp128 %Val)
8223 declare ppc_fp128 @llvm.ceil.ppcf128(ppc_fp128 %Val)
8228 The '``llvm.ceil.*``' intrinsics return the ceiling of the operand.
8233 The argument and return value are floating point numbers of the same
8239 This function returns the same values as the libm ``ceil`` functions
8240 would, and handles error conditions in the same way.
8242 '``llvm.trunc.*``' Intrinsic
8243 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8248 This is an overloaded intrinsic. You can use ``llvm.trunc`` on any
8249 floating point or vector of floating point type. Not all targets support
8254 declare float @llvm.trunc.f32(float %Val)
8255 declare double @llvm.trunc.f64(double %Val)
8256 declare x86_fp80 @llvm.trunc.f80(x86_fp80 %Val)
8257 declare fp128 @llvm.trunc.f128(fp128 %Val)
8258 declare ppc_fp128 @llvm.trunc.ppcf128(ppc_fp128 %Val)
8263 The '``llvm.trunc.*``' intrinsics returns the operand rounded to the
8264 nearest integer not larger in magnitude than the operand.
8269 The argument and return value are floating point numbers of the same
8275 This function returns the same values as the libm ``trunc`` functions
8276 would, and handles error conditions in the same way.
8278 '``llvm.rint.*``' Intrinsic
8279 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8284 This is an overloaded intrinsic. You can use ``llvm.rint`` on any
8285 floating point or vector of floating point type. Not all targets support
8290 declare float @llvm.rint.f32(float %Val)
8291 declare double @llvm.rint.f64(double %Val)
8292 declare x86_fp80 @llvm.rint.f80(x86_fp80 %Val)
8293 declare fp128 @llvm.rint.f128(fp128 %Val)
8294 declare ppc_fp128 @llvm.rint.ppcf128(ppc_fp128 %Val)
8299 The '``llvm.rint.*``' intrinsics returns the operand rounded to the
8300 nearest integer. It may raise an inexact floating-point exception if the
8301 operand isn't an integer.
8306 The argument and return value are floating point numbers of the same
8312 This function returns the same values as the libm ``rint`` functions
8313 would, and handles error conditions in the same way.
8315 '``llvm.nearbyint.*``' Intrinsic
8316 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8321 This is an overloaded intrinsic. You can use ``llvm.nearbyint`` on any
8322 floating point or vector of floating point type. Not all targets support
8327 declare float @llvm.nearbyint.f32(float %Val)
8328 declare double @llvm.nearbyint.f64(double %Val)
8329 declare x86_fp80 @llvm.nearbyint.f80(x86_fp80 %Val)
8330 declare fp128 @llvm.nearbyint.f128(fp128 %Val)
8331 declare ppc_fp128 @llvm.nearbyint.ppcf128(ppc_fp128 %Val)
8336 The '``llvm.nearbyint.*``' intrinsics returns the operand rounded to the
8342 The argument and return value are floating point numbers of the same
8348 This function returns the same values as the libm ``nearbyint``
8349 functions would, and handles error conditions in the same way.
8351 '``llvm.round.*``' Intrinsic
8352 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8357 This is an overloaded intrinsic. You can use ``llvm.round`` on any
8358 floating point or vector of floating point type. Not all targets support
8363 declare float @llvm.round.f32(float %Val)
8364 declare double @llvm.round.f64(double %Val)
8365 declare x86_fp80 @llvm.round.f80(x86_fp80 %Val)
8366 declare fp128 @llvm.round.f128(fp128 %Val)
8367 declare ppc_fp128 @llvm.round.ppcf128(ppc_fp128 %Val)
8372 The '``llvm.round.*``' intrinsics returns the operand rounded to the
8378 The argument and return value are floating point numbers of the same
8384 This function returns the same values as the libm ``round``
8385 functions would, and handles error conditions in the same way.
8387 Bit Manipulation Intrinsics
8388 ---------------------------
8390 LLVM provides intrinsics for a few important bit manipulation
8391 operations. These allow efficient code generation for some algorithms.
8393 '``llvm.bswap.*``' Intrinsics
8394 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8399 This is an overloaded intrinsic function. You can use bswap on any
8400 integer type that is an even number of bytes (i.e. BitWidth % 16 == 0).
8404 declare i16 @llvm.bswap.i16(i16 <id>)
8405 declare i32 @llvm.bswap.i32(i32 <id>)
8406 declare i64 @llvm.bswap.i64(i64 <id>)
8411 The '``llvm.bswap``' family of intrinsics is used to byte swap integer
8412 values with an even number of bytes (positive multiple of 16 bits).
8413 These are useful for performing operations on data that is not in the
8414 target's native byte order.
