1 ==============================
2 LLVM Language Reference Manual
3 ==============================
12 This document is a reference manual for the LLVM assembly language. LLVM
13 is a Static Single Assignment (SSA) based representation that provides
14 type safety, low-level operations, flexibility, and the capability of
15 representing 'all' high-level languages cleanly. It is the common code
16 representation used throughout all phases of the LLVM compilation
22 The LLVM code representation is designed to be used in three different
23 forms: as an in-memory compiler IR, as an on-disk bitcode representation
24 (suitable for fast loading by a Just-In-Time compiler), and as a human
25 readable assembly language representation. This allows LLVM to provide a
26 powerful intermediate representation for efficient compiler
27 transformations and analysis, while providing a natural means to debug
28 and visualize the transformations. The three different forms of LLVM are
29 all equivalent. This document describes the human readable
30 representation and notation.
32 The LLVM representation aims to be light-weight and low-level while
33 being expressive, typed, and extensible at the same time. It aims to be
34 a "universal IR" of sorts, by being at a low enough level that
35 high-level ideas may be cleanly mapped to it (similar to how
36 microprocessors are "universal IR's", allowing many source languages to
37 be mapped to them). By providing type information, LLVM can be used as
38 the target of optimizations: for example, through pointer analysis, it
39 can be proven that a C automatic variable is never accessed outside of
40 the current function, allowing it to be promoted to a simple SSA value
41 instead of a memory location.
48 It is important to note that this document describes 'well formed' LLVM
49 assembly language. There is a difference between what the parser accepts
50 and what is considered 'well formed'. For example, the following
51 instruction is syntactically okay, but not well formed:
57 because the definition of ``%x`` does not dominate all of its uses. The
58 LLVM infrastructure provides a verification pass that may be used to
59 verify that an LLVM module is well formed. This pass is automatically
60 run by the parser after parsing input assembly and by the optimizer
61 before it outputs bitcode. The violations pointed out by the verifier
62 pass indicate bugs in transformation passes or input to the parser.
69 LLVM identifiers come in two basic types: global and local. Global
70 identifiers (functions, global variables) begin with the ``'@'``
71 character. Local identifiers (register names, types) begin with the
72 ``'%'`` character. Additionally, there are three different formats for
73 identifiers, for different purposes:
75 #. Named values are represented as a string of characters with their
76 prefix. For example, ``%foo``, ``@DivisionByZero``,
77 ``%a.really.long.identifier``. The actual regular expression used is
78 '``[%@][a-zA-Z$._][a-zA-Z$._0-9]*``'. Identifiers which require other
79 characters in their names can be surrounded with quotes. Special
80 characters may be escaped using ``"\xx"`` where ``xx`` is the ASCII
81 code for the character in hexadecimal. In this way, any character can
82 be used in a name value, even quotes themselves.
83 #. Unnamed values are represented as an unsigned numeric value with
84 their prefix. For example, ``%12``, ``@2``, ``%44``.
85 #. Constants, which are described in the section Constants_ below.
87 LLVM requires that values start with a prefix for two reasons: Compilers
88 don't need to worry about name clashes with reserved words, and the set
89 of reserved words may be expanded in the future without penalty.
90 Additionally, unnamed identifiers allow a compiler to quickly come up
91 with a temporary variable without having to avoid symbol table
94 Reserved words in LLVM are very similar to reserved words in other
95 languages. There are keywords for different opcodes ('``add``',
96 '``bitcast``', '``ret``', etc...), for primitive type names ('``void``',
97 '``i32``', etc...), and others. These reserved words cannot conflict
98 with variable names, because none of them start with a prefix character
101 Here is an example of LLVM code to multiply the integer variable
108 %result = mul i32 %X, 8
110 After strength reduction:
114 %result = shl i32 %X, 3
120 %0 = add i32 %X, %X ; yields {i32}:%0
121 %1 = add i32 %0, %0 ; yields {i32}:%1
122 %result = add i32 %1, %1
124 This last way of multiplying ``%X`` by 8 illustrates several important
125 lexical features of LLVM:
127 #. Comments are delimited with a '``;``' and go until the end of line.
128 #. Unnamed temporaries are created when the result of a computation is
129 not assigned to a named value.
130 #. Unnamed temporaries are numbered sequentially (using a per-function
131 incrementing counter, starting with 0). Note that basic blocks are
132 included in this numbering. For example, if the entry basic block is not
133 given a label name, then it will get number 0.
135 It also shows a convention that we follow in this document. When
136 demonstrating instructions, we will follow an instruction with a comment
137 that defines the type and name of value produced.
145 LLVM programs are composed of ``Module``'s, each of which is a
146 translation unit of the input programs. Each module consists of
147 functions, global variables, and symbol table entries. Modules may be
148 combined together with the LLVM linker, which merges function (and
149 global variable) definitions, resolves forward declarations, and merges
150 symbol table entries. Here is an example of the "hello world" module:
154 ; Declare the string constant as a global constant.
155 @.str = private unnamed_addr constant [13 x i8] c"hello world\0A\00"
157 ; External declaration of the puts function
158 declare i32 @puts(i8* nocapture) nounwind
160 ; Definition of main function
161 define i32 @main() { ; i32()*
162 ; Convert [13 x i8]* to i8 *...
163 %cast210 = getelementptr [13 x i8]* @.str, i64 0, i64 0
165 ; Call puts function to write out the string to stdout.
166 call i32 @puts(i8* %cast210)
171 !1 = metadata !{i32 42}
174 This example is made up of a :ref:`global variable <globalvars>` named
175 "``.str``", an external declaration of the "``puts``" function, a
176 :ref:`function definition <functionstructure>` for "``main``" and
177 :ref:`named metadata <namedmetadatastructure>` "``foo``".
179 In general, a module is made up of a list of global values (where both
180 functions and global variables are global values). Global values are
181 represented by a pointer to a memory location (in this case, a pointer
182 to an array of char, and a pointer to a function), and have one of the
183 following :ref:`linkage types <linkage>`.
190 All Global Variables and Functions have one of the following types of
194 Global values with "``private``" linkage are only directly
195 accessible by objects in the current module. In particular, linking
196 code into a module with an private global value may cause the
197 private to be renamed as necessary to avoid collisions. Because the
198 symbol is private to the module, all references can be updated. This
199 doesn't show up in any symbol table in the object file.
201 Similar to private, but the value shows as a local symbol
202 (``STB_LOCAL`` in the case of ELF) in the object file. This
203 corresponds to the notion of the '``static``' keyword in C.
204 ``available_externally``
205 Globals with "``available_externally``" linkage are never emitted
206 into the object file corresponding to the LLVM module. They exist to
207 allow inlining and other optimizations to take place given knowledge
208 of the definition of the global, which is known to be somewhere
209 outside the module. Globals with ``available_externally`` linkage
210 are allowed to be discarded at will, and are otherwise the same as
211 ``linkonce_odr``. This linkage type is only allowed on definitions,
214 Globals with "``linkonce``" linkage are merged with other globals of
215 the same name when linkage occurs. This can be used to implement
216 some forms of inline functions, templates, or other code which must
217 be generated in each translation unit that uses it, but where the
218 body may be overridden with a more definitive definition later.
219 Unreferenced ``linkonce`` globals are allowed to be discarded. Note
220 that ``linkonce`` linkage does not actually allow the optimizer to
221 inline the body of this function into callers because it doesn't
222 know if this definition of the function is the definitive definition
223 within the program or whether it will be overridden by a stronger
224 definition. To enable inlining and other optimizations, use
225 "``linkonce_odr``" linkage.
227 "``weak``" linkage has the same merging semantics as ``linkonce``
228 linkage, except that unreferenced globals with ``weak`` linkage may
229 not be discarded. This is used for globals that are declared "weak"
232 "``common``" linkage is most similar to "``weak``" linkage, but they
233 are used for tentative definitions in C, such as "``int X;``" at
234 global scope. Symbols with "``common``" linkage are merged in the
235 same way as ``weak symbols``, and they may not be deleted if
236 unreferenced. ``common`` symbols may not have an explicit section,
237 must have a zero initializer, and may not be marked
238 ':ref:`constant <globalvars>`'. Functions and aliases may not have
241 .. _linkage_appending:
244 "``appending``" linkage may only be applied to global variables of
245 pointer to array type. When two global variables with appending
246 linkage are linked together, the two global arrays are appended
247 together. This is the LLVM, typesafe, equivalent of having the
248 system linker append together "sections" with identical names when
251 The semantics of this linkage follow the ELF object file model: the
252 symbol is weak until linked, if not linked, the symbol becomes null
253 instead of being an undefined reference.
254 ``linkonce_odr``, ``weak_odr``
255 Some languages allow differing globals to be merged, such as two
256 functions with different semantics. Other languages, such as
257 ``C++``, ensure that only equivalent globals are ever merged (the
258 "one definition rule" --- "ODR"). Such languages can use the
259 ``linkonce_odr`` and ``weak_odr`` linkage types to indicate that the
260 global will only be merged with equivalent globals. These linkage
261 types are otherwise the same as their non-``odr`` versions.
263 If none of the above identifiers are used, the global is externally
264 visible, meaning that it participates in linkage and can be used to
265 resolve external symbol references.
267 It is illegal for a function *declaration* to have any linkage type
268 other than ``external`` or ``extern_weak``.
275 LLVM :ref:`functions <functionstructure>`, :ref:`calls <i_call>` and
276 :ref:`invokes <i_invoke>` can all have an optional calling convention
277 specified for the call. The calling convention of any pair of dynamic
278 caller/callee must match, or the behavior of the program is undefined.
