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 = !{i32 42, null, !"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>`,
637 an optional :ref:`prologue <prologuedata>`, an opening
638 curly brace, a list of basic blocks, and a closing curly brace.
640 LLVM function declarations consist of the "``declare``" keyword, an
641 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
642 style <visibility>`, an optional :ref:`DLL storage class <dllstorageclass>`,
643 an optional :ref:`calling convention <callingconv>`,
644 an optional ``unnamed_addr`` attribute, a return type, an optional
645 :ref:`parameter attribute <paramattrs>` for the return type, a function
646 name, a possibly empty list of arguments, an optional alignment, an optional
647 :ref:`garbage collector name <gc>`, an optional :ref:`prefix <prefixdata>`,
648 and an optional :ref:`prologue <prologuedata>`.
650 A function definition contains a list of basic blocks, forming the CFG (Control
651 Flow Graph) for the function. Each basic block may optionally start with a label
652 (giving the basic block a symbol table entry), contains a list of instructions,
653 and ends with a :ref:`terminator <terminators>` instruction (such as a branch or
654 function return). If an explicit label is not provided, a block is assigned an
655 implicit numbered label, using the next value from the same counter as used for
656 unnamed temporaries (:ref:`see above<identifiers>`). For example, if a function
657 entry block does not have an explicit label, it will be assigned label "%0",
658 then the first unnamed temporary in that block will be "%1", etc.
660 The first basic block in a function is special in two ways: it is
661 immediately executed on entrance to the function, and it is not allowed
662 to have predecessor basic blocks (i.e. there can not be any branches to
663 the entry block of a function). Because the block can have no
664 predecessors, it also cannot have any :ref:`PHI nodes <i_phi>`.
666 LLVM allows an explicit section to be specified for functions. If the
667 target supports it, it will emit functions to the section specified.
668 Additionally, the function can placed in a COMDAT.
670 An explicit alignment may be specified for a function. If not present,
671 or if the alignment is set to zero, the alignment of the function is set
672 by the target to whatever it feels convenient. If an explicit alignment
673 is specified, the function is forced to have at least that much
674 alignment. All alignments must be a power of 2.
676 If the ``unnamed_addr`` attribute is given, the address is know to not
677 be significant and two identical functions can be merged.
681 define [linkage] [visibility] [DLLStorageClass]
683 <ResultType> @<FunctionName> ([argument list])
684 [unnamed_addr] [fn Attrs] [section "name"] [comdat $<ComdatName>]
685 [align N] [gc] [prefix Constant] [prologue Constant] { ... }
687 The argument list is a comma seperated sequence of arguments where each
688 argument is of the following form
692 <type> [parameter Attrs] [name]
700 Aliases, unlike function or variables, don't create any new data. They
701 are just a new symbol and metadata for an existing position.
703 Aliases have a name and an aliasee that is either a global value or a
706 Aliases may have an optional :ref:`linkage type <linkage>`, an optional
707 :ref:`visibility style <visibility>`, an optional :ref:`DLL storage class
708 <dllstorageclass>` and an optional :ref:`tls model <tls_model>`.
712 @<Name> = [Linkage] [Visibility] [DLLStorageClass] [ThreadLocal] [unnamed_addr] alias <AliaseeTy> @<Aliasee>
714 The linkage must be one of ``private``, ``internal``, ``linkonce``, ``weak``,
715 ``linkonce_odr``, ``weak_odr``, ``external``. Note that some system linkers
716 might not correctly handle dropping a weak symbol that is aliased.
718 Alias that are not ``unnamed_addr`` are guaranteed to have the same address as
719 the aliasee expression. ``unnamed_addr`` ones are only guaranteed to point
722 Since aliases are only a second name, some restrictions apply, of which
723 some can only be checked when producing an object file:
725 * The expression defining the aliasee must be computable at assembly
726 time. Since it is just a name, no relocations can be used.
728 * No alias in the expression can be weak as the possibility of the
729 intermediate alias being overridden cannot be represented in an
732 * No global value in the expression can be a declaration, since that
733 would require a relocation, which is not possible.
740 Comdat IR provides access to COFF and ELF object file COMDAT functionality.
742 Comdats have a name which represents the COMDAT key. All global objects that
743 specify this key will only end up in the final object file if the linker chooses
744 that key over some other key. Aliases are placed in the same COMDAT that their
745 aliasee computes to, if any.
747 Comdats have a selection kind to provide input on how the linker should
748 choose between keys in two different object files.
752 $<Name> = comdat SelectionKind
754 The selection kind must be one of the following:
757 The linker may choose any COMDAT key, the choice is arbitrary.
759 The linker may choose any COMDAT key but the sections must contain the
762 The linker will choose the section containing the largest COMDAT key.
764 The linker requires that only section with this COMDAT key exist.
766 The linker may choose any COMDAT key but the sections must contain the
769 Note that the Mach-O platform doesn't support COMDATs and ELF only supports
770 ``any`` as a selection kind.
772 Here is an example of a COMDAT group where a function will only be selected if
773 the COMDAT key's section is the largest:
777 $foo = comdat largest
778 @foo = global i32 2, comdat $foo
780 define void @bar() comdat $foo {
784 In a COFF object file, this will create a COMDAT section with selection kind
785 ``IMAGE_COMDAT_SELECT_LARGEST`` containing the contents of the ``@foo`` symbol
786 and another COMDAT section with selection kind
787 ``IMAGE_COMDAT_SELECT_ASSOCIATIVE`` which is associated with the first COMDAT
788 section and contains the contents of the ``@bar`` symbol.
790 There are some restrictions on the properties of the global object.
791 It, or an alias to it, must have the same name as the COMDAT group when
793 The contents and size of this object may be used during link-time to determine
794 which COMDAT groups get selected depending on the selection kind.
795 Because the name of the object must match the name of the COMDAT group, the
796 linkage of the global object must not be local; local symbols can get renamed
797 if a collision occurs in the symbol table.
799 The combined use of COMDATS and section attributes may yield surprising results.
806 @g1 = global i32 42, section "sec", comdat $foo
807 @g2 = global i32 42, section "sec", comdat $bar
809 From the object file perspective, this requires the creation of two sections
810 with the same name. This is necessary because both globals belong to different
811 COMDAT groups and COMDATs, at the object file level, are represented by
814 Note that certain IR constructs like global variables and functions may create
815 COMDATs in the object file in addition to any which are specified using COMDAT
816 IR. This arises, for example, when a global variable has linkonce_odr linkage.
818 .. _namedmetadatastructure:
823 Named metadata is a collection of metadata. :ref:`Metadata
824 nodes <metadata>` (but not metadata strings) are the only valid
825 operands for a named metadata.
829 ; Some unnamed metadata nodes, which are referenced by the named metadata.
834 !name = !{!0, !1, !2}
841 The return type and each parameter of a function type may have a set of
842 *parameter attributes* associated with them. Parameter attributes are
843 used to communicate additional information about the result or
844 parameters of a function. Parameter attributes are considered to be part
845 of the function, not of the function type, so functions with different
846 parameter attributes can have the same function type.
848 Parameter attributes are simple keywords that follow the type specified.
849 If multiple parameter attributes are needed, they are space separated.
854 declare i32 @printf(i8* noalias nocapture, ...)
855 declare i32 @atoi(i8 zeroext)
856 declare signext i8 @returns_signed_char()
858 Note that any attributes for the function result (``nounwind``,
859 ``readonly``) come immediately after the argument list.
861 Currently, only the following parameter attributes are defined:
864 This indicates to the code generator that the parameter or return
865 value should be zero-extended to the extent required by the target's
866 ABI (which is usually 32-bits, but is 8-bits for a i1 on x86-64) by
867 the caller (for a parameter) or the callee (for a return value).
869 This indicates to the code generator that the parameter or return
870 value should be sign-extended to the extent required by the target's
871 ABI (which is usually 32-bits) by the caller (for a parameter) or
872 the callee (for a return value).
874 This indicates that this parameter or return value should be treated
875 in a special target-dependent fashion during while emitting code for
876 a function call or return (usually, by putting it in a register as
877 opposed to memory, though some targets use it to distinguish between
878 two different kinds of registers). Use of this attribute is
881 This indicates that the pointer parameter should really be passed by
882 value to the function. The attribute implies that a hidden copy of
883 the pointee is made between the caller and the callee, so the callee
884 is unable to modify the value in the caller. This attribute is only
885 valid on LLVM pointer arguments. It is generally used to pass
886 structs and arrays by value, but is also valid on pointers to
887 scalars. The copy is considered to belong to the caller not the
888 callee (for example, ``readonly`` functions should not write to
889 ``byval`` parameters). This is not a valid attribute for return
892 The byval attribute also supports specifying an alignment with the
893 align attribute. It indicates the alignment of the stack slot to
894 form and the known alignment of the pointer specified to the call
895 site. If the alignment is not specified, then the code generator
896 makes a target-specific assumption.
902 The ``inalloca`` argument attribute allows the caller to take the
903 address of outgoing stack arguments. An ``inalloca`` argument must
904 be a pointer to stack memory produced by an ``alloca`` instruction.
905 The alloca, or argument allocation, must also be tagged with the
906 inalloca keyword. Only the last argument may have the ``inalloca``
907 attribute, and that argument is guaranteed to be passed in memory.
909 An argument allocation may be used by a call at most once because
910 the call may deallocate it. The ``inalloca`` attribute cannot be
911 used in conjunction with other attributes that affect argument
912 storage, like ``inreg``, ``nest``, ``sret``, or ``byval``. The
913 ``inalloca`` attribute also disables LLVM's implicit lowering of
914 large aggregate return values, which means that frontend authors
915 must lower them with ``sret`` pointers.
917 When the call site is reached, the argument allocation must have
918 been the most recent stack allocation that is still live, or the
919 results are undefined. It is possible to allocate additional stack
920 space after an argument allocation and before its call site, but it
921 must be cleared off with :ref:`llvm.stackrestore
924 See :doc:`InAlloca` for more information on how to use this
928 This indicates that the pointer parameter specifies the address of a
929 structure that is the return value of the function in the source
930 program. This pointer must be guaranteed by the caller to be valid:
931 loads and stores to the structure may be assumed by the callee
932 not to trap and to be properly aligned. This may only be applied to
933 the first parameter. This is not a valid attribute for return
937 This indicates that the pointer value may be assumed by the optimizer to
938 have the specified alignment.
940 Note that this attribute has additional semantics when combined with the
946 This indicates that objects accessed via pointer values
947 :ref:`based <pointeraliasing>` on the argument or return value are not also
948 accessed, during the execution of the function, via pointer values not
949 *based* on the argument or return value. The attribute on a return value
950 also has additional semantics described below. The caller shares the
951 responsibility with the callee for ensuring that these requirements are met.
952 For further details, please see the discussion of the NoAlias response in
953 :ref:`alias 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.
958 For function return values, C99's ``restrict`` is not meaningful,
959 while LLVM's ``noalias`` is. Furthermore, the semantics of the ``noalias``
960 attribute on return values are stronger than the semantics of the attribute
961 when used on function arguments. On function return values, the ``noalias``
962 attribute indicates that the function acts like a system memory allocation
963 function, returning a pointer to allocated storage disjoint from the
964 storage for any other object accessible to the caller.
967 This indicates that the callee does not make any copies of the
968 pointer that outlive the callee itself. This is not a valid
969 attribute for return values.
974 This indicates that the pointer parameter can be excised using the
975 :ref:`trampoline intrinsics <int_trampoline>`. This is not a valid
976 attribute for return values and can only be applied to one parameter.
979 This indicates that the function always returns the argument as its return
980 value. This is an optimization hint to the code generator when generating
981 the caller, allowing tail call optimization and omission of register saves
982 and restores in some cases; it is not checked or enforced when generating
983 the callee. The parameter and the function return type must be valid
984 operands for the :ref:`bitcast instruction <i_bitcast>`. This is not a
985 valid attribute for return values and can only be applied to one parameter.
988 This indicates that the parameter or return pointer is not null. This
989 attribute may only be applied to pointer typed parameters. This is not
990 checked or enforced by LLVM, the caller must ensure that the pointer
991 passed in is non-null, or the callee must ensure that the returned pointer
994 ``dereferenceable(<n>)``
995 This indicates that the parameter or return pointer is dereferenceable. This
996 attribute may only be applied to pointer typed parameters. A pointer that
997 is dereferenceable can be loaded from speculatively without a risk of
998 trapping. The number of bytes known to be dereferenceable must be provided
999 in parentheses. It is legal for the number of bytes to be less than the
1000 size of the pointee type. The ``nonnull`` attribute does not imply
1001 dereferenceability (consider a pointer to one element past the end of an
1002 array), however ``dereferenceable(<n>)`` does imply ``nonnull`` in
1003 ``addrspace(0)`` (which is the default address space).
1007 Garbage Collector Names
1008 -----------------------
1010 Each function may specify a garbage collector name, which is simply a
1013 .. code-block:: llvm
1015 define void @f() gc "name" { ... }
1017 The compiler declares the supported values of *name*. Specifying a
1018 collector will cause the compiler to alter its output in order to
1019 support the named garbage collection algorithm.
1026 Prefix data is data associated with a function which the code
1027 generator will emit immediately before the function's entrypoint.
1028 The purpose of this feature is to allow frontends to associate
1029 language-specific runtime metadata with specific functions and make it
1030 available through the function pointer while still allowing the
1031 function pointer to be called.
1033 To access the data for a given function, a program may bitcast the
1034 function pointer to a pointer to the constant's type and dereference
1035 index -1. This implies that the IR symbol points just past the end of
1036 the prefix data. For instance, take the example of a function annotated
1037 with a single ``i32``,
1039 .. code-block:: llvm
1041 define void @f() prefix i32 123 { ... }
1043 The prefix data can be referenced as,
1045 .. code-block:: llvm
1047 %0 = bitcast *void () @f to *i32
1048 %a = getelementptr inbounds *i32 %0, i32 -1
1051 Prefix data is laid out as if it were an initializer for a global variable
1052 of the prefix data's type. The function will be placed such that the
1053 beginning of the prefix data is aligned. This means that if the size
1054 of the prefix data is not a multiple of the alignment size, the
1055 function's entrypoint will not be aligned. If alignment of the
1056 function's entrypoint is desired, padding must be added to the prefix
1059 A function may have prefix data but no body. This has similar semantics
1060 to the ``available_externally`` linkage in that the data may be used by the
1061 optimizers but will not be emitted in the object file.
1068 The ``prologue`` attribute allows arbitrary code (encoded as bytes) to
1069 be inserted prior to the function body. This can be used for enabling
1070 function hot-patching and instrumentation.
1072 To maintain the semantics of ordinary function calls, the prologue data must
1073 have a particular format. Specifically, it must begin with a sequence of
1074 bytes which decode to a sequence of machine instructions, valid for the
1075 module's target, which transfer control to the point immediately succeeding
1076 the prologue data, without performing any other visible action. This allows
1077 the inliner and other passes to reason about the semantics of the function
1078 definition without needing to reason about the prologue data. Obviously this
1079 makes the format of the prologue data highly target dependent.
1081 A trivial example of valid prologue data for the x86 architecture is ``i8 144``,
1082 which encodes the ``nop`` instruction:
1084 .. code-block:: llvm
1086 define void @f() prologue i8 144 { ... }
1088 Generally prologue data can be formed by encoding a relative branch instruction
1089 which skips the metadata, as in this example of valid prologue data for the
1090 x86_64 architecture, where the first two bytes encode ``jmp .+10``:
1092 .. code-block:: llvm
1094 %0 = type <{ i8, i8, i8* }>
1096 define void @f() prologue %0 <{ i8 235, i8 8, i8* @md}> { ... }
1098 A function may have prologue data but no body. This has similar semantics
1099 to the ``available_externally`` linkage in that the data may be used by the
1100 optimizers but will not be emitted in the object file.
1107 Attribute groups are groups of attributes that are referenced by objects within
1108 the IR. They are important for keeping ``.ll`` files readable, because a lot of
1109 functions will use the same set of attributes. In the degenerative case of a
1110 ``.ll`` file that corresponds to a single ``.c`` file, the single attribute
1111 group will capture the important command line flags used to build that file.
1113 An attribute group is a module-level object. To use an attribute group, an
1114 object references the attribute group's ID (e.g. ``#37``). An object may refer
1115 to more than one attribute group. In that situation, the attributes from the
1116 different groups are merged.
1118 Here is an example of attribute groups for a function that should always be
1119 inlined, has a stack alignment of 4, and which shouldn't use SSE instructions:
1121 .. code-block:: llvm
1123 ; Target-independent attributes:
1124 attributes #0 = { alwaysinline alignstack=4 }
1126 ; Target-dependent attributes:
1127 attributes #1 = { "no-sse" }
1129 ; Function @f has attributes: alwaysinline, alignstack=4, and "no-sse".
1130 define void @f() #0 #1 { ... }
1137 Function attributes are set to communicate additional information about
1138 a function. Function attributes are considered to be part of the
1139 function, not of the function type, so functions with different function
1140 attributes can have the same function type.
1142 Function attributes are simple keywords that follow the type specified.
1143 If multiple attributes are needed, they are space separated. For
1146 .. code-block:: llvm
1148 define void @f() noinline { ... }
1149 define void @f() alwaysinline { ... }
1150 define void @f() alwaysinline optsize { ... }
1151 define void @f() optsize { ... }
1154 This attribute indicates that, when emitting the prologue and
1155 epilogue, the backend should forcibly align the stack pointer.
1156 Specify the desired alignment, which must be a power of two, in
1159 This attribute indicates that the inliner should attempt to inline
1160 this function into callers whenever possible, ignoring any active
1161 inlining size threshold for this caller.
1163 This indicates that the callee function at a call site should be
1164 recognized as a built-in function, even though the function's declaration
1165 uses the ``nobuiltin`` attribute. This is only valid at call sites for
1166 direct calls to functions that are declared with the ``nobuiltin``
1169 This attribute indicates that this function is rarely called. When
1170 computing edge weights, basic blocks post-dominated by a cold
1171 function call are also considered to be cold; and, thus, given low
1174 This attribute indicates that the source code contained a hint that
1175 inlining this function is desirable (such as the "inline" keyword in
1176 C/C++). It is just a hint; it imposes no requirements on the
1179 This attribute indicates that the function should be added to a
1180 jump-instruction table at code-generation time, and that all address-taken
1181 references to this function should be replaced with a reference to the
1182 appropriate jump-instruction-table function pointer. Note that this creates
1183 a new pointer for the original function, which means that code that depends
1184 on function-pointer identity can break. So, any function annotated with
1185 ``jumptable`` must also be ``unnamed_addr``.
1187 This attribute suggests that optimization passes and code generator
1188 passes make choices that keep the code size of this function as small
1189 as possible and perform optimizations that may sacrifice runtime
1190 performance in order to minimize the size of the generated code.
1192 This attribute disables prologue / epilogue emission for the
1193 function. This can have very system-specific consequences.
1195 This indicates that the callee function at a call site is not recognized as
1196 a built-in function. LLVM will retain the original call and not replace it
1197 with equivalent code based on the semantics of the built-in function, unless
1198 the call site uses the ``builtin`` attribute. This is valid at call sites
1199 and on function declarations and definitions.
1201 This attribute indicates that calls to the function cannot be
1202 duplicated. A call to a ``noduplicate`` function may be moved
1203 within its parent function, but may not be duplicated within
1204 its parent function.
1206 A function containing a ``noduplicate`` call may still
1207 be an inlining candidate, provided that the call is not
1208 duplicated by inlining. That implies that the function has
1209 internal linkage and only has one call site, so the original
1210 call is dead after inlining.
1212 This attributes disables implicit floating point instructions.
1214 This attribute indicates that the inliner should never inline this
1215 function in any situation. This attribute may not be used together
1216 with the ``alwaysinline`` attribute.
1218 This attribute suppresses lazy symbol binding for the function. This
1219 may make calls to the function faster, at the cost of extra program
1220 startup time if the function is not called during program startup.
1222 This attribute indicates that the code generator should not use a
1223 red zone, even if the target-specific ABI normally permits it.
1225 This function attribute indicates that the function never returns
1226 normally. This produces undefined behavior at runtime if the
1227 function ever does dynamically return.
1229 This function attribute indicates that the function never returns
1230 with an unwind or exceptional control flow. If the function does
1231 unwind, its runtime behavior is undefined.
1233 This function attribute indicates that the function is not optimized
1234 by any optimization or code generator passes with the
1235 exception of interprocedural optimization passes.
1236 This attribute cannot be used together with the ``alwaysinline``
1237 attribute; this attribute is also incompatible
1238 with the ``minsize`` attribute and the ``optsize`` attribute.
1240 This attribute requires the ``noinline`` attribute to be specified on
1241 the function as well, so the function is never inlined into any caller.
1242 Only functions with the ``alwaysinline`` attribute are valid
1243 candidates for inlining into the body of this function.
1245 This attribute suggests that optimization passes and code generator
1246 passes make choices that keep the code size of this function low,
1247 and otherwise do optimizations specifically to reduce code size as
1248 long as they do not significantly impact runtime performance.
1250 On a function, this attribute indicates that the function computes its
1251 result (or decides to unwind an exception) based strictly on its arguments,
1252 without dereferencing any pointer arguments or otherwise accessing
1253 any mutable state (e.g. memory, control registers, etc) visible to
1254 caller functions. It does not write through any pointer arguments
1255 (including ``byval`` arguments) and never changes any state visible
1256 to callers. This means that it cannot unwind exceptions by calling
1257 the ``C++`` exception throwing methods.
1259 On an argument, this attribute indicates that the function does not
1260 dereference that pointer argument, even though it may read or write the
1261 memory that the pointer points to if accessed through other pointers.
1263 On a function, this attribute indicates that the function does not write
1264 through any pointer arguments (including ``byval`` arguments) or otherwise
1265 modify any state (e.g. memory, control registers, etc) visible to
1266 caller functions. It may dereference pointer arguments and read
1267 state that may be set in the caller. A readonly function always
1268 returns the same value (or unwinds an exception identically) when
1269 called with the same set of arguments and global state. It cannot
1270 unwind an exception by calling the ``C++`` exception throwing
1273 On an argument, this attribute indicates that the function does not write
1274 through this pointer argument, even though it may write to the memory that
1275 the pointer points to.
1277 This attribute indicates that this function can return twice. The C
1278 ``setjmp`` is an example of such a function. The compiler disables
1279 some optimizations (like tail calls) in the caller of these
1281 ``sanitize_address``
1282 This attribute indicates that AddressSanitizer checks
1283 (dynamic address safety analysis) are enabled for this function.
1285 This attribute indicates that MemorySanitizer checks (dynamic detection
1286 of accesses to uninitialized memory) are enabled for this function.
1288 This attribute indicates that ThreadSanitizer checks
1289 (dynamic thread safety analysis) are enabled for this function.
1291 This attribute indicates that the function should emit a stack
1292 smashing protector. It is in the form of a "canary" --- a random value
1293 placed on the stack before the local variables that's checked upon
1294 return from the function to see if it has been overwritten. A
1295 heuristic is used to determine if a function needs stack protectors
1296 or not. The heuristic used will enable protectors for functions with:
1298 - Character arrays larger than ``ssp-buffer-size`` (default 8).
1299 - Aggregates containing character arrays larger than ``ssp-buffer-size``.
1300 - Calls to alloca() with variable sizes or constant sizes greater than
1301 ``ssp-buffer-size``.
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.
1306 If a function that has an ``ssp`` attribute is inlined into a
1307 function that doesn't have an ``ssp`` attribute, then the resulting
1308 function will have an ``ssp`` attribute.
1310 This attribute indicates that the function should *always* emit a
1311 stack smashing protector. This overrides the ``ssp`` function
1314 Variables that are identified as requiring a protector will be arranged
1315 on the stack such that they are adjacent to the stack protector guard.
1316 The specific layout rules are:
1318 #. Large arrays and structures containing large arrays
1319 (``>= ssp-buffer-size``) are closest to the stack protector.
1320 #. Small arrays and structures containing small arrays
1321 (``< ssp-buffer-size``) are 2nd closest to the protector.
1322 #. Variables that have had their address taken are 3rd closest to the
1325 If a function that has an ``sspreq`` attribute is inlined into a
1326 function that doesn't have an ``sspreq`` attribute or which has an
1327 ``ssp`` or ``sspstrong`` attribute, then the resulting function will have
1328 an ``sspreq`` attribute.
1330 This attribute indicates that the function should emit a stack smashing
1331 protector. This attribute causes a strong heuristic to be used when
1332 determining if a function needs stack protectors. The strong heuristic
1333 will enable protectors for functions with:
1335 - Arrays of any size and type
1336 - Aggregates containing an array of any size and type.
1337 - Calls to alloca().
1338 - Local variables that have had their address taken.
1340 Variables that are identified as requiring a protector will be arranged
1341 on the stack such that they are adjacent to the stack protector guard.
1342 The specific layout rules are:
1344 #. Large arrays and structures containing large arrays
1345 (``>= ssp-buffer-size``) are closest to the stack protector.
1346 #. Small arrays and structures containing small arrays
1347 (``< ssp-buffer-size``) are 2nd closest to the protector.
1348 #. Variables that have had their address taken are 3rd closest to the
1351 This overrides the ``ssp`` function attribute.
1353 If a function that has an ``sspstrong`` attribute is inlined into a
1354 function that doesn't have an ``sspstrong`` attribute, then the
1355 resulting function will have an ``sspstrong`` attribute.
1357 This attribute indicates that the ABI being targeted requires that
1358 an unwind table entry be produce for this function even if we can
1359 show that no exceptions passes by it. This is normally the case for
1360 the ELF x86-64 abi, but it can be disabled for some compilation
1365 Module-Level Inline Assembly
1366 ----------------------------
1368 Modules may contain "module-level inline asm" blocks, which corresponds
1369 to the GCC "file scope inline asm" blocks. These blocks are internally
1370 concatenated by LLVM and treated as a single unit, but may be separated
1371 in the ``.ll`` file if desired. The syntax is very simple:
1373 .. code-block:: llvm
1375 module asm "inline asm code goes here"
1376 module asm "more can go here"
1378 The strings can contain any character by escaping non-printable
1379 characters. The escape sequence used is simply "\\xx" where "xx" is the
1380 two digit hex code for the number.
1382 The inline asm code is simply printed to the machine code .s file when
1383 assembly code is generated.
1385 .. _langref_datalayout:
1390 A module may specify a target specific data layout string that specifies
1391 how data is to be laid out in memory. The syntax for the data layout is
1394 .. code-block:: llvm
1396 target datalayout = "layout specification"
1398 The *layout specification* consists of a list of specifications
1399 separated by the minus sign character ('-'). Each specification starts
1400 with a letter and may include other information after the letter to
1401 define some aspect of the data layout. The specifications accepted are
1405 Specifies that the target lays out data in big-endian form. That is,
1406 the bits with the most significance have the lowest address
1409 Specifies that the target lays out data in little-endian form. That
1410 is, the bits with the least significance have the lowest address
1413 Specifies the natural alignment of the stack in bits. Alignment
1414 promotion of stack variables is limited to the natural stack
1415 alignment to avoid dynamic stack realignment. The stack alignment
1416 must be a multiple of 8-bits. If omitted, the natural stack
1417 alignment defaults to "unspecified", which does not prevent any
1418 alignment promotions.
1419 ``p[n]:<size>:<abi>:<pref>``
1420 This specifies the *size* of a pointer and its ``<abi>`` and
1421 ``<pref>``\erred alignments for address space ``n``. All sizes are in
1422 bits. The address space, ``n`` is optional, and if not specified,
1423 denotes the default address space 0. The value of ``n`` must be
1424 in the range [1,2^23).
1425 ``i<size>:<abi>:<pref>``
1426 This specifies the alignment for an integer type of a given bit
1427 ``<size>``. The value of ``<size>`` must be in the range [1,2^23).
1428 ``v<size>:<abi>:<pref>``
1429 This specifies the alignment for a vector type of a given bit
1431 ``f<size>:<abi>:<pref>``
1432 This specifies the alignment for a floating point type of a given bit
1433 ``<size>``. Only values of ``<size>`` that are supported by the target
1434 will work. 32 (float) and 64 (double) are supported on all targets; 80
1435 or 128 (different flavors of long double) are also supported on some
1438 This specifies the alignment for an object of aggregate type.
1440 If present, specifies that llvm names are mangled in the output. The
1443 * ``e``: ELF mangling: Private symbols get a ``.L`` prefix.
1444 * ``m``: Mips mangling: Private symbols get a ``$`` prefix.
1445 * ``o``: Mach-O mangling: Private symbols get ``L`` prefix. Other
1446 symbols get a ``_`` prefix.
1447 * ``w``: Windows COFF prefix: Similar to Mach-O, but stdcall and fastcall
1448 functions also get a suffix based on the frame size.
1449 ``n<size1>:<size2>:<size3>...``
1450 This specifies a set of native integer widths for the target CPU in
1451 bits. For example, it might contain ``n32`` for 32-bit PowerPC,
1452 ``n32:64`` for PowerPC 64, or ``n8:16:32:64`` for X86-64. Elements of
1453 this set are considered to support most general arithmetic operations
1456 On every specification that takes a ``<abi>:<pref>``, specifying the
1457 ``<pref>`` alignment is optional. If omitted, the preceding ``:``
1458 should be omitted too and ``<pref>`` will be equal to ``<abi>``.
1460 When constructing the data layout for a given target, LLVM starts with a
1461 default set of specifications which are then (possibly) overridden by
1462 the specifications in the ``datalayout`` keyword. The default
1463 specifications are given in this list:
1465 - ``E`` - big endian
1466 - ``p:64:64:64`` - 64-bit pointers with 64-bit alignment.
1467 - ``p[n]:64:64:64`` - Other address spaces are assumed to be the
1468 same as the default address space.
1469 - ``S0`` - natural stack alignment is unspecified
1470 - ``i1:8:8`` - i1 is 8-bit (byte) aligned
1471 - ``i8:8:8`` - i8 is 8-bit (byte) aligned
1472 - ``i16:16:16`` - i16 is 16-bit aligned
1473 - ``i32:32:32`` - i32 is 32-bit aligned
1474 - ``i64:32:64`` - i64 has ABI alignment of 32-bits but preferred
1475 alignment of 64-bits
1476 - ``f16:16:16`` - half is 16-bit aligned
1477 - ``f32:32:32`` - float is 32-bit aligned
1478 - ``f64:64:64`` - double is 64-bit aligned
1479 - ``f128:128:128`` - quad is 128-bit aligned
1480 - ``v64:64:64`` - 64-bit vector is 64-bit aligned
1481 - ``v128:128:128`` - 128-bit vector is 128-bit aligned
1482 - ``a:0:64`` - aggregates are 64-bit aligned
1484 When LLVM is determining the alignment for a given type, it uses the
1487 #. If the type sought is an exact match for one of the specifications,
1488 that specification is used.
1489 #. If no match is found, and the type sought is an integer type, then
1490 the smallest integer type that is larger than the bitwidth of the
1491 sought type is used. If none of the specifications are larger than
1492 the bitwidth then the largest integer type is used. For example,
1493 given the default specifications above, the i7 type will use the
1494 alignment of i8 (next largest) while both i65 and i256 will use the
1495 alignment of i64 (largest specified).
1496 #. If no match is found, and the type sought is a vector type, then the
1497 largest vector type that is smaller than the sought vector type will
1498 be used as a fall back. This happens because <128 x double> can be
1499 implemented in terms of 64 <2 x double>, for example.
1501 The function of the data layout string may not be what you expect.
1502 Notably, this is not a specification from the frontend of what alignment
1503 the code generator should use.
1505 Instead, if specified, the target data layout is required to match what
1506 the ultimate *code generator* expects. This string is used by the
1507 mid-level optimizers to improve code, and this only works if it matches
1508 what the ultimate code generator uses. If you would like to generate IR
1509 that does not embed this target-specific detail into the IR, then you
1510 don't have to specify the string. This will disable some optimizations
1511 that require precise layout information, but this also prevents those
1512 optimizations from introducing target specificity into the IR.
1519 A module may specify a target triple string that describes the target
1520 host. The syntax for the target triple is simply:
1522 .. code-block:: llvm
1524 target triple = "x86_64-apple-macosx10.7.0"
1526 The *target triple* string consists of a series of identifiers delimited
1527 by the minus sign character ('-'). The canonical forms are:
1531 ARCHITECTURE-VENDOR-OPERATING_SYSTEM
1532 ARCHITECTURE-VENDOR-OPERATING_SYSTEM-ENVIRONMENT
1534 This information is passed along to the backend so that it generates
1535 code for the proper architecture. It's possible to override this on the
1536 command line with the ``-mtriple`` command line option.
1538 .. _pointeraliasing:
1540 Pointer Aliasing Rules
1541 ----------------------
1543 Any memory access must be done through a pointer value associated with
1544 an address range of the memory access, otherwise the behavior is
1545 undefined. Pointer values are associated with address ranges according
1546 to the following rules:
1548 - A pointer value is associated with the addresses associated with any
1549 value it is *based* on.
1550 - An address of a global variable is associated with the address range
1551 of the variable's storage.
1552 - The result value of an allocation instruction is associated with the
1553 address range of the allocated storage.
1554 - A null pointer in the default address-space is associated with no
1556 - An integer constant other than zero or a pointer value returned from
1557 a function not defined within LLVM may be associated with address
1558 ranges allocated through mechanisms other than those provided by
1559 LLVM. Such ranges shall not overlap with any ranges of addresses
1560 allocated by mechanisms provided by LLVM.
1562 A pointer value is *based* on another pointer value according to the
1565 - A pointer value formed from a ``getelementptr`` operation is *based*
1566 on the first operand of the ``getelementptr``.
1567 - The result value of a ``bitcast`` is *based* on the operand of the
1569 - A pointer value formed by an ``inttoptr`` is *based* on all pointer
1570 values that contribute (directly or indirectly) to the computation of
1571 the pointer's value.
1572 - The "*based* on" relationship is transitive.
1574 Note that this definition of *"based"* is intentionally similar to the
1575 definition of *"based"* in C99, though it is slightly weaker.