8419 The ``llvm.bswap.i16`` intrinsic returns an i16 value that has the high
8420 and low byte of the input i16 swapped. Similarly, the ``llvm.bswap.i32``
8421 intrinsic returns an i32 value that has the four bytes of the input i32
8422 swapped, so that if the input bytes are numbered 0, 1, 2, 3 then the
8423 returned i32 will have its bytes in 3, 2, 1, 0 order. The
8424 ``llvm.bswap.i48``, ``llvm.bswap.i64`` and other intrinsics extend this
8425 concept to additional even-byte lengths (6 bytes, 8 bytes and more,
8428 '``llvm.ctpop.*``' Intrinsic
8429 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8434 This is an overloaded intrinsic. You can use llvm.ctpop on any integer
8435 bit width, or on any vector with integer elements. Not all targets
8436 support all bit widths or vector types, however.
8440 declare i8 @llvm.ctpop.i8(i8 <src>)
8441 declare i16 @llvm.ctpop.i16(i16 <src>)
8442 declare i32 @llvm.ctpop.i32(i32 <src>)
8443 declare i64 @llvm.ctpop.i64(i64 <src>)
8444 declare i256 @llvm.ctpop.i256(i256 <src>)
8445 declare <2 x i32> @llvm.ctpop.v2i32(<2 x i32> <src>)
8450 The '``llvm.ctpop``' family of intrinsics counts the number of bits set
8456 The only argument is the value to be counted. The argument may be of any
8457 integer type, or a vector with integer elements. The return type must
8458 match the argument type.
8463 The '``llvm.ctpop``' intrinsic counts the 1's in a variable, or within
8464 each element of a vector.
8466 '``llvm.ctlz.*``' Intrinsic
8467 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8472 This is an overloaded intrinsic. You can use ``llvm.ctlz`` on any
8473 integer bit width, or any vector whose elements are integers. Not all
8474 targets support all bit widths or vector types, however.
8478 declare i8 @llvm.ctlz.i8 (i8 <src>, i1 <is_zero_undef>)
8479 declare i16 @llvm.ctlz.i16 (i16 <src>, i1 <is_zero_undef>)
8480 declare i32 @llvm.ctlz.i32 (i32 <src>, i1 <is_zero_undef>)
8481 declare i64 @llvm.ctlz.i64 (i64 <src>, i1 <is_zero_undef>)
8482 declare i256 @llvm.ctlz.i256(i256 <src>, i1 <is_zero_undef>)
8483 declase <2 x i32> @llvm.ctlz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
8488 The '``llvm.ctlz``' family of intrinsic functions counts the number of
8489 leading zeros in a variable.
8494 The first argument is the value to be counted. This argument may be of
8495 any integer type, or a vectory with integer element type. The return
8496 type must match the first argument type.
8498 The second argument must be a constant and is a flag to indicate whether
8499 the intrinsic should ensure that a zero as the first argument produces a
8500 defined result. Historically some architectures did not provide a
8501 defined result for zero values as efficiently, and many algorithms are
8502 now predicated on avoiding zero-value inputs.
8507 The '``llvm.ctlz``' intrinsic counts the leading (most significant)
8508 zeros in a variable, or within each element of the vector. If
8509 ``src == 0`` then the result is the size in bits of the type of ``src``
8510 if ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
8511 ``llvm.ctlz(i32 2) = 30``.
8513 '``llvm.cttz.*``' Intrinsic
8514 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8519 This is an overloaded intrinsic. You can use ``llvm.cttz`` on any
8520 integer bit width, or any vector of integer elements. Not all targets
8521 support all bit widths or vector types, however.
8525 declare i8 @llvm.cttz.i8 (i8 <src>, i1 <is_zero_undef>)
8526 declare i16 @llvm.cttz.i16 (i16 <src>, i1 <is_zero_undef>)
8527 declare i32 @llvm.cttz.i32 (i32 <src>, i1 <is_zero_undef>)
8528 declare i64 @llvm.cttz.i64 (i64 <src>, i1 <is_zero_undef>)
8529 declare i256 @llvm.cttz.i256(i256 <src>, i1 <is_zero_undef>)
8530 declase <2 x i32> @llvm.cttz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
8535 The '``llvm.cttz``' family of intrinsic functions counts the number of
8541 The first argument is the value to be counted. This argument may be of
8542 any integer type, or a vectory with integer element type. The return
8543 type must match the first argument type.
8545 The second argument must be a constant and is a flag to indicate whether
8546 the intrinsic should ensure that a zero as the first argument produces a
8547 defined result. Historically some architectures did not provide a
8548 defined result for zero values as efficiently, and many algorithms are
8549 now predicated on avoiding zero-value inputs.