279 The following calling conventions are supported by LLVM, and more may be
282 "``ccc``" - The C calling convention
283 This calling convention (the default if no other calling convention
284 is specified) matches the target C calling conventions. This calling
285 convention supports varargs function calls and tolerates some
286 mismatch in the declared prototype and implemented declaration of
287 the function (as does normal C).
288 "``fastcc``" - The fast calling convention
289 This calling convention attempts to make calls as fast as possible
290 (e.g. by passing things in registers). This calling convention
291 allows the target to use whatever tricks it wants to produce fast
292 code for the target, without having to conform to an externally
293 specified ABI (Application Binary Interface). `Tail calls can only
294 be optimized when this, the GHC or the HiPE convention is
295 used. <CodeGenerator.html#id80>`_ This calling convention does not
296 support varargs and requires the prototype of all callees to exactly
297 match the prototype of the function definition.
298 "``coldcc``" - The cold calling convention
299 This calling convention attempts to make code in the caller as
300 efficient as possible under the assumption that the call is not
301 commonly executed. As such, these calls often preserve all registers
302 so that the call does not break any live ranges in the caller side.
303 This calling convention does not support varargs and requires the
304 prototype of all callees to exactly match the prototype of the
305 function definition. Furthermore the inliner doesn't consider such function
307 "``cc 10``" - GHC convention
308 This calling convention has been implemented specifically for use by
309 the `Glasgow Haskell Compiler (GHC) <http://www.haskell.org/ghc>`_.
310 It passes everything in registers, going to extremes to achieve this
311 by disabling callee save registers. This calling convention should
312 not be used lightly but only for specific situations such as an
313 alternative to the *register pinning* performance technique often
314 used when implementing functional programming languages. At the
315 moment only X86 supports this convention and it has the following
318 - On *X86-32* only supports up to 4 bit type parameters. No
319 floating point types are supported.
320 - On *X86-64* only supports up to 10 bit type parameters and 6
321 floating point parameters.
323 This calling convention supports `tail call
324 optimization <CodeGenerator.html#id80>`_ but requires both the
325 caller and callee are using it.
326 "``cc 11``" - The HiPE calling convention
327 This calling convention has been implemented specifically for use by
328 the `High-Performance Erlang
329 (HiPE) <http://www.it.uu.se/research/group/hipe/>`_ compiler, *the*
330 native code compiler of the `Ericsson's Open Source Erlang/OTP
331 system <http://www.erlang.org/download.shtml>`_. It uses more
332 registers for argument passing than the ordinary C calling
333 convention and defines no callee-saved registers. The calling
334 convention properly supports `tail call
335 optimization <CodeGenerator.html#id80>`_ but requires that both the
336 caller and the callee use it. It uses a *register pinning*
337 mechanism, similar to GHC's convention, for keeping frequently
338 accessed runtime components pinned to specific hardware registers.
339 At the moment only X86 supports this convention (both 32 and 64
341 "``webkit_jscc``" - WebKit's JavaScript calling convention
342 This calling convention has been implemented for `WebKit FTL JIT
343 <https://trac.webkit.org/wiki/FTLJIT>`_. It passes arguments on the
344 stack right to left (as cdecl does), and returns a value in the
345 platform's customary return register.
346 "``anyregcc``" - Dynamic calling convention for code patching
347 This is a special convention that supports patching an arbitrary code
348 sequence in place of a call site. This convention forces the call
349 arguments into registers but allows them to be dynamcially
350 allocated. This can currently only be used with calls to
351 llvm.experimental.patchpoint because only this intrinsic records
352 the location of its arguments in a side table. See :doc:`StackMaps`.
353 "``preserve_mostcc``" - The `PreserveMost` calling convention
354 This calling convention attempts to make the code in the caller as little
355 intrusive as possible. This calling convention behaves identical to the `C`
356 calling convention on how arguments and return values are passed, but it
357 uses a different set of caller/callee-saved registers. This alleviates the
358 burden of saving and recovering a large register set before and after the
359 call in the caller. If the arguments are passed in callee-saved registers,
360 then they will be preserved by the callee across the call. This doesn't
361 apply for values returned in callee-saved registers.
363 - On X86-64 the callee preserves all general purpose registers, except for
364 R11. R11 can be used as a scratch register. Floating-point registers
365 (XMMs/YMMs) are not preserved and need to be saved by the caller.
367 The idea behind this convention is to support calls to runtime functions
368 that have a hot path and a cold path. The hot path is usually a small piece
369 of code that doesn't many registers. The cold path might need to call out to
370 another function and therefore only needs to preserve the caller-saved
371 registers, which haven't already been saved by the caller. The
372 `PreserveMost` calling convention is very similar to the `cold` calling
373 convention in terms of caller/callee-saved registers, but they are used for
374 different types of function calls. `coldcc` is for function calls that are
375 rarely executed, whereas `preserve_mostcc` function calls are intended to be
376 on the hot path and definitely executed a lot. Furthermore `preserve_mostcc`
377 doesn't prevent the inliner from inlining the function call.
379 This calling convention will be used by a future version of the ObjectiveC
380 runtime and should therefore still be considered experimental at this time.
381 Although this convention was created to optimize certain runtime calls to
382 the ObjectiveC runtime, it is not limited to this runtime and might be used
383 by other runtimes in the future too. The current implementation only
384 supports X86-64, but the intention is to support more architectures in the
386 "``preserve_allcc``" - The `PreserveAll` calling convention
387 This calling convention attempts to make the code in the caller even less
388 intrusive than the `PreserveMost` calling convention. This calling
389 convention also behaves identical to the `C` calling convention on how
390 arguments and return values are passed, but it uses a different set of
391 caller/callee-saved registers. This removes the burden of saving and
392 recovering a large register set before and after the call in the caller. If
393 the arguments are passed in callee-saved registers, then they will be
394 preserved by the callee across the call. This doesn't apply for values
395 returned in callee-saved registers.
397 - On X86-64 the callee preserves all general purpose registers, except for
398 R11. R11 can be used as a scratch register. Furthermore it also preserves
399 all floating-point registers (XMMs/YMMs).
401 The idea behind this convention is to support calls to runtime functions
402 that don't need to call out to any other functions.
404 This calling convention, like the `PreserveMost` calling convention, will be
405 used by a future version of the ObjectiveC runtime and should be considered
406 experimental at this time.
407 "``cc <n>``" - Numbered convention
408 Any calling convention may be specified by number, allowing
409 target-specific calling conventions to be used. Target specific
410 calling conventions start at 64.
412 More calling conventions can be added/defined on an as-needed basis, to
413 support Pascal conventions or any other well-known target-independent
416 .. _visibilitystyles:
421 All Global Variables and Functions have one of the following visibility
424 "``default``" - Default style
425 On targets that use the ELF object file format, default visibility
426 means that the declaration is visible to other modules and, in
427 shared libraries, means that the declared entity may be overridden.
428 On Darwin, default visibility means that the declaration is visible
429 to other modules. Default visibility corresponds to "external
430 linkage" in the language.
431 "``hidden``" - Hidden style
432 Two declarations of an object with hidden visibility refer to the
433 same object if they are in the same shared object. Usually, hidden
434 visibility indicates that the symbol will not be placed into the
435 dynamic symbol table, so no other module (executable or shared
436 library) can reference it directly.
437 "``protected``" - Protected style
438 On ELF, protected visibility indicates that the symbol will be
439 placed in the dynamic symbol table, but that references within the
440 defining module will bind to the local symbol. That is, the symbol
441 cannot be overridden by another module.
443 A symbol with ``internal`` or ``private`` linkage must have ``default``
451 All Global Variables, Functions and Aliases can have one of the following
455 "``dllimport``" causes the compiler to reference a function or variable via
456 a global pointer to a pointer that is set up by the DLL exporting the
457 symbol. On Microsoft Windows targets, the pointer name is formed by
458 combining ``__imp_`` and the function or variable name.
460 "``dllexport``" causes the compiler to provide a global pointer to a pointer
461 in a DLL, so that it can be referenced with the ``dllimport`` attribute. On
462 Microsoft Windows targets, the pointer name is formed by combining
463 ``__imp_`` and the function or variable name. Since this storage class
464 exists for defining a dll interface, the compiler, assembler and linker know
465 it is externally referenced and must refrain from deleting the symbol.
469 Thread Local Storage Models
470 ---------------------------
472 A variable may be defined as ``thread_local``, which means that it will
473 not be shared by threads (each thread will have a separated copy of the
474 variable). Not all targets support thread-local variables. Optionally, a
475 TLS model may be specified:
478 For variables that are only used within the current shared library.
480 For variables in modules that will not be loaded dynamically.
482 For variables defined in the executable and only used within it.
484 If no explicit model is given, the "general dynamic" model is used.
486 The models correspond to the ELF TLS models; see `ELF Handling For
487 Thread-Local Storage <http://people.redhat.com/drepper/tls.pdf>`_ for
488 more information on under which circumstances the different models may
489 be used. The target may choose a different TLS model if the specified
490 model is not supported, or if a better choice of model can be made.
492 A model can also be specified in a alias, but then it only governs how
493 the alias is accessed. It will not have any effect in the aliasee.
500 LLVM IR allows you to specify both "identified" and "literal" :ref:`structure
501 types <t_struct>`. Literal types are uniqued structurally, but identified types
502 are never uniqued. An :ref:`opaque structural type <t_opaque>` can also be used
503 to forward declare a type which is not yet available.