1577 LLVM IR does not associate types with memory. The result type of a
1578 ``load`` merely indicates the size and alignment of the memory from
1579 which to load, as well as the interpretation of the value. The first
1580 operand type of a ``store`` similarly only indicates the size and
1581 alignment of the store.
1583 Consequently, type-based alias analysis, aka TBAA, aka
1584 ``-fstrict-aliasing``, is not applicable to general unadorned LLVM IR.
1585 :ref:`Metadata <metadata>` may be used to encode additional information
1586 which specialized optimization passes may use to implement type-based
1591 Volatile Memory Accesses
1592 ------------------------
1594 Certain memory accesses, such as :ref:`load <i_load>`'s,
1595 :ref:`store <i_store>`'s, and :ref:`llvm.memcpy <int_memcpy>`'s may be
1596 marked ``volatile``. The optimizers must not change the number of
1597 volatile operations or change their order of execution relative to other
1598 volatile operations. The optimizers *may* change the order of volatile
1599 operations relative to non-volatile operations. This is not Java's
1600 "volatile" and has no cross-thread synchronization behavior.
1602 IR-level volatile loads and stores cannot safely be optimized into
1603 llvm.memcpy or llvm.memmove intrinsics even when those intrinsics are
1604 flagged volatile. Likewise, the backend should never split or merge
1605 target-legal volatile load/store instructions.
1607 .. admonition:: Rationale
1609 Platforms may rely on volatile loads and stores of natively supported
1610 data width to be executed as single instruction. For example, in C
1611 this holds for an l-value of volatile primitive type with native
1612 hardware support, but not necessarily for aggregate types. The
1613 frontend upholds these expectations, which are intentionally
1614 unspecified in the IR. The rules above ensure that IR transformation
1615 do not violate the frontend's contract with the language.
1619 Memory Model for Concurrent Operations
1620 --------------------------------------
1622 The LLVM IR does not define any way to start parallel threads of
1623 execution or to register signal handlers. Nonetheless, there are
1624 platform-specific ways to create them, and we define LLVM IR's behavior
1625 in their presence. This model is inspired by the C++0x memory model.
1627 For a more informal introduction to this model, see the :doc:`Atomics`.
1629 We define a *happens-before* partial order as the least partial order
1632 - Is a superset of single-thread program order, and
1633 - When a *synchronizes-with* ``b``, includes an edge from ``a`` to
1634 ``b``. *Synchronizes-with* pairs are introduced by platform-specific
1635 techniques, like pthread locks, thread creation, thread joining,
1636 etc., and by atomic instructions. (See also :ref:`Atomic Memory Ordering
1637 Constraints <ordering>`).
1639 Note that program order does not introduce *happens-before* edges
1640 between a thread and signals executing inside that thread.
1642 Every (defined) read operation (load instructions, memcpy, atomic
1643 loads/read-modify-writes, etc.) R reads a series of bytes written by
1644 (defined) write operations (store instructions, atomic
1645 stores/read-modify-writes, memcpy, etc.). For the purposes of this
1646 section, initialized globals are considered to have a write of the
1647 initializer which is atomic and happens before any other read or write
1648 of the memory in question. For each byte of a read R, R\ :sub:`byte`
1649 may see any write to the same byte, except:
1651 - If write\ :sub:`1` happens before write\ :sub:`2`, and
1652 write\ :sub:`2` happens before R\ :sub:`byte`, then
1653 R\ :sub:`byte` does not see write\ :sub:`1`.
1654 - If R\ :sub:`byte` happens before write\ :sub:`3`, then
1655 R\ :sub:`byte` does not see write\ :sub:`3`.
1657 Given that definition, R\ :sub:`byte` is defined as follows:
1659 - If R is volatile, the result is target-dependent. (Volatile is
1660 supposed to give guarantees which can support ``sig_atomic_t`` in
1661 C/C++, and may be used for accesses to addresses that do not behave
1662 like normal memory. It does not generally provide cross-thread
1664 - Otherwise, if there is no write to the same byte that happens before
1665 R\ :sub:`byte`, R\ :sub:`byte` returns ``undef`` for that byte.
1666 - Otherwise, if R\ :sub:`byte` may see exactly one write,
1667 R\ :sub:`byte` returns the value written by that write.
1668 - Otherwise, if R is atomic, and all the writes R\ :sub:`byte` may
1669 see are atomic, it chooses one of the values written. See the :ref:`Atomic
1670 Memory Ordering Constraints <ordering>` section for additional
1671 constraints on how the choice is made.
1672 - Otherwise R\ :sub:`byte` returns ``undef``.
1674 R returns the value composed of the series of bytes it read. This
1675 implies that some bytes within the value may be ``undef`` **without**
1676 the entire value being ``undef``. Note that this only defines the
1677 semantics of the operation; it doesn't mean that targets will emit more
1678 than one instruction to read the series of bytes.
1680 Note that in cases where none of the atomic intrinsics are used, this
1681 model places only one restriction on IR transformations on top of what
1682 is required for single-threaded execution: introducing a store to a byte
1683 which might not otherwise be stored is not allowed in general.
1684 (Specifically, in the case where another thread might write to and read
1685 from an address, introducing a store can change a load that may see
1686 exactly one write into a load that may see multiple writes.)
1690 Atomic Memory Ordering Constraints
1691 ----------------------------------
1693 Atomic instructions (:ref:`cmpxchg <i_cmpxchg>`,
1694 :ref:`atomicrmw <i_atomicrmw>`, :ref:`fence <i_fence>`,
1695 :ref:`atomic load <i_load>`, and :ref:`atomic store <i_store>`) take
1696 ordering parameters that determine which other atomic instructions on
1697 the same address they *synchronize with*. These semantics are borrowed
1698 from Java and C++0x, but are somewhat more colloquial. If these
1699 descriptions aren't precise enough, check those specs (see spec
1700 references in the :doc:`atomics guide <Atomics>`).
1701 :ref:`fence <i_fence>` instructions treat these orderings somewhat
1702 differently since they don't take an address. See that instruction's
1703 documentation for details.
1705 For a simpler introduction to the ordering constraints, see the
1709 The set of values that can be read is governed by the happens-before
1710 partial order. A value cannot be read unless some operation wrote
1711 it. This is intended to provide a guarantee strong enough to model
1712 Java's non-volatile shared variables. This ordering cannot be
1713 specified for read-modify-write operations; it is not strong enough
1714 to make them atomic in any interesting way.
1716 In addition to the guarantees of ``unordered``, there is a single
1717 total order for modifications by ``monotonic`` operations on each
1718 address. All modification orders must be compatible with the
1719 happens-before order. There is no guarantee that the modification
1720 orders can be combined to a global total order for the whole program
1721 (and this often will not be possible). The read in an atomic
1722 read-modify-write operation (:ref:`cmpxchg <i_cmpxchg>` and
1723 :ref:`atomicrmw <i_atomicrmw>`) reads the value in the modification
1724 order immediately before the value it writes. If one atomic read
1725 happens before another atomic read of the same address, the later
1726 read must see the same value or a later value in the address's
1727 modification order. This disallows reordering of ``monotonic`` (or
1728 stronger) operations on the same address. If an address is written
1729 ``monotonic``-ally by one thread, and other threads ``monotonic``-ally
1730 read that address repeatedly, the other threads must eventually see
1731 the write. This corresponds to the C++0x/C1x
1732 ``memory_order_relaxed``.
1734 In addition to the guarantees of ``monotonic``, a
1735 *synchronizes-with* edge may be formed with a ``release`` operation.
1736 This is intended to model C++'s ``memory_order_acquire``.
1738 In addition to the guarantees of ``monotonic``, if this operation
1739 writes a value which is subsequently read by an ``acquire``
1740 operation, it *synchronizes-with* that operation. (This isn't a
1741 complete description; see the C++0x definition of a release
1742 sequence.) This corresponds to the C++0x/C1x
1743 ``memory_order_release``.
1744 ``acq_rel`` (acquire+release)
1745 Acts as both an ``acquire`` and ``release`` operation on its
1746 address. This corresponds to the C++0x/C1x ``memory_order_acq_rel``.
1747 ``seq_cst`` (sequentially consistent)
1748 In addition to the guarantees of ``acq_rel`` (``acquire`` for an
1749 operation that only reads, ``release`` for an operation that only
1750 writes), there is a global total order on all
1751 sequentially-consistent operations on all addresses, which is
1752 consistent with the *happens-before* partial order and with the
1753 modification orders of all the affected addresses. Each
1754 sequentially-consistent read sees the last preceding write to the
1755 same address in this global order. This corresponds to the C++0x/C1x
1756 ``memory_order_seq_cst`` and Java volatile.
1760 If an atomic operation is marked ``singlethread``, it only *synchronizes
1761 with* or participates in modification and seq\_cst total orderings with
1762 other operations running in the same thread (for example, in signal
1770 LLVM IR floating-point binary ops (:ref:`fadd <i_fadd>`,
1771 :ref:`fsub <i_fsub>`, :ref:`fmul <i_fmul>`, :ref:`fdiv <i_fdiv>`,
1772 :ref:`frem <i_frem>`) have the following flags that can set to enable
1773 otherwise unsafe floating point operations
1776 No NaNs - Allow optimizations to assume the arguments and result are not
1777 NaN. Such optimizations are required to retain defined behavior over
1778 NaNs, but the value of the result is undefined.
1781 No Infs - Allow optimizations to assume the arguments and result are not
1782 +/-Inf. Such optimizations are required to retain defined behavior over
1783 +/-Inf, but the value of the result is undefined.
1786 No Signed Zeros - Allow optimizations to treat the sign of a zero
1787 argument or result as insignificant.
1790 Allow Reciprocal - Allow optimizations to use the reciprocal of an
1791 argument rather than perform division.
1794 Fast - Allow algebraically equivalent transformations that may
1795 dramatically change results in floating point (e.g. reassociate). This
1796 flag implies all the others.
1800 Use-list Order Directives
1801 -------------------------
1803 Use-list directives encode the in-memory order of each use-list, allowing the
1804 order to be recreated. ``<order-indexes>`` is a comma-separated list of
1805 indexes that are assigned to the referenced value's uses. The referenced
1806 value's use-list is immediately sorted by these indexes.
1808 Use-list directives may appear at function scope or global scope. They are not
1809 instructions, and have no effect on the semantics of the IR. When they're at
1810 function scope, they must appear after the terminator of the final basic block.
1812 If basic blocks have their address taken via ``blockaddress()`` expressions,
1813 ``uselistorder_bb`` can be used to reorder their use-lists from outside their
1820 uselistorder <ty> <value>, { <order-indexes> }
1821 uselistorder_bb @function, %block { <order-indexes> }
1827 define void @foo(i32 %arg1, i32 %arg2) {
1829 ; ... instructions ...
1831 ; ... instructions ...
1833 ; At function scope.
1834 uselistorder i32 %arg1, { 1, 0, 2 }
1835 uselistorder label %bb, { 1, 0 }
1839 uselistorder i32* @global, { 1, 2, 0 }
1840 uselistorder i32 7, { 1, 0 }
1841 uselistorder i32 (i32) @bar, { 1, 0 }
1842 uselistorder_bb @foo, %bb, { 5, 1, 3, 2, 0, 4 }
1849 The LLVM type system is one of the most important features of the
1850 intermediate representation. Being typed enables a number of
1851 optimizations to be performed on the intermediate representation
1852 directly, without having to do extra analyses on the side before the
1853 transformation. A strong type system makes it easier to read the
1854 generated code and enables novel analyses and transformations that are
1855 not feasible to perform on normal three address code representations.
1865 The void type does not represent any value and has no size.
1883 The function type can be thought of as a function signature. It consists of a
1884 return type and a list of formal parameter types. The return type of a function
1885 type is a void type or first class type --- except for :ref:`label <t_label>`
1886 and :ref:`metadata <t_metadata>` types.
1892 <returntype> (<parameter list>)
1894 ...where '``<parameter list>``' is a comma-separated list of type
1895 specifiers. Optionally, the parameter list may include a type ``...``, which
1896 indicates that the function takes a variable number of arguments. Variable
1897 argument functions can access their arguments with the :ref:`variable argument
1898 handling intrinsic <int_varargs>` functions. '``<returntype>``' is any type
1899 except :ref:`label <t_label>` and :ref:`metadata <t_metadata>`.
1903 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1904 | ``i32 (i32)`` | function taking an ``i32``, returning an ``i32`` |
1905 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1906 | ``float (i16, i32 *) *`` | :ref:`Pointer <t_pointer>` to a function that takes an ``i16`` and a :ref:`pointer <t_pointer>` to ``i32``, returning ``float``. |
1907 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1908 | ``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. |
1909 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1910 | ``{i32, i32} (i32)`` | A function taking an ``i32``, returning a :ref:`structure <t_struct>` containing two ``i32`` values |
1911 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1918 The :ref:`first class <t_firstclass>` types are perhaps the most important.
1919 Values of these types are the only ones which can be produced by
1927 These are the types that are valid in registers from CodeGen's perspective.
1936 The integer type is a very simple type that simply specifies an
1937 arbitrary bit width for the integer type desired. Any bit width from 1
1938 bit to 2\ :sup:`23`\ -1 (about 8 million) can be specified.
1946 The number of bits the integer will occupy is specified by the ``N``
1952 +----------------+------------------------------------------------+
1953 | ``i1`` | a single-bit integer. |
1954 +----------------+------------------------------------------------+
1955 | ``i32`` | a 32-bit integer. |
1956 +----------------+------------------------------------------------+
1957 | ``i1942652`` | a really big integer of over 1 million bits. |
1958 +----------------+------------------------------------------------+
1962 Floating Point Types
1963 """"""""""""""""""""
1972 - 16-bit floating point value
1975 - 32-bit floating point value
1978 - 64-bit floating point value
1981 - 128-bit floating point value (112-bit mantissa)
1984 - 80-bit floating point value (X87)
1987 - 128-bit floating point value (two 64-bits)
1994 The x86_mmx type represents a value held in an MMX register on an x86
1995 machine. The operations allowed on it are quite limited: parameters and
1996 return values, load and store, and bitcast. User-specified MMX
1997 instructions are represented as intrinsic or asm calls with arguments
1998 and/or results of this type. There are no arrays, vectors or constants
2015 The pointer type is used to specify memory locations. Pointers are
2016 commonly used to reference objects in memory.
2018 Pointer types may have an optional address space attribute defining the
2019 numbered address space where the pointed-to object resides. The default
2020 address space is number zero. The semantics of non-zero address spaces
2021 are target-specific.
2023 Note that LLVM does not permit pointers to void (``void*``) nor does it
2024 permit pointers to labels (``label*``). Use ``i8*`` instead.
2034 +-------------------------+--------------------------------------------------------------------------------------------------------------+
2035 | ``[4 x i32]*`` | A :ref:`pointer <t_pointer>` to :ref:`array <t_array>` of four ``i32`` values. |
2036 +-------------------------+--------------------------------------------------------------------------------------------------------------+
2037 | ``i32 (i32*) *`` | A :ref:`pointer <t_pointer>` to a :ref:`function <t_function>` that takes an ``i32*``, returning an ``i32``. |
2038 +-------------------------+--------------------------------------------------------------------------------------------------------------+
2039 | ``i32 addrspace(5)*`` | A :ref:`pointer <t_pointer>` to an ``i32`` value that resides in address space #5. |
2040 +-------------------------+--------------------------------------------------------------------------------------------------------------+
2049 A vector type is a simple derived type that represents a vector of
2050 elements. Vector types are used when multiple primitive data are
2051 operated in parallel using a single instruction (SIMD). A vector type
2052 requires a size (number of elements) and an underlying primitive data
2053 type. Vector types are considered :ref:`first class <t_firstclass>`.
2059 < <# elements> x <elementtype> >
2061 The number of elements is a constant integer value larger than 0;
2062 elementtype may be any integer, floating point or pointer type. Vectors
2063 of size zero are not allowed.
2067 +-------------------+--------------------------------------------------+
2068 | ``<4 x i32>`` | Vector of 4 32-bit integer values. |
2069 +-------------------+--------------------------------------------------+
2070 | ``<8 x float>`` | Vector of 8 32-bit floating-point values. |
2071 +-------------------+--------------------------------------------------+
2072 | ``<2 x i64>`` | Vector of 2 64-bit integer values. |
2073 +-------------------+--------------------------------------------------+
2074 | ``<4 x i64*>`` | Vector of 4 pointers to 64-bit integer values. |
2075 +-------------------+--------------------------------------------------+
2084 The label type represents code labels.
2099 The metadata type represents embedded metadata. No derived types may be
2100 created from metadata except for :ref:`function <t_function>` arguments.
2113 Aggregate Types are a subset of derived types that can contain multiple
2114 member types. :ref:`Arrays <t_array>` and :ref:`structs <t_struct>` are
2115 aggregate types. :ref:`Vectors <t_vector>` are not considered to be
2125 The array type is a very simple derived type that arranges elements
2126 sequentially in memory. The array type requires a size (number of
2127 elements) and an underlying data type.
2133 [<# elements> x <elementtype>]
2135 The number of elements is a constant integer value; ``elementtype`` may
2136 be any type with a size.
2140 +------------------+--------------------------------------+
2141 | ``[40 x i32]`` | Array of 40 32-bit integer values. |
2142 +------------------+--------------------------------------+
2143 | ``[41 x i32]`` | Array of 41 32-bit integer values. |
2144 +------------------+--------------------------------------+
2145 | ``[4 x i8]`` | Array of 4 8-bit integer values. |
2146 +------------------+--------------------------------------+
2148 Here are some examples of multidimensional arrays:
2150 +-----------------------------+----------------------------------------------------------+
2151 | ``[3 x [4 x i32]]`` | 3x4 array of 32-bit integer values. |
2152 +-----------------------------+----------------------------------------------------------+
2153 | ``[12 x [10 x float]]`` | 12x10 array of single precision floating point values. |
2154 +-----------------------------+----------------------------------------------------------+
2155 | ``[2 x [3 x [4 x i16]]]`` | 2x3x4 array of 16-bit integer values. |
2156 +-----------------------------+----------------------------------------------------------+
2158 There is no restriction on indexing beyond the end of the array implied
2159 by a static type (though there are restrictions on indexing beyond the
2160 bounds of an allocated object in some cases). This means that
2161 single-dimension 'variable sized array' addressing can be implemented in
2162 LLVM with a zero length array type. An implementation of 'pascal style
2163 arrays' in LLVM could use the type "``{ i32, [0 x float]}``", for
2173 The structure type is used to represent a collection of data members
2174 together in memory. The elements of a structure may be any type that has
2177 Structures in memory are accessed using '``load``' and '``store``' by
2178 getting a pointer to a field with the '``getelementptr``' instruction.
2179 Structures in registers are accessed using the '``extractvalue``' and
2180 '``insertvalue``' instructions.
2182 Structures may optionally be "packed" structures, which indicate that
2183 the alignment of the struct is one byte, and that there is no padding
2184 between the elements. In non-packed structs, padding between field types
2185 is inserted as defined by the DataLayout string in the module, which is
2186 required to match what the underlying code generator expects.
2188 Structures can either be "literal" or "identified". A literal structure
2189 is defined inline with other types (e.g. ``{i32, i32}*``) whereas
2190 identified types are always defined at the top level with a name.
2191 Literal types are uniqued by their contents and can never be recursive
2192 or opaque since there is no way to write one. Identified types can be
2193 recursive, can be opaqued, and are never uniqued.
2199 %T1 = type { <type list> } ; Identified normal struct type
2200 %T2 = type <{ <type list> }> ; Identified packed struct type
2204 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2205 | ``{ i32, i32, i32 }`` | A triple of three ``i32`` values |
2206 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2207 | ``{ 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``. |
2208 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2209 | ``<{ i8, i32 }>`` | A packed struct known to be 5 bytes in size. |
2210 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2214 Opaque Structure Types
2215 """"""""""""""""""""""
2219 Opaque structure types are used to represent named structure types that
2220 do not have a body specified. This corresponds (for example) to the C
2221 notion of a forward declared structure.
2232 +--------------+-------------------+
2233 | ``opaque`` | An opaque type. |
2234 +--------------+-------------------+
2241 LLVM has several different basic types of constants. This section
2242 describes them all and their syntax.
2247 **Boolean constants**
2248 The two strings '``true``' and '``false``' are both valid constants
2250 **Integer constants**
2251 Standard integers (such as '4') are constants of the
2252 :ref:`integer <t_integer>` type. Negative numbers may be used with
2254 **Floating point constants**
2255 Floating point constants use standard decimal notation (e.g.
2256 123.421), exponential notation (e.g. 1.23421e+2), or a more precise
2257 hexadecimal notation (see below). The assembler requires the exact
2258 decimal value of a floating-point constant. For example, the
2259 assembler accepts 1.25 but rejects 1.3 because 1.3 is a repeating
2260 decimal in binary. Floating point constants must have a :ref:`floating
2261 point <t_floating>` type.
2262 **Null pointer constants**
2263 The identifier '``null``' is recognized as a null pointer constant
2264 and must be of :ref:`pointer type <t_pointer>`.
2266 The one non-intuitive notation for constants is the hexadecimal form of
2267 floating point constants. For example, the form
2268 '``double 0x432ff973cafa8000``' is equivalent to (but harder to read
2269 than) '``double 4.5e+15``'. The only time hexadecimal floating point
2270 constants are required (and the only time that they are generated by the
2271 disassembler) is when a floating point constant must be emitted but it
2272 cannot be represented as a decimal floating point number in a reasonable
2273 number of digits. For example, NaN's, infinities, and other special
2274 values are represented in their IEEE hexadecimal format so that assembly
2275 and disassembly do not cause any bits to change in the constants.
2277 When using the hexadecimal form, constants of types half, float, and
2278 double are represented using the 16-digit form shown above (which
2279 matches the IEEE754 representation for double); half and float values
2280 must, however, be exactly representable as IEEE 754 half and single
2281 precision, respectively. Hexadecimal format is always used for long
2282 double, and there are three forms of long double. The 80-bit format used
2283 by x86 is represented as ``0xK`` followed by 20 hexadecimal digits. The
2284 128-bit format used by PowerPC (two adjacent doubles) is represented by
2285 ``0xM`` followed by 32 hexadecimal digits. The IEEE 128-bit format is
2286 represented by ``0xL`` followed by 32 hexadecimal digits. Long doubles
2287 will only work if they match the long double format on your target.
2288 The IEEE 16-bit format (half precision) is represented by ``0xH``
2289 followed by 4 hexadecimal digits. All hexadecimal formats are big-endian
2290 (sign bit at the left).
2292 There are no constants of type x86_mmx.
2294 .. _complexconstants:
2299 Complex constants are a (potentially recursive) combination of simple
2300 constants and smaller complex constants.
2302 **Structure constants**
2303 Structure constants are represented with notation similar to
2304 structure type definitions (a comma separated list of elements,
2305 surrounded by braces (``{}``)). For example:
2306 "``{ i32 4, float 17.0, i32* @G }``", where "``@G``" is declared as
2307 "``@G = external global i32``". Structure constants must have
2308 :ref:`structure type <t_struct>`, and the number and types of elements
2309 must match those specified by the type.
2311 Array constants are represented with notation similar to array type
2312 definitions (a comma separated list of elements, surrounded by
2313 square brackets (``[]``)). For example:
2314 "``[ i32 42, i32 11, i32 74 ]``". Array constants must have
2315 :ref:`array type <t_array>`, and the number and types of elements must
2316 match those specified by the type. As a special case, character array
2317 constants may also be represented as a double-quoted string using the ``c``
2318 prefix. For example: "``c"Hello World\0A\00"``".
2319 **Vector constants**
2320 Vector constants are represented with notation similar to vector
2321 type definitions (a comma separated list of elements, surrounded by
2322 less-than/greater-than's (``<>``)). For example:
2323 "``< i32 42, i32 11, i32 74, i32 100 >``". Vector constants
2324 must have :ref:`vector type <t_vector>`, and the number and types of
2325 elements must match those specified by the type.
2326 **Zero initialization**
2327 The string '``zeroinitializer``' can be used to zero initialize a
2328 value to zero of *any* type, including scalar and
2329 :ref:`aggregate <t_aggregate>` types. This is often used to avoid
2330 having to print large zero initializers (e.g. for large arrays) and
2331 is always exactly equivalent to using explicit zero initializers.
2333 A metadata node is a constant tuple without types. For example:
2334 "``!{!0, !{!2, !0}, !"test"}``". Metadata can reference constant values,
2335 for example: "``!{!0, i32 0, i8* @global, i64 (i64)* @function, !"str"}``".
2336 Unlike other typed constants that are meant to be interpreted as part of
2337 the instruction stream, metadata is a place to attach additional
2338 information such as debug info.
2340 Global Variable and Function Addresses
2341 --------------------------------------
2343 The addresses of :ref:`global variables <globalvars>` and
2344 :ref:`functions <functionstructure>` are always implicitly valid
2345 (link-time) constants. These constants are explicitly referenced when
2346 the :ref:`identifier for the global <identifiers>` is used and always have
2347 :ref:`pointer <t_pointer>` type. For example, the following is a legal LLVM
2350 .. code-block:: llvm
2354 @Z = global [2 x i32*] [ i32* @X, i32* @Y ]
2361 The string '``undef``' can be used anywhere a constant is expected, and
2362 indicates that the user of the value may receive an unspecified
2363 bit-pattern. Undefined values may be of any type (other than '``label``'
2364 or '``void``') and be used anywhere a constant is permitted.
2366 Undefined values are useful because they indicate to the compiler that
2367 the program is well defined no matter what value is used. This gives the
2368 compiler more freedom to optimize. Here are some examples of
2369 (potentially surprising) transformations that are valid (in pseudo IR):
2371 .. code-block:: llvm
2381 This is safe because all of the output bits are affected by the undef
2382 bits. Any output bit can have a zero or one depending on the input bits.
2384 .. code-block:: llvm
2395 These logical operations have bits that are not always affected by the
2396 input. For example, if ``%X`` has a zero bit, then the output of the
2397 '``and``' operation will always be a zero for that bit, no matter what
2398 the corresponding bit from the '``undef``' is. As such, it is unsafe to
2399 optimize or assume that the result of the '``and``' is '``undef``'.
2400 However, it is safe to assume that all bits of the '``undef``' could be
2401 0, and optimize the '``and``' to 0. Likewise, it is safe to assume that
2402 all the bits of the '``undef``' operand to the '``or``' could be set,
2403 allowing the '``or``' to be folded to -1.
2405 .. code-block:: llvm
2407 %A = select undef, %X, %Y
2408 %B = select undef, 42, %Y
2409 %C = select %X, %Y, undef
2419 This set of examples shows that undefined '``select``' (and conditional
2420 branch) conditions can go *either way*, but they have to come from one
2421 of the two operands. In the ``%A`` example, if ``%X`` and ``%Y`` were
2422 both known to have a clear low bit, then ``%A`` would have to have a
2423 cleared low bit. However, in the ``%C`` example, the optimizer is
2424 allowed to assume that the '``undef``' operand could be the same as
2425 ``%Y``, allowing the whole '``select``' to be eliminated.
2427 .. code-block:: llvm
2429 %A = xor undef, undef
2446 This example points out that two '``undef``' operands are not
2447 necessarily the same. This can be surprising to people (and also matches
2448 C semantics) where they assume that "``X^X``" is always zero, even if
2449 ``X`` is undefined. This isn't true for a number of reasons, but the
2450 short answer is that an '``undef``' "variable" can arbitrarily change
2451 its value over its "live range". This is true because the variable
2452 doesn't actually *have a live range*. Instead, the value is logically
2453 read from arbitrary registers that happen to be around when needed, so
2454 the value is not necessarily consistent over time. In fact, ``%A`` and
2455 ``%C`` need to have the same semantics or the core LLVM "replace all
2456 uses with" concept would not hold.
2458 .. code-block:: llvm
2466 These examples show the crucial difference between an *undefined value*
2467 and *undefined behavior*. An undefined value (like '``undef``') is
2468 allowed to have an arbitrary bit-pattern. This means that the ``%A``
2469 operation can be constant folded to '``undef``', because the '``undef``'
2470 could be an SNaN, and ``fdiv`` is not (currently) defined on SNaN's.
2471 However, in the second example, we can make a more aggressive
2472 assumption: because the ``undef`` is allowed to be an arbitrary value,
2473 we are allowed to assume that it could be zero. Since a divide by zero
2474 has *undefined behavior*, we are allowed to assume that the operation
2475 does not execute at all. This allows us to delete the divide and all
2476 code after it. Because the undefined operation "can't happen", the
2477 optimizer can assume that it occurs in dead code.
2479 .. code-block:: llvm
2481 a: store undef -> %X
2482 b: store %X -> undef
2487 These examples reiterate the ``fdiv`` example: a store *of* an undefined
2488 value can be assumed to not have any effect; we can assume that the
2489 value is overwritten with bits that happen to match what was already
2490 there. However, a store *to* an undefined location could clobber
2491 arbitrary memory, therefore, it has undefined behavior.
2498 Poison values are similar to :ref:`undef values <undefvalues>`, however
2499 they also represent the fact that an instruction or constant expression
2500 that cannot evoke side effects has nevertheless detected a condition
2501 that results in undefined behavior.
2503 There is currently no way of representing a poison value in the IR; they
2504 only exist when produced by operations such as :ref:`add <i_add>` with
2507 Poison value behavior is defined in terms of value *dependence*:
2509 - Values other than :ref:`phi <i_phi>` nodes depend on their operands.
2510 - :ref:`Phi <i_phi>` nodes depend on the operand corresponding to
2511 their dynamic predecessor basic block.
2512 - Function arguments depend on the corresponding actual argument values
2513 in the dynamic callers of their functions.
2514 - :ref:`Call <i_call>` instructions depend on the :ref:`ret <i_ret>`
2515 instructions that dynamically transfer control back to them.
2516 - :ref:`Invoke <i_invoke>` instructions depend on the
2517 :ref:`ret <i_ret>`, :ref:`resume <i_resume>`, or exception-throwing
2518 call instructions that dynamically transfer control back to them.
2519 - Non-volatile loads and stores depend on the most recent stores to all
2520 of the referenced memory addresses, following the order in the IR
2521 (including loads and stores implied by intrinsics such as
2522 :ref:`@llvm.memcpy <int_memcpy>`.)
2523 - An instruction with externally visible side effects depends on the
2524 most recent preceding instruction with externally visible side
2525 effects, following the order in the IR. (This includes :ref:`volatile
2526 operations <volatile>`.)
2527 - An instruction *control-depends* on a :ref:`terminator
2528 instruction <terminators>` if the terminator instruction has
2529 multiple successors and the instruction is always executed when
2530 control transfers to one of the successors, and may not be executed
2531 when control is transferred to another.
2532 - Additionally, an instruction also *control-depends* on a terminator
2533 instruction if the set of instructions it otherwise depends on would
2534 be different if the terminator had transferred control to a different
2536 - Dependence is transitive.
2538 Poison values have the same behavior as :ref:`undef values <undefvalues>`,
2539 with the additional effect that any instruction that has a *dependence*
2540 on a poison value has undefined behavior.
2542 Here are some examples:
2544 .. code-block:: llvm
2547 %poison = sub nuw i32 0, 1 ; Results in a poison value.
2548 %still_poison = and i32 %poison, 0 ; 0, but also poison.
2549 %poison_yet_again = getelementptr i32* @h, i32 %still_poison
2550 store i32 0, i32* %poison_yet_again ; memory at @h[0] is poisoned
2552 store i32 %poison, i32* @g ; Poison value stored to memory.
2553 %poison2 = load i32* @g ; Poison value loaded back from memory.
2555 store volatile i32 %poison, i32* @g ; External observation; undefined behavior.
2557 %narrowaddr = bitcast i32* @g to i16*
2558 %wideaddr = bitcast i32* @g to i64*
2559 %poison3 = load i16* %narrowaddr ; Returns a poison value.
2560 %poison4 = load i64* %wideaddr ; Returns a poison value.
2562 %cmp = icmp slt i32 %poison, 0 ; Returns a poison value.
2563 br i1 %cmp, label %true, label %end ; Branch to either destination.
2566 store volatile i32 0, i32* @g ; This is control-dependent on %cmp, so
2567 ; it has undefined behavior.
2571 %p = phi i32 [ 0, %entry ], [ 1, %true ]
2572 ; Both edges into this PHI are
2573 ; control-dependent on %cmp, so this
2574 ; always results in a poison value.
2576 store volatile i32 0, i32* @g ; This would depend on the store in %true
2577 ; if %cmp is true, or the store in %entry
2578 ; otherwise, so this is undefined behavior.
2580 br i1 %cmp, label %second_true, label %second_end
2581 ; The same branch again, but this time the
2582 ; true block doesn't have side effects.
2589 store volatile i32 0, i32* @g ; This time, the instruction always depends
2590 ; on the store in %end. Also, it is
2591 ; control-equivalent to %end, so this is
2592 ; well-defined (ignoring earlier undefined
2593 ; behavior in this example).
2597 Addresses of Basic Blocks
2598 -------------------------
2600 ``blockaddress(@function, %block)``
2602 The '``blockaddress``' constant computes the address of the specified
2603 basic block in the specified function, and always has an ``i8*`` type.
2604 Taking the address of the entry block is illegal.
2606 This value only has defined behavior when used as an operand to the
2607 ':ref:`indirectbr <i_indirectbr>`' instruction, or for comparisons
2608 against null. Pointer equality tests between labels addresses results in
2609 undefined behavior --- though, again, comparison against null is ok, and
2610 no label is equal to the null pointer. This may be passed around as an
2611 opaque pointer sized value as long as the bits are not inspected. This
2612 allows ``ptrtoint`` and arithmetic to be performed on these values so
2613 long as the original value is reconstituted before the ``indirectbr``
2616 Finally, some targets may provide defined semantics when using the value
2617 as the operand to an inline assembly, but that is target specific.
2621 Constant Expressions
2622 --------------------
2624 Constant expressions are used to allow expressions involving other
2625 constants to be used as constants. Constant expressions may be of any
2626 :ref:`first class <t_firstclass>` type and may involve any LLVM operation
2627 that does not have side effects (e.g. load and call are not supported).