8554 The '``llvm.cttz``' intrinsic counts the trailing (least significant)
8555 zeros in a variable, or within each element of a vector. If ``src == 0``
8556 then the result is the size in bits of the type of ``src`` if
8557 ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
8558 ``llvm.cttz(2) = 1``.
8560 Arithmetic with Overflow Intrinsics
8561 -----------------------------------
8563 LLVM provides intrinsics for some arithmetic with overflow operations.
8565 '``llvm.sadd.with.overflow.*``' Intrinsics
8566 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8571 This is an overloaded intrinsic. You can use ``llvm.sadd.with.overflow``
8572 on any integer bit width.
8576 declare {i16, i1} @llvm.sadd.with.overflow.i16(i16 %a, i16 %b)
8577 declare {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
8578 declare {i64, i1} @llvm.sadd.with.overflow.i64(i64 %a, i64 %b)
8583 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
8584 a signed addition of the two arguments, and indicate whether an overflow
8585 occurred during the signed summation.
8590 The arguments (%a and %b) and the first element of the result structure
8591 may be of integer types of any bit width, but they must have the same
8592 bit width. The second element of the result structure must be of type
8593 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
8599 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
8600 a signed addition of the two variables. They return a structure --- the
8601 first element of which is the signed summation, and the second element
8602 of which is a bit specifying if the signed summation resulted in an
8608 .. code-block:: llvm
8610 %res = call {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
8611 %sum = extractvalue {i32, i1} %res, 0
8612 %obit = extractvalue {i32, i1} %res, 1
8613 br i1 %obit, label %overflow, label %normal
8615 '``llvm.uadd.with.overflow.*``' Intrinsics
8616 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8621 This is an overloaded intrinsic. You can use ``llvm.uadd.with.overflow``
8622 on any integer bit width.
8626 declare {i16, i1} @llvm.uadd.with.overflow.i16(i16 %a, i16 %b)
8627 declare {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
8628 declare {i64, i1} @llvm.uadd.with.overflow.i64(i64 %a, i64 %b)
8633 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
8634 an unsigned addition of the two arguments, and indicate whether a carry
8635 occurred during the unsigned summation.
8640 The arguments (%a and %b) and the first element of the result structure
8641 may be of integer types of any bit width, but they must have the same
8642 bit width. The second element of the result structure must be of type
8643 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
8649 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
8650 an unsigned addition of the two arguments. They return a structure --- the
8651 first element of which is the sum, and the second element of which is a
8652 bit specifying if the unsigned summation resulted in a carry.
8657 .. code-block:: llvm
8659 %res = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
8660 %sum = extractvalue {i32, i1} %res, 0
8661 %obit = extractvalue {i32, i1} %res, 1
8662 br i1 %obit, label %carry, label %normal
8664 '``llvm.ssub.with.overflow.*``' Intrinsics
8665 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8670 This is an overloaded intrinsic. You can use ``llvm.ssub.with.overflow``
8671 on any integer bit width.
8675 declare {i16, i1} @llvm.ssub.with.overflow.i16(i16 %a, i16 %b)
8676 declare {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
8677 declare {i64, i1} @llvm.ssub.with.overflow.i64(i64 %a, i64 %b)
8682 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
8683 a signed subtraction of the two arguments, and indicate whether an
8684 overflow occurred during the signed subtraction.
8689 The arguments (%a and %b) and the first element of the result structure
8690 may be of integer types of any bit width, but they must have the same
8691 bit width. The second element of the result structure must be of type
8692 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
8698 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
8699 a signed subtraction of the two arguments. They return a structure --- the
8700 first element of which is the subtraction, and the second element of
8701 which is a bit specifying if the signed subtraction resulted in an
8707 .. code-block:: llvm
8709 %res = call {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
8710 %sum = extractvalue {i32, i1} %res, 0
8711 %obit = extractvalue {i32, i1} %res, 1
8712 br i1 %obit, label %overflow, label %normal
8714 '``llvm.usub.with.overflow.*``' Intrinsics
8715 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8720 This is an overloaded intrinsic. You can use ``llvm.usub.with.overflow``
8721 on any integer bit width.
8725 declare {i16, i1} @llvm.usub.with.overflow.i16(i16 %a, i16 %b)
8726 declare {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
8727 declare {i64, i1} @llvm.usub.with.overflow.i64(i64 %a, i64 %b)
8732 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
8733 an unsigned subtraction of the two arguments, and indicate whether an
8734 overflow occurred during the unsigned subtraction.