505 An example of a identified structure specification is:
509 %mytype = type { %mytype*, i32 }
511 Prior to the LLVM 3.0 release, identified types were structurally uniqued. Only
512 literal types are uniqued in recent versions of LLVM.
519 Global variables define regions of memory allocated at compilation time
522 Global variables definitions must be initialized, may have an explicit section
523 to be placed in, and may have an optional explicit alignment specified.
525 Global variables in other translation units can also be declared, in which
526 case they don't have an initializer.
528 A variable may be defined as a global ``constant``, which indicates that
529 the contents of the variable will **never** be modified (enabling better
530 optimization, allowing the global data to be placed in the read-only
531 section of an executable, etc). Note that variables that need runtime
532 initialization cannot be marked ``constant`` as there is a store to the
535 LLVM explicitly allows *declarations* of global variables to be marked
536 constant, even if the final definition of the global is not. This
537 capability can be used to enable slightly better optimization of the
538 program, but requires the language definition to guarantee that
539 optimizations based on the 'constantness' are valid for the translation
540 units that do not include the definition.
542 As SSA values, global variables define pointer values that are in scope
543 (i.e. they dominate) all basic blocks in the program. Global variables
544 always define a pointer to their "content" type because they describe a
545 region of memory, and all memory objects in LLVM are accessed through
548 Global variables can be marked with ``unnamed_addr`` which indicates
549 that the address is not significant, only the content. Constants marked
550 like this can be merged with other constants if they have the same
551 initializer. Note that a constant with significant address *can* be
552 merged with a ``unnamed_addr`` constant, the result being a constant
553 whose address is significant.
555 A global variable may be declared to reside in a target-specific
556 numbered address space. For targets that support them, address spaces
557 may affect how optimizations are performed and/or what target
558 instructions are used to access the variable. The default address space
559 is zero. The address space qualifier must precede any other attributes.
561 LLVM allows an explicit section to be specified for globals. If the
562 target supports it, it will emit globals to the section specified.
564 By default, global initializers are optimized by assuming that global
565 variables defined within the module are not modified from their
566 initial values before the start of the global initializer. This is
567 true even for variables potentially accessible from outside the
568 module, including those with external linkage or appearing in
569 ``@llvm.used`` or dllexported variables. This assumption may be suppressed
570 by marking the variable with ``externally_initialized``.
572 An explicit alignment may be specified for a global, which must be a
573 power of 2. If not present, or if the alignment is set to zero, the
574 alignment of the global is set by the target to whatever it feels
575 convenient. If an explicit alignment is specified, the global is forced
576 to have exactly that alignment. Targets and optimizers are not allowed
577 to over-align the global if the global has an assigned section. In this
578 case, the extra alignment could be observable: for example, code could
579 assume that the globals are densely packed in their section and try to
580 iterate over them as an array, alignment padding would break this
583 Globals can also have a :ref:`DLL storage class <dllstorageclass>`.
585 Variables and aliasaes can have a
586 :ref:`Thread Local Storage Model <tls_model>`.
590 [@<GlobalVarName> =] [Linkage] [Visibility] [DLLStorageClass] [ThreadLocal]
591 [AddrSpace] [unnamed_addr] [ExternallyInitialized]
592 <global | constant> <Type>
593 [, section "name"] [, align <Alignment>]
595 For example, the following defines a global in a numbered address space
596 with an initializer, section, and alignment:
600 @G = addrspace(5) constant float 1.0, section "foo", align 4
602 The following example just declares a global variable
606 @G = external global i32
608 The following example defines a thread-local global with the
609 ``initialexec`` TLS model:
613 @G = thread_local(initialexec) global i32 0, align 4
615 .. _functionstructure:
620 LLVM function definitions consist of the "``define``" keyword, an
621 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
622 style <visibility>`, an optional :ref:`DLL storage class <dllstorageclass>`,
623 an optional :ref:`calling convention <callingconv>`,
624 an optional ``unnamed_addr`` attribute, a return type, an optional
625 :ref:`parameter attribute <paramattrs>` for the return type, a function
626 name, a (possibly empty) argument list (each with optional :ref:`parameter
627 attributes <paramattrs>`), optional :ref:`function attributes <fnattrs>`,
628 an optional section, an optional alignment, an optional :ref:`garbage
629 collector name <gc>`, an optional :ref:`prefix <prefixdata>`, an opening
630 curly brace, a list of basic blocks, and a closing curly brace.
632 LLVM function declarations consist of the "``declare``" keyword, an
633 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
634 style <visibility>`, an optional :ref:`DLL storage class <dllstorageclass>`,
635 an optional :ref:`calling convention <callingconv>`,
636 an optional ``unnamed_addr`` attribute, a return type, an optional
637 :ref:`parameter attribute <paramattrs>` for the return type, a function
638 name, a possibly empty list of arguments, an optional alignment, an optional
639 :ref:`garbage collector name <gc>` and an optional :ref:`prefix <prefixdata>`.
641 A function definition contains a list of basic blocks, forming the CFG (Control
642 Flow Graph) for the function. Each basic block may optionally start with a label
643 (giving the basic block a symbol table entry), contains a list of instructions,
644 and ends with a :ref:`terminator <terminators>` instruction (such as a branch or
645 function return). If an explicit label is not provided, a block is assigned an
646 implicit numbered label, using the next value from the same counter as used for
647 unnamed temporaries (:ref:`see above<identifiers>`). For example, if a function
648 entry block does not have an explicit label, it will be assigned label "%0",
649 then the first unnamed temporary in that block will be "%1", etc.
651 The first basic block in a function is special in two ways: it is
652 immediately executed on entrance to the function, and it is not allowed
653 to have predecessor basic blocks (i.e. there can not be any branches to
654 the entry block of a function). Because the block can have no
655 predecessors, it also cannot have any :ref:`PHI nodes <i_phi>`.
657 LLVM allows an explicit section to be specified for functions. If the
658 target supports it, it will emit functions to the section specified.
660 An explicit alignment may be specified for a function. If not present,
661 or if the alignment is set to zero, the alignment of the function is set
662 by the target to whatever it feels convenient. If an explicit alignment
663 is specified, the function is forced to have at least that much
664 alignment. All alignments must be a power of 2.
666 If the ``unnamed_addr`` attribute is given, the address is know to not
667 be significant and two identical functions can be merged.
671 define [linkage] [visibility] [DLLStorageClass]
673 <ResultType> @<FunctionName> ([argument list])
674 [unnamed_addr] [fn Attrs] [section "name"] [align N]
675 [gc] [prefix Constant] { ... }
682 Aliases, unlike function or variables, don't create any new data. They
683 are just a new symbol and metadata for an existing position.
685 Aliases have a name and an aliasee that is either a global value or a
688 Aliases may have an optional :ref:`linkage type <linkage>`, an optional
689 :ref:`visibility style <visibility>`, an optional :ref:`DLL storage class
690 <dllstorageclass>` and an optional :ref:`tls model <tls_model>`.
694 @<Name> = [Visibility] [DLLStorageClass] [ThreadLocal] alias [Linkage] <AliaseeTy> @<Aliasee>
696 The linkage must be one of ``private``, ``internal``, ``linkonce``, ``weak``,
697 ``linkonce_odr``, ``weak_odr``, ``external``. Note that some system linkers
698 might not correctly handle dropping a weak symbol that is aliased.
700 Alias that are not ``unnamed_addr`` are guaranteed to have the same address as
703 Since aliases are only a second name, some restrictions apply, of which
704 some can only be checked when producing an object file:
706 * The expression defining the aliasee must be computable at assembly
707 time. Since it is just a name, no relocations can be used.
709 * No alias in the expression can be weak as the possibility of the
710 intermediate alias being overridden cannot be represented in an
713 * No global value in the expression can be a declaration, since that
714 would require a relocation, which is not possible.
716 .. _namedmetadatastructure:
721 Named metadata is a collection of metadata. :ref:`Metadata
722 nodes <metadata>` (but not metadata strings) are the only valid
723 operands for a named metadata.
727 ; Some unnamed metadata nodes, which are referenced by the named metadata.
728 !0 = metadata !{metadata !"zero"}
729 !1 = metadata !{metadata !"one"}
730 !2 = metadata !{metadata !"two"}
732 !name = !{!0, !1, !2}
739 The return type and each parameter of a function type may have a set of
740 *parameter attributes* associated with them. Parameter attributes are
741 used to communicate additional information about the result or
742 parameters of a function. Parameter attributes are considered to be part
743 of the function, not of the function type, so functions with different
744 parameter attributes can have the same function type.
746 Parameter attributes are simple keywords that follow the type specified.
747 If multiple parameter attributes are needed, they are space separated.
752 declare i32 @printf(i8* noalias nocapture, ...)
753 declare i32 @atoi(i8 zeroext)
754 declare signext i8 @returns_signed_char()
756 Note that any attributes for the function result (``nounwind``,
757 ``readonly``) come immediately after the argument list.
759 Currently, only the following parameter attributes are defined:
762 This indicates to the code generator that the parameter or return
763 value should be zero-extended to the extent required by the target's
764 ABI (which is usually 32-bits, but is 8-bits for a i1 on x86-64) by
765 the caller (for a parameter) or the callee (for a return value).
767 This indicates to the code generator that the parameter or return
768 value should be sign-extended to the extent required by the target's
769 ABI (which is usually 32-bits) by the caller (for a parameter) or
770 the callee (for a return value).