2628 The following is the syntax for constant expressions:
2630 ``trunc (CST to TYPE)``
2631 Truncate a constant to another type. The bit size of CST must be
2632 larger than the bit size of TYPE. Both types must be integers.
2633 ``zext (CST to TYPE)``
2634 Zero extend a constant to another type. The bit size of CST must be
2635 smaller than the bit size of TYPE. Both types must be integers.
2636 ``sext (CST to TYPE)``
2637 Sign extend a constant to another type. The bit size of CST must be
2638 smaller than the bit size of TYPE. Both types must be integers.
2639 ``fptrunc (CST to TYPE)``
2640 Truncate a floating point constant to another floating point type.
2641 The size of CST must be larger than the size of TYPE. Both types
2642 must be floating point.
2643 ``fpext (CST to TYPE)``
2644 Floating point extend a constant to another type. The size of CST
2645 must be smaller or equal to the size of TYPE. Both types must be
2647 ``fptoui (CST to TYPE)``
2648 Convert a floating point constant to the corresponding unsigned
2649 integer constant. TYPE must be a scalar or vector integer type. CST
2650 must be of scalar or vector floating point type. Both CST and TYPE
2651 must be scalars, or vectors of the same number of elements. If the
2652 value won't fit in the integer type, the results are undefined.
2653 ``fptosi (CST to TYPE)``
2654 Convert a floating point constant to the corresponding signed
2655 integer constant. TYPE must be a scalar or vector integer type. CST
2656 must be of scalar or vector floating point type. Both CST and TYPE
2657 must be scalars, or vectors of the same number of elements. If the
2658 value won't fit in the integer type, the results are undefined.
2659 ``uitofp (CST to TYPE)``
2660 Convert an unsigned integer constant to the corresponding floating
2661 point constant. TYPE must be a scalar or vector floating point type.
2662 CST must be of scalar or vector integer type. Both CST and TYPE must
2663 be scalars, or vectors of the same number of elements. If the value
2664 won't fit in the floating point type, the results are undefined.
2665 ``sitofp (CST to TYPE)``
2666 Convert a signed integer constant to the corresponding floating
2667 point constant. TYPE must be a scalar or vector floating point type.
2668 CST must be of scalar or vector integer type. Both CST and TYPE must
2669 be scalars, or vectors of the same number of elements. If the value
2670 won't fit in the floating point type, the results are undefined.
2671 ``ptrtoint (CST to TYPE)``
2672 Convert a pointer typed constant to the corresponding integer
2673 constant. ``TYPE`` must be an integer type. ``CST`` must be of
2674 pointer type. The ``CST`` value is zero extended, truncated, or
2675 unchanged to make it fit in ``TYPE``.
2676 ``inttoptr (CST to TYPE)``
2677 Convert an integer constant to a pointer constant. TYPE must be a
2678 pointer type. CST must be of integer type. The CST value is zero
2679 extended, truncated, or unchanged to make it fit in a pointer size.
2680 This one is *really* dangerous!
2681 ``bitcast (CST to TYPE)``
2682 Convert a constant, CST, to another TYPE. The constraints of the
2683 operands are the same as those for the :ref:`bitcast
2684 instruction <i_bitcast>`.
2685 ``addrspacecast (CST to TYPE)``
2686 Convert a constant pointer or constant vector of pointer, CST, to another
2687 TYPE in a different address space. The constraints of the operands are the
2688 same as those for the :ref:`addrspacecast instruction <i_addrspacecast>`.
2689 ``getelementptr (CSTPTR, IDX0, IDX1, ...)``, ``getelementptr inbounds (CSTPTR, IDX0, IDX1, ...)``
2690 Perform the :ref:`getelementptr operation <i_getelementptr>` on
2691 constants. As with the :ref:`getelementptr <i_getelementptr>`
2692 instruction, the index list may have zero or more indexes, which are
2693 required to make sense for the type of "CSTPTR".
2694 ``select (COND, VAL1, VAL2)``
2695 Perform the :ref:`select operation <i_select>` on constants.
2696 ``icmp COND (VAL1, VAL2)``
2697 Performs the :ref:`icmp operation <i_icmp>` on constants.
2698 ``fcmp COND (VAL1, VAL2)``
2699 Performs the :ref:`fcmp operation <i_fcmp>` on constants.
2700 ``extractelement (VAL, IDX)``
2701 Perform the :ref:`extractelement operation <i_extractelement>` on
2703 ``insertelement (VAL, ELT, IDX)``
2704 Perform the :ref:`insertelement operation <i_insertelement>` on
2706 ``shufflevector (VEC1, VEC2, IDXMASK)``
2707 Perform the :ref:`shufflevector operation <i_shufflevector>` on
2709 ``extractvalue (VAL, IDX0, IDX1, ...)``
2710 Perform the :ref:`extractvalue operation <i_extractvalue>` on
2711 constants. The index list is interpreted in a similar manner as
2712 indices in a ':ref:`getelementptr <i_getelementptr>`' operation. At
2713 least one index value must be specified.
2714 ``insertvalue (VAL, ELT, IDX0, IDX1, ...)``
2715 Perform the :ref:`insertvalue operation <i_insertvalue>` on constants.
2716 The index list is interpreted in a similar manner as indices in a
2717 ':ref:`getelementptr <i_getelementptr>`' operation. At least one index
2718 value must be specified.
2719 ``OPCODE (LHS, RHS)``
2720 Perform the specified operation of the LHS and RHS constants. OPCODE
2721 may be any of the :ref:`binary <binaryops>` or :ref:`bitwise
2722 binary <bitwiseops>` operations. The constraints on operands are
2723 the same as those for the corresponding instruction (e.g. no bitwise
2724 operations on floating point values are allowed).
2731 Inline Assembler Expressions
2732 ----------------------------
2734 LLVM supports inline assembler expressions (as opposed to :ref:`Module-Level
2735 Inline Assembly <moduleasm>`) through the use of a special value. This
2736 value represents the inline assembler as a string (containing the
2737 instructions to emit), a list of operand constraints (stored as a
2738 string), a flag that indicates whether or not the inline asm expression
2739 has side effects, and a flag indicating whether the function containing
2740 the asm needs to align its stack conservatively. An example inline
2741 assembler expression is:
2743 .. code-block:: llvm
2745 i32 (i32) asm "bswap $0", "=r,r"
2747 Inline assembler expressions may **only** be used as the callee operand
2748 of a :ref:`call <i_call>` or an :ref:`invoke <i_invoke>` instruction.
2749 Thus, typically we have:
2751 .. code-block:: llvm
2753 %X = call i32 asm "bswap $0", "=r,r"(i32 %Y)
2755 Inline asms with side effects not visible in the constraint list must be
2756 marked as having side effects. This is done through the use of the
2757 '``sideeffect``' keyword, like so:
2759 .. code-block:: llvm
2761 call void asm sideeffect "eieio", ""()
2763 In some cases inline asms will contain code that will not work unless
2764 the stack is aligned in some way, such as calls or SSE instructions on
2765 x86, yet will not contain code that does that alignment within the asm.
2766 The compiler should make conservative assumptions about what the asm
2767 might contain and should generate its usual stack alignment code in the
2768 prologue if the '``alignstack``' keyword is present:
2770 .. code-block:: llvm
2772 call void asm alignstack "eieio", ""()
2774 Inline asms also support using non-standard assembly dialects. The
2775 assumed dialect is ATT. When the '``inteldialect``' keyword is present,
2776 the inline asm is using the Intel dialect. Currently, ATT and Intel are
2777 the only supported dialects. An example is:
2779 .. code-block:: llvm
2781 call void asm inteldialect "eieio", ""()
2783 If multiple keywords appear the '``sideeffect``' keyword must come
2784 first, the '``alignstack``' keyword second and the '``inteldialect``'
2790 The call instructions that wrap inline asm nodes may have a
2791 "``!srcloc``" MDNode attached to it that contains a list of constant
2792 integers. If present, the code generator will use the integer as the
2793 location cookie value when report errors through the ``LLVMContext``
2794 error reporting mechanisms. This allows a front-end to correlate backend
2795 errors that occur with inline asm back to the source code that produced
2798 .. code-block:: llvm
2800 call void asm sideeffect "something bad", ""(), !srcloc !42
2802 !42 = !{ i32 1234567 }
2804 It is up to the front-end to make sense of the magic numbers it places
2805 in the IR. If the MDNode contains multiple constants, the code generator
2806 will use the one that corresponds to the line of the asm that the error
2814 LLVM IR allows metadata to be attached to instructions in the program
2815 that can convey extra information about the code to the optimizers and
2816 code generator. One example application of metadata is source-level
2817 debug information. There are two metadata primitives: strings and nodes.
2819 Metadata does not have a type, and is not a value. If referenced from a
2820 ``call`` instruction, it uses the ``metadata`` type.
2822 All metadata are identified in syntax by a exclamation point ('``!``').
2824 Metadata Nodes and Metadata Strings
2825 -----------------------------------
2827 A metadata string is a string surrounded by double quotes. It can
2828 contain any character by escaping non-printable characters with
2829 "``\xx``" where "``xx``" is the two digit hex code. For example:
2832 Metadata nodes are represented with notation similar to structure
2833 constants (a comma separated list of elements, surrounded by braces and
2834 preceded by an exclamation point). Metadata nodes can have any values as
2835 their operand. For example:
2837 .. code-block:: llvm
2839 !{ !"test\00", i32 10}
2841 A :ref:`named metadata <namedmetadatastructure>` is a collection of
2842 metadata nodes, which can be looked up in the module symbol table. For
2845 .. code-block:: llvm
2849 Metadata can be used as function arguments. Here ``llvm.dbg.value``
2850 function is using two metadata arguments:
2852 .. code-block:: llvm
2854 call void @llvm.dbg.value(metadata !24, i64 0, metadata !25)
2856 Metadata can be attached with an instruction. Here metadata ``!21`` is
2857 attached to the ``add`` instruction using the ``!dbg`` identifier:
2859 .. code-block:: llvm
2861 %indvar.next = add i64 %indvar, 1, !dbg !21
2863 More information about specific metadata nodes recognized by the
2864 optimizers and code generator is found below.
2869 In LLVM IR, memory does not have types, so LLVM's own type system is not
2870 suitable for doing TBAA. Instead, metadata is added to the IR to
2871 describe a type system of a higher level language. This can be used to
2872 implement typical C/C++ TBAA, but it can also be used to implement
2873 custom alias analysis behavior for other languages.
2875 The current metadata format is very simple. TBAA metadata nodes have up
2876 to three fields, e.g.:
2878 .. code-block:: llvm
2880 !0 = !{ !"an example type tree" }
2881 !1 = !{ !"int", !0 }
2882 !2 = !{ !"float", !0 }
2883 !3 = !{ !"const float", !2, i64 1 }
2885 The first field is an identity field. It can be any value, usually a
2886 metadata string, which uniquely identifies the type. The most important
2887 name in the tree is the name of the root node. Two trees with different
2888 root node names are entirely disjoint, even if they have leaves with
2891 The second field identifies the type's parent node in the tree, or is
2892 null or omitted for a root node. A type is considered to alias all of
2893 its descendants and all of its ancestors in the tree. Also, a type is
2894 considered to alias all types in other trees, so that bitcode produced
2895 from multiple front-ends is handled conservatively.
2897 If the third field is present, it's an integer which if equal to 1
2898 indicates that the type is "constant" (meaning
2899 ``pointsToConstantMemory`` should return true; see `other useful
2900 AliasAnalysis methods <AliasAnalysis.html#OtherItfs>`_).
2902 '``tbaa.struct``' Metadata
2903 ^^^^^^^^^^^^^^^^^^^^^^^^^^
2905 The :ref:`llvm.memcpy <int_memcpy>` is often used to implement
2906 aggregate assignment operations in C and similar languages, however it
2907 is defined to copy a contiguous region of memory, which is more than
2908 strictly necessary for aggregate types which contain holes due to
2909 padding. Also, it doesn't contain any TBAA information about the fields
2912 ``!tbaa.struct`` metadata can describe which memory subregions in a
2913 memcpy are padding and what the TBAA tags of the struct are.
2915 The current metadata format is very simple. ``!tbaa.struct`` metadata
2916 nodes are a list of operands which are in conceptual groups of three.
2917 For each group of three, the first operand gives the byte offset of a
2918 field in bytes, the second gives its size in bytes, and the third gives
2921 .. code-block:: llvm
2923 !4 = !{ i64 0, i64 4, !1, i64 8, i64 4, !2 }
2925 This describes a struct with two fields. The first is at offset 0 bytes
2926 with size 4 bytes, and has tbaa tag !1. The second is at offset 8 bytes
2927 and has size 4 bytes and has tbaa tag !2.
2929 Note that the fields need not be contiguous. In this example, there is a
2930 4 byte gap between the two fields. This gap represents padding which
2931 does not carry useful data and need not be preserved.
2933 '``noalias``' and '``alias.scope``' Metadata
2934 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2936 ``noalias`` and ``alias.scope`` metadata provide the ability to specify generic
2937 noalias memory-access sets. This means that some collection of memory access
2938 instructions (loads, stores, memory-accessing calls, etc.) that carry
2939 ``noalias`` metadata can specifically be specified not to alias with some other
2940 collection of memory access instructions that carry ``alias.scope`` metadata.
2941 Each type of metadata specifies a list of scopes where each scope has an id and
2942 a domain. When evaluating an aliasing query, if for some some domain, the set
2943 of scopes with that domain in one instruction's ``alias.scope`` list is a
2944 subset of (or qual to) the set of scopes for that domain in another
2945 instruction's ``noalias`` list, then the two memory accesses are assumed not to
2948 The metadata identifying each domain is itself a list containing one or two
2949 entries. The first entry is the name of the domain. Note that if the name is a
2950 string then it can be combined accross functions and translation units. A
2951 self-reference can be used to create globally unique domain names. A
2952 descriptive string may optionally be provided as a second list entry.
2954 The metadata identifying each scope is also itself a list containing two or
2955 three entries. The first entry is the name of the scope. Note that if the name
2956 is a string then it can be combined accross functions and translation units. A
2957 self-reference can be used to create globally unique scope names. A metadata
2958 reference to the scope's domain is the second entry. A descriptive string may
2959 optionally be provided as a third list entry.
2963 .. code-block:: llvm
2965 ; Two scope domains:
2969 ; Some scopes in these domains:
2975 !5 = !{!4} ; A list containing only scope !4
2979 ; These two instructions don't alias:
2980 %0 = load float* %c, align 4, !alias.scope !5
2981 store float %0, float* %arrayidx.i, align 4, !noalias !5
2983 ; These two instructions also don't alias (for domain !1, the set of scopes
2984 ; in the !alias.scope equals that in the !noalias list):
2985 %2 = load float* %c, align 4, !alias.scope !5
2986 store float %2, float* %arrayidx.i2, align 4, !noalias !6
2988 ; These two instructions don't alias (for domain !0, the set of scopes in
2989 ; the !noalias list is not a superset of, or equal to, the scopes in the
2990 ; !alias.scope list):
2991 %2 = load float* %c, align 4, !alias.scope !6
2992 store float %0, float* %arrayidx.i, align 4, !noalias !7
2994 '``fpmath``' Metadata
2995 ^^^^^^^^^^^^^^^^^^^^^
2997 ``fpmath`` metadata may be attached to any instruction of floating point
2998 type. It can be used to express the maximum acceptable error in the
2999 result of that instruction, in ULPs, thus potentially allowing the
3000 compiler to use a more efficient but less accurate method of computing
3001 it. ULP is defined as follows:
3003 If ``x`` is a real number that lies between two finite consecutive
3004 floating-point numbers ``a`` and ``b``, without being equal to one
3005 of them, then ``ulp(x) = |b - a|``, otherwise ``ulp(x)`` is the
3006 distance between the two non-equal finite floating-point numbers
3007 nearest ``x``. Moreover, ``ulp(NaN)`` is ``NaN``.
3009 The metadata node shall consist of a single positive floating point
3010 number representing the maximum relative error, for example:
3012 .. code-block:: llvm
3014 !0 = !{ float 2.5 } ; maximum acceptable inaccuracy is 2.5 ULPs
3016 '``range``' Metadata
3017 ^^^^^^^^^^^^^^^^^^^^
3019 ``range`` metadata may be attached only to ``load``, ``call`` and ``invoke`` of
3020 integer types. It expresses the possible ranges the loaded value or the value
3021 returned by the called function at this call site is in. The ranges are
3022 represented with a flattened list of integers. The loaded value or the value
3023 returned is known to be in the union of the ranges defined by each consecutive
3024 pair. Each pair has the following properties:
3026 - The type must match the type loaded by the instruction.
3027 - The pair ``a,b`` represents the range ``[a,b)``.
3028 - Both ``a`` and ``b`` are constants.
3029 - The range is allowed to wrap.
3030 - The range should not represent the full or empty set. That is,
3033 In addition, the pairs must be in signed order of the lower bound and
3034 they must be non-contiguous.
3038 .. code-block:: llvm
3040 %a = load i8* %x, align 1, !range !0 ; Can only be 0 or 1
3041 %b = load i8* %y, align 1, !range !1 ; Can only be 255 (-1), 0 or 1
3042 %c = call i8 @foo(), !range !2 ; Can only be 0, 1, 3, 4 or 5
3043 %d = invoke i8 @bar() to label %cont
3044 unwind label %lpad, !range !3 ; Can only be -2, -1, 3, 4 or 5
3046 !0 = !{ i8 0, i8 2 }
3047 !1 = !{ i8 255, i8 2 }
3048 !2 = !{ i8 0, i8 2, i8 3, i8 6 }
3049 !3 = !{ i8 -2, i8 0, i8 3, i8 6 }
3054 It is sometimes useful to attach information to loop constructs. Currently,
3055 loop metadata is implemented as metadata attached to the branch instruction
3056 in the loop latch block. This type of metadata refer to a metadata node that is
3057 guaranteed to be separate for each loop. The loop identifier metadata is
3058 specified with the name ``llvm.loop``.
3060 The loop identifier metadata is implemented using a metadata that refers to
3061 itself to avoid merging it with any other identifier metadata, e.g.,
3062 during module linkage or function inlining. That is, each loop should refer
3063 to their own identification metadata even if they reside in separate functions.
3064 The following example contains loop identifier metadata for two separate loop
3067 .. code-block:: llvm
3072 The loop identifier metadata can be used to specify additional
3073 per-loop metadata. Any operands after the first operand can be treated
3074 as user-defined metadata. For example the ``llvm.loop.unroll.count``
3075 suggests an unroll factor to the loop unroller:
3077 .. code-block:: llvm
3079 br i1 %exitcond, label %._crit_edge, label %.lr.ph, !llvm.loop !0
3082 !1 = !{!"llvm.loop.unroll.count", i32 4}
3084 '``llvm.loop.vectorize``' and '``llvm.loop.interleave``'
3085 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3087 Metadata prefixed with ``llvm.loop.vectorize`` or ``llvm.loop.interleave`` are
3088 used to control per-loop vectorization and interleaving parameters such as
3089 vectorization width and interleave count. These metadata should be used in
3090 conjunction with ``llvm.loop`` loop identification metadata. The
3091 ``llvm.loop.vectorize`` and ``llvm.loop.interleave`` metadata are only
3092 optimization hints and the optimizer will only interleave and vectorize loops if
3093 it believes it is safe to do so. The ``llvm.mem.parallel_loop_access`` metadata
3094 which contains information about loop-carried memory dependencies can be helpful
3095 in determining the safety of these transformations.
3097 '``llvm.loop.interleave.count``' Metadata
3098 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3100 This metadata suggests an interleave count to the loop interleaver.
3101 The first operand is the string ``llvm.loop.interleave.count`` and the
3102 second operand is an integer specifying the interleave count. For
3105 .. code-block:: llvm
3107 !0 = !{!"llvm.loop.interleave.count", i32 4}
3109 Note that setting ``llvm.loop.interleave.count`` to 1 disables interleaving
3110 multiple iterations of the loop. If ``llvm.loop.interleave.count`` is set to 0
3111 then the interleave count will be determined automatically.
3113 '``llvm.loop.vectorize.enable``' Metadata
3114 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3116 This metadata selectively enables or disables vectorization for the loop. The
3117 first operand is the string ``llvm.loop.vectorize.enable`` and the second operand
3118 is a bit. If the bit operand value is 1 vectorization is enabled. A value of
3119 0 disables vectorization:
3121 .. code-block:: llvm
3123 !0 = !{!"llvm.loop.vectorize.enable", i1 0}
3124 !1 = !{!"llvm.loop.vectorize.enable", i1 1}
3126 '``llvm.loop.vectorize.width``' Metadata
3127 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3129 This metadata sets the target width of the vectorizer. The first
3130 operand is the string ``llvm.loop.vectorize.width`` and the second
3131 operand is an integer specifying the width. For example:
3133 .. code-block:: llvm
3135 !0 = !{!"llvm.loop.vectorize.width", i32 4}
3137 Note that setting ``llvm.loop.vectorize.width`` to 1 disables
3138 vectorization of the loop. If ``llvm.loop.vectorize.width`` is set to
3139 0 or if the loop does not have this metadata the width will be
3140 determined automatically.
3142 '``llvm.loop.unroll``'
3143 ^^^^^^^^^^^^^^^^^^^^^^
3145 Metadata prefixed with ``llvm.loop.unroll`` are loop unrolling
3146 optimization hints such as the unroll factor. ``llvm.loop.unroll``
3147 metadata should be used in conjunction with ``llvm.loop`` loop
3148 identification metadata. The ``llvm.loop.unroll`` metadata are only
3149 optimization hints and the unrolling will only be performed if the
3150 optimizer believes it is safe to do so.
3152 '``llvm.loop.unroll.count``' Metadata
3153 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3155 This metadata suggests an unroll factor to the loop unroller. The
3156 first operand is the string ``llvm.loop.unroll.count`` and the second
3157 operand is a positive integer specifying the unroll factor. For
3160 .. code-block:: llvm
3162 !0 = !{!"llvm.loop.unroll.count", i32 4}
3164 If the trip count of the loop is less than the unroll count the loop
3165 will be partially unrolled.
3167 '``llvm.loop.unroll.disable``' Metadata
3168 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3170 This metadata either disables loop unrolling. The metadata has a single operand
3171 which is the string ``llvm.loop.unroll.disable``. For example:
3173 .. code-block:: llvm
3175 !0 = !{!"llvm.loop.unroll.disable"}
3177 '``llvm.loop.unroll.full``' Metadata
3178 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3180 This metadata either suggests that the loop should be unrolled fully. The
3181 metadata has a single operand which is the string ``llvm.loop.unroll.disable``.
3184 .. code-block:: llvm
3186 !0 = !{!"llvm.loop.unroll.full"}
3191 Metadata types used to annotate memory accesses with information helpful
3192 for optimizations are prefixed with ``llvm.mem``.
3194 '``llvm.mem.parallel_loop_access``' Metadata
3195 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3197 The ``llvm.mem.parallel_loop_access`` metadata refers to a loop identifier,
3198 or metadata containing a list of loop identifiers for nested loops.
3199 The metadata is attached to memory accessing instructions and denotes that
3200 no loop carried memory dependence exist between it and other instructions denoted
3201 with the same loop identifier.
3203 Precisely, given two instructions ``m1`` and ``m2`` that both have the
3204 ``llvm.mem.parallel_loop_access`` metadata, with ``L1`` and ``L2`` being the
3205 set of loops associated with that metadata, respectively, then there is no loop
3206 carried dependence between ``m1`` and ``m2`` for loops in both ``L1`` and
3209 As a special case, if all memory accessing instructions in a loop have
3210 ``llvm.mem.parallel_loop_access`` metadata that refers to that loop, then the
3211 loop has no loop carried memory dependences and is considered to be a parallel
3214 Note that if not all memory access instructions have such metadata referring to
3215 the loop, then the loop is considered not being trivially parallel. Additional
3216 memory dependence analysis is required to make that determination. As a fail
3217 safe mechanism, this causes loops that were originally parallel to be considered
3218 sequential (if optimization passes that are unaware of the parallel semantics
3219 insert new memory instructions into the loop body).
3221 Example of a loop that is considered parallel due to its correct use of
3222 both ``llvm.loop`` and ``llvm.mem.parallel_loop_access``
3223 metadata types that refer to the same loop identifier metadata.
3225 .. code-block:: llvm
3229 %val0 = load i32* %arrayidx, !llvm.mem.parallel_loop_access !0
3231 store i32 %val0, i32* %arrayidx1, !llvm.mem.parallel_loop_access !0
3233 br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !0
3239 It is also possible to have nested parallel loops. In that case the
3240 memory accesses refer to a list of loop identifier metadata nodes instead of
3241 the loop identifier metadata node directly:
3243 .. code-block:: llvm
3247 %val1 = load i32* %arrayidx3, !llvm.mem.parallel_loop_access !2
3249 br label %inner.for.body
3253 %val0 = load i32* %arrayidx1, !llvm.mem.parallel_loop_access !0
3255 store i32 %val0, i32* %arrayidx2, !llvm.mem.parallel_loop_access !0
3257 br i1 %exitcond, label %inner.for.end, label %inner.for.body, !llvm.loop !1
3261 store i32 %val1, i32* %arrayidx4, !llvm.mem.parallel_loop_access !2
3263 br i1 %exitcond, label %outer.for.end, label %outer.for.body, !llvm.loop !2
3265 outer.for.end: ; preds = %for.body
3267 !0 = !{!1, !2} ; a list of loop identifiers
3268 !1 = !{!1} ; an identifier for the inner loop
3269 !2 = !{!2} ; an identifier for the outer loop
3271 Module Flags Metadata
3272 =====================
3274 Information about the module as a whole is difficult to convey to LLVM's
3275 subsystems. The LLVM IR isn't sufficient to transmit this information.
3276 The ``llvm.module.flags`` named metadata exists in order to facilitate
3277 this. These flags are in the form of key / value pairs --- much like a
3278 dictionary --- making it easy for any subsystem who cares about a flag to
3281 The ``llvm.module.flags`` metadata contains a list of metadata triplets.
3282 Each triplet has the following form:
3284 - The first element is a *behavior* flag, which specifies the behavior
3285 when two (or more) modules are merged together, and it encounters two
3286 (or more) metadata with the same ID. The supported behaviors are
3288 - The second element is a metadata string that is a unique ID for the
3289 metadata. Each module may only have one flag entry for each unique ID (not
3290 including entries with the **Require** behavior).
3291 - The third element is the value of the flag.
3293 When two (or more) modules are merged together, the resulting
3294 ``llvm.module.flags`` metadata is the union of the modules' flags. That is, for
3295 each unique metadata ID string, there will be exactly one entry in the merged
3296 modules ``llvm.module.flags`` metadata table, and the value for that entry will
3297 be determined by the merge behavior flag, as described below. The only exception
3298 is that entries with the *Require* behavior are always preserved.
3300 The following behaviors are supported:
3311 Emits an error if two values disagree, otherwise the resulting value
3312 is that of the operands.
3316 Emits a warning if two values disagree. The result value will be the
3317 operand for the flag from the first module being linked.
3321 Adds a requirement that another module flag be present and have a
3322 specified value after linking is performed. The value must be a
3323 metadata pair, where the first element of the pair is the ID of the
3324 module flag to be restricted, and the second element of the pair is
3325 the value the module flag should be restricted to. This behavior can
3326 be used to restrict the allowable results (via triggering of an
3327 error) of linking IDs with the **Override** behavior.
3331 Uses the specified value, regardless of the behavior or value of the
3332 other module. If both modules specify **Override**, but the values
3333 differ, an error will be emitted.
3337 Appends the two values, which are required to be metadata nodes.
3341 Appends the two values, which are required to be metadata
3342 nodes. However, duplicate entries in the second list are dropped
3343 during the append operation.
3345 It is an error for a particular unique flag ID to have multiple behaviors,
3346 except in the case of **Require** (which adds restrictions on another metadata
3347 value) or **Override**.
3349 An example of module flags:
3351 .. code-block:: llvm
3353 !0 = !{ i32 1, !"foo", i32 1 }
3354 !1 = !{ i32 4, !"bar", i32 37 }
3355 !2 = !{ i32 2, !"qux", i32 42 }
3356 !3 = !{ i32 3, !"qux",
3361 !llvm.module.flags = !{ !0, !1, !2, !3 }
3363 - Metadata ``!0`` has the ID ``!"foo"`` and the value '1'. The behavior
3364 if two or more ``!"foo"`` flags are seen is to emit an error if their
3365 values are not equal.
3367 - Metadata ``!1`` has the ID ``!"bar"`` and the value '37'. The
3368 behavior if two or more ``!"bar"`` flags are seen is to use the value
3371 - Metadata ``!2`` has the ID ``!"qux"`` and the value '42'. The
3372 behavior if two or more ``!"qux"`` flags are seen is to emit a
3373 warning if their values are not equal.
3375 - Metadata ``!3`` has the ID ``!"qux"`` and the value:
3381 The behavior is to emit an error if the ``llvm.module.flags`` does not
3382 contain a flag with the ID ``!"foo"`` that has the value '1' after linking is
3385 Objective-C Garbage Collection Module Flags Metadata
3386 ----------------------------------------------------
3388 On the Mach-O platform, Objective-C stores metadata about garbage
3389 collection in a special section called "image info". The metadata
3390 consists of a version number and a bitmask specifying what types of
3391 garbage collection are supported (if any) by the file. If two or more
3392 modules are linked together their garbage collection metadata needs to
3393 be merged rather than appended together.
3395 The Objective-C garbage collection module flags metadata consists of the
3396 following key-value pairs:
3405 * - ``Objective-C Version``
3406 - **[Required]** --- The Objective-C ABI version. Valid values are 1 and 2.
3408 * - ``Objective-C Image Info Version``
3409 - **[Required]** --- The version of the image info section. Currently
3412 * - ``Objective-C Image Info Section``
3413 - **[Required]** --- The section to place the metadata. Valid values are
3414 ``"__OBJC, __image_info, regular"`` for Objective-C ABI version 1, and
3415 ``"__DATA,__objc_imageinfo, regular, no_dead_strip"`` for
3416 Objective-C ABI version 2.
3418 * - ``Objective-C Garbage Collection``
3419 - **[Required]** --- Specifies whether garbage collection is supported or
3420 not. Valid values are 0, for no garbage collection, and 2, for garbage
3421 collection supported.
3423 * - ``Objective-C GC Only``
3424 - **[Optional]** --- Specifies that only garbage collection is supported.
3425 If present, its value must be 6. This flag requires that the
3426 ``Objective-C Garbage Collection`` flag have the value 2.
3428 Some important flag interactions:
3430 - If a module with ``Objective-C Garbage Collection`` set to 0 is
3431 merged with a module with ``Objective-C Garbage Collection`` set to
3432 2, then the resulting module has the
3433 ``Objective-C Garbage Collection`` flag set to 0.
3434 - A module with ``Objective-C Garbage Collection`` set to 0 cannot be
3435 merged with a module with ``Objective-C GC Only`` set to 6.
3437 Automatic Linker Flags Module Flags Metadata
3438 --------------------------------------------
3440 Some targets support embedding flags to the linker inside individual object
3441 files. Typically this is used in conjunction with language extensions which
3442 allow source files to explicitly declare the libraries they depend on, and have
3443 these automatically be transmitted to the linker via object files.
3445 These flags are encoded in the IR using metadata in the module flags section,
3446 using the ``Linker Options`` key. The merge behavior for this flag is required
3447 to be ``AppendUnique``, and the value for the key is expected to be a metadata
3448 node which should be a list of other metadata nodes, each of which should be a
3449 list of metadata strings defining linker options.
3451 For example, the following metadata section specifies two separate sets of
3452 linker options, presumably to link against ``libz`` and the ``Cocoa``
3455 !0 = !{ i32 6, !"Linker Options",
3458 !{ !"-framework", !"Cocoa" } } }
3459 !llvm.module.flags = !{ !0 }
3461 The metadata encoding as lists of lists of options, as opposed to a collapsed
3462 list of options, is chosen so that the IR encoding can use multiple option
3463 strings to specify e.g., a single library, while still having that specifier be
3464 preserved as an atomic element that can be recognized by a target specific
3465 assembly writer or object file emitter.
3467 Each individual option is required to be either a valid option for the target's
3468 linker, or an option that is reserved by the target specific assembly writer or
3469 object file emitter. No other aspect of these options is defined by the IR.
3471 C type width Module Flags Metadata
3472 ----------------------------------
3474 The ARM backend emits a section into each generated object file describing the
3475 options that it was compiled with (in a compiler-independent way) to prevent
3476 linking incompatible objects, and to allow automatic library selection. Some
3477 of these options are not visible at the IR level, namely wchar_t width and enum
3480 To pass this information to the backend, these options are encoded in module
3481 flags metadata, using the following key-value pairs:
3491 - * 0 --- sizeof(wchar_t) == 4
3492 * 1 --- sizeof(wchar_t) == 2
3495 - * 0 --- Enums are at least as large as an ``int``.
3496 * 1 --- Enums are stored in the smallest integer type which can
3497 represent all of its values.
3499 For example, the following metadata section specifies that the module was
3500 compiled with a ``wchar_t`` width of 4 bytes, and the underlying type of an
3501 enum is the smallest type which can represent all of its values::
3503 !llvm.module.flags = !{!0, !1}
3504 !0 = !{i32 1, !"short_wchar", i32 1}
3505 !1 = !{i32 1, !"short_enum", i32 0}
3507 .. _intrinsicglobalvariables:
3509 Intrinsic Global Variables
3510 ==========================
3512 LLVM has a number of "magic" global variables that contain data that
3513 affect code generation or other IR semantics. These are documented here.
3514 All globals of this sort should have a section specified as
3515 "``llvm.metadata``". This section and all globals that start with
3516 "``llvm.``" are reserved for use by LLVM.