8739 The arguments (%a and %b) and the first element of the result structure
8740 may be of integer types of any bit width, but they must have the same
8741 bit width. The second element of the result structure must be of type
8742 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
8748 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
8749 an unsigned subtraction of the two arguments. They return a structure ---
8750 the first element of which is the subtraction, and the second element of
8751 which is a bit specifying if the unsigned subtraction resulted in an
8757 .. code-block:: llvm
8759 %res = call {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
8760 %sum = extractvalue {i32, i1} %res, 0
8761 %obit = extractvalue {i32, i1} %res, 1
8762 br i1 %obit, label %overflow, label %normal
8764 '``llvm.smul.with.overflow.*``' Intrinsics
8765 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8770 This is an overloaded intrinsic. You can use ``llvm.smul.with.overflow``
8771 on any integer bit width.
8775 declare {i16, i1} @llvm.smul.with.overflow.i16(i16 %a, i16 %b)
8776 declare {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
8777 declare {i64, i1} @llvm.smul.with.overflow.i64(i64 %a, i64 %b)
8782 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
8783 a signed multiplication of the two arguments, and indicate whether an
8784 overflow occurred during the signed multiplication.
8789 The arguments (%a and %b) and the first element of the result structure
8790 may be of integer types of any bit width, but they must have the same
8791 bit width. The second element of the result structure must be of type
8792 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
8798 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
8799 a signed multiplication of the two arguments. They return a structure ---
8800 the first element of which is the multiplication, and the second element
8801 of which is a bit specifying if the signed multiplication resulted in an
8807 .. code-block:: llvm
8809 %res = call {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
8810 %sum = extractvalue {i32, i1} %res, 0
8811 %obit = extractvalue {i32, i1} %res, 1
8812 br i1 %obit, label %overflow, label %normal
8814 '``llvm.umul.with.overflow.*``' Intrinsics
8815 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8820 This is an overloaded intrinsic. You can use ``llvm.umul.with.overflow``
8821 on any integer bit width.
8825 declare {i16, i1} @llvm.umul.with.overflow.i16(i16 %a, i16 %b)
8826 declare {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
8827 declare {i64, i1} @llvm.umul.with.overflow.i64(i64 %a, i64 %b)
8832 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
8833 a unsigned multiplication of the two arguments, and indicate whether an
8834 overflow occurred during the unsigned multiplication.
8839 The arguments (%a and %b) and the first element of the result structure
8840 may be of integer types of any bit width, but they must have the same
8841 bit width. The second element of the result structure must be of type
8842 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
8848 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
8849 an unsigned multiplication of the two arguments. They return a structure ---
8850 the first element of which is the multiplication, and the second
8851 element of which is a bit specifying if the unsigned multiplication
8852 resulted in an overflow.
8857 .. code-block:: llvm
8859 %res = call {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
8860 %sum = extractvalue {i32, i1} %res, 0
8861 %obit = extractvalue {i32, i1} %res, 1
8862 br i1 %obit, label %overflow, label %normal
8864 Specialised Arithmetic Intrinsics
8865 ---------------------------------
8867 '``llvm.fmuladd.*``' Intrinsic
8868 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8875 declare float @llvm.fmuladd.f32(float %a, float %b, float %c)
8876 declare double @llvm.fmuladd.f64(double %a, double %b, double %c)
8881 The '``llvm.fmuladd.*``' intrinsic functions represent multiply-add
8882 expressions that can be fused if the code generator determines that (a) the
8883 target instruction set has support for a fused operation, and (b) that the
8884 fused operation is more efficient than the equivalent, separate pair of mul
8885 and add instructions.
8890 The '``llvm.fmuladd.*``' intrinsics each take three arguments: two
8891 multiplicands, a and b, and an addend c.
8900 %0 = call float @llvm.fmuladd.f32(%a, %b, %c)
8902 is equivalent to the expression a \* b + c, except that rounding will
8903 not be performed between the multiplication and addition steps if the
8904 code generator fuses the operations. Fusion is not guaranteed, even if
8905 the target platform supports it. If a fused multiply-add is required the
8906 corresponding llvm.fma.\* intrinsic function should be used
8907 instead. This never sets errno, just as '``llvm.fma.*``'.
8912 .. code-block:: llvm
8914 %r2 = call float @llvm.fmuladd.f32(float %a, float %b, float %c) ; yields float:r2 = (a * b) + c
8916 Half Precision Floating Point Intrinsics
8917 ----------------------------------------
8919 For most target platforms, half precision floating point is a
8920 storage-only format. This means that it is a dense encoding (in memory)
8921 but does not support computation in the format.