772 This indicates that this parameter or return value should be treated
773 in a special target-dependent fashion during while emitting code for
774 a function call or return (usually, by putting it in a register as
775 opposed to memory, though some targets use it to distinguish between
776 two different kinds of registers). Use of this attribute is
779 This indicates that the pointer parameter should really be passed by
780 value to the function. The attribute implies that a hidden copy of
781 the pointee is made between the caller and the callee, so the callee
782 is unable to modify the value in the caller. This attribute is only
783 valid on LLVM pointer arguments. It is generally used to pass
784 structs and arrays by value, but is also valid on pointers to
785 scalars. The copy is considered to belong to the caller not the
786 callee (for example, ``readonly`` functions should not write to
787 ``byval`` parameters). This is not a valid attribute for return
790 The byval attribute also supports specifying an alignment with the
791 align attribute. It indicates the alignment of the stack slot to
792 form and the known alignment of the pointer specified to the call
793 site. If the alignment is not specified, then the code generator
794 makes a target-specific assumption.
800 The ``inalloca`` argument attribute allows the caller to take the
801 address of outgoing stack arguments. An ``inalloca`` argument must
802 be a pointer to stack memory produced by an ``alloca`` instruction.
803 The alloca, or argument allocation, must also be tagged with the
804 inalloca keyword. Only the past argument may have the ``inalloca``
805 attribute, and that argument is guaranteed to be passed in memory.
807 An argument allocation may be used by a call at most once because
808 the call may deallocate it. The ``inalloca`` attribute cannot be
809 used in conjunction with other attributes that affect argument
810 storage, like ``inreg``, ``nest``, ``sret``, or ``byval``. The
811 ``inalloca`` attribute also disables LLVM's implicit lowering of
812 large aggregate return values, which means that frontend authors
813 must lower them with ``sret`` pointers.
815 When the call site is reached, the argument allocation must have
816 been the most recent stack allocation that is still live, or the
817 results are undefined. It is possible to allocate additional stack
818 space after an argument allocation and before its call site, but it
819 must be cleared off with :ref:`llvm.stackrestore
822 See :doc:`InAlloca` for more information on how to use this
826 This indicates that the pointer parameter specifies the address of a
827 structure that is the return value of the function in the source
828 program. This pointer must be guaranteed by the caller to be valid:
829 loads and stores to the structure may be assumed by the callee
830 not to trap and to be properly aligned. This may only be applied to
831 the first parameter. This is not a valid attribute for return
837 This indicates that pointer values :ref:`based <pointeraliasing>` on
838 the argument or return value do not alias pointer values which are
839 not *based* on it, ignoring certain "irrelevant" dependencies. For a
840 call to the parent function, dependencies between memory references
841 from before or after the call and from those during the call are
842 "irrelevant" to the ``noalias`` keyword for the arguments and return
843 value used in that call. The caller shares the responsibility with
844 the callee for ensuring that these requirements are met. For further
845 details, please see the discussion of the NoAlias response in :ref:`alias
846 analysis <Must, May, or No>`.
848 Note that this definition of ``noalias`` is intentionally similar
849 to the definition of ``restrict`` in C99 for function arguments,
850 though it is slightly weaker.
852 For function return values, C99's ``restrict`` is not meaningful,
853 while LLVM's ``noalias`` is.
855 This indicates that the callee does not make any copies of the
856 pointer that outlive the callee itself. This is not a valid
857 attribute for return values.
862 This indicates that the pointer parameter can be excised using the
863 :ref:`trampoline intrinsics <int_trampoline>`. This is not a valid
864 attribute for return values and can only be applied to one parameter.
867 This indicates that the function always returns the argument as its return
868 value. This is an optimization hint to the code generator when generating
869 the caller, allowing tail call optimization and omission of register saves
870 and restores in some cases; it is not checked or enforced when generating
871 the callee. The parameter and the function return type must be valid
872 operands for the :ref:`bitcast instruction <i_bitcast>`. This is not a
873 valid attribute for return values and can only be applied to one parameter.
876 This indicates that the parameter or return pointer is not null. This
877 attribute may only be applied to pointer typed parameters. This is not
878 checked or enforced by LLVM, the caller must ensure that the pointer
879 passed in is non-null, or the callee must ensure that the returned pointer
884 Garbage Collector Names
885 -----------------------
887 Each function may specify a garbage collector name, which is simply a
892 define void @f() gc "name" { ... }
894 The compiler declares the supported values of *name*. Specifying a
895 collector which will cause the compiler to alter its output in order to
896 support the named garbage collection algorithm.
903 Prefix data is data associated with a function which the code generator
904 will emit immediately before the function body. The purpose of this feature
905 is to allow frontends to associate language-specific runtime metadata with
906 specific functions and make it available through the function pointer while
907 still allowing the function pointer to be called. To access the data for a
908 given function, a program may bitcast the function pointer to a pointer to
909 the constant's type. This implies that the IR symbol points to the start
912 To maintain the semantics of ordinary function calls, the prefix data must
913 have a particular format. Specifically, it must begin with a sequence of
914 bytes which decode to a sequence of machine instructions, valid for the
915 module's target, which transfer control to the point immediately succeeding
916 the prefix data, without performing any other visible action. This allows
917 the inliner and other passes to reason about the semantics of the function
918 definition without needing to reason about the prefix data. Obviously this
919 makes the format of the prefix data highly target dependent.
921 Prefix data is laid out as if it were an initializer for a global variable
922 of the prefix data's type. No padding is automatically placed between the
923 prefix data and the function body. If padding is required, it must be part
926 A trivial example of valid prefix data for the x86 architecture is ``i8 144``,
927 which encodes the ``nop`` instruction:
931 define void @f() prefix i8 144 { ... }
933 Generally prefix data can be formed by encoding a relative branch instruction
934 which skips the metadata, as in this example of valid prefix data for the
935 x86_64 architecture, where the first two bytes encode ``jmp .+10``:
939 %0 = type <{ i8, i8, i8* }>
941 define void @f() prefix %0 <{ i8 235, i8 8, i8* @md}> { ... }
943 A function may have prefix data but no body. This has similar semantics
944 to the ``available_externally`` linkage in that the data may be used by the
945 optimizers but will not be emitted in the object file.
952 Attribute groups are groups of attributes that are referenced by objects within
953 the IR. They are important for keeping ``.ll`` files readable, because a lot of
954 functions will use the same set of attributes. In the degenerative case of a
955 ``.ll`` file that corresponds to a single ``.c`` file, the single attribute
956 group will capture the important command line flags used to build that file.
958 An attribute group is a module-level object. To use an attribute group, an
959 object references the attribute group's ID (e.g. ``#37``). An object may refer
960 to more than one attribute group. In that situation, the attributes from the
961 different groups are merged.
963 Here is an example of attribute groups for a function that should always be
964 inlined, has a stack alignment of 4, and which shouldn't use SSE instructions:
968 ; Target-independent attributes:
969 attributes #0 = { alwaysinline alignstack=4 }
971 ; Target-dependent attributes:
972 attributes #1 = { "no-sse" }
974 ; Function @f has attributes: alwaysinline, alignstack=4, and "no-sse".
975 define void @f() #0 #1 { ... }
982 Function attributes are set to communicate additional information about
983 a function. Function attributes are considered to be part of the
984 function, not of the function type, so functions with different function
985 attributes can have the same function type.
987 Function attributes are simple keywords that follow the type specified.
988 If multiple attributes are needed, they are space separated. For
993 define void @f() noinline { ... }
994 define void @f() alwaysinline { ... }
995 define void @f() alwaysinline optsize { ... }
996 define void @f() optsize { ... }
999 This attribute indicates that, when emitting the prologue and
1000 epilogue, the backend should forcibly align the stack pointer.
1001 Specify the desired alignment, which must be a power of two, in
1004 This attribute indicates that the inliner should attempt to inline
1005 this function into callers whenever possible, ignoring any active
1006 inlining size threshold for this caller.
1008 This indicates that the callee function at a call site should be
1009 recognized as a built-in function, even though the function's declaration
1010 uses the ``nobuiltin`` attribute. This is only valid at call sites for
1011 direct calls to functions which are declared with the ``nobuiltin``
1014 This attribute indicates that this function is rarely called. When
1015 computing edge weights, basic blocks post-dominated by a cold
1016 function call are also considered to be cold; and, thus, given low
1019 This attribute indicates that the source code contained a hint that
1020 inlining this function is desirable (such as the "inline" keyword in
1021 C/C++). It is just a hint; it imposes no requirements on the
1024 This attribute indicates that the function should be added to a
1025 jump-instruction table at code-generation time, and that all address-taken
1026 references to this function should be replaced with a reference to the
1027 appropriate jump-instruction-table function pointer. Note that this creates
1028 a new pointer for the original function, which means that code that depends
1029 on function-pointer identity can break. So, any function annotated with
1030 ``jumptable`` must also be ``unnamed_addr``.
1032 This attribute suggests that optimization passes and code generator
1033 passes make choices that keep the code size of this function as small
1034 as possible and perform optimizations that may sacrifice runtime
1035 performance in order to minimize the size of the generated code.
1037 This attribute disables prologue / epilogue emission for the
1038 function. This can have very system-specific consequences.
1040 This indicates that the callee function at a call site is not recognized as
1041 a built-in function. LLVM will retain the original call and not replace it
1042 with equivalent code based on the semantics of the built-in function, unless
1043 the call site uses the ``builtin`` attribute. This is valid at call sites
1044 and on function declarations and definitions.
1046 This attribute indicates that calls to the function cannot be
1047 duplicated. A call to a ``noduplicate`` function may be moved
1048 within its parent function, but may not be duplicated within
1049 its parent function.