3520 The '``llvm.used``' Global Variable
3521 -----------------------------------
3523 The ``@llvm.used`` global is an array which has
3524 :ref:`appending linkage <linkage_appending>`. This array contains a list of
3525 pointers to named global variables, functions and aliases which may optionally
3526 have a pointer cast formed of bitcast or getelementptr. For example, a legal
3529 .. code-block:: llvm
3534 @llvm.used = appending global [2 x i8*] [
3536 i8* bitcast (i32* @Y to i8*)
3537 ], section "llvm.metadata"
3539 If a symbol appears in the ``@llvm.used`` list, then the compiler, assembler,
3540 and linker are required to treat the symbol as if there is a reference to the
3541 symbol that it cannot see (which is why they have to be named). For example, if
3542 a variable has internal linkage and no references other than that from the
3543 ``@llvm.used`` list, it cannot be deleted. This is commonly used to represent
3544 references from inline asms and other things the compiler cannot "see", and
3545 corresponds to "``attribute((used))``" in GNU C.
3547 On some targets, the code generator must emit a directive to the
3548 assembler or object file to prevent the assembler and linker from
3549 molesting the symbol.
3551 .. _gv_llvmcompilerused:
3553 The '``llvm.compiler.used``' Global Variable
3554 --------------------------------------------
3556 The ``@llvm.compiler.used`` directive is the same as the ``@llvm.used``
3557 directive, except that it only prevents the compiler from touching the
3558 symbol. On targets that support it, this allows an intelligent linker to
3559 optimize references to the symbol without being impeded as it would be
3562 This is a rare construct that should only be used in rare circumstances,
3563 and should not be exposed to source languages.
3565 .. _gv_llvmglobalctors:
3567 The '``llvm.global_ctors``' Global Variable
3568 -------------------------------------------
3570 .. code-block:: llvm
3572 %0 = type { i32, void ()*, i8* }
3573 @llvm.global_ctors = appending global [1 x %0] [%0 { i32 65535, void ()* @ctor, i8* @data }]
3575 The ``@llvm.global_ctors`` array contains a list of constructor
3576 functions, priorities, and an optional associated global or function.
3577 The functions referenced by this array will be called in ascending order
3578 of priority (i.e. lowest first) when the module is loaded. The order of
3579 functions with the same priority is not defined.
3581 If the third field is present, non-null, and points to a global variable
3582 or function, the initializer function will only run if the associated
3583 data from the current module is not discarded.
3585 .. _llvmglobaldtors:
3587 The '``llvm.global_dtors``' Global Variable
3588 -------------------------------------------
3590 .. code-block:: llvm
3592 %0 = type { i32, void ()*, i8* }
3593 @llvm.global_dtors = appending global [1 x %0] [%0 { i32 65535, void ()* @dtor, i8* @data }]
3595 The ``@llvm.global_dtors`` array contains a list of destructor
3596 functions, priorities, and an optional associated global or function.
3597 The functions referenced by this array will be called in descending
3598 order of priority (i.e. highest first) when the module is unloaded. The
3599 order of functions with the same priority is not defined.
3601 If the third field is present, non-null, and points to a global variable
3602 or function, the destructor function will only run if the associated
3603 data from the current module is not discarded.
3605 Instruction Reference
3606 =====================
3608 The LLVM instruction set consists of several different classifications
3609 of instructions: :ref:`terminator instructions <terminators>`, :ref:`binary
3610 instructions <binaryops>`, :ref:`bitwise binary
3611 instructions <bitwiseops>`, :ref:`memory instructions <memoryops>`, and
3612 :ref:`other instructions <otherops>`.
3616 Terminator Instructions
3617 -----------------------
3619 As mentioned :ref:`previously <functionstructure>`, every basic block in a
3620 program ends with a "Terminator" instruction, which indicates which
3621 block should be executed after the current block is finished. These
3622 terminator instructions typically yield a '``void``' value: they produce
3623 control flow, not values (the one exception being the
3624 ':ref:`invoke <i_invoke>`' instruction).
3626 The terminator instructions are: ':ref:`ret <i_ret>`',
3627 ':ref:`br <i_br>`', ':ref:`switch <i_switch>`',
3628 ':ref:`indirectbr <i_indirectbr>`', ':ref:`invoke <i_invoke>`',
3629 ':ref:`resume <i_resume>`', and ':ref:`unreachable <i_unreachable>`'.
3633 '``ret``' Instruction
3634 ^^^^^^^^^^^^^^^^^^^^^
3641 ret <type> <value> ; Return a value from a non-void function
3642 ret void ; Return from void function
3647 The '``ret``' instruction is used to return control flow (and optionally
3648 a value) from a function back to the caller.
3650 There are two forms of the '``ret``' instruction: one that returns a
3651 value and then causes control flow, and one that just causes control
3657 The '``ret``' instruction optionally accepts a single argument, the
3658 return value. The type of the return value must be a ':ref:`first
3659 class <t_firstclass>`' type.
3661 A function is not :ref:`well formed <wellformed>` if it it has a non-void
3662 return type and contains a '``ret``' instruction with no return value or
3663 a return value with a type that does not match its type, or if it has a
3664 void return type and contains a '``ret``' instruction with a return
3670 When the '``ret``' instruction is executed, control flow returns back to
3671 the calling function's context. If the caller is a
3672 ":ref:`call <i_call>`" instruction, execution continues at the
3673 instruction after the call. If the caller was an
3674 ":ref:`invoke <i_invoke>`" instruction, execution continues at the
3675 beginning of the "normal" destination block. If the instruction returns
3676 a value, that value shall set the call or invoke instruction's return
3682 .. code-block:: llvm
3684 ret i32 5 ; Return an integer value of 5
3685 ret void ; Return from a void function
3686 ret { i32, i8 } { i32 4, i8 2 } ; Return a struct of values 4 and 2
3690 '``br``' Instruction
3691 ^^^^^^^^^^^^^^^^^^^^
3698 br i1 <cond>, label <iftrue>, label <iffalse>
3699 br label <dest> ; Unconditional branch
3704 The '``br``' instruction is used to cause control flow to transfer to a
3705 different basic block in the current function. There are two forms of
3706 this instruction, corresponding to a conditional branch and an
3707 unconditional branch.
3712 The conditional branch form of the '``br``' instruction takes a single
3713 '``i1``' value and two '``label``' values. The unconditional form of the
3714 '``br``' instruction takes a single '``label``' value as a target.
3719 Upon execution of a conditional '``br``' instruction, the '``i1``'
3720 argument is evaluated. If the value is ``true``, control flows to the
3721 '``iftrue``' ``label`` argument. If "cond" is ``false``, control flows
3722 to the '``iffalse``' ``label`` argument.
3727 .. code-block:: llvm
3730 %cond = icmp eq i32 %a, %b
3731 br i1 %cond, label %IfEqual, label %IfUnequal
3739 '``switch``' Instruction
3740 ^^^^^^^^^^^^^^^^^^^^^^^^
3747 switch <intty> <value>, label <defaultdest> [ <intty> <val>, label <dest> ... ]
3752 The '``switch``' instruction is used to transfer control flow to one of
3753 several different places. It is a generalization of the '``br``'
3754 instruction, allowing a branch to occur to one of many possible
3760 The '``switch``' instruction uses three parameters: an integer
3761 comparison value '``value``', a default '``label``' destination, and an
3762 array of pairs of comparison value constants and '``label``'s. The table
3763 is not allowed to contain duplicate constant entries.
3768 The ``switch`` instruction specifies a table of values and destinations.
3769 When the '``switch``' instruction is executed, this table is searched
3770 for the given value. If the value is found, control flow is transferred
3771 to the corresponding destination; otherwise, control flow is transferred
3772 to the default destination.
3777 Depending on properties of the target machine and the particular
3778 ``switch`` instruction, this instruction may be code generated in
3779 different ways. For example, it could be generated as a series of
3780 chained conditional branches or with a lookup table.
3785 .. code-block:: llvm
3787 ; Emulate a conditional br instruction
3788 %Val = zext i1 %value to i32
3789 switch i32 %Val, label %truedest [ i32 0, label %falsedest ]
3791 ; Emulate an unconditional br instruction
3792 switch i32 0, label %dest [ ]
3794 ; Implement a jump table:
3795 switch i32 %val, label %otherwise [ i32 0, label %onzero
3797 i32 2, label %ontwo ]
3801 '``indirectbr``' Instruction
3802 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3809 indirectbr <somety>* <address>, [ label <dest1>, label <dest2>, ... ]
3814 The '``indirectbr``' instruction implements an indirect branch to a
3815 label within the current function, whose address is specified by
3816 "``address``". Address must be derived from a
3817 :ref:`blockaddress <blockaddress>` constant.
3822 The '``address``' argument is the address of the label to jump to. The
3823 rest of the arguments indicate the full set of possible destinations
3824 that the address may point to. Blocks are allowed to occur multiple
3825 times in the destination list, though this isn't particularly useful.
3827 This destination list is required so that dataflow analysis has an
3828 accurate understanding of the CFG.
3833 Control transfers to the block specified in the address argument. All
3834 possible destination blocks must be listed in the label list, otherwise
3835 this instruction has undefined behavior. This implies that jumps to
3836 labels defined in other functions have undefined behavior as well.
3841 This is typically implemented with a jump through a register.
3846 .. code-block:: llvm
3848 indirectbr i8* %Addr, [ label %bb1, label %bb2, label %bb3 ]
3852 '``invoke``' Instruction
3853 ^^^^^^^^^^^^^^^^^^^^^^^^
3860 <result> = invoke [cconv] [ret attrs] <ptr to function ty> <function ptr val>(<function args>) [fn attrs]
3861 to label <normal label> unwind label <exception label>
3866 The '``invoke``' instruction causes control to transfer to a specified
3867 function, with the possibility of control flow transfer to either the
3868 '``normal``' label or the '``exception``' label. If the callee function
3869 returns with the "``ret``" instruction, control flow will return to the
3870 "normal" label. If the callee (or any indirect callees) returns via the
3871 ":ref:`resume <i_resume>`" instruction or other exception handling
3872 mechanism, control is interrupted and continued at the dynamically
3873 nearest "exception" label.
3875 The '``exception``' label is a `landing
3876 pad <ExceptionHandling.html#overview>`_ for the exception. As such,
3877 '``exception``' label is required to have the
3878 ":ref:`landingpad <i_landingpad>`" instruction, which contains the
3879 information about the behavior of the program after unwinding happens,
3880 as its first non-PHI instruction. The restrictions on the
3881 "``landingpad``" instruction's tightly couples it to the "``invoke``"
3882 instruction, so that the important information contained within the
3883 "``landingpad``" instruction can't be lost through normal code motion.
3888 This instruction requires several arguments:
3890 #. The optional "cconv" marker indicates which :ref:`calling
3891 convention <callingconv>` the call should use. If none is
3892 specified, the call defaults to using C calling conventions.
3893 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
3894 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
3896 #. '``ptr to function ty``': shall be the signature of the pointer to
3897 function value being invoked. In most cases, this is a direct
3898 function invocation, but indirect ``invoke``'s are just as possible,
3899 branching off an arbitrary pointer to function value.
3900 #. '``function ptr val``': An LLVM value containing a pointer to a
3901 function to be invoked.
3902 #. '``function args``': argument list whose types match the function
3903 signature argument types and parameter attributes. All arguments must
3904 be of :ref:`first class <t_firstclass>` type. If the function signature
3905 indicates the function accepts a variable number of arguments, the
3906 extra arguments can be specified.
3907 #. '``normal label``': the label reached when the called function
3908 executes a '``ret``' instruction.
3909 #. '``exception label``': the label reached when a callee returns via
3910 the :ref:`resume <i_resume>` instruction or other exception handling
3912 #. The optional :ref:`function attributes <fnattrs>` list. Only
3913 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
3914 attributes are valid here.
3919 This instruction is designed to operate as a standard '``call``'
3920 instruction in most regards. The primary difference is that it
3921 establishes an association with a label, which is used by the runtime
3922 library to unwind the stack.
3924 This instruction is used in languages with destructors to ensure that
3925 proper cleanup is performed in the case of either a ``longjmp`` or a
3926 thrown exception. Additionally, this is important for implementation of
3927 '``catch``' clauses in high-level languages that support them.
3929 For the purposes of the SSA form, the definition of the value returned
3930 by the '``invoke``' instruction is deemed to occur on the edge from the
3931 current block to the "normal" label. If the callee unwinds then no
3932 return value is available.
3937 .. code-block:: llvm
3939 %retval = invoke i32 @Test(i32 15) to label %Continue
3940 unwind label %TestCleanup ; i32:retval set
3941 %retval = invoke coldcc i32 %Testfnptr(i32 15) to label %Continue
3942 unwind label %TestCleanup ; i32:retval set
3946 '``resume``' Instruction
3947 ^^^^^^^^^^^^^^^^^^^^^^^^
3954 resume <type> <value>
3959 The '``resume``' instruction is a terminator instruction that has no
3965 The '``resume``' instruction requires one argument, which must have the
3966 same type as the result of any '``landingpad``' instruction in the same
3972 The '``resume``' instruction resumes propagation of an existing
3973 (in-flight) exception whose unwinding was interrupted with a
3974 :ref:`landingpad <i_landingpad>` instruction.
3979 .. code-block:: llvm
3981 resume { i8*, i32 } %exn
3985 '``unreachable``' Instruction
3986 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3998 The '``unreachable``' instruction has no defined semantics. This
3999 instruction is used to inform the optimizer that a particular portion of
4000 the code is not reachable. This can be used to indicate that the code
4001 after a no-return function cannot be reached, and other facts.
4006 The '``unreachable``' instruction has no defined semantics.
4013 Binary operators are used to do most of the computation in a program.
4014 They require two operands of the same type, execute an operation on
4015 them, and produce a single value. The operands might represent multiple
4016 data, as is the case with the :ref:`vector <t_vector>` data type. The
4017 result value has the same type as its operands.
4019 There are several different binary operators:
4023 '``add``' Instruction
4024 ^^^^^^^^^^^^^^^^^^^^^
4031 <result> = add <ty> <op1>, <op2> ; yields ty:result
4032 <result> = add nuw <ty> <op1>, <op2> ; yields ty:result
4033 <result> = add nsw <ty> <op1>, <op2> ; yields ty:result
4034 <result> = add nuw nsw <ty> <op1>, <op2> ; yields ty:result
4039 The '``add``' instruction returns the sum of its two operands.
4044 The two arguments to the '``add``' instruction must be
4045 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4046 arguments must have identical types.
4051 The value produced is the integer sum of the two operands.
4053 If the sum has unsigned overflow, the result returned is the
4054 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
4057 Because LLVM integers use a two's complement representation, this
4058 instruction is appropriate for both signed and unsigned integers.
4060 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
4061 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
4062 result value of the ``add`` is a :ref:`poison value <poisonvalues>` if
4063 unsigned and/or signed overflow, respectively, occurs.
4068 .. code-block:: llvm
4070 <result> = add i32 4, %var ; yields i32:result = 4 + %var
4074 '``fadd``' Instruction
4075 ^^^^^^^^^^^^^^^^^^^^^^
4082 <result> = fadd [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
4087 The '``fadd``' instruction returns the sum of its two operands.
4092 The two arguments to the '``fadd``' instruction must be :ref:`floating
4093 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
4094 Both arguments must have identical types.
4099 The value produced is the floating point sum of the two operands. This
4100 instruction can also take any number of :ref:`fast-math flags <fastmath>`,
4101 which are optimization hints to enable otherwise unsafe floating point
4107 .. code-block:: llvm
4109 <result> = fadd float 4.0, %var ; yields float:result = 4.0 + %var
4111 '``sub``' Instruction
4112 ^^^^^^^^^^^^^^^^^^^^^
4119 <result> = sub <ty> <op1>, <op2> ; yields ty:result
4120 <result> = sub nuw <ty> <op1>, <op2> ; yields ty:result
4121 <result> = sub nsw <ty> <op1>, <op2> ; yields ty:result
4122 <result> = sub nuw nsw <ty> <op1>, <op2> ; yields ty:result
4127 The '``sub``' instruction returns the difference of its two operands.
4129 Note that the '``sub``' instruction is used to represent the '``neg``'
4130 instruction present in most other intermediate representations.
4135 The two arguments to the '``sub``' instruction must be
4136 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4137 arguments must have identical types.
4142 The value produced is the integer difference of the two operands.
4144 If the difference has unsigned overflow, the result returned is the
4145 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
4148 Because LLVM integers use a two's complement representation, this
4149 instruction is appropriate for both signed and unsigned integers.
4151 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
4152 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
4153 result value of the ``sub`` is a :ref:`poison value <poisonvalues>` if
4154 unsigned and/or signed overflow, respectively, occurs.
4159 .. code-block:: llvm
4161 <result> = sub i32 4, %var ; yields i32:result = 4 - %var
4162 <result> = sub i32 0, %val ; yields i32:result = -%var
4166 '``fsub``' Instruction
4167 ^^^^^^^^^^^^^^^^^^^^^^
4174 <result> = fsub [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
4179 The '``fsub``' instruction returns the difference of its two operands.
4181 Note that the '``fsub``' instruction is used to represent the '``fneg``'
4182 instruction present in most other intermediate representations.
4187 The two arguments to the '``fsub``' instruction must be :ref:`floating
4188 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
4189 Both arguments must have identical types.
4194 The value produced is the floating point difference of the two operands.
4195 This instruction can also take any number of :ref:`fast-math
4196 flags <fastmath>`, which are optimization hints to enable otherwise
4197 unsafe floating point optimizations:
4202 .. code-block:: llvm
4204 <result> = fsub float 4.0, %var ; yields float:result = 4.0 - %var
4205 <result> = fsub float -0.0, %val ; yields float:result = -%var
4207 '``mul``' Instruction
4208 ^^^^^^^^^^^^^^^^^^^^^
4215 <result> = mul <ty> <op1>, <op2> ; yields ty:result
4216 <result> = mul nuw <ty> <op1>, <op2> ; yields ty:result
4217 <result> = mul nsw <ty> <op1>, <op2> ; yields ty:result
4218 <result> = mul nuw nsw <ty> <op1>, <op2> ; yields ty:result
4223 The '``mul``' instruction returns the product of its two operands.
4228 The two arguments to the '``mul``' instruction must be
4229 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4230 arguments must have identical types.
4235 The value produced is the integer product of the two operands.
4237 If the result of the multiplication has unsigned overflow, the result
4238 returned is the mathematical result modulo 2\ :sup:`n`\ , where n is the
4239 bit width of the result.
4241 Because LLVM integers use a two's complement representation, and the
4242 result is the same width as the operands, this instruction returns the
4243 correct result for both signed and unsigned integers. If a full product
4244 (e.g. ``i32`` * ``i32`` -> ``i64``) is needed, the operands should be
4245 sign-extended or zero-extended as appropriate to the width of the full
4248 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
4249 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
4250 result value of the ``mul`` is a :ref:`poison value <poisonvalues>` if
4251 unsigned and/or signed overflow, respectively, occurs.
4256 .. code-block:: llvm
4258 <result> = mul i32 4, %var ; yields i32:result = 4 * %var
4262 '``fmul``' Instruction
4263 ^^^^^^^^^^^^^^^^^^^^^^
4270 <result> = fmul [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
4275 The '``fmul``' instruction returns the product of its two operands.
4280 The two arguments to the '``fmul``' instruction must be :ref:`floating
4281 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
4282 Both arguments must have identical types.
4287 The value produced is the floating point product of the two operands.
4288 This instruction can also take any number of :ref:`fast-math
4289 flags <fastmath>`, which are optimization hints to enable otherwise
4290 unsafe floating point optimizations:
4295 .. code-block:: llvm
4297 <result> = fmul float 4.0, %var ; yields float:result = 4.0 * %var
4299 '``udiv``' Instruction
4300 ^^^^^^^^^^^^^^^^^^^^^^
4307 <result> = udiv <ty> <op1>, <op2> ; yields ty:result
4308 <result> = udiv exact <ty> <op1>, <op2> ; yields ty:result
4313 The '``udiv``' instruction returns the quotient of its two operands.
4318 The two arguments to the '``udiv``' instruction must be
4319 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4320 arguments must have identical types.
4325 The value produced is the unsigned integer quotient of the two operands.
4327 Note that unsigned integer division and signed integer division are
4328 distinct operations; for signed integer division, use '``sdiv``'.
4330 Division by zero leads to undefined behavior.
4332 If the ``exact`` keyword is present, the result value of the ``udiv`` is
4333 a :ref:`poison value <poisonvalues>` if %op1 is not a multiple of %op2 (as
4334 such, "((a udiv exact b) mul b) == a").
4339 .. code-block:: llvm
4341 <result> = udiv i32 4, %var ; yields i32:result = 4 / %var
4343 '``sdiv``' Instruction
4344 ^^^^^^^^^^^^^^^^^^^^^^
4351 <result> = sdiv <ty> <op1>, <op2> ; yields ty:result
4352 <result> = sdiv exact <ty> <op1>, <op2> ; yields ty:result
4357 The '``sdiv``' instruction returns the quotient of its two operands.
4362 The two arguments to the '``sdiv``' instruction must be
4363 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4364 arguments must have identical types.
4369 The value produced is the signed integer quotient of the two operands
4370 rounded towards zero.
4372 Note that signed integer division and unsigned integer division are
4373 distinct operations; for unsigned integer division, use '``udiv``'.
4375 Division by zero leads to undefined behavior. Overflow also leads to
4376 undefined behavior; this is a rare case, but can occur, for example, by
4377 doing a 32-bit division of -2147483648 by -1.
4379 If the ``exact`` keyword is present, the result value of the ``sdiv`` is
4380 a :ref:`poison value <poisonvalues>` if the result would be rounded.
4385 .. code-block:: llvm
4387 <result> = sdiv i32 4, %var ; yields i32:result = 4 / %var
4391 '``fdiv``' Instruction
4392 ^^^^^^^^^^^^^^^^^^^^^^
4399 <result> = fdiv [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
4404 The '``fdiv``' instruction returns the quotient of its two operands.
4409 The two arguments to the '``fdiv``' instruction must be :ref:`floating
4410 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
4411 Both arguments must have identical types.
4416 The value produced is the floating point quotient of the two operands.
4417 This instruction can also take any number of :ref:`fast-math
4418 flags <fastmath>`, which are optimization hints to enable otherwise
4419 unsafe floating point optimizations:
4424 .. code-block:: llvm
4426 <result> = fdiv float 4.0, %var ; yields float:result = 4.0 / %var
4428 '``urem``' Instruction
4429 ^^^^^^^^^^^^^^^^^^^^^^
4436 <result> = urem <ty> <op1>, <op2> ; yields ty:result
4441 The '``urem``' instruction returns the remainder from the unsigned
4442 division of its two arguments.
4447 The two arguments to the '``urem``' instruction must be
4448 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4449 arguments must have identical types.
4454 This instruction returns the unsigned integer *remainder* of a division.
4455 This instruction always performs an unsigned division to get the
4458 Note that unsigned integer remainder and signed integer remainder are
4459 distinct operations; for signed integer remainder, use '``srem``'.
4461 Taking the remainder of a division by zero leads to undefined behavior.
4466 .. code-block:: llvm
4468 <result> = urem i32 4, %var ; yields i32:result = 4 % %var
4470 '``srem``' Instruction
4471 ^^^^^^^^^^^^^^^^^^^^^^
4478 <result> = srem <ty> <op1>, <op2> ; yields ty:result
4483 The '``srem``' instruction returns the remainder from the signed
4484 division of its two operands. This instruction can also take
4485 :ref:`vector <t_vector>` versions of the values in which case the elements
4491 The two arguments to the '``srem``' instruction must be
4492 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4493 arguments must have identical types.
4498 This instruction returns the *remainder* of a division (where the result
4499 is either zero or has the same sign as the dividend, ``op1``), not the
4500 *modulo* operator (where the result is either zero or has the same sign
4501 as the divisor, ``op2``) of a value. For more information about the
4502 difference, see `The Math
4503 Forum <http://mathforum.org/dr.math/problems/anne.4.28.99.html>`_. For a
4504 table of how this is implemented in various languages, please see
4506 operation <http://en.wikipedia.org/wiki/Modulo_operation>`_.
4508 Note that signed integer remainder and unsigned integer remainder are
4509 distinct operations; for unsigned integer remainder, use '``urem``'.
4511 Taking the remainder of a division by zero leads to undefined behavior.
4512 Overflow also leads to undefined behavior; this is a rare case, but can
4513 occur, for example, by taking the remainder of a 32-bit division of
4514 -2147483648 by -1. (The remainder doesn't actually overflow, but this
4515 rule lets srem be implemented using instructions that return both the
4516 result of the division and the remainder.)
4521 .. code-block:: llvm
4523 <result> = srem i32 4, %var ; yields i32:result = 4 % %var
4527 '``frem``' Instruction
4528 ^^^^^^^^^^^^^^^^^^^^^^
4535 <result> = frem [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
4540 The '``frem``' instruction returns the remainder from the division of
4546 The two arguments to the '``frem``' instruction must be :ref:`floating
4547 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
4548 Both arguments must have identical types.
4553 This instruction returns the *remainder* of a division. The remainder
4554 has the same sign as the dividend. This instruction can also take any
4555 number of :ref:`fast-math flags <fastmath>`, which are optimization hints
4556 to enable otherwise unsafe floating point optimizations:
4561 .. code-block:: llvm
4563 <result> = frem float 4.0, %var ; yields float:result = 4.0 % %var
4567 Bitwise Binary Operations
4568 -------------------------
4570 Bitwise binary operators are used to do various forms of bit-twiddling
4571 in a program. They are generally very efficient instructions and can
4572 commonly be strength reduced from other instructions. They require two
4573 operands of the same type, execute an operation on them, and produce a
4574 single value. The resulting value is the same type as its operands.
4576 '``shl``' Instruction
4577 ^^^^^^^^^^^^^^^^^^^^^
4584 <result> = shl <ty> <op1>, <op2> ; yields ty:result
4585 <result> = shl nuw <ty> <op1>, <op2> ; yields ty:result
4586 <result> = shl nsw <ty> <op1>, <op2> ; yields ty:result
4587 <result> = shl nuw nsw <ty> <op1>, <op2> ; yields ty:result
4592 The '``shl``' instruction returns the first operand shifted to the left
4593 a specified number of bits.
4598 Both arguments to the '``shl``' instruction must be the same
4599 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4600 '``op2``' is treated as an unsigned value.
4605 The value produced is ``op1`` \* 2\ :sup:`op2` mod 2\ :sup:`n`,
4606 where ``n`` is the width of the result. If ``op2`` is (statically or
4607 dynamically) negative or equal to or larger than the number of bits in
4608 ``op1``, the result is undefined. If the arguments are vectors, each
4609 vector element of ``op1`` is shifted by the corresponding shift amount
4612 If the ``nuw`` keyword is present, then the shift produces a :ref:`poison
4613 value <poisonvalues>` if it shifts out any non-zero bits. If the
4614 ``nsw`` keyword is present, then the shift produces a :ref:`poison
4615 value <poisonvalues>` if it shifts out any bits that disagree with the
4616 resultant sign bit. As such, NUW/NSW have the same semantics as they
4617 would if the shift were expressed as a mul instruction with the same
4618 nsw/nuw bits in (mul %op1, (shl 1, %op2)).
4623 .. code-block:: llvm
4625 <result> = shl i32 4, %var ; yields i32: 4 << %var
4626 <result> = shl i32 4, 2 ; yields i32: 16
4627 <result> = shl i32 1, 10 ; yields i32: 1024
4628 <result> = shl i32 1, 32 ; undefined
4629 <result> = shl <2 x i32> < i32 1, i32 1>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 2, i32 4>
4631 '``lshr``' Instruction
4632 ^^^^^^^^^^^^^^^^^^^^^^
4639 <result> = lshr <ty> <op1>, <op2> ; yields ty:result
4640 <result> = lshr exact <ty> <op1>, <op2> ; yields ty:result
4645 The '``lshr``' instruction (logical shift right) returns the first
4646 operand shifted to the right a specified number of bits with zero fill.
4651 Both arguments to the '``lshr``' instruction must be the same
4652 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4653 '``op2``' is treated as an unsigned value.
4658 This instruction always performs a logical shift right operation. The
4659 most significant bits of the result will be filled with zero bits after
4660 the shift. If ``op2`` is (statically or dynamically) equal to or larger
4661 than the number of bits in ``op1``, the result is undefined. If the
4662 arguments are vectors, each vector element of ``op1`` is shifted by the
4663 corresponding shift amount in ``op2``.
4665 If the ``exact`` keyword is present, the result value of the ``lshr`` is
4666 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
4672 .. code-block:: llvm
4674 <result> = lshr i32 4, 1 ; yields i32:result = 2
4675 <result> = lshr i32 4, 2 ; yields i32:result = 1
4676 <result> = lshr i8 4, 3 ; yields i8:result = 0
4677 <result> = lshr i8 -2, 1 ; yields i8:result = 0x7F
4678 <result> = lshr i32 1, 32 ; undefined
4679 <result> = lshr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 0x7FFFFFFF, i32 1>
4681 '``ashr``' Instruction
4682 ^^^^^^^^^^^^^^^^^^^^^^
4689 <result> = ashr <ty> <op1>, <op2> ; yields ty:result
4690 <result> = ashr exact <ty> <op1>, <op2> ; yields ty:result
4695 The '``ashr``' instruction (arithmetic shift right) returns the first
4696 operand shifted to the right a specified number of bits with sign
4702 Both arguments to the '``ashr``' instruction must be the same
4703 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4704 '``op2``' is treated as an unsigned value.
4709 This instruction always performs an arithmetic shift right operation,
4710 The most significant bits of the result will be filled with the sign bit
4711 of ``op1``. If ``op2`` is (statically or dynamically) equal to or larger
4712 than the number of bits in ``op1``, the result is undefined. If the
4713 arguments are vectors, each vector element of ``op1`` is shifted by the
4714 corresponding shift amount in ``op2``.
4716 If the ``exact`` keyword is present, the result value of the ``ashr`` is
4717 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
4723 .. code-block:: llvm
4725 <result> = ashr i32 4, 1 ; yields i32:result = 2
4726 <result> = ashr i32 4, 2 ; yields i32:result = 1
4727 <result> = ashr i8 4, 3 ; yields i8:result = 0
4728 <result> = ashr i8 -2, 1 ; yields i8:result = -1
4729 <result> = ashr i32 1, 32 ; undefined
4730 <result> = ashr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 3> ; yields: result=<2 x i32> < i32 -1, i32 0>
4732 '``and``' Instruction
4733 ^^^^^^^^^^^^^^^^^^^^^
4740 <result> = and <ty> <op1>, <op2> ; yields ty:result
4745 The '``and``' instruction returns the bitwise logical and of its two
4751 The two arguments to the '``and``' instruction must be
4752 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4753 arguments must have identical types.
4758 The truth table used for the '``and``' instruction is:
4775 .. code-block:: llvm
4777 <result> = and i32 4, %var ; yields i32:result = 4 & %var
4778 <result> = and i32 15, 40 ; yields i32:result = 8
4779 <result> = and i32 4, 8 ; yields i32:result = 0
4781 '``or``' Instruction
4782 ^^^^^^^^^^^^^^^^^^^^
4789 <result> = or <ty> <op1>, <op2> ; yields ty:result
4794 The '``or``' instruction returns the bitwise logical inclusive or of its
4800 The two arguments to the '``or``' instruction must be
4801 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4802 arguments must have identical types.
4807 The truth table used for the '``or``' instruction is:
4826 <result> = or i32 4, %var ; yields i32:result = 4 | %var
4827 <result> = or i32 15, 40 ; yields i32:result = 47
4828 <result> = or i32 4, 8 ; yields i32:result = 12
4830 '``xor``' Instruction
4831 ^^^^^^^^^^^^^^^^^^^^^
4838 <result> = xor <ty> <op1>, <op2> ; yields ty:result
4843 The '``xor``' instruction returns the bitwise logical exclusive or of
4844 its two operands. The ``xor`` is used to implement the "one's
4845 complement" operation, which is the "~" operator in C.
4850 The two arguments to the '``xor``' instruction must be
4851 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4852 arguments must have identical types.
4857 The truth table used for the '``xor``' instruction is:
4874 .. code-block:: llvm
4876 <result> = xor i32 4, %var ; yields i32:result = 4 ^ %var
4877 <result> = xor i32 15, 40 ; yields i32:result = 39
4878 <result> = xor i32 4, 8 ; yields i32:result = 12
4879 <result> = xor i32 %V, -1 ; yields i32:result = ~%V
4884 LLVM supports several instructions to represent vector operations in a
4885 target-independent manner. These instructions cover the element-access
4886 and vector-specific operations needed to process vectors effectively.
4887 While LLVM does directly support these vector operations, many
4888 sophisticated algorithms will want to use target-specific intrinsics to
4889 take full advantage of a specific target.
4891 .. _i_extractelement:
4893 '``extractelement``' Instruction
4894 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4901 <result> = extractelement <n x <ty>> <val>, <ty2> <idx> ; yields <ty>
4906 The '``extractelement``' instruction extracts a single scalar element
4907 from a vector at a specified index.
4912 The first operand of an '``extractelement``' instruction is a value of
4913 :ref:`vector <t_vector>` type. The second operand is an index indicating
4914 the position from which to extract the element. The index may be a
4915 variable of any integer type.
4920 The result is a scalar of the same type as the element type of ``val``.
4921 Its value is the value at position ``idx`` of ``val``. If ``idx``
4922 exceeds the length of ``val``, the results are undefined.
4927 .. code-block:: llvm
4929 <result> = extractelement <4 x i32> %vec, i32 0 ; yields i32
4931 .. _i_insertelement:
4933 '``insertelement``' Instruction
4934 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4941 <result> = insertelement <n x <ty>> <val>, <ty> <elt>, <ty2> <idx> ; yields <n x <ty>>
4946 The '``insertelement``' instruction inserts a scalar element into a
4947 vector at a specified index.
4952 The first operand of an '``insertelement``' instruction is a value of
4953 :ref:`vector <t_vector>` type. The second operand is a scalar value whose
4954 type must equal the element type of the first operand. The third operand
4955 is an index indicating the position at which to insert the value. The
4956 index may be a variable of any integer type.