8923 This means that code must first load the half-precision floating point
8924 value as an i16, then convert it to float with
8925 :ref:`llvm.convert.from.fp16 <int_convert_from_fp16>`. Computation can
8926 then be performed on the float value (including extending to double
8927 etc). To store the value back to memory, it is first converted to float
8928 if needed, then converted to i16 with
8929 :ref:`llvm.convert.to.fp16 <int_convert_to_fp16>`, then storing as an
8932 .. _int_convert_to_fp16:
8934 '``llvm.convert.to.fp16``' Intrinsic
8935 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8942 declare i16 @llvm.convert.to.fp16.f32(float %a)
8943 declare i16 @llvm.convert.to.fp16.f64(double %a)
8948 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion from a
8949 conventional floating point type to half precision floating point format.
8954 The intrinsic function contains single argument - the value to be
8960 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion from a
8961 conventional floating point format to half precision floating point format. The
8962 return value is an ``i16`` which contains the converted number.
8967 .. code-block:: llvm
8969 %res = call i16 @llvm.convert.to.fp16.f32(float %a)
8970 store i16 %res, i16* @x, align 2
8972 .. _int_convert_from_fp16:
8974 '``llvm.convert.from.fp16``' Intrinsic
8975 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8982 declare float @llvm.convert.from.fp16.f32(i16 %a)
8983 declare double @llvm.convert.from.fp16.f64(i16 %a)
8988 The '``llvm.convert.from.fp16``' intrinsic function performs a
8989 conversion from half precision floating point format to single precision
8990 floating point format.
8995 The intrinsic function contains single argument - the value to be
9001 The '``llvm.convert.from.fp16``' intrinsic function performs a
9002 conversion from half single precision floating point format to single
9003 precision floating point format. The input half-float value is
9004 represented by an ``i16`` value.
9009 .. code-block:: llvm
9011 %a = load i16* @x, align 2
9012 %res = call float @llvm.convert.from.fp16(i16 %a)
9017 The LLVM debugger intrinsics (which all start with ``llvm.dbg.``
9018 prefix), are described in the `LLVM Source Level
9019 Debugging <SourceLevelDebugging.html#format_common_intrinsics>`_
9022 Exception Handling Intrinsics
9023 -----------------------------
9025 The LLVM exception handling intrinsics (which all start with
9026 ``llvm.eh.`` prefix), are described in the `LLVM Exception
9027 Handling <ExceptionHandling.html#format_common_intrinsics>`_ document.
9031 Trampoline Intrinsics
9032 ---------------------
9034 These intrinsics make it possible to excise one parameter, marked with
9035 the :ref:`nest <nest>` attribute, from a function. The result is a
9036 callable function pointer lacking the nest parameter - the caller does
9037 not need to provide a value for it. Instead, the value to use is stored
9038 in advance in a "trampoline", a block of memory usually allocated on the
9039 stack, which also contains code to splice the nest value into the
9040 argument list. This is used to implement the GCC nested function address
9043 For example, if the function is ``i32 f(i8* nest %c, i32 %x, i32 %y)``
9044 then the resulting function pointer has signature ``i32 (i32, i32)*``.
9045 It can be created as follows:
9047 .. code-block:: llvm
9049 %tramp = alloca [10 x i8], align 4 ; size and alignment only correct for X86
9050 %tramp1 = getelementptr [10 x i8]* %tramp, i32 0, i32 0
9051 call i8* @llvm.init.trampoline(i8* %tramp1, i8* bitcast (i32 (i8*, i32, i32)* @f to i8*), i8* %nval)
9052 %p = call i8* @llvm.adjust.trampoline(i8* %tramp1)
9053 %fp = bitcast i8* %p to i32 (i32, i32)*
9055 The call ``%val = call i32 %fp(i32 %x, i32 %y)`` is then equivalent to
9056 ``%val = call i32 %f(i8* %nval, i32 %x, i32 %y)``.
9060 '``llvm.init.trampoline``' Intrinsic
9061 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9068 declare void @llvm.init.trampoline(i8* <tramp>, i8* <func>, i8* <nval>)
9073 This fills the memory pointed to by ``tramp`` with executable code,
9074 turning it into a trampoline.
9079 The ``llvm.init.trampoline`` intrinsic takes three arguments, all
9080 pointers. The ``tramp`` argument must point to a sufficiently large and
9081 sufficiently aligned block of memory; this memory is written to by the
9082 intrinsic. Note that the size and the alignment are target-specific -
9083 LLVM currently provides no portable way of determining them, so a
9084 front-end that generates this intrinsic needs to have some
9085 target-specific knowledge. The ``func`` argument must hold a function
9086 bitcast to an ``i8*``.