1051 A function containing a ``noduplicate`` call may still
1052 be an inlining candidate, provided that the call is not
1053 duplicated by inlining. That implies that the function has
1054 internal linkage and only has one call site, so the original
1055 call is dead after inlining.
1057 This attributes disables implicit floating point instructions.
1059 This attribute indicates that the inliner should never inline this
1060 function in any situation. This attribute may not be used together
1061 with the ``alwaysinline`` attribute.
1063 This attribute suppresses lazy symbol binding for the function. This
1064 may make calls to the function faster, at the cost of extra program
1065 startup time if the function is not called during program startup.
1067 This attribute indicates that the code generator should not use a
1068 red zone, even if the target-specific ABI normally permits it.
1070 This function attribute indicates that the function never returns
1071 normally. This produces undefined behavior at runtime if the
1072 function ever does dynamically return.
1074 This function attribute indicates that the function never returns
1075 with an unwind or exceptional control flow. If the function does
1076 unwind, its runtime behavior is undefined.
1078 This function attribute indicates that the function is not optimized
1079 by any optimization or code generator passes with the
1080 exception of interprocedural optimization passes.
1081 This attribute cannot be used together with the ``alwaysinline``
1082 attribute; this attribute is also incompatible
1083 with the ``minsize`` attribute and the ``optsize`` attribute.
1085 This attribute requires the ``noinline`` attribute to be specified on
1086 the function as well, so the function is never inlined into any caller.
1087 Only functions with the ``alwaysinline`` attribute are valid
1088 candidates for inlining into the body of this function.
1090 This attribute suggests that optimization passes and code generator
1091 passes make choices that keep the code size of this function low,
1092 and otherwise do optimizations specifically to reduce code size as
1093 long as they do not significantly impact runtime performance.
1095 On a function, this attribute indicates that the function computes its
1096 result (or decides to unwind an exception) based strictly on its arguments,
1097 without dereferencing any pointer arguments or otherwise accessing
1098 any mutable state (e.g. memory, control registers, etc) visible to
1099 caller functions. It does not write through any pointer arguments
1100 (including ``byval`` arguments) and never changes any state visible
1101 to callers. This means that it cannot unwind exceptions by calling
1102 the ``C++`` exception throwing methods.
1104 On an argument, this attribute indicates that the function does not
1105 dereference that pointer argument, even though it may read or write the
1106 memory that the pointer points to if accessed through other pointers.
1108 On a function, this attribute indicates that the function does not write
1109 through any pointer arguments (including ``byval`` arguments) or otherwise
1110 modify any state (e.g. memory, control registers, etc) visible to
1111 caller functions. It may dereference pointer arguments and read
1112 state that may be set in the caller. A readonly function always
1113 returns the same value (or unwinds an exception identically) when
1114 called with the same set of arguments and global state. It cannot
1115 unwind an exception by calling the ``C++`` exception throwing
1118 On an argument, this attribute indicates that the function does not write
1119 through this pointer argument, even though it may write to the memory that
1120 the pointer points to.
1122 This attribute indicates that this function can return twice. The C
1123 ``setjmp`` is an example of such a function. The compiler disables
1124 some optimizations (like tail calls) in the caller of these
1126 ``sanitize_address``
1127 This attribute indicates that AddressSanitizer checks
1128 (dynamic address safety analysis) are enabled for this function.
1130 This attribute indicates that MemorySanitizer checks (dynamic detection
1131 of accesses to uninitialized memory) are enabled for this function.
1133 This attribute indicates that ThreadSanitizer checks
1134 (dynamic thread safety analysis) are enabled for this function.
1136 This attribute indicates that the function should emit a stack
1137 smashing protector. It is in the form of a "canary" --- a random value
1138 placed on the stack before the local variables that's checked upon
1139 return from the function to see if it has been overwritten. A
1140 heuristic is used to determine if a function needs stack protectors
1141 or not. The heuristic used will enable protectors for functions with:
1143 - Character arrays larger than ``ssp-buffer-size`` (default 8).
1144 - Aggregates containing character arrays larger than ``ssp-buffer-size``.
1145 - Calls to alloca() with variable sizes or constant sizes greater than
1146 ``ssp-buffer-size``.
1148 Variables that are identified as requiring a protector will be arranged
1149 on the stack such that they are adjacent to the stack protector guard.
1151 If a function that has an ``ssp`` attribute is inlined into a
1152 function that doesn't have an ``ssp`` attribute, then the resulting
1153 function will have an ``ssp`` attribute.
1155 This attribute indicates that the function should *always* emit a
1156 stack smashing protector. This overrides the ``ssp`` function
1159 Variables that are identified as requiring a protector will be arranged
1160 on the stack such that they are adjacent to the stack protector guard.
1161 The specific layout rules are:
1163 #. Large arrays and structures containing large arrays
1164 (``>= ssp-buffer-size``) are closest to the stack protector.
1165 #. Small arrays and structures containing small arrays
1166 (``< ssp-buffer-size``) are 2nd closest to the protector.
1167 #. Variables that have had their address taken are 3rd closest to the
1170 If a function that has an ``sspreq`` attribute is inlined into a
1171 function that doesn't have an ``sspreq`` attribute or which has an
1172 ``ssp`` or ``sspstrong`` attribute, then the resulting function will have
1173 an ``sspreq`` attribute.
1175 This attribute indicates that the function should emit a stack smashing
1176 protector. This attribute causes a strong heuristic to be used when
1177 determining if a function needs stack protectors. The strong heuristic
1178 will enable protectors for functions with:
1180 - Arrays of any size and type
1181 - Aggregates containing an array of any size and type.
1182 - Calls to alloca().
1183 - Local variables that have had their address taken.
1185 Variables that are identified as requiring a protector will be arranged
1186 on the stack such that they are adjacent to the stack protector guard.
1187 The specific layout rules are:
1189 #. Large arrays and structures containing large arrays
1190 (``>= ssp-buffer-size``) are closest to the stack protector.
1191 #. Small arrays and structures containing small arrays
1192 (``< ssp-buffer-size``) are 2nd closest to the protector.
1193 #. Variables that have had their address taken are 3rd closest to the
1196 This overrides the ``ssp`` function attribute.
1198 If a function that has an ``sspstrong`` attribute is inlined into a
1199 function that doesn't have an ``sspstrong`` attribute, then the
1200 resulting function will have an ``sspstrong`` attribute.
1202 This attribute indicates that the ABI being targeted requires that
1203 an unwind table entry be produce for this function even if we can
1204 show that no exceptions passes by it. This is normally the case for
1205 the ELF x86-64 abi, but it can be disabled for some compilation
1210 Module-Level Inline Assembly
1211 ----------------------------
1213 Modules may contain "module-level inline asm" blocks, which corresponds
1214 to the GCC "file scope inline asm" blocks. These blocks are internally
1215 concatenated by LLVM and treated as a single unit, but may be separated
1216 in the ``.ll`` file if desired. The syntax is very simple:
1218 .. code-block:: llvm
1220 module asm "inline asm code goes here"
1221 module asm "more can go here"
1223 The strings can contain any character by escaping non-printable
1224 characters. The escape sequence used is simply "\\xx" where "xx" is the
1225 two digit hex code for the number.
1227 The inline asm code is simply printed to the machine code .s file when
1228 assembly code is generated.
1230 .. _langref_datalayout:
1235 A module may specify a target specific data layout string that specifies
1236 how data is to be laid out in memory. The syntax for the data layout is
1239 .. code-block:: llvm
1241 target datalayout = "layout specification"
1243 The *layout specification* consists of a list of specifications
1244 separated by the minus sign character ('-'). Each specification starts
1245 with a letter and may include other information after the letter to
1246 define some aspect of the data layout. The specifications accepted are
1250 Specifies that the target lays out data in big-endian form. That is,
1251 the bits with the most significance have the lowest address
1254 Specifies that the target lays out data in little-endian form. That
1255 is, the bits with the least significance have the lowest address
1258 Specifies the natural alignment of the stack in bits. Alignment
1259 promotion of stack variables is limited to the natural stack
1260 alignment to avoid dynamic stack realignment. The stack alignment
1261 must be a multiple of 8-bits. If omitted, the natural stack
1262 alignment defaults to "unspecified", which does not prevent any
1263 alignment promotions.
1264 ``p[n]:<size>:<abi>:<pref>``
1265 This specifies the *size* of a pointer and its ``<abi>`` and
1266 ``<pref>``\erred alignments for address space ``n``. All sizes are in
1267 bits. The address space, ``n`` is optional, and if not specified,
1268 denotes the default address space 0. The value of ``n`` must be
1269 in the range [1,2^23).
1270 ``i<size>:<abi>:<pref>``
1271 This specifies the alignment for an integer type of a given bit
1272 ``<size>``. The value of ``<size>`` must be in the range [1,2^23).
1273 ``v<size>:<abi>:<pref>``
1274 This specifies the alignment for a vector type of a given bit
1276 ``f<size>:<abi>:<pref>``
1277 This specifies the alignment for a floating point type of a given bit
1278 ``<size>``. Only values of ``<size>`` that are supported by the target
1279 will work. 32 (float) and 64 (double) are supported on all targets; 80
1280 or 128 (different flavors of long double) are also supported on some
1283 This specifies the alignment for an object of aggregate type.
1285 If present, specifies that llvm names are mangled in the output. The
1288 * ``e``: ELF mangling: Private symbols get a ``.L`` prefix.
1289 * ``m``: Mips mangling: Private symbols get a ``$`` prefix.
1290 * ``o``: Mach-O mangling: Private symbols get ``L`` prefix. Other
1291 symbols get a ``_`` prefix.