4961 The result is a vector of the same type as ``val``. Its element values
4962 are those of ``val`` except at position ``idx``, where it gets the value
4963 ``elt``. If ``idx`` exceeds the length of ``val``, the results are
4969 .. code-block:: llvm
4971 <result> = insertelement <4 x i32> %vec, i32 1, i32 0 ; yields <4 x i32>
4973 .. _i_shufflevector:
4975 '``shufflevector``' Instruction
4976 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4983 <result> = shufflevector <n x <ty>> <v1>, <n x <ty>> <v2>, <m x i32> <mask> ; yields <m x <ty>>
4988 The '``shufflevector``' instruction constructs a permutation of elements
4989 from two input vectors, returning a vector with the same element type as
4990 the input and length that is the same as the shuffle mask.
4995 The first two operands of a '``shufflevector``' instruction are vectors
4996 with the same type. The third argument is a shuffle mask whose element
4997 type is always 'i32'. The result of the instruction is a vector whose
4998 length is the same as the shuffle mask and whose element type is the
4999 same as the element type of the first two operands.
5001 The shuffle mask operand is required to be a constant vector with either
5002 constant integer or undef values.
5007 The elements of the two input vectors are numbered from left to right
5008 across both of the vectors. The shuffle mask operand specifies, for each
5009 element of the result vector, which element of the two input vectors the
5010 result element gets. The element selector may be undef (meaning "don't
5011 care") and the second operand may be undef if performing a shuffle from
5017 .. code-block:: llvm
5019 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
5020 <4 x i32> <i32 0, i32 4, i32 1, i32 5> ; yields <4 x i32>
5021 <result> = shufflevector <4 x i32> %v1, <4 x i32> undef,
5022 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32> - Identity shuffle.
5023 <result> = shufflevector <8 x i32> %v1, <8 x i32> undef,
5024 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32>
5025 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
5026 <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7 > ; yields <8 x i32>
5028 Aggregate Operations
5029 --------------------
5031 LLVM supports several instructions for working with
5032 :ref:`aggregate <t_aggregate>` values.
5036 '``extractvalue``' Instruction
5037 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5044 <result> = extractvalue <aggregate type> <val>, <idx>{, <idx>}*
5049 The '``extractvalue``' instruction extracts the value of a member field
5050 from an :ref:`aggregate <t_aggregate>` value.
5055 The first operand of an '``extractvalue``' instruction is a value of
5056 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The operands are
5057 constant indices to specify which value to extract in a similar manner
5058 as indices in a '``getelementptr``' instruction.
5060 The major differences to ``getelementptr`` indexing are:
5062 - Since the value being indexed is not a pointer, the first index is
5063 omitted and assumed to be zero.
5064 - At least one index must be specified.
5065 - Not only struct indices but also array indices must be in bounds.
5070 The result is the value at the position in the aggregate specified by
5076 .. code-block:: llvm
5078 <result> = extractvalue {i32, float} %agg, 0 ; yields i32
5082 '``insertvalue``' Instruction
5083 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5090 <result> = insertvalue <aggregate type> <val>, <ty> <elt>, <idx>{, <idx>}* ; yields <aggregate type>
5095 The '``insertvalue``' instruction inserts a value into a member field in
5096 an :ref:`aggregate <t_aggregate>` value.
5101 The first operand of an '``insertvalue``' instruction is a value of
5102 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The second operand is
5103 a first-class value to insert. The following operands are constant
5104 indices indicating the position at which to insert the value in a
5105 similar manner as indices in a '``extractvalue``' instruction. The value
5106 to insert must have the same type as the value identified by the
5112 The result is an aggregate of the same type as ``val``. Its value is
5113 that of ``val`` except that the value at the position specified by the
5114 indices is that of ``elt``.
5119 .. code-block:: llvm
5121 %agg1 = insertvalue {i32, float} undef, i32 1, 0 ; yields {i32 1, float undef}
5122 %agg2 = insertvalue {i32, float} %agg1, float %val, 1 ; yields {i32 1, float %val}
5123 %agg3 = insertvalue {i32, {float}} undef, float %val, 1, 0 ; yields {i32 undef, {float %val}}
5127 Memory Access and Addressing Operations
5128 ---------------------------------------
5130 A key design point of an SSA-based representation is how it represents
5131 memory. In LLVM, no memory locations are in SSA form, which makes things
5132 very simple. This section describes how to read, write, and allocate
5137 '``alloca``' Instruction
5138 ^^^^^^^^^^^^^^^^^^^^^^^^
5145 <result> = alloca [inalloca] <type> [, <ty> <NumElements>] [, align <alignment>] ; yields type*:result
5150 The '``alloca``' instruction allocates memory on the stack frame of the
5151 currently executing function, to be automatically released when this
5152 function returns to its caller. The object is always allocated in the
5153 generic address space (address space zero).
5158 The '``alloca``' instruction allocates ``sizeof(<type>)*NumElements``
5159 bytes of memory on the runtime stack, returning a pointer of the
5160 appropriate type to the program. If "NumElements" is specified, it is
5161 the number of elements allocated, otherwise "NumElements" is defaulted
5162 to be one. If a constant alignment is specified, the value result of the
5163 allocation is guaranteed to be aligned to at least that boundary. The
5164 alignment may not be greater than ``1 << 29``. If not specified, or if
5165 zero, the target can choose to align the allocation on any convenient
5166 boundary compatible with the type.
5168 '``type``' may be any sized type.
5173 Memory is allocated; a pointer is returned. The operation is undefined
5174 if there is insufficient stack space for the allocation. '``alloca``'d
5175 memory is automatically released when the function returns. The
5176 '``alloca``' instruction is commonly used to represent automatic
5177 variables that must have an address available. When the function returns
5178 (either with the ``ret`` or ``resume`` instructions), the memory is
5179 reclaimed. Allocating zero bytes is legal, but the result is undefined.
5180 The order in which memory is allocated (ie., which way the stack grows)
5186 .. code-block:: llvm
5188 %ptr = alloca i32 ; yields i32*:ptr
5189 %ptr = alloca i32, i32 4 ; yields i32*:ptr
5190 %ptr = alloca i32, i32 4, align 1024 ; yields i32*:ptr
5191 %ptr = alloca i32, align 1024 ; yields i32*:ptr
5195 '``load``' Instruction
5196 ^^^^^^^^^^^^^^^^^^^^^^
5203 <result> = load [volatile] <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>][, !invariant.load !<index>][, !nonnull !<index>]
5204 <result> = load atomic [volatile] <ty>* <pointer> [singlethread] <ordering>, align <alignment>
5205 !<index> = !{ i32 1 }
5210 The '``load``' instruction is used to read from memory.
5215 The argument to the ``load`` instruction specifies the memory address
5216 from which to load. The pointer must point to a :ref:`first
5217 class <t_firstclass>` type. If the ``load`` is marked as ``volatile``,
5218 then the optimizer is not allowed to modify the number or order of
5219 execution of this ``load`` with other :ref:`volatile
5220 operations <volatile>`.
5222 If the ``load`` is marked as ``atomic``, it takes an extra
5223 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
5224 ``release`` and ``acq_rel`` orderings are not valid on ``load``
5225 instructions. Atomic loads produce :ref:`defined <memmodel>` results
5226 when they may see multiple atomic stores. The type of the pointee must
5227 be an integer type whose bit width is a power of two greater than or
5228 equal to eight and less than or equal to a target-specific size limit.
5229 ``align`` must be explicitly specified on atomic loads, and the load has
5230 undefined behavior if the alignment is not set to a value which is at
5231 least the size in bytes of the pointee. ``!nontemporal`` does not have
5232 any defined semantics for atomic loads.
5234 The optional constant ``align`` argument specifies the alignment of the
5235 operation (that is, the alignment of the memory address). A value of 0
5236 or an omitted ``align`` argument means that the operation has the ABI
5237 alignment for the target. It is the responsibility of the code emitter
5238 to ensure that the alignment information is correct. Overestimating the
5239 alignment results in undefined behavior. Underestimating the alignment
5240 may produce less efficient code. An alignment of 1 is always safe. The
5241 maximum possible alignment is ``1 << 29``.
5243 The optional ``!nontemporal`` metadata must reference a single
5244 metadata name ``<index>`` corresponding to a metadata node with one
5245 ``i32`` entry of value 1. The existence of the ``!nontemporal``
5246 metadata on the instruction tells the optimizer and code generator
5247 that this load is not expected to be reused in the cache. The code
5248 generator may select special instructions to save cache bandwidth, such
5249 as the ``MOVNT`` instruction on x86.
5251 The optional ``!invariant.load`` metadata must reference a single
5252 metadata name ``<index>`` corresponding to a metadata node with no
5253 entries. The existence of the ``!invariant.load`` metadata on the
5254 instruction tells the optimizer and code generator that the address
5255 operand to this load points to memory which can be assumed unchanged.
5256 Being invariant does not imply that a location is dereferenceable,
5257 but it does imply that once the location is known dereferenceable
5258 its value is henceforth unchanging.
5260 The optional ``!nonnull`` metadata must reference a single
5261 metadata name ``<index>`` corresponding to a metadata node with no
5262 entries. The existence of the ``!nonnull`` metadata on the
5263 instruction tells the optimizer that the value loaded is known to
5264 never be null. This is analogous to the ''nonnull'' attribute
5265 on parameters and return values. This metadata can only be applied
5266 to loads of a pointer type.
5271 The location of memory pointed to is loaded. If the value being loaded
5272 is of scalar type then the number of bytes read does not exceed the
5273 minimum number of bytes needed to hold all bits of the type. For
5274 example, loading an ``i24`` reads at most three bytes. When loading a
5275 value of a type like ``i20`` with a size that is not an integral number
5276 of bytes, the result is undefined if the value was not originally
5277 written using a store of the same type.
5282 .. code-block:: llvm
5284 %ptr = alloca i32 ; yields i32*:ptr
5285 store i32 3, i32* %ptr ; yields void
5286 %val = load i32* %ptr ; yields i32:val = i32 3
5290 '``store``' Instruction
5291 ^^^^^^^^^^^^^^^^^^^^^^^
5298 store [volatile] <ty> <value>, <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>] ; yields void
5299 store atomic [volatile] <ty> <value>, <ty>* <pointer> [singlethread] <ordering>, align <alignment> ; yields void
5304 The '``store``' instruction is used to write to memory.
5309 There are two arguments to the ``store`` instruction: a value to store
5310 and an address at which to store it. The type of the ``<pointer>``
5311 operand must be a pointer to the :ref:`first class <t_firstclass>` type of
5312 the ``<value>`` operand. If the ``store`` is marked as ``volatile``,
5313 then the optimizer is not allowed to modify the number or order of
5314 execution of this ``store`` with other :ref:`volatile
5315 operations <volatile>`.
5317 If the ``store`` is marked as ``atomic``, it takes an extra
5318 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
5319 ``acquire`` and ``acq_rel`` orderings aren't valid on ``store``
5320 instructions. Atomic loads produce :ref:`defined <memmodel>` results
5321 when they may see multiple atomic stores. The type of the pointee must
5322 be an integer type whose bit width is a power of two greater than or
5323 equal to eight and less than or equal to a target-specific size limit.
5324 ``align`` must be explicitly specified on atomic stores, and the store
5325 has undefined behavior if the alignment is not set to a value which is
5326 at least the size in bytes of the pointee. ``!nontemporal`` does not
5327 have any defined semantics for atomic stores.
5329 The optional constant ``align`` argument specifies the alignment of the
5330 operation (that is, the alignment of the memory address). A value of 0
5331 or an omitted ``align`` argument means that the operation has the ABI
5332 alignment for the target. It is the responsibility of the code emitter
5333 to ensure that the alignment information is correct. Overestimating the
5334 alignment results in undefined behavior. Underestimating the
5335 alignment may produce less efficient code. An alignment of 1 is always
5336 safe. The maximum possible alignment is ``1 << 29``.
5338 The optional ``!nontemporal`` metadata must reference a single metadata
5339 name ``<index>`` corresponding to a metadata node with one ``i32`` entry of
5340 value 1. The existence of the ``!nontemporal`` metadata on the instruction
5341 tells the optimizer and code generator that this load is not expected to
5342 be reused in the cache. The code generator may select special
5343 instructions to save cache bandwidth, such as the MOVNT instruction on
5349 The contents of memory are updated to contain ``<value>`` at the
5350 location specified by the ``<pointer>`` operand. If ``<value>`` is
5351 of scalar type then the number of bytes written does not exceed the
5352 minimum number of bytes needed to hold all bits of the type. For
5353 example, storing an ``i24`` writes at most three bytes. When writing a
5354 value of a type like ``i20`` with a size that is not an integral number
5355 of bytes, it is unspecified what happens to the extra bits that do not
5356 belong to the type, but they will typically be overwritten.
5361 .. code-block:: llvm
5363 %ptr = alloca i32 ; yields i32*:ptr
5364 store i32 3, i32* %ptr ; yields void
5365 %val = load i32* %ptr ; yields i32:val = i32 3
5369 '``fence``' Instruction
5370 ^^^^^^^^^^^^^^^^^^^^^^^
5377 fence [singlethread] <ordering> ; yields void
5382 The '``fence``' instruction is used to introduce happens-before edges
5388 '``fence``' instructions take an :ref:`ordering <ordering>` argument which
5389 defines what *synchronizes-with* edges they add. They can only be given
5390 ``acquire``, ``release``, ``acq_rel``, and ``seq_cst`` orderings.
5395 A fence A which has (at least) ``release`` ordering semantics
5396 *synchronizes with* a fence B with (at least) ``acquire`` ordering
5397 semantics if and only if there exist atomic operations X and Y, both
5398 operating on some atomic object M, such that A is sequenced before X, X
5399 modifies M (either directly or through some side effect of a sequence
5400 headed by X), Y is sequenced before B, and Y observes M. This provides a
5401 *happens-before* dependency between A and B. Rather than an explicit
5402 ``fence``, one (but not both) of the atomic operations X or Y might
5403 provide a ``release`` or ``acquire`` (resp.) ordering constraint and
5404 still *synchronize-with* the explicit ``fence`` and establish the
5405 *happens-before* edge.
5407 A ``fence`` which has ``seq_cst`` ordering, in addition to having both
5408 ``acquire`` and ``release`` semantics specified above, participates in
5409 the global program order of other ``seq_cst`` operations and/or fences.
5411 The optional ":ref:`singlethread <singlethread>`" argument specifies
5412 that the fence only synchronizes with other fences in the same thread.
5413 (This is useful for interacting with signal handlers.)
5418 .. code-block:: llvm
5420 fence acquire ; yields void
5421 fence singlethread seq_cst ; yields void
5425 '``cmpxchg``' Instruction
5426 ^^^^^^^^^^^^^^^^^^^^^^^^^
5433 cmpxchg [weak] [volatile] <ty>* <pointer>, <ty> <cmp>, <ty> <new> [singlethread] <success ordering> <failure ordering> ; yields { ty, i1 }
5438 The '``cmpxchg``' instruction is used to atomically modify memory. It
5439 loads a value in memory and compares it to a given value. If they are
5440 equal, it tries to store a new value into the memory.
5445 There are three arguments to the '``cmpxchg``' instruction: an address
5446 to operate on, a value to compare to the value currently be at that
5447 address, and a new value to place at that address if the compared values
5448 are equal. The type of '<cmp>' must be an integer type whose bit width
5449 is a power of two greater than or equal to eight and less than or equal
5450 to a target-specific size limit. '<cmp>' and '<new>' must have the same
5451 type, and the type of '<pointer>' must be a pointer to that type. If the
5452 ``cmpxchg`` is marked as ``volatile``, then the optimizer is not allowed
5453 to modify the number or order of execution of this ``cmpxchg`` with
5454 other :ref:`volatile operations <volatile>`.
5456 The success and failure :ref:`ordering <ordering>` arguments specify how this
5457 ``cmpxchg`` synchronizes with other atomic operations. Both ordering parameters
5458 must be at least ``monotonic``, the ordering constraint on failure must be no
5459 stronger than that on success, and the failure ordering cannot be either
5460 ``release`` or ``acq_rel``.
5462 The optional "``singlethread``" argument declares that the ``cmpxchg``
5463 is only atomic with respect to code (usually signal handlers) running in
5464 the same thread as the ``cmpxchg``. Otherwise the cmpxchg is atomic with
5465 respect to all other code in the system.
5467 The pointer passed into cmpxchg must have alignment greater than or
5468 equal to the size in memory of the operand.
5473 The contents of memory at the location specified by the '``<pointer>``' operand
5474 is read and compared to '``<cmp>``'; if the read value is the equal, the
5475 '``<new>``' is written. The original value at the location is returned, together
5476 with a flag indicating success (true) or failure (false).
5478 If the cmpxchg operation is marked as ``weak`` then a spurious failure is
5479 permitted: the operation may not write ``<new>`` even if the comparison
5482 If the cmpxchg operation is strong (the default), the i1 value is 1 if and only
5483 if the value loaded equals ``cmp``.
5485 A successful ``cmpxchg`` is a read-modify-write instruction for the purpose of
5486 identifying release sequences. A failed ``cmpxchg`` is equivalent to an atomic
5487 load with an ordering parameter determined the second ordering parameter.
5492 .. code-block:: llvm
5495 %orig = atomic load i32* %ptr unordered ; yields i32
5499 %cmp = phi i32 [ %orig, %entry ], [%old, %loop]
5500 %squared = mul i32 %cmp, %cmp
5501 %val_success = cmpxchg i32* %ptr, i32 %cmp, i32 %squared acq_rel monotonic ; yields { i32, i1 }
5502 %value_loaded = extractvalue { i32, i1 } %val_success, 0
5503 %success = extractvalue { i32, i1 } %val_success, 1
5504 br i1 %success, label %done, label %loop
5511 '``atomicrmw``' Instruction
5512 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
5519 atomicrmw [volatile] <operation> <ty>* <pointer>, <ty> <value> [singlethread] <ordering> ; yields ty
5524 The '``atomicrmw``' instruction is used to atomically modify memory.
5529 There are three arguments to the '``atomicrmw``' instruction: an
5530 operation to apply, an address whose value to modify, an argument to the
5531 operation. The operation must be one of the following keywords:
5545 The type of '<value>' must be an integer type whose bit width is a power
5546 of two greater than or equal to eight and less than or equal to a
5547 target-specific size limit. The type of the '``<pointer>``' operand must
5548 be a pointer to that type. If the ``atomicrmw`` is marked as
5549 ``volatile``, then the optimizer is not allowed to modify the number or
5550 order of execution of this ``atomicrmw`` with other :ref:`volatile
5551 operations <volatile>`.
5556 The contents of memory at the location specified by the '``<pointer>``'
5557 operand are atomically read, modified, and written back. The original
5558 value at the location is returned. The modification is specified by the
5561 - xchg: ``*ptr = val``
5562 - add: ``*ptr = *ptr + val``
5563 - sub: ``*ptr = *ptr - val``
5564 - and: ``*ptr = *ptr & val``
5565 - nand: ``*ptr = ~(*ptr & val)``
5566 - or: ``*ptr = *ptr | val``
5567 - xor: ``*ptr = *ptr ^ val``
5568 - max: ``*ptr = *ptr > val ? *ptr : val`` (using a signed comparison)
5569 - min: ``*ptr = *ptr < val ? *ptr : val`` (using a signed comparison)
5570 - umax: ``*ptr = *ptr > val ? *ptr : val`` (using an unsigned
5572 - umin: ``*ptr = *ptr < val ? *ptr : val`` (using an unsigned
5578 .. code-block:: llvm
5580 %old = atomicrmw add i32* %ptr, i32 1 acquire ; yields i32
5582 .. _i_getelementptr:
5584 '``getelementptr``' Instruction
5585 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5592 <result> = getelementptr <pty>* <ptrval>{, <ty> <idx>}*
5593 <result> = getelementptr inbounds <pty>* <ptrval>{, <ty> <idx>}*
5594 <result> = getelementptr <ptr vector> ptrval, <vector index type> idx
5599 The '``getelementptr``' instruction is used to get the address of a
5600 subelement of an :ref:`aggregate <t_aggregate>` data structure. It performs
5601 address calculation only and does not access memory.
5606 The first argument is always a pointer or a vector of pointers, and
5607 forms the basis of the calculation. The remaining arguments are indices
5608 that indicate which of the elements of the aggregate object are indexed.
5609 The interpretation of each index is dependent on the type being indexed
5610 into. The first index always indexes the pointer value given as the
5611 first argument, the second index indexes a value of the type pointed to
5612 (not necessarily the value directly pointed to, since the first index
5613 can be non-zero), etc. The first type indexed into must be a pointer
5614 value, subsequent types can be arrays, vectors, and structs. Note that
5615 subsequent types being indexed into can never be pointers, since that
5616 would require loading the pointer before continuing calculation.
5618 The type of each index argument depends on the type it is indexing into.
5619 When indexing into a (optionally packed) structure, only ``i32`` integer
5620 **constants** are allowed (when using a vector of indices they must all
5621 be the **same** ``i32`` integer constant). When indexing into an array,
5622 pointer or vector, integers of any width are allowed, and they are not
5623 required to be constant. These integers are treated as signed values
5626 For example, let's consider a C code fragment and how it gets compiled
5642 int *foo(struct ST *s) {
5643 return &s[1].Z.B[5][13];
5646 The LLVM code generated by Clang is:
5648 .. code-block:: llvm
5650 %struct.RT = type { i8, [10 x [20 x i32]], i8 }
5651 %struct.ST = type { i32, double, %struct.RT }
5653 define i32* @foo(%struct.ST* %s) nounwind uwtable readnone optsize ssp {
5655 %arrayidx = getelementptr inbounds %struct.ST* %s, i64 1, i32 2, i32 1, i64 5, i64 13
5662 In the example above, the first index is indexing into the
5663 '``%struct.ST*``' type, which is a pointer, yielding a '``%struct.ST``'
5664 = '``{ i32, double, %struct.RT }``' type, a structure. The second index
5665 indexes into the third element of the structure, yielding a
5666 '``%struct.RT``' = '``{ i8 , [10 x [20 x i32]], i8 }``' type, another
5667 structure. The third index indexes into the second element of the
5668 structure, yielding a '``[10 x [20 x i32]]``' type, an array. The two
5669 dimensions of the array are subscripted into, yielding an '``i32``'
5670 type. The '``getelementptr``' instruction returns a pointer to this
5671 element, thus computing a value of '``i32*``' type.
5673 Note that it is perfectly legal to index partially through a structure,
5674 returning a pointer to an inner element. Because of this, the LLVM code
5675 for the given testcase is equivalent to:
5677 .. code-block:: llvm
5679 define i32* @foo(%struct.ST* %s) {
5680 %t1 = getelementptr %struct.ST* %s, i32 1 ; yields %struct.ST*:%t1
5681 %t2 = getelementptr %struct.ST* %t1, i32 0, i32 2 ; yields %struct.RT*:%t2
5682 %t3 = getelementptr %struct.RT* %t2, i32 0, i32 1 ; yields [10 x [20 x i32]]*:%t3
5683 %t4 = getelementptr [10 x [20 x i32]]* %t3, i32 0, i32 5 ; yields [20 x i32]*:%t4
5684 %t5 = getelementptr [20 x i32]* %t4, i32 0, i32 13 ; yields i32*:%t5
5688 If the ``inbounds`` keyword is present, the result value of the
5689 ``getelementptr`` is a :ref:`poison value <poisonvalues>` if the base
5690 pointer is not an *in bounds* address of an allocated object, or if any
5691 of the addresses that would be formed by successive addition of the
5692 offsets implied by the indices to the base address with infinitely
5693 precise signed arithmetic are not an *in bounds* address of that
5694 allocated object. The *in bounds* addresses for an allocated object are
5695 all the addresses that point into the object, plus the address one byte
5696 past the end. In cases where the base is a vector of pointers the
5697 ``inbounds`` keyword applies to each of the computations element-wise.
5699 If the ``inbounds`` keyword is not present, the offsets are added to the
5700 base address with silently-wrapping two's complement arithmetic. If the
5701 offsets have a different width from the pointer, they are sign-extended
5702 or truncated to the width of the pointer. The result value of the
5703 ``getelementptr`` may be outside the object pointed to by the base
5704 pointer. The result value may not necessarily be used to access memory
5705 though, even if it happens to point into allocated storage. See the
5706 :ref:`Pointer Aliasing Rules <pointeraliasing>` section for more
5709 The getelementptr instruction is often confusing. For some more insight
5710 into how it works, see :doc:`the getelementptr FAQ <GetElementPtr>`.
5715 .. code-block:: llvm
5717 ; yields [12 x i8]*:aptr
5718 %aptr = getelementptr {i32, [12 x i8]}* %saptr, i64 0, i32 1
5720 %vptr = getelementptr {i32, <2 x i8>}* %svptr, i64 0, i32 1, i32 1
5722 %eptr = getelementptr [12 x i8]* %aptr, i64 0, i32 1
5724 %iptr = getelementptr [10 x i32]* @arr, i16 0, i16 0
5726 In cases where the pointer argument is a vector of pointers, each index
5727 must be a vector with the same number of elements. For example:
5729 .. code-block:: llvm
5731 %A = getelementptr <4 x i8*> %ptrs, <4 x i64> %offsets,
5733 Conversion Operations
5734 ---------------------
5736 The instructions in this category are the conversion instructions
5737 (casting) which all take a single operand and a type. They perform
5738 various bit conversions on the operand.
5740 '``trunc .. to``' Instruction
5741 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5748 <result> = trunc <ty> <value> to <ty2> ; yields ty2
5753 The '``trunc``' instruction truncates its operand to the type ``ty2``.
5758 The '``trunc``' instruction takes a value to trunc, and a type to trunc
5759 it to. Both types must be of :ref:`integer <t_integer>` types, or vectors
5760 of the same number of integers. The bit size of the ``value`` must be
5761 larger than the bit size of the destination type, ``ty2``. Equal sized
5762 types are not allowed.
5767 The '``trunc``' instruction truncates the high order bits in ``value``
5768 and converts the remaining bits to ``ty2``. Since the source size must
5769 be larger than the destination size, ``trunc`` cannot be a *no-op cast*.
5770 It will always truncate bits.
5775 .. code-block:: llvm
5777 %X = trunc i32 257 to i8 ; yields i8:1
5778 %Y = trunc i32 123 to i1 ; yields i1:true
5779 %Z = trunc i32 122 to i1 ; yields i1:false
5780 %W = trunc <2 x i16> <i16 8, i16 7> to <2 x i8> ; yields <i8 8, i8 7>
5782 '``zext .. to``' Instruction
5783 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5790 <result> = zext <ty> <value> to <ty2> ; yields ty2
5795 The '``zext``' instruction zero extends its operand to type ``ty2``.
5800 The '``zext``' instruction takes a value to cast, and a type to cast it
5801 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
5802 the same number of integers. The bit size of the ``value`` must be
5803 smaller than the bit size of the destination type, ``ty2``.
5808 The ``zext`` fills the high order bits of the ``value`` with zero bits
5809 until it reaches the size of the destination type, ``ty2``.
5811 When zero extending from i1, the result will always be either 0 or 1.
5816 .. code-block:: llvm
5818 %X = zext i32 257 to i64 ; yields i64:257
5819 %Y = zext i1 true to i32 ; yields i32:1
5820 %Z = zext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
5822 '``sext .. to``' Instruction
5823 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5830 <result> = sext <ty> <value> to <ty2> ; yields ty2
5835 The '``sext``' sign extends ``value`` to the type ``ty2``.
5840 The '``sext``' instruction takes a value to cast, and a type to cast it
5841 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
5842 the same number of integers. The bit size of the ``value`` must be
5843 smaller than the bit size of the destination type, ``ty2``.
5848 The '``sext``' instruction performs a sign extension by copying the sign
5849 bit (highest order bit) of the ``value`` until it reaches the bit size
5850 of the type ``ty2``.
5852 When sign extending from i1, the extension always results in -1 or 0.
5857 .. code-block:: llvm
5859 %X = sext i8 -1 to i16 ; yields i16 :65535
5860 %Y = sext i1 true to i32 ; yields i32:-1
5861 %Z = sext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
5863 '``fptrunc .. to``' Instruction
5864 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5871 <result> = fptrunc <ty> <value> to <ty2> ; yields ty2
5876 The '``fptrunc``' instruction truncates ``value`` to type ``ty2``.
5881 The '``fptrunc``' instruction takes a :ref:`floating point <t_floating>`
5882 value to cast and a :ref:`floating point <t_floating>` type to cast it to.
5883 The size of ``value`` must be larger than the size of ``ty2``. This
5884 implies that ``fptrunc`` cannot be used to make a *no-op cast*.
5889 The '``fptrunc``' instruction truncates a ``value`` from a larger
5890 :ref:`floating point <t_floating>` type to a smaller :ref:`floating
5891 point <t_floating>` type. If the value cannot fit within the
5892 destination type, ``ty2``, then the results are undefined.
5897 .. code-block:: llvm
5899 %X = fptrunc double 123.0 to float ; yields float:123.0
5900 %Y = fptrunc double 1.0E+300 to float ; yields undefined
5902 '``fpext .. to``' Instruction
5903 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5910 <result> = fpext <ty> <value> to <ty2> ; yields ty2
5915 The '``fpext``' extends a floating point ``value`` to a larger floating
5921 The '``fpext``' instruction takes a :ref:`floating point <t_floating>`
5922 ``value`` to cast, and a :ref:`floating point <t_floating>` type to cast it
5923 to. The source type must be smaller than the destination type.
5928 The '``fpext``' instruction extends the ``value`` from a smaller
5929 :ref:`floating point <t_floating>` type to a larger :ref:`floating
5930 point <t_floating>` type. The ``fpext`` cannot be used to make a
5931 *no-op cast* because it always changes bits. Use ``bitcast`` to make a
5932 *no-op cast* for a floating point cast.
5937 .. code-block:: llvm
5939 %X = fpext float 3.125 to double ; yields double:3.125000e+00
5940 %Y = fpext double %X to fp128 ; yields fp128:0xL00000000000000004000900000000000
5942 '``fptoui .. to``' Instruction
5943 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5950 <result> = fptoui <ty> <value> to <ty2> ; yields ty2
5955 The '``fptoui``' converts a floating point ``value`` to its unsigned
5956 integer equivalent of type ``ty2``.
5961 The '``fptoui``' instruction takes a value to cast, which must be a
5962 scalar or vector :ref:`floating point <t_floating>` value, and a type to
5963 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
5964 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
5965 type with the same number of elements as ``ty``
5970 The '``fptoui``' instruction converts its :ref:`floating
5971 point <t_floating>` operand into the nearest (rounding towards zero)
5972 unsigned integer value. If the value cannot fit in ``ty2``, the results
5978 .. code-block:: llvm
5980 %X = fptoui double 123.0 to i32 ; yields i32:123
5981 %Y = fptoui float 1.0E+300 to i1 ; yields undefined:1
5982 %Z = fptoui float 1.04E+17 to i8 ; yields undefined:1
5984 '``fptosi .. to``' Instruction
5985 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5992 <result> = fptosi <ty> <value> to <ty2> ; yields ty2
5997 The '``fptosi``' instruction converts :ref:`floating point <t_floating>`
5998 ``value`` to type ``ty2``.
6003 The '``fptosi``' instruction takes a value to cast, which must be a
6004 scalar or vector :ref:`floating point <t_floating>` value, and a type to
6005 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
6006 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
6007 type with the same number of elements as ``ty``
6012 The '``fptosi``' instruction converts its :ref:`floating
6013 point <t_floating>` operand into the nearest (rounding towards zero)
6014 signed integer value. If the value cannot fit in ``ty2``, the results
6020 .. code-block:: llvm
6022 %X = fptosi double -123.0 to i32 ; yields i32:-123
6023 %Y = fptosi float 1.0E-247 to i1 ; yields undefined:1
6024 %Z = fptosi float 1.04E+17 to i8 ; yields undefined:1
6026 '``uitofp .. to``' Instruction
6027 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6034 <result> = uitofp <ty> <value> to <ty2> ; yields ty2
6039 The '``uitofp``' instruction regards ``value`` as an unsigned integer
6040 and converts that value to the ``ty2`` type.
6045 The '``uitofp``' instruction takes a value to cast, which must be a
6046 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
6047 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
6048 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
6049 type with the same number of elements as ``ty``
6054 The '``uitofp``' instruction interprets its operand as an unsigned
6055 integer quantity and converts it to the corresponding floating point
6056 value. If the value cannot fit in the floating point value, the results
6062 .. code-block:: llvm
6064 %X = uitofp i32 257 to float ; yields float:257.0
6065 %Y = uitofp i8 -1 to double ; yields double:255.0
6067 '``sitofp .. to``' Instruction
6068 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6075 <result> = sitofp <ty> <value> to <ty2> ; yields ty2
6080 The '``sitofp``' instruction regards ``value`` as a signed integer and
6081 converts that value to the ``ty2`` type.
6086 The '``sitofp``' instruction takes a value to cast, which must be a
6087 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
6088 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
6089 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
6090 type with the same number of elements as ``ty``
6095 The '``sitofp``' instruction interprets its operand as a signed integer
6096 quantity and converts it to the corresponding floating point value. If
6097 the value cannot fit in the floating point value, the results are
6103 .. code-block:: llvm
6105 %X = sitofp i32 257 to float ; yields float:257.0
6106 %Y = sitofp i8 -1 to double ; yields double:-1.0
6110 '``ptrtoint .. to``' Instruction
6111 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6118 <result> = ptrtoint <ty> <value> to <ty2> ; yields ty2
6123 The '``ptrtoint``' instruction converts the pointer or a vector of
6124 pointers ``value`` to the integer (or vector of integers) type ``ty2``.
6129 The '``ptrtoint``' instruction takes a ``value`` to cast, which must be
6130 a a value of type :ref:`pointer <t_pointer>` or a vector of pointers, and a
6131 type to cast it to ``ty2``, which must be an :ref:`integer <t_integer>` or
6132 a vector of integers type.
6137 The '``ptrtoint``' instruction converts ``value`` to integer type
6138 ``ty2`` by interpreting the pointer value as an integer and either
6139 truncating or zero extending that value to the size of the integer type.