9091 The block of memory pointed to by ``tramp`` is filled with target
9092 dependent code, turning it into a function. Then ``tramp`` needs to be
9093 passed to :ref:`llvm.adjust.trampoline <int_at>` to get a pointer which can
9094 be :ref:`bitcast (to a new function) and called <int_trampoline>`. The new
9095 function's signature is the same as that of ``func`` with any arguments
9096 marked with the ``nest`` attribute removed. At most one such ``nest``
9097 argument is allowed, and it must be of pointer type. Calling the new
9098 function is equivalent to calling ``func`` with the same argument list,
9099 but with ``nval`` used for the missing ``nest`` argument. If, after
9100 calling ``llvm.init.trampoline``, the memory pointed to by ``tramp`` is
9101 modified, then the effect of any later call to the returned function
9102 pointer is undefined.
9106 '``llvm.adjust.trampoline``' Intrinsic
9107 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9114 declare i8* @llvm.adjust.trampoline(i8* <tramp>)
9119 This performs any required machine-specific adjustment to the address of
9120 a trampoline (passed as ``tramp``).
9125 ``tramp`` must point to a block of memory which already has trampoline
9126 code filled in by a previous call to
9127 :ref:`llvm.init.trampoline <int_it>`.
9132 On some architectures the address of the code to be executed needs to be
9133 different than the address where the trampoline is actually stored. This
9134 intrinsic returns the executable address corresponding to ``tramp``
9135 after performing the required machine specific adjustments. The pointer
9136 returned can then be :ref:`bitcast and executed <int_trampoline>`.
9141 This class of intrinsics provides information about the lifetime of
9142 memory objects and ranges where variables are immutable.
9146 '``llvm.lifetime.start``' Intrinsic
9147 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9154 declare void @llvm.lifetime.start(i64 <size>, i8* nocapture <ptr>)
9159 The '``llvm.lifetime.start``' intrinsic specifies the start of a memory
9165 The first argument is a constant integer representing the size of the
9166 object, or -1 if it is variable sized. The second argument is a pointer
9172 This intrinsic indicates that before this point in the code, the value
9173 of the memory pointed to by ``ptr`` is dead. This means that it is known
9174 to never be used and has an undefined value. A load from the pointer
9175 that precedes this intrinsic can be replaced with ``'undef'``.
9179 '``llvm.lifetime.end``' Intrinsic
9180 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9187 declare void @llvm.lifetime.end(i64 <size>, i8* nocapture <ptr>)
9192 The '``llvm.lifetime.end``' intrinsic specifies the end of a memory
9198 The first argument is a constant integer representing the size of the
9199 object, or -1 if it is variable sized. The second argument is a pointer
9205 This intrinsic indicates that after this point in the code, the value of
9206 the memory pointed to by ``ptr`` is dead. This means that it is known to
9207 never be used and has an undefined value. Any stores into the memory
9208 object following this intrinsic may be removed as dead.
9210 '``llvm.invariant.start``' Intrinsic
9211 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9218 declare {}* @llvm.invariant.start(i64 <size>, i8* nocapture <ptr>)
9223 The '``llvm.invariant.start``' intrinsic specifies that the contents of
9224 a memory object will not change.
9229 The first argument is a constant integer representing the size of the
9230 object, or -1 if it is variable sized. The second argument is a pointer
9236 This intrinsic indicates that until an ``llvm.invariant.end`` that uses
9237 the return value, the referenced memory location is constant and
9240 '``llvm.invariant.end``' Intrinsic
9241 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9248 declare void @llvm.invariant.end({}* <start>, i64 <size>, i8* nocapture <ptr>)
9253 The '``llvm.invariant.end``' intrinsic specifies that the contents of a
9254 memory object are mutable.
9259 The first argument is the matching ``llvm.invariant.start`` intrinsic.
9260 The second argument is a constant integer representing the size of the
9261 object, or -1 if it is variable sized and the third argument is a
9262 pointer to the object.
9267 This intrinsic indicates that the memory is mutable again.
9272 This class of intrinsics is designed to be generic and has no specific
9275 '``llvm.var.annotation``' Intrinsic
9276 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9283 declare void @llvm.var.annotation(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
9288 The '``llvm.var.annotation``' intrinsic.
9293 The first argument is a pointer to a value, the second is a pointer to a
9294 global string, the third is a pointer to a global string which is the
9295 source file name, and the last argument is the line number.
9300 This intrinsic allows annotation of local variables with arbitrary
9301 strings. This can be useful for special purpose optimizations that want
9302 to look for these annotations. These have no other defined use; they are
9303 ignored by code generation and optimization.