1292 * ``w``: Windows COFF prefix: Similar to Mach-O, but stdcall and fastcall
1293 functions also get a suffix based on the frame size.
1294 ``n<size1>:<size2>:<size3>...``
1295 This specifies a set of native integer widths for the target CPU in
1296 bits. For example, it might contain ``n32`` for 32-bit PowerPC,
1297 ``n32:64`` for PowerPC 64, or ``n8:16:32:64`` for X86-64. Elements of
1298 this set are considered to support most general arithmetic operations
1301 On every specification that takes a ``<abi>:<pref>``, specifying the
1302 ``<pref>`` alignment is optional. If omitted, the preceding ``:``
1303 should be omitted too and ``<pref>`` will be equal to ``<abi>``.
1305 When constructing the data layout for a given target, LLVM starts with a
1306 default set of specifications which are then (possibly) overridden by
1307 the specifications in the ``datalayout`` keyword. The default
1308 specifications are given in this list:
1310 - ``E`` - big endian
1311 - ``p:64:64:64`` - 64-bit pointers with 64-bit alignment.
1312 - ``p[n]:64:64:64`` - Other address spaces are assumed to be the
1313 same as the default address space.
1314 - ``S0`` - natural stack alignment is unspecified
1315 - ``i1:8:8`` - i1 is 8-bit (byte) aligned
1316 - ``i8:8:8`` - i8 is 8-bit (byte) aligned
1317 - ``i16:16:16`` - i16 is 16-bit aligned
1318 - ``i32:32:32`` - i32 is 32-bit aligned
1319 - ``i64:32:64`` - i64 has ABI alignment of 32-bits but preferred
1320 alignment of 64-bits
1321 - ``f16:16:16`` - half is 16-bit aligned
1322 - ``f32:32:32`` - float is 32-bit aligned
1323 - ``f64:64:64`` - double is 64-bit aligned
1324 - ``f128:128:128`` - quad is 128-bit aligned
1325 - ``v64:64:64`` - 64-bit vector is 64-bit aligned
1326 - ``v128:128:128`` - 128-bit vector is 128-bit aligned
1327 - ``a:0:64`` - aggregates are 64-bit aligned
1329 When LLVM is determining the alignment for a given type, it uses the
1332 #. If the type sought is an exact match for one of the specifications,
1333 that specification is used.
1334 #. If no match is found, and the type sought is an integer type, then
1335 the smallest integer type that is larger than the bitwidth of the
1336 sought type is used. If none of the specifications are larger than
1337 the bitwidth then the largest integer type is used. For example,
1338 given the default specifications above, the i7 type will use the
1339 alignment of i8 (next largest) while both i65 and i256 will use the
1340 alignment of i64 (largest specified).
1341 #. If no match is found, and the type sought is a vector type, then the
1342 largest vector type that is smaller than the sought vector type will
1343 be used as a fall back. This happens because <128 x double> can be
1344 implemented in terms of 64 <2 x double>, for example.
1346 The function of the data layout string may not be what you expect.
1347 Notably, this is not a specification from the frontend of what alignment
1348 the code generator should use.
1350 Instead, if specified, the target data layout is required to match what
1351 the ultimate *code generator* expects. This string is used by the
1352 mid-level optimizers to improve code, and this only works if it matches
1353 what the ultimate code generator uses. If you would like to generate IR
1354 that does not embed this target-specific detail into the IR, then you
1355 don't have to specify the string. This will disable some optimizations
1356 that require precise layout information, but this also prevents those
1357 optimizations from introducing target specificity into the IR.
1364 A module may specify a target triple string that describes the target
1365 host. The syntax for the target triple is simply:
1367 .. code-block:: llvm
1369 target triple = "x86_64-apple-macosx10.7.0"
1371 The *target triple* string consists of a series of identifiers delimited
1372 by the minus sign character ('-'). The canonical forms are:
1376 ARCHITECTURE-VENDOR-OPERATING_SYSTEM
1377 ARCHITECTURE-VENDOR-OPERATING_SYSTEM-ENVIRONMENT
1379 This information is passed along to the backend so that it generates
1380 code for the proper architecture. It's possible to override this on the
1381 command line with the ``-mtriple`` command line option.
1383 .. _pointeraliasing:
1385 Pointer Aliasing Rules
1386 ----------------------
1388 Any memory access must be done through a pointer value associated with
1389 an address range of the memory access, otherwise the behavior is
1390 undefined. Pointer values are associated with address ranges according
1391 to the following rules:
1393 - A pointer value is associated with the addresses associated with any
1394 value it is *based* on.
1395 - An address of a global variable is associated with the address range
1396 of the variable's storage.
1397 - The result value of an allocation instruction is associated with the
1398 address range of the allocated storage.
1399 - A null pointer in the default address-space is associated with no
1401 - An integer constant other than zero or a pointer value returned from
1402 a function not defined within LLVM may be associated with address
1403 ranges allocated through mechanisms other than those provided by
1404 LLVM. Such ranges shall not overlap with any ranges of addresses
1405 allocated by mechanisms provided by LLVM.
1407 A pointer value is *based* on another pointer value according to the
1410 - A pointer value formed from a ``getelementptr`` operation is *based*
1411 on the first operand of the ``getelementptr``.
1412 - The result value of a ``bitcast`` is *based* on the operand of the
1414 - A pointer value formed by an ``inttoptr`` is *based* on all pointer
1415 values that contribute (directly or indirectly) to the computation of
1416 the pointer's value.
1417 - The "*based* on" relationship is transitive.
1419 Note that this definition of *"based"* is intentionally similar to the
1420 definition of *"based"* in C99, though it is slightly weaker.
1422 LLVM IR does not associate types with memory. The result type of a
1423 ``load`` merely indicates the size and alignment of the memory from
1424 which to load, as well as the interpretation of the value. The first
1425 operand type of a ``store`` similarly only indicates the size and
1426 alignment of the store.
1428 Consequently, type-based alias analysis, aka TBAA, aka
1429 ``-fstrict-aliasing``, is not applicable to general unadorned LLVM IR.
1430 :ref:`Metadata <metadata>` may be used to encode additional information
1431 which specialized optimization passes may use to implement type-based
1436 Volatile Memory Accesses
1437 ------------------------
1439 Certain memory accesses, such as :ref:`load <i_load>`'s,
1440 :ref:`store <i_store>`'s, and :ref:`llvm.memcpy <int_memcpy>`'s may be
1441 marked ``volatile``. The optimizers must not change the number of
1442 volatile operations or change their order of execution relative to other
1443 volatile operations. The optimizers *may* change the order of volatile
1444 operations relative to non-volatile operations. This is not Java's
1445 "volatile" and has no cross-thread synchronization behavior.
1447 IR-level volatile loads and stores cannot safely be optimized into
1448 llvm.memcpy or llvm.memmove intrinsics even when those intrinsics are
1449 flagged volatile. Likewise, the backend should never split or merge
1450 target-legal volatile load/store instructions.
1452 .. admonition:: Rationale
1454 Platforms may rely on volatile loads and stores of natively supported
1455 data width to be executed as single instruction. For example, in C
1456 this holds for an l-value of volatile primitive type with native
1457 hardware support, but not necessarily for aggregate types. The
1458 frontend upholds these expectations, which are intentionally
1459 unspecified in the IR. The rules above ensure that IR transformation
1460 do not violate the frontend's contract with the language.
1464 Memory Model for Concurrent Operations
1465 --------------------------------------
1467 The LLVM IR does not define any way to start parallel threads of
1468 execution or to register signal handlers. Nonetheless, there are
1469 platform-specific ways to create them, and we define LLVM IR's behavior
1470 in their presence. This model is inspired by the C++0x memory model.
1472 For a more informal introduction to this model, see the :doc:`Atomics`.
1474 We define a *happens-before* partial order as the least partial order
1477 - Is a superset of single-thread program order, and
1478 - When a *synchronizes-with* ``b``, includes an edge from ``a`` to
1479 ``b``. *Synchronizes-with* pairs are introduced by platform-specific
1480 techniques, like pthread locks, thread creation, thread joining,
1481 etc., and by atomic instructions. (See also :ref:`Atomic Memory Ordering
1482 Constraints <ordering>`).
1484 Note that program order does not introduce *happens-before* edges
1485 between a thread and signals executing inside that thread.
1487 Every (defined) read operation (load instructions, memcpy, atomic
1488 loads/read-modify-writes, etc.) R reads a series of bytes written by
1489 (defined) write operations (store instructions, atomic
1490 stores/read-modify-writes, memcpy, etc.). For the purposes of this
1491 section, initialized globals are considered to have a write of the
1492 initializer which is atomic and happens before any other read or write
1493 of the memory in question. For each byte of a read R, R\ :sub:`byte`
1494 may see any write to the same byte, except:
1496 - If write\ :sub:`1` happens before write\ :sub:`2`, and
1497 write\ :sub:`2` happens before R\ :sub:`byte`, then
1498 R\ :sub:`byte` does not see write\ :sub:`1`.
1499 - If R\ :sub:`byte` happens before write\ :sub:`3`, then
1500 R\ :sub:`byte` does not see write\ :sub:`3`.
1502 Given that definition, R\ :sub:`byte` is defined as follows:
1504 - If R is volatile, the result is target-dependent. (Volatile is
1505 supposed to give guarantees which can support ``sig_atomic_t`` in
1506 C/C++, and may be used for accesses to addresses which do not behave
1507 like normal memory. It does not generally provide cross-thread
1509 - Otherwise, if there is no write to the same byte that happens before
1510 R\ :sub:`byte`, R\ :sub:`byte` returns ``undef`` for that byte.
1511 - Otherwise, if R\ :sub:`byte` may see exactly one write,
1512 R\ :sub:`byte` returns the value written by that write.
1513 - Otherwise, if R is atomic, and all the writes R\ :sub:`byte` may
1514 see are atomic, it chooses one of the values written. See the :ref:`Atomic
1515 Memory Ordering Constraints <ordering>` section for additional
1516 constraints on how the choice is made.