6140 If ``value`` is smaller than ``ty2`` then a zero extension is done. If
6141 ``value`` is larger than ``ty2`` then a truncation is done. If they are
6142 the same size, then nothing is done (*no-op cast*) other than a type
6148 .. code-block:: llvm
6150 %X = ptrtoint i32* %P to i8 ; yields truncation on 32-bit architecture
6151 %Y = ptrtoint i32* %P to i64 ; yields zero extension on 32-bit architecture
6152 %Z = ptrtoint <4 x i32*> %P to <4 x i64>; yields vector zero extension for a vector of addresses on 32-bit architecture
6156 '``inttoptr .. to``' Instruction
6157 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6164 <result> = inttoptr <ty> <value> to <ty2> ; yields ty2
6169 The '``inttoptr``' instruction converts an integer ``value`` to a
6170 pointer type, ``ty2``.
6175 The '``inttoptr``' instruction takes an :ref:`integer <t_integer>` value to
6176 cast, and a type to cast it to, which must be a :ref:`pointer <t_pointer>`
6182 The '``inttoptr``' instruction converts ``value`` to type ``ty2`` by
6183 applying either a zero extension or a truncation depending on the size
6184 of the integer ``value``. If ``value`` is larger than the size of a
6185 pointer then a truncation is done. If ``value`` is smaller than the size
6186 of a pointer then a zero extension is done. If they are the same size,
6187 nothing is done (*no-op cast*).
6192 .. code-block:: llvm
6194 %X = inttoptr i32 255 to i32* ; yields zero extension on 64-bit architecture
6195 %Y = inttoptr i32 255 to i32* ; yields no-op on 32-bit architecture
6196 %Z = inttoptr i64 0 to i32* ; yields truncation on 32-bit architecture
6197 %Z = inttoptr <4 x i32> %G to <4 x i8*>; yields truncation of vector G to four pointers
6201 '``bitcast .. to``' Instruction
6202 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6209 <result> = bitcast <ty> <value> to <ty2> ; yields ty2
6214 The '``bitcast``' instruction converts ``value`` to type ``ty2`` without
6220 The '``bitcast``' instruction takes a value to cast, which must be a
6221 non-aggregate first class value, and a type to cast it to, which must
6222 also be a non-aggregate :ref:`first class <t_firstclass>` type. The
6223 bit sizes of ``value`` and the destination type, ``ty2``, must be
6224 identical. If the source type is a pointer, the destination type must
6225 also be a pointer of the same size. This instruction supports bitwise
6226 conversion of vectors to integers and to vectors of other types (as
6227 long as they have the same size).
6232 The '``bitcast``' instruction converts ``value`` to type ``ty2``. It
6233 is always a *no-op cast* because no bits change with this
6234 conversion. The conversion is done as if the ``value`` had been stored
6235 to memory and read back as type ``ty2``. Pointer (or vector of
6236 pointers) types may only be converted to other pointer (or vector of
6237 pointers) types with the same address space through this instruction.
6238 To convert pointers to other types, use the :ref:`inttoptr <i_inttoptr>`
6239 or :ref:`ptrtoint <i_ptrtoint>` instructions first.
6244 .. code-block:: llvm
6246 %X = bitcast i8 255 to i8 ; yields i8 :-1
6247 %Y = bitcast i32* %x to sint* ; yields sint*:%x
6248 %Z = bitcast <2 x int> %V to i64; ; yields i64: %V
6249 %Z = bitcast <2 x i32*> %V to <2 x i64*> ; yields <2 x i64*>
6251 .. _i_addrspacecast:
6253 '``addrspacecast .. to``' Instruction
6254 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6261 <result> = addrspacecast <pty> <ptrval> to <pty2> ; yields pty2
6266 The '``addrspacecast``' instruction converts ``ptrval`` from ``pty`` in
6267 address space ``n`` to type ``pty2`` in address space ``m``.
6272 The '``addrspacecast``' instruction takes a pointer or vector of pointer value
6273 to cast and a pointer type to cast it to, which must have a different
6279 The '``addrspacecast``' instruction converts the pointer value
6280 ``ptrval`` to type ``pty2``. It can be a *no-op cast* or a complex
6281 value modification, depending on the target and the address space
6282 pair. Pointer conversions within the same address space must be
6283 performed with the ``bitcast`` instruction. Note that if the address space
6284 conversion is legal then both result and operand refer to the same memory
6290 .. code-block:: llvm
6292 %X = addrspacecast i32* %x to i32 addrspace(1)* ; yields i32 addrspace(1)*:%x
6293 %Y = addrspacecast i32 addrspace(1)* %y to i64 addrspace(2)* ; yields i64 addrspace(2)*:%y
6294 %Z = addrspacecast <4 x i32*> %z to <4 x float addrspace(3)*> ; yields <4 x float addrspace(3)*>:%z
6301 The instructions in this category are the "miscellaneous" instructions,
6302 which defy better classification.
6306 '``icmp``' Instruction
6307 ^^^^^^^^^^^^^^^^^^^^^^
6314 <result> = icmp <cond> <ty> <op1>, <op2> ; yields i1 or <N x i1>:result
6319 The '``icmp``' instruction returns a boolean value or a vector of
6320 boolean values based on comparison of its two integer, integer vector,
6321 pointer, or pointer vector operands.
6326 The '``icmp``' instruction takes three operands. The first operand is
6327 the condition code indicating the kind of comparison to perform. It is
6328 not a value, just a keyword. The possible condition code are:
6331 #. ``ne``: not equal
6332 #. ``ugt``: unsigned greater than
6333 #. ``uge``: unsigned greater or equal
6334 #. ``ult``: unsigned less than
6335 #. ``ule``: unsigned less or equal
6336 #. ``sgt``: signed greater than
6337 #. ``sge``: signed greater or equal
6338 #. ``slt``: signed less than
6339 #. ``sle``: signed less or equal
6341 The remaining two arguments must be :ref:`integer <t_integer>` or
6342 :ref:`pointer <t_pointer>` or integer :ref:`vector <t_vector>` typed. They
6343 must also be identical types.
6348 The '``icmp``' compares ``op1`` and ``op2`` according to the condition
6349 code given as ``cond``. The comparison performed always yields either an
6350 :ref:`i1 <t_integer>` or vector of ``i1`` result, as follows:
6352 #. ``eq``: yields ``true`` if the operands are equal, ``false``
6353 otherwise. No sign interpretation is necessary or performed.
6354 #. ``ne``: yields ``true`` if the operands are unequal, ``false``
6355 otherwise. No sign interpretation is necessary or performed.
6356 #. ``ugt``: interprets the operands as unsigned values and yields
6357 ``true`` if ``op1`` is greater than ``op2``.
6358 #. ``uge``: interprets the operands as unsigned values and yields
6359 ``true`` if ``op1`` is greater than or equal to ``op2``.
6360 #. ``ult``: interprets the operands as unsigned values and yields
6361 ``true`` if ``op1`` is less than ``op2``.
6362 #. ``ule``: interprets the operands as unsigned values and yields
6363 ``true`` if ``op1`` is less than or equal to ``op2``.
6364 #. ``sgt``: interprets the operands as signed values and yields ``true``
6365 if ``op1`` is greater than ``op2``.
6366 #. ``sge``: interprets the operands as signed values and yields ``true``
6367 if ``op1`` is greater than or equal to ``op2``.
6368 #. ``slt``: interprets the operands as signed values and yields ``true``
6369 if ``op1`` is less than ``op2``.
6370 #. ``sle``: interprets the operands as signed values and yields ``true``
6371 if ``op1`` is less than or equal to ``op2``.
6373 If the operands are :ref:`pointer <t_pointer>` typed, the pointer values
6374 are compared as if they were integers.
6376 If the operands are integer vectors, then they are compared element by
6377 element. The result is an ``i1`` vector with the same number of elements
6378 as the values being compared. Otherwise, the result is an ``i1``.
6383 .. code-block:: llvm
6385 <result> = icmp eq i32 4, 5 ; yields: result=false
6386 <result> = icmp ne float* %X, %X ; yields: result=false
6387 <result> = icmp ult i16 4, 5 ; yields: result=true
6388 <result> = icmp sgt i16 4, 5 ; yields: result=false
6389 <result> = icmp ule i16 -4, 5 ; yields: result=false
6390 <result> = icmp sge i16 4, 5 ; yields: result=false
6392 Note that the code generator does not yet support vector types with the
6393 ``icmp`` instruction.
6397 '``fcmp``' Instruction
6398 ^^^^^^^^^^^^^^^^^^^^^^
6405 <result> = fcmp <cond> <ty> <op1>, <op2> ; yields i1 or <N x i1>:result
6410 The '``fcmp``' instruction returns a boolean value or vector of boolean
6411 values based on comparison of its operands.
6413 If the operands are floating point scalars, then the result type is a
6414 boolean (:ref:`i1 <t_integer>`).
6416 If the operands are floating point vectors, then the result type is a
6417 vector of boolean with the same number of elements as the operands being
6423 The '``fcmp``' instruction takes three operands. The first operand is
6424 the condition code indicating the kind of comparison to perform. It is
6425 not a value, just a keyword. The possible condition code are:
6427 #. ``false``: no comparison, always returns false
6428 #. ``oeq``: ordered and equal
6429 #. ``ogt``: ordered and greater than
6430 #. ``oge``: ordered and greater than or equal
6431 #. ``olt``: ordered and less than
6432 #. ``ole``: ordered and less than or equal
6433 #. ``one``: ordered and not equal
6434 #. ``ord``: ordered (no nans)
6435 #. ``ueq``: unordered or equal
6436 #. ``ugt``: unordered or greater than
6437 #. ``uge``: unordered or greater than or equal
6438 #. ``ult``: unordered or less than
6439 #. ``ule``: unordered or less than or equal
6440 #. ``une``: unordered or not equal
6441 #. ``uno``: unordered (either nans)
6442 #. ``true``: no comparison, always returns true
6444 *Ordered* means that neither operand is a QNAN while *unordered* means
6445 that either operand may be a QNAN.
6447 Each of ``val1`` and ``val2`` arguments must be either a :ref:`floating
6448 point <t_floating>` type or a :ref:`vector <t_vector>` of floating point
6449 type. They must have identical types.
6454 The '``fcmp``' instruction compares ``op1`` and ``op2`` according to the
6455 condition code given as ``cond``. If the operands are vectors, then the
6456 vectors are compared element by element. Each comparison performed
6457 always yields an :ref:`i1 <t_integer>` result, as follows:
6459 #. ``false``: always yields ``false``, regardless of operands.
6460 #. ``oeq``: yields ``true`` if both operands are not a QNAN and ``op1``
6461 is equal to ``op2``.
6462 #. ``ogt``: yields ``true`` if both operands are not a QNAN and ``op1``
6463 is greater than ``op2``.
6464 #. ``oge``: yields ``true`` if both operands are not a QNAN and ``op1``
6465 is greater than or equal to ``op2``.
6466 #. ``olt``: yields ``true`` if both operands are not a QNAN and ``op1``
6467 is less than ``op2``.
6468 #. ``ole``: yields ``true`` if both operands are not a QNAN and ``op1``
6469 is less than or equal to ``op2``.
6470 #. ``one``: yields ``true`` if both operands are not a QNAN and ``op1``
6471 is not equal to ``op2``.
6472 #. ``ord``: yields ``true`` if both operands are not a QNAN.
6473 #. ``ueq``: yields ``true`` if either operand is a QNAN or ``op1`` is
6475 #. ``ugt``: yields ``true`` if either operand is a QNAN or ``op1`` is
6476 greater than ``op2``.
6477 #. ``uge``: yields ``true`` if either operand is a QNAN or ``op1`` is
6478 greater than or equal to ``op2``.
6479 #. ``ult``: yields ``true`` if either operand is a QNAN or ``op1`` is
6481 #. ``ule``: yields ``true`` if either operand is a QNAN or ``op1`` is
6482 less than or equal to ``op2``.
6483 #. ``une``: yields ``true`` if either operand is a QNAN or ``op1`` is
6484 not equal to ``op2``.
6485 #. ``uno``: yields ``true`` if either operand is a QNAN.
6486 #. ``true``: always yields ``true``, regardless of operands.
6491 .. code-block:: llvm
6493 <result> = fcmp oeq float 4.0, 5.0 ; yields: result=false
6494 <result> = fcmp one float 4.0, 5.0 ; yields: result=true
6495 <result> = fcmp olt float 4.0, 5.0 ; yields: result=true
6496 <result> = fcmp ueq double 1.0, 2.0 ; yields: result=false
6498 Note that the code generator does not yet support vector types with the
6499 ``fcmp`` instruction.
6503 '``phi``' Instruction
6504 ^^^^^^^^^^^^^^^^^^^^^
6511 <result> = phi <ty> [ <val0>, <label0>], ...
6516 The '``phi``' instruction is used to implement the φ node in the SSA
6517 graph representing the function.
6522 The type of the incoming values is specified with the first type field.
6523 After this, the '``phi``' instruction takes a list of pairs as
6524 arguments, with one pair for each predecessor basic block of the current
6525 block. Only values of :ref:`first class <t_firstclass>` type may be used as
6526 the value arguments to the PHI node. Only labels may be used as the
6529 There must be no non-phi instructions between the start of a basic block
6530 and the PHI instructions: i.e. PHI instructions must be first in a basic
6533 For the purposes of the SSA form, the use of each incoming value is
6534 deemed to occur on the edge from the corresponding predecessor block to
6535 the current block (but after any definition of an '``invoke``'
6536 instruction's return value on the same edge).
6541 At runtime, the '``phi``' instruction logically takes on the value
6542 specified by the pair corresponding to the predecessor basic block that
6543 executed just prior to the current block.
6548 .. code-block:: llvm
6550 Loop: ; Infinite loop that counts from 0 on up...
6551 %indvar = phi i32 [ 0, %LoopHeader ], [ %nextindvar, %Loop ]
6552 %nextindvar = add i32 %indvar, 1
6557 '``select``' Instruction
6558 ^^^^^^^^^^^^^^^^^^^^^^^^
6565 <result> = select selty <cond>, <ty> <val1>, <ty> <val2> ; yields ty
6567 selty is either i1 or {<N x i1>}
6572 The '``select``' instruction is used to choose one value based on a
6573 condition, without IR-level branching.
6578 The '``select``' instruction requires an 'i1' value or a vector of 'i1'
6579 values indicating the condition, and two values of the same :ref:`first
6580 class <t_firstclass>` type. If the val1/val2 are vectors and the
6581 condition is a scalar, then entire vectors are selected, not individual
6587 If the condition is an i1 and it evaluates to 1, the instruction returns
6588 the first value argument; otherwise, it returns the second value
6591 If the condition is a vector of i1, then the value arguments must be
6592 vectors of the same size, and the selection is done element by element.
6597 .. code-block:: llvm
6599 %X = select i1 true, i8 17, i8 42 ; yields i8:17
6603 '``call``' Instruction
6604 ^^^^^^^^^^^^^^^^^^^^^^
6611 <result> = [tail | musttail] call [cconv] [ret attrs] <ty> [<fnty>*] <fnptrval>(<function args>) [fn attrs]
6616 The '``call``' instruction represents a simple function call.
6621 This instruction requires several arguments:
6623 #. The optional ``tail`` and ``musttail`` markers indicate that the optimizers
6624 should perform tail call optimization. The ``tail`` marker is a hint that
6625 `can be ignored <CodeGenerator.html#sibcallopt>`_. The ``musttail`` marker
6626 means that the call must be tail call optimized in order for the program to
6627 be correct. The ``musttail`` marker provides these guarantees:
6629 #. The call will not cause unbounded stack growth if it is part of a
6630 recursive cycle in the call graph.
6631 #. Arguments with the :ref:`inalloca <attr_inalloca>` attribute are
6634 Both markers imply that the callee does not access allocas or varargs from
6635 the caller. Calls marked ``musttail`` must obey the following additional
6638 - The call must immediately precede a :ref:`ret <i_ret>` instruction,
6639 or a pointer bitcast followed by a ret instruction.
6640 - The ret instruction must return the (possibly bitcasted) value
6641 produced by the call or void.
6642 - The caller and callee prototypes must match. Pointer types of
6643 parameters or return types may differ in pointee type, but not
6645 - The calling conventions of the caller and callee must match.
6646 - All ABI-impacting function attributes, such as sret, byval, inreg,
6647 returned, and inalloca, must match.
6648 - The callee must be varargs iff the caller is varargs. Bitcasting a
6649 non-varargs function to the appropriate varargs type is legal so
6650 long as the non-varargs prefixes obey the other rules.
6652 Tail call optimization for calls marked ``tail`` is guaranteed to occur if
6653 the following conditions are met:
6655 - Caller and callee both have the calling convention ``fastcc``.
6656 - The call is in tail position (ret immediately follows call and ret
6657 uses value of call or is void).
6658 - Option ``-tailcallopt`` is enabled, or
6659 ``llvm::GuaranteedTailCallOpt`` is ``true``.
6660 - `Platform-specific constraints are
6661 met. <CodeGenerator.html#tailcallopt>`_
6663 #. The optional "cconv" marker indicates which :ref:`calling
6664 convention <callingconv>` the call should use. If none is
6665 specified, the call defaults to using C calling conventions. The
6666 calling convention of the call must match the calling convention of
6667 the target function, or else the behavior is undefined.
6668 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
6669 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
6671 #. '``ty``': the type of the call instruction itself which is also the
6672 type of the return value. Functions that return no value are marked
6674 #. '``fnty``': shall be the signature of the pointer to function value
6675 being invoked. The argument types must match the types implied by
6676 this signature. This type can be omitted if the function is not
6677 varargs and if the function type does not return a pointer to a
6679 #. '``fnptrval``': An LLVM value containing a pointer to a function to
6680 be invoked. In most cases, this is a direct function invocation, but
6681 indirect ``call``'s are just as possible, calling an arbitrary pointer
6683 #. '``function args``': argument list whose types match the function
6684 signature argument types and parameter attributes. All arguments must
6685 be of :ref:`first class <t_firstclass>` type. If the function signature
6686 indicates the function accepts a variable number of arguments, the
6687 extra arguments can be specified.
6688 #. The optional :ref:`function attributes <fnattrs>` list. Only
6689 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
6690 attributes are valid here.
6695 The '``call``' instruction is used to cause control flow to transfer to
6696 a specified function, with its incoming arguments bound to the specified
6697 values. Upon a '``ret``' instruction in the called function, control
6698 flow continues with the instruction after the function call, and the
6699 return value of the function is bound to the result argument.
6704 .. code-block:: llvm
6706 %retval = call i32 @test(i32 %argc)
6707 call i32 (i8*, ...)* @printf(i8* %msg, i32 12, i8 42) ; yields i32
6708 %X = tail call i32 @foo() ; yields i32
6709 %Y = tail call fastcc i32 @foo() ; yields i32
6710 call void %foo(i8 97 signext)
6712 %struct.A = type { i32, i8 }
6713 %r = call %struct.A @foo() ; yields { i32, i8 }
6714 %gr = extractvalue %struct.A %r, 0 ; yields i32
6715 %gr1 = extractvalue %struct.A %r, 1 ; yields i8
6716 %Z = call void @foo() noreturn ; indicates that %foo never returns normally
6717 %ZZ = call zeroext i32 @bar() ; Return value is %zero extended
6719 llvm treats calls to some functions with names and arguments that match
6720 the standard C99 library as being the C99 library functions, and may
6721 perform optimizations or generate code for them under that assumption.
6722 This is something we'd like to change in the future to provide better
6723 support for freestanding environments and non-C-based languages.
6727 '``va_arg``' Instruction
6728 ^^^^^^^^^^^^^^^^^^^^^^^^
6735 <resultval> = va_arg <va_list*> <arglist>, <argty>
6740 The '``va_arg``' instruction is used to access arguments passed through
6741 the "variable argument" area of a function call. It is used to implement
6742 the ``va_arg`` macro in C.
6747 This instruction takes a ``va_list*`` value and the type of the
6748 argument. It returns a value of the specified argument type and
6749 increments the ``va_list`` to point to the next argument. The actual
6750 type of ``va_list`` is target specific.
6755 The '``va_arg``' instruction loads an argument of the specified type
6756 from the specified ``va_list`` and causes the ``va_list`` to point to
6757 the next argument. For more information, see the variable argument
6758 handling :ref:`Intrinsic Functions <int_varargs>`.
6760 It is legal for this instruction to be called in a function which does
6761 not take a variable number of arguments, for example, the ``vfprintf``
6764 ``va_arg`` is an LLVM instruction instead of an :ref:`intrinsic
6765 function <intrinsics>` because it takes a type as an argument.
6770 See the :ref:`variable argument processing <int_varargs>` section.
6772 Note that the code generator does not yet fully support va\_arg on many
6773 targets. Also, it does not currently support va\_arg with aggregate
6774 types on any target.
6778 '``landingpad``' Instruction
6779 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6786 <resultval> = landingpad <resultty> personality <type> <pers_fn> <clause>+
6787 <resultval> = landingpad <resultty> personality <type> <pers_fn> cleanup <clause>*
6789 <clause> := catch <type> <value>
6790 <clause> := filter <array constant type> <array constant>
6795 The '``landingpad``' instruction is used by `LLVM's exception handling
6796 system <ExceptionHandling.html#overview>`_ to specify that a basic block
6797 is a landing pad --- one where the exception lands, and corresponds to the
6798 code found in the ``catch`` portion of a ``try``/``catch`` sequence. It
6799 defines values supplied by the personality function (``pers_fn``) upon
6800 re-entry to the function. The ``resultval`` has the type ``resultty``.
6805 This instruction takes a ``pers_fn`` value. This is the personality
6806 function associated with the unwinding mechanism. The optional
6807 ``cleanup`` flag indicates that the landing pad block is a cleanup.
6809 A ``clause`` begins with the clause type --- ``catch`` or ``filter`` --- and
6810 contains the global variable representing the "type" that may be caught
6811 or filtered respectively. Unlike the ``catch`` clause, the ``filter``
6812 clause takes an array constant as its argument. Use
6813 "``[0 x i8**] undef``" for a filter which cannot throw. The
6814 '``landingpad``' instruction must contain *at least* one ``clause`` or
6815 the ``cleanup`` flag.
6820 The '``landingpad``' instruction defines the values which are set by the
6821 personality function (``pers_fn``) upon re-entry to the function, and
6822 therefore the "result type" of the ``landingpad`` instruction. As with
6823 calling conventions, how the personality function results are
6824 represented in LLVM IR is target specific.
6826 The clauses are applied in order from top to bottom. If two
6827 ``landingpad`` instructions are merged together through inlining, the
6828 clauses from the calling function are appended to the list of clauses.
6829 When the call stack is being unwound due to an exception being thrown,
6830 the exception is compared against each ``clause`` in turn. If it doesn't
6831 match any of the clauses, and the ``cleanup`` flag is not set, then
6832 unwinding continues further up the call stack.
6834 The ``landingpad`` instruction has several restrictions:
6836 - A landing pad block is a basic block which is the unwind destination
6837 of an '``invoke``' instruction.
6838 - A landing pad block must have a '``landingpad``' instruction as its
6839 first non-PHI instruction.
6840 - There can be only one '``landingpad``' instruction within the landing
6842 - A basic block that is not a landing pad block may not include a
6843 '``landingpad``' instruction.
6844 - All '``landingpad``' instructions in a function must have the same
6845 personality function.
6850 .. code-block:: llvm
6852 ;; A landing pad which can catch an integer.
6853 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6855 ;; A landing pad that is a cleanup.
6856 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6858 ;; A landing pad which can catch an integer and can only throw a double.
6859 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6861 filter [1 x i8**] [@_ZTId]
6868 LLVM supports the notion of an "intrinsic function". These functions
6869 have well known names and semantics and are required to follow certain
6870 restrictions. Overall, these intrinsics represent an extension mechanism
6871 for the LLVM language that does not require changing all of the
6872 transformations in LLVM when adding to the language (or the bitcode
6873 reader/writer, the parser, etc...).
6875 Intrinsic function names must all start with an "``llvm.``" prefix. This
6876 prefix is reserved in LLVM for intrinsic names; thus, function names may
6877 not begin with this prefix. Intrinsic functions must always be external
6878 functions: you cannot define the body of intrinsic functions. Intrinsic
6879 functions may only be used in call or invoke instructions: it is illegal
6880 to take the address of an intrinsic function. Additionally, because
6881 intrinsic functions are part of the LLVM language, it is required if any
6882 are added that they be documented here.
6884 Some intrinsic functions can be overloaded, i.e., the intrinsic
6885 represents a family of functions that perform the same operation but on
6886 different data types. Because LLVM can represent over 8 million
6887 different integer types, overloading is used commonly to allow an
6888 intrinsic function to operate on any integer type. One or more of the
6889 argument types or the result type can be overloaded to accept any
6890 integer type. Argument types may also be defined as exactly matching a
6891 previous argument's type or the result type. This allows an intrinsic
6892 function which accepts multiple arguments, but needs all of them to be
6893 of the same type, to only be overloaded with respect to a single
6894 argument or the result.
6896 Overloaded intrinsics will have the names of its overloaded argument
6897 types encoded into its function name, each preceded by a period. Only
6898 those types which are overloaded result in a name suffix. Arguments
6899 whose type is matched against another type do not. For example, the
6900 ``llvm.ctpop`` function can take an integer of any width and returns an
6901 integer of exactly the same integer width. This leads to a family of
6902 functions such as ``i8 @llvm.ctpop.i8(i8 %val)`` and
6903 ``i29 @llvm.ctpop.i29(i29 %val)``. Only one type, the return type, is
6904 overloaded, and only one type suffix is required. Because the argument's
6905 type is matched against the return type, it does not require its own
6908 To learn how to add an intrinsic function, please see the `Extending
6909 LLVM Guide <ExtendingLLVM.html>`_.
6913 Variable Argument Handling Intrinsics
6914 -------------------------------------
6916 Variable argument support is defined in LLVM with the
6917 :ref:`va_arg <i_va_arg>` instruction and these three intrinsic
6918 functions. These functions are related to the similarly named macros
6919 defined in the ``<stdarg.h>`` header file.
6921 All of these functions operate on arguments that use a target-specific
6922 value type "``va_list``". The LLVM assembly language reference manual
6923 does not define what this type is, so all transformations should be
6924 prepared to handle these functions regardless of the type used.
6926 This example shows how the :ref:`va_arg <i_va_arg>` instruction and the
6927 variable argument handling intrinsic functions are used.
6929 .. code-block:: llvm
6931 ; This struct is different for every platform. For most platforms,
6932 ; it is merely an i8*.
6933 %struct.va_list = type { i8* }
6935 ; For Unix x86_64 platforms, va_list is the following struct:
6936 ; %struct.va_list = type { i32, i32, i8*, i8* }
6938 define i32 @test(i32 %X, ...) {
6939 ; Initialize variable argument processing
6940 %ap = alloca %struct.va_list
6941 %ap2 = bitcast %struct.va_list* %ap to i8*
6942 call void @llvm.va_start(i8* %ap2)
6944 ; Read a single integer argument
6945 %tmp = va_arg i8* %ap2, i32
6947 ; Demonstrate usage of llvm.va_copy and llvm.va_end
6949 %aq2 = bitcast i8** %aq to i8*
6950 call void @llvm.va_copy(i8* %aq2, i8* %ap2)
6951 call void @llvm.va_end(i8* %aq2)
6953 ; Stop processing of arguments.
6954 call void @llvm.va_end(i8* %ap2)
6958 declare void @llvm.va_start(i8*)
6959 declare void @llvm.va_copy(i8*, i8*)
6960 declare void @llvm.va_end(i8*)
6964 '``llvm.va_start``' Intrinsic
6965 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6972 declare void @llvm.va_start(i8* <arglist>)
6977 The '``llvm.va_start``' intrinsic initializes ``*<arglist>`` for
6978 subsequent use by ``va_arg``.
6983 The argument is a pointer to a ``va_list`` element to initialize.
6988 The '``llvm.va_start``' intrinsic works just like the ``va_start`` macro
6989 available in C. In a target-dependent way, it initializes the
6990 ``va_list`` element to which the argument points, so that the next call
6991 to ``va_arg`` will produce the first variable argument passed to the
6992 function. Unlike the C ``va_start`` macro, this intrinsic does not need
6993 to know the last argument of the function as the compiler can figure
6996 '``llvm.va_end``' Intrinsic
6997 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7004 declare void @llvm.va_end(i8* <arglist>)
7009 The '``llvm.va_end``' intrinsic destroys ``*<arglist>``, which has been
7010 initialized previously with ``llvm.va_start`` or ``llvm.va_copy``.
7015 The argument is a pointer to a ``va_list`` to destroy.
7020 The '``llvm.va_end``' intrinsic works just like the ``va_end`` macro
7021 available in C. In a target-dependent way, it destroys the ``va_list``
7022 element to which the argument points. Calls to
7023 :ref:`llvm.va_start <int_va_start>` and
7024 :ref:`llvm.va_copy <int_va_copy>` must be matched exactly with calls to
7029 '``llvm.va_copy``' Intrinsic
7030 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7037 declare void @llvm.va_copy(i8* <destarglist>, i8* <srcarglist>)
7042 The '``llvm.va_copy``' intrinsic copies the current argument position
7043 from the source argument list to the destination argument list.
7048 The first argument is a pointer to a ``va_list`` element to initialize.
7049 The second argument is a pointer to a ``va_list`` element to copy from.
7054 The '``llvm.va_copy``' intrinsic works just like the ``va_copy`` macro
7055 available in C. In a target-dependent way, it copies the source
7056 ``va_list`` element into the destination ``va_list`` element. This
7057 intrinsic is necessary because the `` llvm.va_start`` intrinsic may be
7058 arbitrarily complex and require, for example, memory allocation.
7060 Accurate Garbage Collection Intrinsics
7061 --------------------------------------
7063 LLVM support for `Accurate Garbage Collection <GarbageCollection.html>`_
7064 (GC) requires the implementation and generation of these intrinsics.
7065 These intrinsics allow identification of :ref:`GC roots on the
7066 stack <int_gcroot>`, as well as garbage collector implementations that
7067 require :ref:`read <int_gcread>` and :ref:`write <int_gcwrite>` barriers.
7068 Front-ends for type-safe garbage collected languages should generate
7069 these intrinsics to make use of the LLVM garbage collectors. For more
7070 details, see `Accurate Garbage Collection with
7071 LLVM <GarbageCollection.html>`_.
7073 The garbage collection intrinsics only operate on objects in the generic
7074 address space (address space zero).
7078 '``llvm.gcroot``' Intrinsic
7079 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7086 declare void @llvm.gcroot(i8** %ptrloc, i8* %metadata)
7091 The '``llvm.gcroot``' intrinsic declares the existence of a GC root to
7092 the code generator, and allows some metadata to be associated with it.
7097 The first argument specifies the address of a stack object that contains
7098 the root pointer. The second pointer (which must be either a constant or
7099 a global value address) contains the meta-data to be associated with the
7105 At runtime, a call to this intrinsic stores a null pointer into the
7106 "ptrloc" location. At compile-time, the code generator generates
7107 information to allow the runtime to find the pointer at GC safe points.
7108 The '``llvm.gcroot``' intrinsic may only be used in a function which
7109 :ref:`specifies a GC algorithm <gc>`.
7113 '``llvm.gcread``' Intrinsic
7114 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7121 declare i8* @llvm.gcread(i8* %ObjPtr, i8** %Ptr)
7126 The '``llvm.gcread``' intrinsic identifies reads of references from heap
7127 locations, allowing garbage collector implementations that require read
7133 The second argument is the address to read from, which should be an
7134 address allocated from the garbage collector. The first object is a
7135 pointer to the start of the referenced object, if needed by the language
7136 runtime (otherwise null).
7141 The '``llvm.gcread``' intrinsic has the same semantics as a load
7142 instruction, but may be replaced with substantially more complex code by
7143 the garbage collector runtime, as needed. The '``llvm.gcread``'
7144 intrinsic may only be used in a function which :ref:`specifies a GC
7149 '``llvm.gcwrite``' Intrinsic
7150 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7157 declare void @llvm.gcwrite(i8* %P1, i8* %Obj, i8** %P2)
7162 The '``llvm.gcwrite``' intrinsic identifies writes of references to heap
7163 locations, allowing garbage collector implementations that require write
7164 barriers (such as generational or reference counting collectors).
7169 The first argument is the reference to store, the second is the start of
7170 the object to store it to, and the third is the address of the field of
7171 Obj to store to. If the runtime does not require a pointer to the
7172 object, Obj may be null.
7177 The '``llvm.gcwrite``' intrinsic has the same semantics as a store
7178 instruction, but may be replaced with substantially more complex code by
7179 the garbage collector runtime, as needed. The '``llvm.gcwrite``'
7180 intrinsic may only be used in a function which :ref:`specifies a GC
7183 Code Generator Intrinsics
7184 -------------------------
7186 These intrinsics are provided by LLVM to expose special features that
7187 may only be implemented with code generator support.
7189 '``llvm.returnaddress``' Intrinsic
7190 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7197 declare i8 *@llvm.returnaddress(i32 <level>)
7202 The '``llvm.returnaddress``' intrinsic attempts to compute a
7203 target-specific value indicating the return address of the current
7204 function or one of its callers.
7209 The argument to this intrinsic indicates which function to return the
7210 address for. Zero indicates the calling function, one indicates its
7211 caller, etc. The argument is **required** to be a constant integer
7217 The '``llvm.returnaddress``' intrinsic either returns a pointer
7218 indicating the return address of the specified call frame, or zero if it
7219 cannot be identified. The value returned by this intrinsic is likely to
7220 be incorrect or 0 for arguments other than zero, so it should only be
7221 used for debugging purposes.
7223 Note that calling this intrinsic does not prevent function inlining or
7224 other aggressive transformations, so the value returned may not be that
7225 of the obvious source-language caller.
7227 '``llvm.frameaddress``' Intrinsic
7228 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7235 declare i8* @llvm.frameaddress(i32 <level>)
7240 The '``llvm.frameaddress``' intrinsic attempts to return the
7241 target-specific frame pointer value for the specified stack frame.