9305 '``llvm.ptr.annotation.*``' Intrinsic
9306 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9311 This is an overloaded intrinsic. You can use '``llvm.ptr.annotation``' on a
9312 pointer to an integer of any width. *NOTE* you must specify an address space for
9313 the pointer. The identifier for the default address space is the integer
9318 declare i8* @llvm.ptr.annotation.p<address space>i8(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
9319 declare i16* @llvm.ptr.annotation.p<address space>i16(i16* <val>, i8* <str>, i8* <str>, i32 <int>)
9320 declare i32* @llvm.ptr.annotation.p<address space>i32(i32* <val>, i8* <str>, i8* <str>, i32 <int>)
9321 declare i64* @llvm.ptr.annotation.p<address space>i64(i64* <val>, i8* <str>, i8* <str>, i32 <int>)
9322 declare i256* @llvm.ptr.annotation.p<address space>i256(i256* <val>, i8* <str>, i8* <str>, i32 <int>)
9327 The '``llvm.ptr.annotation``' intrinsic.
9332 The first argument is a pointer to an integer value of arbitrary bitwidth
9333 (result of some expression), the second is a pointer to a global string, the
9334 third is a pointer to a global string which is the source file name, and the
9335 last argument is the line number. It returns the value of the first argument.
9340 This intrinsic allows annotation of a pointer to an integer with arbitrary
9341 strings. This can be useful for special purpose optimizations that want to look
9342 for these annotations. These have no other defined use; they are ignored by code
9343 generation and optimization.
9345 '``llvm.annotation.*``' Intrinsic
9346 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9351 This is an overloaded intrinsic. You can use '``llvm.annotation``' on
9352 any integer bit width.
9356 declare i8 @llvm.annotation.i8(i8 <val>, i8* <str>, i8* <str>, i32 <int>)
9357 declare i16 @llvm.annotation.i16(i16 <val>, i8* <str>, i8* <str>, i32 <int>)
9358 declare i32 @llvm.annotation.i32(i32 <val>, i8* <str>, i8* <str>, i32 <int>)
9359 declare i64 @llvm.annotation.i64(i64 <val>, i8* <str>, i8* <str>, i32 <int>)
9360 declare i256 @llvm.annotation.i256(i256 <val>, i8* <str>, i8* <str>, i32 <int>)
9365 The '``llvm.annotation``' intrinsic.
9370 The first argument is an integer value (result of some expression), the
9371 second is a pointer to a global string, the third is a pointer to a
9372 global string which is the source file name, and the last argument is
9373 the line number. It returns the value of the first argument.
9378 This intrinsic allows annotations to be put on arbitrary expressions
9379 with arbitrary strings. This can be useful for special purpose
9380 optimizations that want to look for these annotations. These have no
9381 other defined use; they are ignored by code generation and optimization.
9383 '``llvm.trap``' Intrinsic
9384 ^^^^^^^^^^^^^^^^^^^^^^^^^
9391 declare void @llvm.trap() noreturn nounwind
9396 The '``llvm.trap``' intrinsic.
9406 This intrinsic is lowered to the target dependent trap instruction. If
9407 the target does not have a trap instruction, this intrinsic will be
9408 lowered to a call of the ``abort()`` function.
9410 '``llvm.debugtrap``' Intrinsic
9411 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9418 declare void @llvm.debugtrap() nounwind
9423 The '``llvm.debugtrap``' intrinsic.
9433 This intrinsic is lowered to code which is intended to cause an
9434 execution trap with the intention of requesting the attention of a
9437 '``llvm.stackprotector``' Intrinsic
9438 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9445 declare void @llvm.stackprotector(i8* <guard>, i8** <slot>)
9450 The ``llvm.stackprotector`` intrinsic takes the ``guard`` and stores it
9451 onto the stack at ``slot``. The stack slot is adjusted to ensure that it
9452 is placed on the stack before local variables.
9457 The ``llvm.stackprotector`` intrinsic requires two pointer arguments.
9458 The first argument is the value loaded from the stack guard
9459 ``@__stack_chk_guard``. The second variable is an ``alloca`` that has
9460 enough space to hold the value of the guard.
9465 This intrinsic causes the prologue/epilogue inserter to force the position of
9466 the ``AllocaInst`` stack slot to be before local variables on the stack. This is
9467 to ensure that if a local variable on the stack is overwritten, it will destroy
9468 the value of the guard. When the function exits, the guard on the stack is
9469 checked against the original guard by ``llvm.stackprotectorcheck``. If they are
9470 different, then ``llvm.stackprotectorcheck`` causes the program to abort by
9471 calling the ``__stack_chk_fail()`` function.