1517 - Otherwise R\ :sub:`byte` returns ``undef``.
1519 R returns the value composed of the series of bytes it read. This
1520 implies that some bytes within the value may be ``undef`` **without**
1521 the entire value being ``undef``. Note that this only defines the
1522 semantics of the operation; it doesn't mean that targets will emit more
1523 than one instruction to read the series of bytes.
1525 Note that in cases where none of the atomic intrinsics are used, this
1526 model places only one restriction on IR transformations on top of what
1527 is required for single-threaded execution: introducing a store to a byte
1528 which might not otherwise be stored is not allowed in general.
1529 (Specifically, in the case where another thread might write to and read
1530 from an address, introducing a store can change a load that may see
1531 exactly one write into a load that may see multiple writes.)
1535 Atomic Memory Ordering Constraints
1536 ----------------------------------
1538 Atomic instructions (:ref:`cmpxchg <i_cmpxchg>`,
1539 :ref:`atomicrmw <i_atomicrmw>`, :ref:`fence <i_fence>`,
1540 :ref:`atomic load <i_load>`, and :ref:`atomic store <i_store>`) take
1541 ordering parameters that determine which other atomic instructions on
1542 the same address they *synchronize with*. These semantics are borrowed
1543 from Java and C++0x, but are somewhat more colloquial. If these
1544 descriptions aren't precise enough, check those specs (see spec
1545 references in the :doc:`atomics guide <Atomics>`).
1546 :ref:`fence <i_fence>` instructions treat these orderings somewhat
1547 differently since they don't take an address. See that instruction's
1548 documentation for details.
1550 For a simpler introduction to the ordering constraints, see the
1554 The set of values that can be read is governed by the happens-before
1555 partial order. A value cannot be read unless some operation wrote
1556 it. This is intended to provide a guarantee strong enough to model
1557 Java's non-volatile shared variables. This ordering cannot be
1558 specified for read-modify-write operations; it is not strong enough
1559 to make them atomic in any interesting way.
1561 In addition to the guarantees of ``unordered``, there is a single
1562 total order for modifications by ``monotonic`` operations on each
1563 address. All modification orders must be compatible with the
1564 happens-before order. There is no guarantee that the modification
1565 orders can be combined to a global total order for the whole program
1566 (and this often will not be possible). The read in an atomic
1567 read-modify-write operation (:ref:`cmpxchg <i_cmpxchg>` and
1568 :ref:`atomicrmw <i_atomicrmw>`) reads the value in the modification
1569 order immediately before the value it writes. If one atomic read
1570 happens before another atomic read of the same address, the later
1571 read must see the same value or a later value in the address's
1572 modification order. This disallows reordering of ``monotonic`` (or
1573 stronger) operations on the same address. If an address is written
1574 ``monotonic``-ally by one thread, and other threads ``monotonic``-ally
1575 read that address repeatedly, the other threads must eventually see
1576 the write. This corresponds to the C++0x/C1x
1577 ``memory_order_relaxed``.
1579 In addition to the guarantees of ``monotonic``, a
1580 *synchronizes-with* edge may be formed with a ``release`` operation.
1581 This is intended to model C++'s ``memory_order_acquire``.
1583 In addition to the guarantees of ``monotonic``, if this operation
1584 writes a value which is subsequently read by an ``acquire``
1585 operation, it *synchronizes-with* that operation. (This isn't a
1586 complete description; see the C++0x definition of a release
1587 sequence.) This corresponds to the C++0x/C1x
1588 ``memory_order_release``.
1589 ``acq_rel`` (acquire+release)
1590 Acts as both an ``acquire`` and ``release`` operation on its
1591 address. This corresponds to the C++0x/C1x ``memory_order_acq_rel``.
1592 ``seq_cst`` (sequentially consistent)
1593 In addition to the guarantees of ``acq_rel`` (``acquire`` for an
1594 operation which only reads, ``release`` for an operation which only
1595 writes), there is a global total order on all
1596 sequentially-consistent operations on all addresses, which is
1597 consistent with the *happens-before* partial order and with the
1598 modification orders of all the affected addresses. Each
1599 sequentially-consistent read sees the last preceding write to the
1600 same address in this global order. This corresponds to the C++0x/C1x
1601 ``memory_order_seq_cst`` and Java volatile.
1605 If an atomic operation is marked ``singlethread``, it only *synchronizes
1606 with* or participates in modification and seq\_cst total orderings with
1607 other operations running in the same thread (for example, in signal
1615 LLVM IR floating-point binary ops (:ref:`fadd <i_fadd>`,
1616 :ref:`fsub <i_fsub>`, :ref:`fmul <i_fmul>`, :ref:`fdiv <i_fdiv>`,
1617 :ref:`frem <i_frem>`) have the following flags that can set to enable
1618 otherwise unsafe floating point operations
1621 No NaNs - Allow optimizations to assume the arguments and result are not
1622 NaN. Such optimizations are required to retain defined behavior over
1623 NaNs, but the value of the result is undefined.
1626 No Infs - Allow optimizations to assume the arguments and result are not
1627 +/-Inf. Such optimizations are required to retain defined behavior over
1628 +/-Inf, but the value of the result is undefined.
1631 No Signed Zeros - Allow optimizations to treat the sign of a zero
1632 argument or result as insignificant.
1635 Allow Reciprocal - Allow optimizations to use the reciprocal of an
1636 argument rather than perform division.
1639 Fast - Allow algebraically equivalent transformations that may
1640 dramatically change results in floating point (e.g. reassociate). This
1641 flag implies all the others.
1648 The LLVM type system is one of the most important features of the
1649 intermediate representation. Being typed enables a number of
1650 optimizations to be performed on the intermediate representation
1651 directly, without having to do extra analyses on the side before the
1652 transformation. A strong type system makes it easier to read the
1653 generated code and enables novel analyses and transformations that are
1654 not feasible to perform on normal three address code representations.
1664 The void type does not represent any value and has no size.
1682 The function type can be thought of as a function signature. It consists of a
1683 return type and a list of formal parameter types. The return type of a function
1684 type is a void type or first class type --- except for :ref:`label <t_label>`
1685 and :ref:`metadata <t_metadata>` types.
1691 <returntype> (<parameter list>)
1693 ...where '``<parameter list>``' is a comma-separated list of type
1694 specifiers. Optionally, the parameter list may include a type ``...``, which
1695 indicates that the function takes a variable number of arguments. Variable
1696 argument functions can access their arguments with the :ref:`variable argument
1697 handling intrinsic <int_varargs>` functions. '``<returntype>``' is any type
1698 except :ref:`label <t_label>` and :ref:`metadata <t_metadata>`.
1702 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1703 | ``i32 (i32)`` | function taking an ``i32``, returning an ``i32`` |
1704 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1705 | ``float (i16, i32 *) *`` | :ref:`Pointer <t_pointer>` to a function that takes an ``i16`` and a :ref:`pointer <t_pointer>` to ``i32``, returning ``float``. |
1706 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1707 | ``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. |
1708 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1709 | ``{i32, i32} (i32)`` | A function taking an ``i32``, returning a :ref:`structure <t_struct>` containing two ``i32`` values |
1710 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1717 The :ref:`first class <t_firstclass>` types are perhaps the most important.
1718 Values of these types are the only ones which can be produced by
1726 These are the types that are valid in registers from CodeGen's perspective.
1735 The integer type is a very simple type that simply specifies an
1736 arbitrary bit width for the integer type desired. Any bit width from 1
1737 bit to 2\ :sup:`23`\ -1 (about 8 million) can be specified.
1745 The number of bits the integer will occupy is specified by the ``N``
1751 +----------------+------------------------------------------------+
1752 | ``i1`` | a single-bit integer. |
1753 +----------------+------------------------------------------------+
1754 | ``i32`` | a 32-bit integer. |
1755 +----------------+------------------------------------------------+
1756 | ``i1942652`` | a really big integer of over 1 million bits. |
1757 +----------------+------------------------------------------------+
1761 Floating Point Types
1762 """"""""""""""""""""
1771 - 16-bit floating point value
1774 - 32-bit floating point value
1777 - 64-bit floating point value
1780 - 128-bit floating point value (112-bit mantissa)
1783 - 80-bit floating point value (X87)
1786 - 128-bit floating point value (two 64-bits)
1793 The x86_mmx type represents a value held in an MMX register on an x86
1794 machine. The operations allowed on it are quite limited: parameters and
1795 return values, load and store, and bitcast. User-specified MMX
1796 instructions are represented as intrinsic or asm calls with arguments
1797 and/or results of this type. There are no arrays, vectors or constants
1814 The pointer type is used to specify memory locations. Pointers are
1815 commonly used to reference objects in memory.
1817 Pointer types may have an optional address space attribute defining the
1818 numbered address space where the pointed-to object resides. The default
1819 address space is number zero. The semantics of non-zero address spaces
1820 are target-specific.
1822 Note that LLVM does not permit pointers to void (``void*``) nor does it
1823 permit pointers to labels (``label*``). Use ``i8*`` instead.