7246 The argument to this intrinsic indicates which function to return the
7247 frame pointer for. Zero indicates the calling function, one indicates
7248 its caller, etc. The argument is **required** to be a constant integer
7254 The '``llvm.frameaddress``' intrinsic either returns a pointer
7255 indicating the frame address of the specified call frame, or zero if it
7256 cannot be identified. The value returned by this intrinsic is likely to
7257 be incorrect or 0 for arguments other than zero, so it should only be
7258 used for debugging purposes.
7260 Note that calling this intrinsic does not prevent function inlining or
7261 other aggressive transformations, so the value returned may not be that
7262 of the obvious source-language caller.
7264 .. _int_read_register:
7265 .. _int_write_register:
7267 '``llvm.read_register``' and '``llvm.write_register``' Intrinsics
7268 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7275 declare i32 @llvm.read_register.i32(metadata)
7276 declare i64 @llvm.read_register.i64(metadata)
7277 declare void @llvm.write_register.i32(metadata, i32 @value)
7278 declare void @llvm.write_register.i64(metadata, i64 @value)
7284 The '``llvm.read_register``' and '``llvm.write_register``' intrinsics
7285 provides access to the named register. The register must be valid on
7286 the architecture being compiled to. The type needs to be compatible
7287 with the register being read.
7292 The '``llvm.read_register``' intrinsic returns the current value of the
7293 register, where possible. The '``llvm.write_register``' intrinsic sets
7294 the current value of the register, where possible.
7296 This is useful to implement named register global variables that need
7297 to always be mapped to a specific register, as is common practice on
7298 bare-metal programs including OS kernels.
7300 The compiler doesn't check for register availability or use of the used
7301 register in surrounding code, including inline assembly. Because of that,
7302 allocatable registers are not supported.
7304 Warning: So far it only works with the stack pointer on selected
7305 architectures (ARM, AArch64, PowerPC and x86_64). Significant amount of
7306 work is needed to support other registers and even more so, allocatable
7311 '``llvm.stacksave``' Intrinsic
7312 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7319 declare i8* @llvm.stacksave()
7324 The '``llvm.stacksave``' intrinsic is used to remember the current state
7325 of the function stack, for use with
7326 :ref:`llvm.stackrestore <int_stackrestore>`. This is useful for
7327 implementing language features like scoped automatic variable sized
7333 This intrinsic returns a opaque pointer value that can be passed to
7334 :ref:`llvm.stackrestore <int_stackrestore>`. When an
7335 ``llvm.stackrestore`` intrinsic is executed with a value saved from
7336 ``llvm.stacksave``, it effectively restores the state of the stack to
7337 the state it was in when the ``llvm.stacksave`` intrinsic executed. In
7338 practice, this pops any :ref:`alloca <i_alloca>` blocks from the stack that
7339 were allocated after the ``llvm.stacksave`` was executed.
7341 .. _int_stackrestore:
7343 '``llvm.stackrestore``' Intrinsic
7344 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7351 declare void @llvm.stackrestore(i8* %ptr)
7356 The '``llvm.stackrestore``' intrinsic is used to restore the state of
7357 the function stack to the state it was in when the corresponding
7358 :ref:`llvm.stacksave <int_stacksave>` intrinsic executed. This is
7359 useful for implementing language features like scoped automatic variable
7360 sized arrays in C99.
7365 See the description for :ref:`llvm.stacksave <int_stacksave>`.
7367 '``llvm.prefetch``' Intrinsic
7368 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7375 declare void @llvm.prefetch(i8* <address>, i32 <rw>, i32 <locality>, i32 <cache type>)
7380 The '``llvm.prefetch``' intrinsic is a hint to the code generator to
7381 insert a prefetch instruction if supported; otherwise, it is a noop.
7382 Prefetches have no effect on the behavior of the program but can change
7383 its performance characteristics.
7388 ``address`` is the address to be prefetched, ``rw`` is the specifier
7389 determining if the fetch should be for a read (0) or write (1), and
7390 ``locality`` is a temporal locality specifier ranging from (0) - no
7391 locality, to (3) - extremely local keep in cache. The ``cache type``
7392 specifies whether the prefetch is performed on the data (1) or
7393 instruction (0) cache. The ``rw``, ``locality`` and ``cache type``
7394 arguments must be constant integers.
7399 This intrinsic does not modify the behavior of the program. In
7400 particular, prefetches cannot trap and do not produce a value. On
7401 targets that support this intrinsic, the prefetch can provide hints to
7402 the processor cache for better performance.
7404 '``llvm.pcmarker``' Intrinsic
7405 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7412 declare void @llvm.pcmarker(i32 <id>)
7417 The '``llvm.pcmarker``' intrinsic is a method to export a Program
7418 Counter (PC) in a region of code to simulators and other tools. The
7419 method is target specific, but it is expected that the marker will use
7420 exported symbols to transmit the PC of the marker. The marker makes no
7421 guarantees that it will remain with any specific instruction after
7422 optimizations. It is possible that the presence of a marker will inhibit
7423 optimizations. The intended use is to be inserted after optimizations to
7424 allow correlations of simulation runs.
7429 ``id`` is a numerical id identifying the marker.
7434 This intrinsic does not modify the behavior of the program. Backends
7435 that do not support this intrinsic may ignore it.
7437 '``llvm.readcyclecounter``' Intrinsic
7438 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7445 declare i64 @llvm.readcyclecounter()
7450 The '``llvm.readcyclecounter``' intrinsic provides access to the cycle
7451 counter register (or similar low latency, high accuracy clocks) on those
7452 targets that support it. On X86, it should map to RDTSC. On Alpha, it
7453 should map to RPCC. As the backing counters overflow quickly (on the
7454 order of 9 seconds on alpha), this should only be used for small
7460 When directly supported, reading the cycle counter should not modify any
7461 memory. Implementations are allowed to either return a application
7462 specific value or a system wide value. On backends without support, this
7463 is lowered to a constant 0.
7465 Note that runtime support may be conditional on the privilege-level code is
7466 running at and the host platform.
7468 '``llvm.clear_cache``' Intrinsic
7469 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7476 declare void @llvm.clear_cache(i8*, i8*)
7481 The '``llvm.clear_cache``' intrinsic ensures visibility of modifications
7482 in the specified range to the execution unit of the processor. On
7483 targets with non-unified instruction and data cache, the implementation
7484 flushes the instruction cache.
7489 On platforms with coherent instruction and data caches (e.g. x86), this
7490 intrinsic is a nop. On platforms with non-coherent instruction and data
7491 cache (e.g. ARM, MIPS), the intrinsic is lowered either to appropriate
7492 instructions or a system call, if cache flushing requires special
7495 The default behavior is to emit a call to ``__clear_cache`` from the run
7498 This instrinsic does *not* empty the instruction pipeline. Modifications
7499 of the current function are outside the scope of the intrinsic.
7501 '``llvm.instrprof_increment``' Intrinsic
7502 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7509 declare void @llvm.instrprof_increment(i8* <name>, i64 <hash>,
7510 i32 <num-counters>, i32 <index>)
7515 The '``llvm.instrprof_increment``' intrinsic can be emitted by a
7516 frontend for use with instrumentation based profiling. These will be
7517 lowered by the ``-instrprof`` pass to generate execution counts of a
7523 The first argument is a pointer to a global variable containing the
7524 name of the entity being instrumented. This should generally be the
7525 (mangled) function name for a set of counters.
7527 The second argument is a hash value that can be used by the consumer
7528 of the profile data to detect changes to the instrumented source, and
7529 the third is the number of counters associated with ``name``. It is an
7530 error if ``hash`` or ``num-counters`` differ between two instances of
7531 ``instrprof_increment`` that refer to the same name.
7533 The last argument refers to which of the counters for ``name`` should
7534 be incremented. It should be a value between 0 and ``num-counters``.
7539 This intrinsic represents an increment of a profiling counter. It will
7540 cause the ``-instrprof`` pass to generate the appropriate data
7541 structures and the code to increment the appropriate value, in a
7542 format that can be written out by a compiler runtime and consumed via
7543 the ``llvm-profdata`` tool.
7545 Standard C Library Intrinsics
7546 -----------------------------
7548 LLVM provides intrinsics for a few important standard C library
7549 functions. These intrinsics allow source-language front-ends to pass
7550 information about the alignment of the pointer arguments to the code
7551 generator, providing opportunity for more efficient code generation.
7555 '``llvm.memcpy``' Intrinsic
7556 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7561 This is an overloaded intrinsic. You can use ``llvm.memcpy`` on any
7562 integer bit width and for different address spaces. Not all targets
7563 support all bit widths however.
7567 declare void @llvm.memcpy.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
7568 i32 <len>, i32 <align>, i1 <isvolatile>)
7569 declare void @llvm.memcpy.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
7570 i64 <len>, i32 <align>, i1 <isvolatile>)
7575 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
7576 source location to the destination location.
7578 Note that, unlike the standard libc function, the ``llvm.memcpy.*``
7579 intrinsics do not return a value, takes extra alignment/isvolatile
7580 arguments and the pointers can be in specified address spaces.
7585 The first argument is a pointer to the destination, the second is a
7586 pointer to the source. The third argument is an integer argument
7587 specifying the number of bytes to copy, the fourth argument is the
7588 alignment of the source and destination locations, and the fifth is a
7589 boolean indicating a volatile access.
7591 If the call to this intrinsic has an alignment value that is not 0 or 1,
7592 then the caller guarantees that both the source and destination pointers
7593 are aligned to that boundary.
7595 If the ``isvolatile`` parameter is ``true``, the ``llvm.memcpy`` call is
7596 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
7597 very cleanly specified and it is unwise to depend on it.
7602 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
7603 source location to the destination location, which are not allowed to
7604 overlap. It copies "len" bytes of memory over. If the argument is known
7605 to be aligned to some boundary, this can be specified as the fourth
7606 argument, otherwise it should be set to 0 or 1 (both meaning no alignment).
7608 '``llvm.memmove``' Intrinsic
7609 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7614 This is an overloaded intrinsic. You can use llvm.memmove on any integer
7615 bit width and for different address space. Not all targets support all
7620 declare void @llvm.memmove.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
7621 i32 <len>, i32 <align>, i1 <isvolatile>)
7622 declare void @llvm.memmove.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
7623 i64 <len>, i32 <align>, i1 <isvolatile>)
7628 The '``llvm.memmove.*``' intrinsics move a block of memory from the
7629 source location to the destination location. It is similar to the
7630 '``llvm.memcpy``' intrinsic but allows the two memory locations to
7633 Note that, unlike the standard libc function, the ``llvm.memmove.*``
7634 intrinsics do not return a value, takes extra alignment/isvolatile
7635 arguments and the pointers can be in specified address spaces.
7640 The first argument is a pointer to the destination, the second is a
7641 pointer to the source. The third argument is an integer argument
7642 specifying the number of bytes to copy, the fourth argument is the
7643 alignment of the source and destination locations, and the fifth is a
7644 boolean indicating a volatile access.
7646 If the call to this intrinsic has an alignment value that is not 0 or 1,
7647 then the caller guarantees that the source and destination pointers are
7648 aligned to that boundary.
7650 If the ``isvolatile`` parameter is ``true``, the ``llvm.memmove`` call
7651 is a :ref:`volatile operation <volatile>`. The detailed access behavior is
7652 not very cleanly specified and it is unwise to depend on it.
7657 The '``llvm.memmove.*``' intrinsics copy a block of memory from the
7658 source location to the destination location, which may overlap. It
7659 copies "len" bytes of memory over. If the argument is known to be
7660 aligned to some boundary, this can be specified as the fourth argument,
7661 otherwise it should be set to 0 or 1 (both meaning no alignment).
7663 '``llvm.memset.*``' Intrinsics
7664 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7669 This is an overloaded intrinsic. You can use llvm.memset on any integer
7670 bit width and for different address spaces. However, not all targets
7671 support all bit widths.
7675 declare void @llvm.memset.p0i8.i32(i8* <dest>, i8 <val>,
7676 i32 <len>, i32 <align>, i1 <isvolatile>)
7677 declare void @llvm.memset.p0i8.i64(i8* <dest>, i8 <val>,
7678 i64 <len>, i32 <align>, i1 <isvolatile>)
7683 The '``llvm.memset.*``' intrinsics fill a block of memory with a
7684 particular byte value.
7686 Note that, unlike the standard libc function, the ``llvm.memset``
7687 intrinsic does not return a value and takes extra alignment/volatile
7688 arguments. Also, the destination can be in an arbitrary address space.
7693 The first argument is a pointer to the destination to fill, the second
7694 is the byte value with which to fill it, the third argument is an
7695 integer argument specifying the number of bytes to fill, and the fourth
7696 argument is the known alignment of the destination location.
7698 If the call to this intrinsic has an alignment value that is not 0 or 1,
7699 then the caller guarantees that the destination pointer is aligned to
7702 If the ``isvolatile`` parameter is ``true``, the ``llvm.memset`` call is
7703 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
7704 very cleanly specified and it is unwise to depend on it.
7709 The '``llvm.memset.*``' intrinsics fill "len" bytes of memory starting
7710 at the destination location. If the argument is known to be aligned to
7711 some boundary, this can be specified as the fourth argument, otherwise
7712 it should be set to 0 or 1 (both meaning no alignment).
7714 '``llvm.sqrt.*``' Intrinsic
7715 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7720 This is an overloaded intrinsic. You can use ``llvm.sqrt`` on any
7721 floating point or vector of floating point type. Not all targets support
7726 declare float @llvm.sqrt.f32(float %Val)
7727 declare double @llvm.sqrt.f64(double %Val)
7728 declare x86_fp80 @llvm.sqrt.f80(x86_fp80 %Val)
7729 declare fp128 @llvm.sqrt.f128(fp128 %Val)
7730 declare ppc_fp128 @llvm.sqrt.ppcf128(ppc_fp128 %Val)
7735 The '``llvm.sqrt``' intrinsics return the sqrt of the specified operand,
7736 returning the same value as the libm '``sqrt``' functions would. Unlike
7737 ``sqrt`` in libm, however, ``llvm.sqrt`` has undefined behavior for
7738 negative numbers other than -0.0 (which allows for better optimization,
7739 because there is no need to worry about errno being set).
7740 ``llvm.sqrt(-0.0)`` is defined to return -0.0 like IEEE sqrt.
7745 The argument and return value are floating point numbers of the same
7751 This function returns the sqrt of the specified operand if it is a
7752 nonnegative floating point number.
7754 '``llvm.powi.*``' Intrinsic
7755 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7760 This is an overloaded intrinsic. You can use ``llvm.powi`` on any
7761 floating point or vector of floating point type. Not all targets support
7766 declare float @llvm.powi.f32(float %Val, i32 %power)
7767 declare double @llvm.powi.f64(double %Val, i32 %power)
7768 declare x86_fp80 @llvm.powi.f80(x86_fp80 %Val, i32 %power)
7769 declare fp128 @llvm.powi.f128(fp128 %Val, i32 %power)
7770 declare ppc_fp128 @llvm.powi.ppcf128(ppc_fp128 %Val, i32 %power)
7775 The '``llvm.powi.*``' intrinsics return the first operand raised to the
7776 specified (positive or negative) power. The order of evaluation of
7777 multiplications is not defined. When a vector of floating point type is
7778 used, the second argument remains a scalar integer value.
7783 The second argument is an integer power, and the first is a value to
7784 raise to that power.
7789 This function returns the first value raised to the second power with an
7790 unspecified sequence of rounding operations.
7792 '``llvm.sin.*``' Intrinsic
7793 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7798 This is an overloaded intrinsic. You can use ``llvm.sin`` on any
7799 floating point or vector of floating point type. Not all targets support
7804 declare float @llvm.sin.f32(float %Val)
7805 declare double @llvm.sin.f64(double %Val)
7806 declare x86_fp80 @llvm.sin.f80(x86_fp80 %Val)
7807 declare fp128 @llvm.sin.f128(fp128 %Val)
7808 declare ppc_fp128 @llvm.sin.ppcf128(ppc_fp128 %Val)
7813 The '``llvm.sin.*``' intrinsics return the sine of the operand.
7818 The argument and return value are floating point numbers of the same
7824 This function returns the sine of the specified operand, returning the
7825 same values as the libm ``sin`` functions would, and handles error
7826 conditions in the same way.
7828 '``llvm.cos.*``' Intrinsic
7829 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7834 This is an overloaded intrinsic. You can use ``llvm.cos`` on any
7835 floating point or vector of floating point type. Not all targets support
7840 declare float @llvm.cos.f32(float %Val)
7841 declare double @llvm.cos.f64(double %Val)
7842 declare x86_fp80 @llvm.cos.f80(x86_fp80 %Val)
7843 declare fp128 @llvm.cos.f128(fp128 %Val)
7844 declare ppc_fp128 @llvm.cos.ppcf128(ppc_fp128 %Val)
7849 The '``llvm.cos.*``' intrinsics return the cosine of the operand.
7854 The argument and return value are floating point numbers of the same
7860 This function returns the cosine of the specified operand, returning the
7861 same values as the libm ``cos`` functions would, and handles error
7862 conditions in the same way.
7864 '``llvm.pow.*``' Intrinsic
7865 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7870 This is an overloaded intrinsic. You can use ``llvm.pow`` on any
7871 floating point or vector of floating point type. Not all targets support
7876 declare float @llvm.pow.f32(float %Val, float %Power)
7877 declare double @llvm.pow.f64(double %Val, double %Power)
7878 declare x86_fp80 @llvm.pow.f80(x86_fp80 %Val, x86_fp80 %Power)
7879 declare fp128 @llvm.pow.f128(fp128 %Val, fp128 %Power)
7880 declare ppc_fp128 @llvm.pow.ppcf128(ppc_fp128 %Val, ppc_fp128 Power)
7885 The '``llvm.pow.*``' intrinsics return the first operand raised to the
7886 specified (positive or negative) power.
7891 The second argument is a floating point power, and the first is a value
7892 to raise to that power.
7897 This function returns the first value raised to the second power,
7898 returning the same values as the libm ``pow`` functions would, and
7899 handles error conditions in the same way.
7901 '``llvm.exp.*``' Intrinsic
7902 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7907 This is an overloaded intrinsic. You can use ``llvm.exp`` on any
7908 floating point or vector of floating point type. Not all targets support
7913 declare float @llvm.exp.f32(float %Val)
7914 declare double @llvm.exp.f64(double %Val)
7915 declare x86_fp80 @llvm.exp.f80(x86_fp80 %Val)
7916 declare fp128 @llvm.exp.f128(fp128 %Val)
7917 declare ppc_fp128 @llvm.exp.ppcf128(ppc_fp128 %Val)
7922 The '``llvm.exp.*``' intrinsics perform the exp function.
7927 The argument and return value are floating point numbers of the same
7933 This function returns the same values as the libm ``exp`` functions
7934 would, and handles error conditions in the same way.
7936 '``llvm.exp2.*``' Intrinsic
7937 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7942 This is an overloaded intrinsic. You can use ``llvm.exp2`` on any
7943 floating point or vector of floating point type. Not all targets support
7948 declare float @llvm.exp2.f32(float %Val)
7949 declare double @llvm.exp2.f64(double %Val)
7950 declare x86_fp80 @llvm.exp2.f80(x86_fp80 %Val)
7951 declare fp128 @llvm.exp2.f128(fp128 %Val)
7952 declare ppc_fp128 @llvm.exp2.ppcf128(ppc_fp128 %Val)
7957 The '``llvm.exp2.*``' intrinsics perform the exp2 function.
7962 The argument and return value are floating point numbers of the same
7968 This function returns the same values as the libm ``exp2`` functions
7969 would, and handles error conditions in the same way.
7971 '``llvm.log.*``' Intrinsic
7972 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7977 This is an overloaded intrinsic. You can use ``llvm.log`` on any
7978 floating point or vector of floating point type. Not all targets support
7983 declare float @llvm.log.f32(float %Val)
7984 declare double @llvm.log.f64(double %Val)
7985 declare x86_fp80 @llvm.log.f80(x86_fp80 %Val)
7986 declare fp128 @llvm.log.f128(fp128 %Val)
7987 declare ppc_fp128 @llvm.log.ppcf128(ppc_fp128 %Val)
7992 The '``llvm.log.*``' intrinsics perform the log function.
7997 The argument and return value are floating point numbers of the same
8003 This function returns the same values as the libm ``log`` functions
8004 would, and handles error conditions in the same way.
8006 '``llvm.log10.*``' Intrinsic
8007 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8012 This is an overloaded intrinsic. You can use ``llvm.log10`` on any
8013 floating point or vector of floating point type. Not all targets support
8018 declare float @llvm.log10.f32(float %Val)
8019 declare double @llvm.log10.f64(double %Val)
8020 declare x86_fp80 @llvm.log10.f80(x86_fp80 %Val)
8021 declare fp128 @llvm.log10.f128(fp128 %Val)
8022 declare ppc_fp128 @llvm.log10.ppcf128(ppc_fp128 %Val)
8027 The '``llvm.log10.*``' intrinsics perform the log10 function.
8032 The argument and return value are floating point numbers of the same
8038 This function returns the same values as the libm ``log10`` functions
8039 would, and handles error conditions in the same way.
8041 '``llvm.log2.*``' Intrinsic
8042 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8047 This is an overloaded intrinsic. You can use ``llvm.log2`` on any
8048 floating point or vector of floating point type. Not all targets support
8053 declare float @llvm.log2.f32(float %Val)
8054 declare double @llvm.log2.f64(double %Val)
8055 declare x86_fp80 @llvm.log2.f80(x86_fp80 %Val)
8056 declare fp128 @llvm.log2.f128(fp128 %Val)
8057 declare ppc_fp128 @llvm.log2.ppcf128(ppc_fp128 %Val)
8062 The '``llvm.log2.*``' intrinsics perform the log2 function.
8067 The argument and return value are floating point numbers of the same
8073 This function returns the same values as the libm ``log2`` functions
8074 would, and handles error conditions in the same way.
8076 '``llvm.fma.*``' Intrinsic
8077 ^^^^^^^^^^^^^^^^^^^^^^^^^^
8082 This is an overloaded intrinsic. You can use ``llvm.fma`` on any
8083 floating point or vector of floating point type. Not all targets support
8088 declare float @llvm.fma.f32(float %a, float %b, float %c)
8089 declare double @llvm.fma.f64(double %a, double %b, double %c)
8090 declare x86_fp80 @llvm.fma.f80(x86_fp80 %a, x86_fp80 %b, x86_fp80 %c)
8091 declare fp128 @llvm.fma.f128(fp128 %a, fp128 %b, fp128 %c)
8092 declare ppc_fp128 @llvm.fma.ppcf128(ppc_fp128 %a, ppc_fp128 %b, ppc_fp128 %c)
8097 The '``llvm.fma.*``' intrinsics perform the fused multiply-add
8103 The argument and return value are floating point numbers of the same
8109 This function returns the same values as the libm ``fma`` functions
8110 would, and does not set errno.
8112 '``llvm.fabs.*``' Intrinsic
8113 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8118 This is an overloaded intrinsic. You can use ``llvm.fabs`` on any
8119 floating point or vector of floating point type. Not all targets support
8124 declare float @llvm.fabs.f32(float %Val)
8125 declare double @llvm.fabs.f64(double %Val)
8126 declare x86_fp80 @llvm.fabs.f80(x86_fp80 %Val)
8127 declare fp128 @llvm.fabs.f128(fp128 %Val)
8128 declare ppc_fp128 @llvm.fabs.ppcf128(ppc_fp128 %Val)
8133 The '``llvm.fabs.*``' intrinsics return the absolute value of the
8139 The argument and return value are floating point numbers of the same
8145 This function returns the same values as the libm ``fabs`` functions
8146 would, and handles error conditions in the same way.
8148 '``llvm.minnum.*``' Intrinsic
8149 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8154 This is an overloaded intrinsic. You can use ``llvm.minnum`` on any
8155 floating point or vector of floating point type. Not all targets support
8160 declare float @llvm.minnum.f32(float %Val0, float %Val1)
8161 declare double @llvm.minnum.f64(double %Val0, double %Val1)
8162 declare x86_fp80 @llvm.minnum.f80(x86_fp80 %Val0, x86_fp80 %Val1)
8163 declare fp128 @llvm.minnum.f128(fp128 %Val0, fp128 %Val1)
8164 declare ppc_fp128 @llvm.minnum.ppcf128(ppc_fp128 %Val0, ppc_fp128 %Val1)
8169 The '``llvm.minnum.*``' intrinsics return the minimum of the two
8176 The arguments and return value are floating point numbers of the same
8182 Follows the IEEE-754 semantics for minNum, which also match for libm's
8185 If either operand is a NaN, returns the other non-NaN operand. Returns
8186 NaN only if both operands are NaN. If the operands compare equal,
8187 returns a value that compares equal to both operands. This means that
8188 fmin(+/-0.0, +/-0.0) could return either -0.0 or 0.0.
8190 '``llvm.maxnum.*``' Intrinsic
8191 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8196 This is an overloaded intrinsic. You can use ``llvm.maxnum`` on any
8197 floating point or vector of floating point type. Not all targets support
8202 declare float @llvm.maxnum.f32(float %Val0, float %Val1l)
8203 declare double @llvm.maxnum.f64(double %Val0, double %Val1)
8204 declare x86_fp80 @llvm.maxnum.f80(x86_fp80 %Val0, x86_fp80 %Val1)
8205 declare fp128 @llvm.maxnum.f128(fp128 %Val0, fp128 %Val1)
8206 declare ppc_fp128 @llvm.maxnum.ppcf128(ppc_fp128 %Val0, ppc_fp128 %Val1)
8211 The '``llvm.maxnum.*``' intrinsics return the maximum of the two
8218 The arguments and return value are floating point numbers of the same
8223 Follows the IEEE-754 semantics for maxNum, which also match for libm's
8226 If either operand is a NaN, returns the other non-NaN operand. Returns
8227 NaN only if both operands are NaN. If the operands compare equal,
8228 returns a value that compares equal to both operands. This means that
8229 fmax(+/-0.0, +/-0.0) could return either -0.0 or 0.0.
8231 '``llvm.copysign.*``' Intrinsic
8232 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8237 This is an overloaded intrinsic. You can use ``llvm.copysign`` on any
8238 floating point or vector of floating point type. Not all targets support
8243 declare float @llvm.copysign.f32(float %Mag, float %Sgn)
8244 declare double @llvm.copysign.f64(double %Mag, double %Sgn)
8245 declare x86_fp80 @llvm.copysign.f80(x86_fp80 %Mag, x86_fp80 %Sgn)
8246 declare fp128 @llvm.copysign.f128(fp128 %Mag, fp128 %Sgn)
8247 declare ppc_fp128 @llvm.copysign.ppcf128(ppc_fp128 %Mag, ppc_fp128 %Sgn)
8252 The '``llvm.copysign.*``' intrinsics return a value with the magnitude of the
8253 first operand and the sign of the second operand.
8258 The arguments and return value are floating point numbers of the same
8264 This function returns the same values as the libm ``copysign``
8265 functions would, and handles error conditions in the same way.
8267 '``llvm.floor.*``' Intrinsic
8268 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8273 This is an overloaded intrinsic. You can use ``llvm.floor`` on any
8274 floating point or vector of floating point type. Not all targets support
8279 declare float @llvm.floor.f32(float %Val)
8280 declare double @llvm.floor.f64(double %Val)
8281 declare x86_fp80 @llvm.floor.f80(x86_fp80 %Val)
8282 declare fp128 @llvm.floor.f128(fp128 %Val)
8283 declare ppc_fp128 @llvm.floor.ppcf128(ppc_fp128 %Val)
8288 The '``llvm.floor.*``' intrinsics return the floor of the operand.
8293 The argument and return value are floating point numbers of the same
8299 This function returns the same values as the libm ``floor`` functions
8300 would, and handles error conditions in the same way.
8302 '``llvm.ceil.*``' Intrinsic
8303 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8308 This is an overloaded intrinsic. You can use ``llvm.ceil`` on any
8309 floating point or vector of floating point type. Not all targets support
8314 declare float @llvm.ceil.f32(float %Val)
8315 declare double @llvm.ceil.f64(double %Val)
8316 declare x86_fp80 @llvm.ceil.f80(x86_fp80 %Val)
8317 declare fp128 @llvm.ceil.f128(fp128 %Val)
8318 declare ppc_fp128 @llvm.ceil.ppcf128(ppc_fp128 %Val)
8323 The '``llvm.ceil.*``' intrinsics return the ceiling of the operand.
8328 The argument and return value are floating point numbers of the same
8334 This function returns the same values as the libm ``ceil`` functions
8335 would, and handles error conditions in the same way.
8337 '``llvm.trunc.*``' Intrinsic
8338 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8343 This is an overloaded intrinsic. You can use ``llvm.trunc`` on any
8344 floating point or vector of floating point type. Not all targets support
8349 declare float @llvm.trunc.f32(float %Val)
8350 declare double @llvm.trunc.f64(double %Val)
8351 declare x86_fp80 @llvm.trunc.f80(x86_fp80 %Val)
8352 declare fp128 @llvm.trunc.f128(fp128 %Val)
8353 declare ppc_fp128 @llvm.trunc.ppcf128(ppc_fp128 %Val)
8358 The '``llvm.trunc.*``' intrinsics returns the operand rounded to the
8359 nearest integer not larger in magnitude than the operand.
8364 The argument and return value are floating point numbers of the same
8370 This function returns the same values as the libm ``trunc`` functions
8371 would, and handles error conditions in the same way.
8373 '``llvm.rint.*``' Intrinsic
8374 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8379 This is an overloaded intrinsic. You can use ``llvm.rint`` on any
8380 floating point or vector of floating point type. Not all targets support
8385 declare float @llvm.rint.f32(float %Val)
8386 declare double @llvm.rint.f64(double %Val)
8387 declare x86_fp80 @llvm.rint.f80(x86_fp80 %Val)
8388 declare fp128 @llvm.rint.f128(fp128 %Val)
8389 declare ppc_fp128 @llvm.rint.ppcf128(ppc_fp128 %Val)
8394 The '``llvm.rint.*``' intrinsics returns the operand rounded to the
8395 nearest integer. It may raise an inexact floating-point exception if the
8396 operand isn't an integer.
8401 The argument and return value are floating point numbers of the same
8407 This function returns the same values as the libm ``rint`` functions
8408 would, and handles error conditions in the same way.
8410 '``llvm.nearbyint.*``' Intrinsic
8411 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8416 This is an overloaded intrinsic. You can use ``llvm.nearbyint`` on any
8417 floating point or vector of floating point type. Not all targets support
8422 declare float @llvm.nearbyint.f32(float %Val)
8423 declare double @llvm.nearbyint.f64(double %Val)
8424 declare x86_fp80 @llvm.nearbyint.f80(x86_fp80 %Val)
8425 declare fp128 @llvm.nearbyint.f128(fp128 %Val)
8426 declare ppc_fp128 @llvm.nearbyint.ppcf128(ppc_fp128 %Val)
8431 The '``llvm.nearbyint.*``' intrinsics returns the operand rounded to the
8437 The argument and return value are floating point numbers of the same
8443 This function returns the same values as the libm ``nearbyint``
8444 functions would, and handles error conditions in the same way.
8446 '``llvm.round.*``' Intrinsic
8447 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8452 This is an overloaded intrinsic. You can use ``llvm.round`` on any
8453 floating point or vector of floating point type. Not all targets support
8458 declare float @llvm.round.f32(float %Val)
8459 declare double @llvm.round.f64(double %Val)
8460 declare x86_fp80 @llvm.round.f80(x86_fp80 %Val)
8461 declare fp128 @llvm.round.f128(fp128 %Val)
8462 declare ppc_fp128 @llvm.round.ppcf128(ppc_fp128 %Val)
8467 The '``llvm.round.*``' intrinsics returns the operand rounded to the
8473 The argument and return value are floating point numbers of the same
8479 This function returns the same values as the libm ``round``
8480 functions would, and handles error conditions in the same way.
8482 Bit Manipulation Intrinsics
8483 ---------------------------
8485 LLVM provides intrinsics for a few important bit manipulation
8486 operations. These allow efficient code generation for some algorithms.
8488 '``llvm.bswap.*``' Intrinsics
8489 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8494 This is an overloaded intrinsic function. You can use bswap on any
8495 integer type that is an even number of bytes (i.e. BitWidth % 16 == 0).
8499 declare i16 @llvm.bswap.i16(i16 <id>)
8500 declare i32 @llvm.bswap.i32(i32 <id>)
8501 declare i64 @llvm.bswap.i64(i64 <id>)
8506 The '``llvm.bswap``' family of intrinsics is used to byte swap integer
8507 values with an even number of bytes (positive multiple of 16 bits).
8508 These are useful for performing operations on data that is not in the
8509 target's native byte order.
8514 The ``llvm.bswap.i16`` intrinsic returns an i16 value that has the high
8515 and low byte of the input i16 swapped. Similarly, the ``llvm.bswap.i32``
8516 intrinsic returns an i32 value that has the four bytes of the input i32
8517 swapped, so that if the input bytes are numbered 0, 1, 2, 3 then the
8518 returned i32 will have its bytes in 3, 2, 1, 0 order. The
8519 ``llvm.bswap.i48``, ``llvm.bswap.i64`` and other intrinsics extend this
8520 concept to additional even-byte lengths (6 bytes, 8 bytes and more,
8523 '``llvm.ctpop.*``' Intrinsic
8524 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8529 This is an overloaded intrinsic. You can use llvm.ctpop on any integer
8530 bit width, or on any vector with integer elements. Not all targets
8531 support all bit widths or vector types, however.
8535 declare i8 @llvm.ctpop.i8(i8 <src>)
8536 declare i16 @llvm.ctpop.i16(i16 <src>)
8537 declare i32 @llvm.ctpop.i32(i32 <src>)
8538 declare i64 @llvm.ctpop.i64(i64 <src>)
8539 declare i256 @llvm.ctpop.i256(i256 <src>)
8540 declare <2 x i32> @llvm.ctpop.v2i32(<2 x i32> <src>)
8545 The '``llvm.ctpop``' family of intrinsics counts the number of bits set
8551 The only argument is the value to be counted. The argument may be of any
8552 integer type, or a vector with integer elements. The return type must
8553 match the argument type.