9473 '``llvm.stackprotectorcheck``' Intrinsic
9474 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9481 declare void @llvm.stackprotectorcheck(i8** <guard>)
9486 The ``llvm.stackprotectorcheck`` intrinsic compares ``guard`` against an already
9487 created stack protector and if they are not equal calls the
9488 ``__stack_chk_fail()`` function.
9493 The ``llvm.stackprotectorcheck`` intrinsic requires one pointer argument, the
9494 the variable ``@__stack_chk_guard``.
9499 This intrinsic is provided to perform the stack protector check by comparing
9500 ``guard`` with the stack slot created by ``llvm.stackprotector`` and if the
9501 values do not match call the ``__stack_chk_fail()`` function.
9503 The reason to provide this as an IR level intrinsic instead of implementing it
9504 via other IR operations is that in order to perform this operation at the IR
9505 level without an intrinsic, one would need to create additional basic blocks to
9506 handle the success/failure cases. This makes it difficult to stop the stack
9507 protector check from disrupting sibling tail calls in Codegen. With this
9508 intrinsic, we are able to generate the stack protector basic blocks late in
9509 codegen after the tail call decision has occurred.
9511 '``llvm.objectsize``' Intrinsic
9512 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9519 declare i32 @llvm.objectsize.i32(i8* <object>, i1 <min>)
9520 declare i64 @llvm.objectsize.i64(i8* <object>, i1 <min>)
9525 The ``llvm.objectsize`` intrinsic is designed to provide information to
9526 the optimizers to determine at compile time whether a) an operation
9527 (like memcpy) will overflow a buffer that corresponds to an object, or
9528 b) that a runtime check for overflow isn't necessary. An object in this
9529 context means an allocation of a specific class, structure, array, or
9535 The ``llvm.objectsize`` intrinsic takes two arguments. The first
9536 argument is a pointer to or into the ``object``. The second argument is
9537 a boolean and determines whether ``llvm.objectsize`` returns 0 (if true)
9538 or -1 (if false) when the object size is unknown. The second argument
9539 only accepts constants.
9544 The ``llvm.objectsize`` intrinsic is lowered to a constant representing
9545 the size of the object concerned. If the size cannot be determined at
9546 compile time, ``llvm.objectsize`` returns ``i32/i64 -1 or 0`` (depending
9547 on the ``min`` argument).
9549 '``llvm.expect``' Intrinsic
9550 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9555 This is an overloaded intrinsic. You can use ``llvm.expect`` on any
9560 declare i1 @llvm.expect.i1(i1 <val>, i1 <expected_val>)
9561 declare i32 @llvm.expect.i32(i32 <val>, i32 <expected_val>)
9562 declare i64 @llvm.expect.i64(i64 <val>, i64 <expected_val>)
9567 The ``llvm.expect`` intrinsic provides information about expected (the
9568 most probable) value of ``val``, which can be used by optimizers.
9573 The ``llvm.expect`` intrinsic takes two arguments. The first argument is
9574 a value. The second argument is an expected value, this needs to be a
9575 constant value, variables are not allowed.
9580 This intrinsic is lowered to the ``val``.
9582 '``llvm.assume``' Intrinsic
9583 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9590 declare void @llvm.assume(i1 %cond)
9595 The ``llvm.assume`` allows the optimizer to assume that the provided
9596 condition is true. This information can then be used in simplifying other parts
9602 The condition which the optimizer may assume is always true.
9607 The intrinsic allows the optimizer to assume that the provided condition is
9608 always true whenever the control flow reaches the intrinsic call. No code is
9609 generated for this intrinsic, and instructions that contribute only to the
9610 provided condition are not used for code generation. If the condition is
9611 violated during execution, the behavior is undefined.
9613 Please note that optimizer might limit the transformations performed on values
9614 used by the ``llvm.assume`` intrinsic in order to preserve the instructions
9615 only used to form the intrinsic's input argument. This might prove undesirable
9616 if the extra information provided by the ``llvm.assume`` intrinsic does cause
9617 sufficient overall improvement in code quality. For this reason,
9618 ``llvm.assume`` should not be used to document basic mathematical invariants
9619 that the optimizer can otherwise deduce or facts that are of little use to the
9622 '``llvm.donothing``' Intrinsic
9623 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9630 declare void @llvm.donothing() nounwind readnone
9635 The ``llvm.donothing`` intrinsic doesn't perform any operation. It's the
9636 only intrinsic that can be called with an invoke instruction.
9646 This intrinsic does nothing, and it's removed by optimizers and ignored
9649 Stack Map Intrinsics
9650 --------------------
9652 LLVM provides experimental intrinsics to support runtime patching
9653 mechanisms commonly desired in dynamic language JITs. These intrinsics
9654 are described in :doc:`StackMaps`.