8558 The '``llvm.ctpop``' intrinsic counts the 1's in a variable, or within
8559 each element of a vector.
8561 '``llvm.ctlz.*``' Intrinsic
8562 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8567 This is an overloaded intrinsic. You can use ``llvm.ctlz`` on any
8568 integer bit width, or any vector whose elements are integers. Not all
8569 targets support all bit widths or vector types, however.
8573 declare i8 @llvm.ctlz.i8 (i8 <src>, i1 <is_zero_undef>)
8574 declare i16 @llvm.ctlz.i16 (i16 <src>, i1 <is_zero_undef>)
8575 declare i32 @llvm.ctlz.i32 (i32 <src>, i1 <is_zero_undef>)
8576 declare i64 @llvm.ctlz.i64 (i64 <src>, i1 <is_zero_undef>)
8577 declare i256 @llvm.ctlz.i256(i256 <src>, i1 <is_zero_undef>)
8578 declase <2 x i32> @llvm.ctlz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
8583 The '``llvm.ctlz``' family of intrinsic functions counts the number of
8584 leading zeros in a variable.
8589 The first argument is the value to be counted. This argument may be of
8590 any integer type, or a vectory with integer element type. The return
8591 type must match the first argument type.
8593 The second argument must be a constant and is a flag to indicate whether
8594 the intrinsic should ensure that a zero as the first argument produces a
8595 defined result. Historically some architectures did not provide a
8596 defined result for zero values as efficiently, and many algorithms are
8597 now predicated on avoiding zero-value inputs.
8602 The '``llvm.ctlz``' intrinsic counts the leading (most significant)
8603 zeros in a variable, or within each element of the vector. If
8604 ``src == 0`` then the result is the size in bits of the type of ``src``
8605 if ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
8606 ``llvm.ctlz(i32 2) = 30``.
8608 '``llvm.cttz.*``' Intrinsic
8609 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8614 This is an overloaded intrinsic. You can use ``llvm.cttz`` on any
8615 integer bit width, or any vector of integer elements. Not all targets
8616 support all bit widths or vector types, however.
8620 declare i8 @llvm.cttz.i8 (i8 <src>, i1 <is_zero_undef>)
8621 declare i16 @llvm.cttz.i16 (i16 <src>, i1 <is_zero_undef>)
8622 declare i32 @llvm.cttz.i32 (i32 <src>, i1 <is_zero_undef>)
8623 declare i64 @llvm.cttz.i64 (i64 <src>, i1 <is_zero_undef>)
8624 declare i256 @llvm.cttz.i256(i256 <src>, i1 <is_zero_undef>)
8625 declase <2 x i32> @llvm.cttz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
8630 The '``llvm.cttz``' family of intrinsic functions counts the number of
8636 The first argument is the value to be counted. This argument may be of
8637 any integer type, or a vectory with integer element type. The return
8638 type must match the first argument type.
8640 The second argument must be a constant and is a flag to indicate whether
8641 the intrinsic should ensure that a zero as the first argument produces a
8642 defined result. Historically some architectures did not provide a
8643 defined result for zero values as efficiently, and many algorithms are
8644 now predicated on avoiding zero-value inputs.
8649 The '``llvm.cttz``' intrinsic counts the trailing (least significant)
8650 zeros in a variable, or within each element of a vector. If ``src == 0``
8651 then the result is the size in bits of the type of ``src`` if
8652 ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
8653 ``llvm.cttz(2) = 1``.
8655 Arithmetic with Overflow Intrinsics
8656 -----------------------------------
8658 LLVM provides intrinsics for some arithmetic with overflow operations.
8660 '``llvm.sadd.with.overflow.*``' Intrinsics
8661 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8666 This is an overloaded intrinsic. You can use ``llvm.sadd.with.overflow``
8667 on any integer bit width.
8671 declare {i16, i1} @llvm.sadd.with.overflow.i16(i16 %a, i16 %b)
8672 declare {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
8673 declare {i64, i1} @llvm.sadd.with.overflow.i64(i64 %a, i64 %b)
8678 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
8679 a signed addition of the two arguments, and indicate whether an overflow
8680 occurred during the signed summation.
8685 The arguments (%a and %b) and the first element of the result structure
8686 may be of integer types of any bit width, but they must have the same
8687 bit width. The second element of the result structure must be of type
8688 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
8694 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
8695 a signed addition of the two variables. They return a structure --- the
8696 first element of which is the signed summation, and the second element
8697 of which is a bit specifying if the signed summation resulted in an
8703 .. code-block:: llvm
8705 %res = call {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
8706 %sum = extractvalue {i32, i1} %res, 0
8707 %obit = extractvalue {i32, i1} %res, 1
8708 br i1 %obit, label %overflow, label %normal
8710 '``llvm.uadd.with.overflow.*``' Intrinsics
8711 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8716 This is an overloaded intrinsic. You can use ``llvm.uadd.with.overflow``
8717 on any integer bit width.
8721 declare {i16, i1} @llvm.uadd.with.overflow.i16(i16 %a, i16 %b)
8722 declare {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
8723 declare {i64, i1} @llvm.uadd.with.overflow.i64(i64 %a, i64 %b)
8728 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
8729 an unsigned addition of the two arguments, and indicate whether a carry
8730 occurred during the unsigned summation.
8735 The arguments (%a and %b) and the first element of the result structure
8736 may be of integer types of any bit width, but they must have the same
8737 bit width. The second element of the result structure must be of type
8738 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
8744 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
8745 an unsigned addition of the two arguments. They return a structure --- the
8746 first element of which is the sum, and the second element of which is a
8747 bit specifying if the unsigned summation resulted in a carry.
8752 .. code-block:: llvm
8754 %res = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
8755 %sum = extractvalue {i32, i1} %res, 0
8756 %obit = extractvalue {i32, i1} %res, 1
8757 br i1 %obit, label %carry, label %normal
8759 '``llvm.ssub.with.overflow.*``' Intrinsics
8760 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8765 This is an overloaded intrinsic. You can use ``llvm.ssub.with.overflow``
8766 on any integer bit width.
8770 declare {i16, i1} @llvm.ssub.with.overflow.i16(i16 %a, i16 %b)
8771 declare {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
8772 declare {i64, i1} @llvm.ssub.with.overflow.i64(i64 %a, i64 %b)
8777 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
8778 a signed subtraction of the two arguments, and indicate whether an
8779 overflow occurred during the signed subtraction.
8784 The arguments (%a and %b) and the first element of the result structure
8785 may be of integer types of any bit width, but they must have the same
8786 bit width. The second element of the result structure must be of type
8787 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
8793 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
8794 a signed subtraction of the two arguments. They return a structure --- the
8795 first element of which is the subtraction, and the second element of
8796 which is a bit specifying if the signed subtraction resulted in an
8802 .. code-block:: llvm
8804 %res = call {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
8805 %sum = extractvalue {i32, i1} %res, 0
8806 %obit = extractvalue {i32, i1} %res, 1
8807 br i1 %obit, label %overflow, label %normal
8809 '``llvm.usub.with.overflow.*``' Intrinsics
8810 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8815 This is an overloaded intrinsic. You can use ``llvm.usub.with.overflow``
8816 on any integer bit width.
8820 declare {i16, i1} @llvm.usub.with.overflow.i16(i16 %a, i16 %b)
8821 declare {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
8822 declare {i64, i1} @llvm.usub.with.overflow.i64(i64 %a, i64 %b)
8827 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
8828 an unsigned subtraction of the two arguments, and indicate whether an
8829 overflow occurred during the unsigned subtraction.
8834 The arguments (%a and %b) and the first element of the result structure
8835 may be of integer types of any bit width, but they must have the same
8836 bit width. The second element of the result structure must be of type
8837 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
8843 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
8844 an unsigned subtraction of the two arguments. They return a structure ---
8845 the first element of which is the subtraction, and the second element of
8846 which is a bit specifying if the unsigned subtraction resulted in an
8852 .. code-block:: llvm
8854 %res = call {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
8855 %sum = extractvalue {i32, i1} %res, 0
8856 %obit = extractvalue {i32, i1} %res, 1
8857 br i1 %obit, label %overflow, label %normal
8859 '``llvm.smul.with.overflow.*``' Intrinsics
8860 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8865 This is an overloaded intrinsic. You can use ``llvm.smul.with.overflow``
8866 on any integer bit width.
8870 declare {i16, i1} @llvm.smul.with.overflow.i16(i16 %a, i16 %b)
8871 declare {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
8872 declare {i64, i1} @llvm.smul.with.overflow.i64(i64 %a, i64 %b)
8877 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
8878 a signed multiplication of the two arguments, and indicate whether an
8879 overflow occurred during the signed multiplication.
8884 The arguments (%a and %b) and the first element of the result structure
8885 may be of integer types of any bit width, but they must have the same
8886 bit width. The second element of the result structure must be of type
8887 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
8893 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
8894 a signed multiplication of the two arguments. They return a structure ---
8895 the first element of which is the multiplication, and the second element
8896 of which is a bit specifying if the signed multiplication resulted in an
8902 .. code-block:: llvm
8904 %res = call {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
8905 %sum = extractvalue {i32, i1} %res, 0
8906 %obit = extractvalue {i32, i1} %res, 1
8907 br i1 %obit, label %overflow, label %normal
8909 '``llvm.umul.with.overflow.*``' Intrinsics
8910 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8915 This is an overloaded intrinsic. You can use ``llvm.umul.with.overflow``
8916 on any integer bit width.
8920 declare {i16, i1} @llvm.umul.with.overflow.i16(i16 %a, i16 %b)
8921 declare {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
8922 declare {i64, i1} @llvm.umul.with.overflow.i64(i64 %a, i64 %b)
8927 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
8928 a unsigned multiplication of the two arguments, and indicate whether an
8929 overflow occurred during the unsigned multiplication.
8934 The arguments (%a and %b) and the first element of the result structure
8935 may be of integer types of any bit width, but they must have the same
8936 bit width. The second element of the result structure must be of type
8937 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
8943 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
8944 an unsigned multiplication of the two arguments. They return a structure ---
8945 the first element of which is the multiplication, and the second
8946 element of which is a bit specifying if the unsigned multiplication
8947 resulted in an overflow.
8952 .. code-block:: llvm
8954 %res = call {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
8955 %sum = extractvalue {i32, i1} %res, 0
8956 %obit = extractvalue {i32, i1} %res, 1
8957 br i1 %obit, label %overflow, label %normal
8959 Specialised Arithmetic Intrinsics
8960 ---------------------------------
8962 '``llvm.fmuladd.*``' Intrinsic
8963 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8970 declare float @llvm.fmuladd.f32(float %a, float %b, float %c)
8971 declare double @llvm.fmuladd.f64(double %a, double %b, double %c)
8976 The '``llvm.fmuladd.*``' intrinsic functions represent multiply-add
8977 expressions that can be fused if the code generator determines that (a) the
8978 target instruction set has support for a fused operation, and (b) that the
8979 fused operation is more efficient than the equivalent, separate pair of mul
8980 and add instructions.
8985 The '``llvm.fmuladd.*``' intrinsics each take three arguments: two
8986 multiplicands, a and b, and an addend c.
8995 %0 = call float @llvm.fmuladd.f32(%a, %b, %c)
8997 is equivalent to the expression a \* b + c, except that rounding will
8998 not be performed between the multiplication and addition steps if the
8999 code generator fuses the operations. Fusion is not guaranteed, even if
9000 the target platform supports it. If a fused multiply-add is required the
9001 corresponding llvm.fma.\* intrinsic function should be used
9002 instead. This never sets errno, just as '``llvm.fma.*``'.
9007 .. code-block:: llvm
9009 %r2 = call float @llvm.fmuladd.f32(float %a, float %b, float %c) ; yields float:r2 = (a * b) + c
9011 Half Precision Floating Point Intrinsics
9012 ----------------------------------------
9014 For most target platforms, half precision floating point is a
9015 storage-only format. This means that it is a dense encoding (in memory)
9016 but does not support computation in the format.
9018 This means that code must first load the half-precision floating point
9019 value as an i16, then convert it to float with
9020 :ref:`llvm.convert.from.fp16 <int_convert_from_fp16>`. Computation can
9021 then be performed on the float value (including extending to double
9022 etc). To store the value back to memory, it is first converted to float
9023 if needed, then converted to i16 with
9024 :ref:`llvm.convert.to.fp16 <int_convert_to_fp16>`, then storing as an
9027 .. _int_convert_to_fp16:
9029 '``llvm.convert.to.fp16``' Intrinsic
9030 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9037 declare i16 @llvm.convert.to.fp16.f32(float %a)
9038 declare i16 @llvm.convert.to.fp16.f64(double %a)
9043 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion from a
9044 conventional floating point type to half precision floating point format.
9049 The intrinsic function contains single argument - the value to be
9055 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion from a
9056 conventional floating point format to half precision floating point format. The
9057 return value is an ``i16`` which contains the converted number.
9062 .. code-block:: llvm
9064 %res = call i16 @llvm.convert.to.fp16.f32(float %a)
9065 store i16 %res, i16* @x, align 2
9067 .. _int_convert_from_fp16:
9069 '``llvm.convert.from.fp16``' Intrinsic
9070 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9077 declare float @llvm.convert.from.fp16.f32(i16 %a)
9078 declare double @llvm.convert.from.fp16.f64(i16 %a)
9083 The '``llvm.convert.from.fp16``' intrinsic function performs a
9084 conversion from half precision floating point format to single precision
9085 floating point format.
9090 The intrinsic function contains single argument - the value to be
9096 The '``llvm.convert.from.fp16``' intrinsic function performs a
9097 conversion from half single precision floating point format to single
9098 precision floating point format. The input half-float value is
9099 represented by an ``i16`` value.
9104 .. code-block:: llvm
9106 %a = load i16* @x, align 2
9107 %res = call float @llvm.convert.from.fp16(i16 %a)
9112 The LLVM debugger intrinsics (which all start with ``llvm.dbg.``
9113 prefix), are described in the `LLVM Source Level
9114 Debugging <SourceLevelDebugging.html#format_common_intrinsics>`_
9117 Exception Handling Intrinsics
9118 -----------------------------
9120 The LLVM exception handling intrinsics (which all start with
9121 ``llvm.eh.`` prefix), are described in the `LLVM Exception
9122 Handling <ExceptionHandling.html#format_common_intrinsics>`_ document.
9126 Trampoline Intrinsics
9127 ---------------------
9129 These intrinsics make it possible to excise one parameter, marked with
9130 the :ref:`nest <nest>` attribute, from a function. The result is a
9131 callable function pointer lacking the nest parameter - the caller does
9132 not need to provide a value for it. Instead, the value to use is stored
9133 in advance in a "trampoline", a block of memory usually allocated on the
9134 stack, which also contains code to splice the nest value into the
9135 argument list. This is used to implement the GCC nested function address
9138 For example, if the function is ``i32 f(i8* nest %c, i32 %x, i32 %y)``
9139 then the resulting function pointer has signature ``i32 (i32, i32)*``.
9140 It can be created as follows:
9142 .. code-block:: llvm
9144 %tramp = alloca [10 x i8], align 4 ; size and alignment only correct for X86
9145 %tramp1 = getelementptr [10 x i8]* %tramp, i32 0, i32 0
9146 call i8* @llvm.init.trampoline(i8* %tramp1, i8* bitcast (i32 (i8*, i32, i32)* @f to i8*), i8* %nval)
9147 %p = call i8* @llvm.adjust.trampoline(i8* %tramp1)
9148 %fp = bitcast i8* %p to i32 (i32, i32)*
9150 The call ``%val = call i32 %fp(i32 %x, i32 %y)`` is then equivalent to
9151 ``%val = call i32 %f(i8* %nval, i32 %x, i32 %y)``.
9155 '``llvm.init.trampoline``' Intrinsic
9156 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9163 declare void @llvm.init.trampoline(i8* <tramp>, i8* <func>, i8* <nval>)
9168 This fills the memory pointed to by ``tramp`` with executable code,
9169 turning it into a trampoline.
9174 The ``llvm.init.trampoline`` intrinsic takes three arguments, all
9175 pointers. The ``tramp`` argument must point to a sufficiently large and
9176 sufficiently aligned block of memory; this memory is written to by the
9177 intrinsic. Note that the size and the alignment are target-specific -
9178 LLVM currently provides no portable way of determining them, so a
9179 front-end that generates this intrinsic needs to have some
9180 target-specific knowledge. The ``func`` argument must hold a function
9181 bitcast to an ``i8*``.
9186 The block of memory pointed to by ``tramp`` is filled with target
9187 dependent code, turning it into a function. Then ``tramp`` needs to be
9188 passed to :ref:`llvm.adjust.trampoline <int_at>` to get a pointer which can
9189 be :ref:`bitcast (to a new function) and called <int_trampoline>`. The new
9190 function's signature is the same as that of ``func`` with any arguments
9191 marked with the ``nest`` attribute removed. At most one such ``nest``
9192 argument is allowed, and it must be of pointer type. Calling the new
9193 function is equivalent to calling ``func`` with the same argument list,
9194 but with ``nval`` used for the missing ``nest`` argument. If, after
9195 calling ``llvm.init.trampoline``, the memory pointed to by ``tramp`` is
9196 modified, then the effect of any later call to the returned function
9197 pointer is undefined.
9201 '``llvm.adjust.trampoline``' Intrinsic
9202 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9209 declare i8* @llvm.adjust.trampoline(i8* <tramp>)
9214 This performs any required machine-specific adjustment to the address of
9215 a trampoline (passed as ``tramp``).
9220 ``tramp`` must point to a block of memory which already has trampoline
9221 code filled in by a previous call to
9222 :ref:`llvm.init.trampoline <int_it>`.
9227 On some architectures the address of the code to be executed needs to be
9228 different than the address where the trampoline is actually stored. This
9229 intrinsic returns the executable address corresponding to ``tramp``
9230 after performing the required machine specific adjustments. The pointer
9231 returned can then be :ref:`bitcast and executed <int_trampoline>`.
9233 Masked Vector Load and Store Intrinsics
9234 ---------------------------------------
9236 LLVM provides intrinsics for predicated vector load and store operations. The predicate is specified by a mask operand, which holds one bit per vector element, switching the associated vector lane on or off. The memory addresses corresponding to the "off" lanes are not accessed. When all bits of the mask are on, the intrinsic is identical to a regular vector load or store. When all bits are off, no memory is accessed.
9240 '``llvm.masked.load.*``' Intrinsics
9241 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9245 This is an overloaded intrinsic. The loaded data is a vector of any integer or floating point data type.
9249 declare <16 x float> @llvm.masked.load.v16f32 (<16 x float>* <ptr>, i32 <alignment>, <16 x i1> <mask>, <16 x float> <passthru>)
9250 declare <2 x double> @llvm.masked.load.v2f64 (<2 x double>* <ptr>, i32 <alignment>, <2 x i1> <mask>, <2 x double> <passthru>)
9255 Reads a vector from memory according to the provided mask. The mask holds a bit for each vector lane, and is used to prevent memory accesses to the masked-off lanes. The masked-off lanes in the result vector are taken from the corresponding lanes in the passthru operand.
9261 The first operand is the base pointer for the load. The second operand is the alignment of the source location. It must be a constant integer value. The third operand, mask, is a vector of boolean 'i1' values with the same number of elements as the return type. The fourth is a pass-through value that is used to fill the masked-off lanes of the result. The return type, underlying type of the base pointer and the type of passthru operand are the same vector types.
9267 The '``llvm.masked.load``' intrinsic is designed for conditional reading of selected vector elements in a single IR operation. It is useful for targets that support vector masked loads and allows vectorizing predicated basic blocks on these targets. Other targets may support this intrinsic differently, for example by lowering it into a sequence of branches that guard scalar load operations.
9268 The result of this operation is equivalent to a regular vector load instruction followed by a 'select' between the loaded and the passthru values, predicated on the same mask. However, using this intrinsic prevents exceptions on memory access to masked-off lanes.
9273 %res = call <16 x float> @llvm.masked.load.v16f32 (<16 x float>* %ptr, i32 4, <16 x i1>%mask, <16 x float> %passthru)
9275 ;; The result of the two following instructions is identical aside from potential memory access exception
9276 %loadlal = load <16 x float>* %ptr, align 4
9277 %res = select <16 x i1> %Mask, <16 x float> %loadlal, <16 x float> %passthru
9281 '``llvm.masked.store.*``' Intrinsics
9282 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9286 This is an overloaded intrinsic. The data stored in memory is a vector of any integer or floating point data type.
9290 declare void @llvm.masked.store.v8i32 (<8 x i32> <value>, <8 x i32> * <ptr>, i32 <alignment>, <8 x i1> <mask>)
9291 declare void @llvm.masked.store.v16f32(<16 x i32> <value>, <16 x i32>* <ptr>, i32 <alignment>, <16 x i1> <mask>)
9296 Writes a vector to memory according to the provided mask. The mask holds a bit for each vector lane, and is used to prevent memory accesses to the masked-off lanes.
9301 The first operand is the vector value to be written to memory. The second operand is the base pointer for the store, it has the same underlying type as the value operand. The third operand is the alignment of the destination location. The fourth operand, mask, is a vector of boolean values. The types of the mask and the value operand must have the same number of vector elements.
9307 The '``llvm.masked.store``' intrinsics is designed for conditional writing of selected vector elements in a single IR operation. It is useful for targets that support vector masked store and allows vectorizing predicated basic blocks on these targets. Other targets may support this intrinsic differently, for example by lowering it into a sequence of branches that guard scalar store operations.
9308 The result of this operation is equivalent to a load-modify-store sequence. However, using this intrinsic prevents exceptions and data races on memory access to masked-off lanes.
9312 call void @llvm.masked.store.v16f32(<16 x float> %value, <16 x float>* %ptr, i32 4, <16 x i1> %mask)
9314 ;; The result of the following instructions is identcal aside from potential data races and memory access exceptions
9315 %oldval = load <16 x float>* %ptr, align 4
9316 %res = select <16 x i1> %mask, <16 x float> %value, <16 x float> %oldval
9317 store <16 x float> %res, <16 x float>* %ptr, align 4
9323 This class of intrinsics provides information about the lifetime of
9324 memory objects and ranges where variables are immutable.
9328 '``llvm.lifetime.start``' Intrinsic
9329 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9336 declare void @llvm.lifetime.start(i64 <size>, i8* nocapture <ptr>)
9341 The '``llvm.lifetime.start``' intrinsic specifies the start of a memory
9347 The first argument is a constant integer representing the size of the
9348 object, or -1 if it is variable sized. The second argument is a pointer
9354 This intrinsic indicates that before this point in the code, the value
9355 of the memory pointed to by ``ptr`` is dead. This means that it is known
9356 to never be used and has an undefined value. A load from the pointer
9357 that precedes this intrinsic can be replaced with ``'undef'``.
9361 '``llvm.lifetime.end``' Intrinsic
9362 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9369 declare void @llvm.lifetime.end(i64 <size>, i8* nocapture <ptr>)
9374 The '``llvm.lifetime.end``' intrinsic specifies the end of a memory
9380 The first argument is a constant integer representing the size of the
9381 object, or -1 if it is variable sized. The second argument is a pointer
9387 This intrinsic indicates that after this point in the code, the value of
9388 the memory pointed to by ``ptr`` is dead. This means that it is known to
9389 never be used and has an undefined value. Any stores into the memory
9390 object following this intrinsic may be removed as dead.
9392 '``llvm.invariant.start``' Intrinsic
9393 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9400 declare {}* @llvm.invariant.start(i64 <size>, i8* nocapture <ptr>)
9405 The '``llvm.invariant.start``' intrinsic specifies that the contents of
9406 a memory object will not change.
9411 The first argument is a constant integer representing the size of the
9412 object, or -1 if it is variable sized. The second argument is a pointer
9418 This intrinsic indicates that until an ``llvm.invariant.end`` that uses
9419 the return value, the referenced memory location is constant and
9422 '``llvm.invariant.end``' Intrinsic
9423 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9430 declare void @llvm.invariant.end({}* <start>, i64 <size>, i8* nocapture <ptr>)
9435 The '``llvm.invariant.end``' intrinsic specifies that the contents of a
9436 memory object are mutable.
9441 The first argument is the matching ``llvm.invariant.start`` intrinsic.
9442 The second argument is a constant integer representing the size of the
9443 object, or -1 if it is variable sized and the third argument is a
9444 pointer to the object.
9449 This intrinsic indicates that the memory is mutable again.
9454 This class of intrinsics is designed to be generic and has no specific
9457 '``llvm.var.annotation``' Intrinsic
9458 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9465 declare void @llvm.var.annotation(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
9470 The '``llvm.var.annotation``' intrinsic.
9475 The first argument is a pointer to a value, the second is a pointer to a
9476 global string, the third is a pointer to a global string which is the
9477 source file name, and the last argument is the line number.
9482 This intrinsic allows annotation of local variables with arbitrary
9483 strings. This can be useful for special purpose optimizations that want
9484 to look for these annotations. These have no other defined use; they are
9485 ignored by code generation and optimization.
9487 '``llvm.ptr.annotation.*``' Intrinsic
9488 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9493 This is an overloaded intrinsic. You can use '``llvm.ptr.annotation``' on a
9494 pointer to an integer of any width. *NOTE* you must specify an address space for
9495 the pointer. The identifier for the default address space is the integer
9500 declare i8* @llvm.ptr.annotation.p<address space>i8(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
9501 declare i16* @llvm.ptr.annotation.p<address space>i16(i16* <val>, i8* <str>, i8* <str>, i32 <int>)
9502 declare i32* @llvm.ptr.annotation.p<address space>i32(i32* <val>, i8* <str>, i8* <str>, i32 <int>)
9503 declare i64* @llvm.ptr.annotation.p<address space>i64(i64* <val>, i8* <str>, i8* <str>, i32 <int>)
9504 declare i256* @llvm.ptr.annotation.p<address space>i256(i256* <val>, i8* <str>, i8* <str>, i32 <int>)
9509 The '``llvm.ptr.annotation``' intrinsic.
9514 The first argument is a pointer to an integer value of arbitrary bitwidth
9515 (result of some expression), the second is a pointer to a global string, the
9516 third is a pointer to a global string which is the source file name, and the
9517 last argument is the line number. It returns the value of the first argument.
9522 This intrinsic allows annotation of a pointer to an integer with arbitrary
9523 strings. This can be useful for special purpose optimizations that want to look
9524 for these annotations. These have no other defined use; they are ignored by code
9525 generation and optimization.
9527 '``llvm.annotation.*``' Intrinsic
9528 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9533 This is an overloaded intrinsic. You can use '``llvm.annotation``' on
9534 any integer bit width.
9538 declare i8 @llvm.annotation.i8(i8 <val>, i8* <str>, i8* <str>, i32 <int>)
9539 declare i16 @llvm.annotation.i16(i16 <val>, i8* <str>, i8* <str>, i32 <int>)
9540 declare i32 @llvm.annotation.i32(i32 <val>, i8* <str>, i8* <str>, i32 <int>)
9541 declare i64 @llvm.annotation.i64(i64 <val>, i8* <str>, i8* <str>, i32 <int>)
9542 declare i256 @llvm.annotation.i256(i256 <val>, i8* <str>, i8* <str>, i32 <int>)
9547 The '``llvm.annotation``' intrinsic.
9552 The first argument is an integer value (result of some expression), the
9553 second is a pointer to a global string, the third is a pointer to a
9554 global string which is the source file name, and the last argument is
9555 the line number. It returns the value of the first argument.
9560 This intrinsic allows annotations to be put on arbitrary expressions
9561 with arbitrary strings. This can be useful for special purpose
9562 optimizations that want to look for these annotations. These have no
9563 other defined use; they are ignored by code generation and optimization.
9565 '``llvm.trap``' Intrinsic
9566 ^^^^^^^^^^^^^^^^^^^^^^^^^
9573 declare void @llvm.trap() noreturn nounwind
9578 The '``llvm.trap``' intrinsic.
9588 This intrinsic is lowered to the target dependent trap instruction. If
9589 the target does not have a trap instruction, this intrinsic will be
9590 lowered to a call of the ``abort()`` function.
9592 '``llvm.debugtrap``' Intrinsic
9593 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9600 declare void @llvm.debugtrap() nounwind
9605 The '``llvm.debugtrap``' intrinsic.
9615 This intrinsic is lowered to code which is intended to cause an
9616 execution trap with the intention of requesting the attention of a
9619 '``llvm.stackprotector``' Intrinsic
9620 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9627 declare void @llvm.stackprotector(i8* <guard>, i8** <slot>)
9632 The ``llvm.stackprotector`` intrinsic takes the ``guard`` and stores it
9633 onto the stack at ``slot``. The stack slot is adjusted to ensure that it
9634 is placed on the stack before local variables.
9639 The ``llvm.stackprotector`` intrinsic requires two pointer arguments.
9640 The first argument is the value loaded from the stack guard
9641 ``@__stack_chk_guard``. The second variable is an ``alloca`` that has
9642 enough space to hold the value of the guard.
9647 This intrinsic causes the prologue/epilogue inserter to force the position of
9648 the ``AllocaInst`` stack slot to be before local variables on the stack. This is
9649 to ensure that if a local variable on the stack is overwritten, it will destroy
9650 the value of the guard. When the function exits, the guard on the stack is
9651 checked against the original guard by ``llvm.stackprotectorcheck``. If they are
9652 different, then ``llvm.stackprotectorcheck`` causes the program to abort by
9653 calling the ``__stack_chk_fail()`` function.
9655 '``llvm.stackprotectorcheck``' Intrinsic
9656 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9663 declare void @llvm.stackprotectorcheck(i8** <guard>)
9668 The ``llvm.stackprotectorcheck`` intrinsic compares ``guard`` against an already
9669 created stack protector and if they are not equal calls the
9670 ``__stack_chk_fail()`` function.
9675 The ``llvm.stackprotectorcheck`` intrinsic requires one pointer argument, the
9676 the variable ``@__stack_chk_guard``.
9681 This intrinsic is provided to perform the stack protector check by comparing
9682 ``guard`` with the stack slot created by ``llvm.stackprotector`` and if the
9683 values do not match call the ``__stack_chk_fail()`` function.
9685 The reason to provide this as an IR level intrinsic instead of implementing it
9686 via other IR operations is that in order to perform this operation at the IR
9687 level without an intrinsic, one would need to create additional basic blocks to
9688 handle the success/failure cases. This makes it difficult to stop the stack
9689 protector check from disrupting sibling tail calls in Codegen. With this
9690 intrinsic, we are able to generate the stack protector basic blocks late in
9691 codegen after the tail call decision has occurred.
9693 '``llvm.objectsize``' Intrinsic
9694 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9701 declare i32 @llvm.objectsize.i32(i8* <object>, i1 <min>)
9702 declare i64 @llvm.objectsize.i64(i8* <object>, i1 <min>)
9707 The ``llvm.objectsize`` intrinsic is designed to provide information to
9708 the optimizers to determine at compile time whether a) an operation
9709 (like memcpy) will overflow a buffer that corresponds to an object, or
9710 b) that a runtime check for overflow isn't necessary. An object in this
9711 context means an allocation of a specific class, structure, array, or
9717 The ``llvm.objectsize`` intrinsic takes two arguments. The first
9718 argument is a pointer to or into the ``object``. The second argument is
9719 a boolean and determines whether ``llvm.objectsize`` returns 0 (if true)
9720 or -1 (if false) when the object size is unknown. The second argument
9721 only accepts constants.
9726 The ``llvm.objectsize`` intrinsic is lowered to a constant representing
9727 the size of the object concerned. If the size cannot be determined at
9728 compile time, ``llvm.objectsize`` returns ``i32/i64 -1 or 0`` (depending
9729 on the ``min`` argument).
9731 '``llvm.expect``' Intrinsic
9732 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9737 This is an overloaded intrinsic. You can use ``llvm.expect`` on any
9742 declare i1 @llvm.expect.i1(i1 <val>, i1 <expected_val>)
9743 declare i32 @llvm.expect.i32(i32 <val>, i32 <expected_val>)
9744 declare i64 @llvm.expect.i64(i64 <val>, i64 <expected_val>)
9749 The ``llvm.expect`` intrinsic provides information about expected (the
9750 most probable) value of ``val``, which can be used by optimizers.
9755 The ``llvm.expect`` intrinsic takes two arguments. The first argument is
9756 a value. The second argument is an expected value, this needs to be a
9757 constant value, variables are not allowed.
9762 This intrinsic is lowered to the ``val``.
9764 '``llvm.assume``' Intrinsic
9765 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9772 declare void @llvm.assume(i1 %cond)
9777 The ``llvm.assume`` allows the optimizer to assume that the provided
9778 condition is true. This information can then be used in simplifying other parts
9784 The condition which the optimizer may assume is always true.
9789 The intrinsic allows the optimizer to assume that the provided condition is
9790 always true whenever the control flow reaches the intrinsic call. No code is
9791 generated for this intrinsic, and instructions that contribute only to the
9792 provided condition are not used for code generation. If the condition is
9793 violated during execution, the behavior is undefined.
9795 Please note that optimizer might limit the transformations performed on values
9796 used by the ``llvm.assume`` intrinsic in order to preserve the instructions
9797 only used to form the intrinsic's input argument. This might prove undesirable
9798 if the extra information provided by the ``llvm.assume`` intrinsic does cause
9799 sufficient overall improvement in code quality. For this reason,
9800 ``llvm.assume`` should not be used to document basic mathematical invariants
9801 that the optimizer can otherwise deduce or facts that are of little use to the
9804 '``llvm.donothing``' Intrinsic
9805 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9812 declare void @llvm.donothing() nounwind readnone
9817 The ``llvm.donothing`` intrinsic doesn't perform any operation. It's one of only
9818 two intrinsics (besides ``llvm.experimental.patchpoint``) that can be called
9819 with an invoke instruction.
9829 This intrinsic does nothing, and it's removed by optimizers and ignored
9832 Stack Map Intrinsics
9833 --------------------
9835 LLVM provides experimental intrinsics to support runtime patching
9836 mechanisms commonly desired in dynamic language JITs. These intrinsics
9837 are described in :doc:`StackMaps`.