1833 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1834 | ``[4 x i32]*`` | A :ref:`pointer <t_pointer>` to :ref:`array <t_array>` of four ``i32`` values. |
1835 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1836 | ``i32 (i32*) *`` | A :ref:`pointer <t_pointer>` to a :ref:`function <t_function>` that takes an ``i32*``, returning an ``i32``. |
1837 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1838 | ``i32 addrspace(5)*`` | A :ref:`pointer <t_pointer>` to an ``i32`` value that resides in address space #5. |
1839 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1848 A vector type is a simple derived type that represents a vector of
1849 elements. Vector types are used when multiple primitive data are
1850 operated in parallel using a single instruction (SIMD). A vector type
1851 requires a size (number of elements) and an underlying primitive data
1852 type. Vector types are considered :ref:`first class <t_firstclass>`.
1858 < <# elements> x <elementtype> >
1860 The number of elements is a constant integer value larger than 0;
1861 elementtype may be any integer or floating point type, or a pointer to
1862 these types. Vectors of size zero are not allowed.
1866 +-------------------+--------------------------------------------------+
1867 | ``<4 x i32>`` | Vector of 4 32-bit integer values. |
1868 +-------------------+--------------------------------------------------+
1869 | ``<8 x float>`` | Vector of 8 32-bit floating-point values. |
1870 +-------------------+--------------------------------------------------+
1871 | ``<2 x i64>`` | Vector of 2 64-bit integer values. |
1872 +-------------------+--------------------------------------------------+
1873 | ``<4 x i64*>`` | Vector of 4 pointers to 64-bit integer values. |
1874 +-------------------+--------------------------------------------------+
1883 The label type represents code labels.
1898 The metadata type represents embedded metadata. No derived types may be
1899 created from metadata except for :ref:`function <t_function>` arguments.
1912 Aggregate Types are a subset of derived types that can contain multiple
1913 member types. :ref:`Arrays <t_array>` and :ref:`structs <t_struct>` are
1914 aggregate types. :ref:`Vectors <t_vector>` are not considered to be
1924 The array type is a very simple derived type that arranges elements
1925 sequentially in memory. The array type requires a size (number of
1926 elements) and an underlying data type.
1932 [<# elements> x <elementtype>]
1934 The number of elements is a constant integer value; ``elementtype`` may
1935 be any type with a size.
1939 +------------------+--------------------------------------+
1940 | ``[40 x i32]`` | Array of 40 32-bit integer values. |
1941 +------------------+--------------------------------------+
1942 | ``[41 x i32]`` | Array of 41 32-bit integer values. |
1943 +------------------+--------------------------------------+
1944 | ``[4 x i8]`` | Array of 4 8-bit integer values. |
1945 +------------------+--------------------------------------+
1947 Here are some examples of multidimensional arrays:
1949 +-----------------------------+----------------------------------------------------------+
1950 | ``[3 x [4 x i32]]`` | 3x4 array of 32-bit integer values. |
1951 +-----------------------------+----------------------------------------------------------+
1952 | ``[12 x [10 x float]]`` | 12x10 array of single precision floating point values. |
1953 +-----------------------------+----------------------------------------------------------+
1954 | ``[2 x [3 x [4 x i16]]]`` | 2x3x4 array of 16-bit integer values. |
1955 +-----------------------------+----------------------------------------------------------+
1957 There is no restriction on indexing beyond the end of the array implied
1958 by a static type (though there are restrictions on indexing beyond the
1959 bounds of an allocated object in some cases). This means that
1960 single-dimension 'variable sized array' addressing can be implemented in
1961 LLVM with a zero length array type. An implementation of 'pascal style
1962 arrays' in LLVM could use the type "``{ i32, [0 x float]}``", for
1972 The structure type is used to represent a collection of data members
1973 together in memory. The elements of a structure may be any type that has
1976 Structures in memory are accessed using '``load``' and '``store``' by
1977 getting a pointer to a field with the '``getelementptr``' instruction.
1978 Structures in registers are accessed using the '``extractvalue``' and
1979 '``insertvalue``' instructions.
1981 Structures may optionally be "packed" structures, which indicate that
1982 the alignment of the struct is one byte, and that there is no padding
1983 between the elements. In non-packed structs, padding between field types
1984 is inserted as defined by the DataLayout string in the module, which is
1985 required to match what the underlying code generator expects.
1987 Structures can either be "literal" or "identified". A literal structure
1988 is defined inline with other types (e.g. ``{i32, i32}*``) whereas
1989 identified types are always defined at the top level with a name.
1990 Literal types are uniqued by their contents and can never be recursive
1991 or opaque since there is no way to write one. Identified types can be
1992 recursive, can be opaqued, and are never uniqued.
1998 %T1 = type { <type list> } ; Identified normal struct type
1999 %T2 = type <{ <type list> }> ; Identified packed struct type
2003 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2004 | ``{ i32, i32, i32 }`` | A triple of three ``i32`` values |
2005 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2006 | ``{ 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``. |
2007 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2008 | ``<{ i8, i32 }>`` | A packed struct known to be 5 bytes in size. |
2009 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2013 Opaque Structure Types
2014 """"""""""""""""""""""
2018 Opaque structure types are used to represent named structure types that
2019 do not have a body specified. This corresponds (for example) to the C
2020 notion of a forward declared structure.
2031 +--------------+-------------------+
2032 | ``opaque`` | An opaque type. |
2033 +--------------+-------------------+
2040 LLVM has several different basic types of constants. This section
2041 describes them all and their syntax.
2046 **Boolean constants**
2047 The two strings '``true``' and '``false``' are both valid constants
2049 **Integer constants**
2050 Standard integers (such as '4') are constants of the
2051 :ref:`integer <t_integer>` type. Negative numbers may be used with
2053 **Floating point constants**
2054 Floating point constants use standard decimal notation (e.g.
2055 123.421), exponential notation (e.g. 1.23421e+2), or a more precise
2056 hexadecimal notation (see below). The assembler requires the exact
2057 decimal value of a floating-point constant. For example, the
2058 assembler accepts 1.25 but rejects 1.3 because 1.3 is a repeating
2059 decimal in binary. Floating point constants must have a :ref:`floating
2060 point <t_floating>` type.
2061 **Null pointer constants**
2062 The identifier '``null``' is recognized as a null pointer constant
2063 and must be of :ref:`pointer type <t_pointer>`.
2065 The one non-intuitive notation for constants is the hexadecimal form of
2066 floating point constants. For example, the form
2067 '``double 0x432ff973cafa8000``' is equivalent to (but harder to read
2068 than) '``double 4.5e+15``'. The only time hexadecimal floating point
2069 constants are required (and the only time that they are generated by the
2070 disassembler) is when a floating point constant must be emitted but it
2071 cannot be represented as a decimal floating point number in a reasonable
2072 number of digits. For example, NaN's, infinities, and other special
2073 values are represented in their IEEE hexadecimal format so that assembly
2074 and disassembly do not cause any bits to change in the constants.
2076 When using the hexadecimal form, constants of types half, float, and
2077 double are represented using the 16-digit form shown above (which
2078 matches the IEEE754 representation for double); half and float values
2079 must, however, be exactly representable as IEEE 754 half and single
2080 precision, respectively. Hexadecimal format is always used for long
2081 double, and there are three forms of long double. The 80-bit format used
2082 by x86 is represented as ``0xK`` followed by 20 hexadecimal digits. The
2083 128-bit format used by PowerPC (two adjacent doubles) is represented by
2084 ``0xM`` followed by 32 hexadecimal digits. The IEEE 128-bit format is
2085 represented by ``0xL`` followed by 32 hexadecimal digits. Long doubles
2086 will only work if they match the long double format on your target.
2087 The IEEE 16-bit format (half precision) is represented by ``0xH``
2088 followed by 4 hexadecimal digits. All hexadecimal formats are big-endian
2089 (sign bit at the left).
2091 There are no constants of type x86_mmx.
2093 .. _complexconstants:
2098 Complex constants are a (potentially recursive) combination of simple
2099 constants and smaller complex constants.
2101 **Structure constants**
2102 Structure constants are represented with notation similar to
2103 structure type definitions (a comma separated list of elements,
2104 surrounded by braces (``{}``)). For example:
2105 "``{ i32 4, float 17.0, i32* @G }``", where "``@G``" is declared as
2106 "``@G = external global i32``". Structure constants must have
2107 :ref:`structure type <t_struct>`, and the number and types of elements
2108 must match those specified by the type.
2110 Array constants are represented with notation similar to array type
2111 definitions (a comma separated list of elements, surrounded by
2112 square brackets (``[]``)). For example:
2113 "``[ i32 42, i32 11, i32 74 ]``". Array constants must have
2114 :ref:`array type <t_array>`, and the number and types of elements must
2115 match those specified by the type.
2116 **Vector constants**
2117 Vector constants are represented with notation similar to vector
2118 type definitions (a comma separated list of elements, surrounded by
2119 less-than/greater-than's (``<>``)). For example:
2120 "``< i32 42, i32 11, i32 74, i32 100 >``". Vector constants
2121 must have :ref:`vector type <t_vector>`, and the number and types of
2122 elements must match those specified by the type.
2123 **Zero initialization**
2124 The string '``zeroinitializer``' can be used to zero initialize a
2125 value to zero of *any* type, including scalar and
2126 :ref:`aggregate <t_aggregate>` types. This is often used to avoid
2127 having to print large zero initializers (e.g. for large arrays) and
2128 is always exactly equivalent to using explicit zero initializers.
2130 A metadata node is a structure-like constant with :ref:`metadata
2131 type <t_metadata>`. For example:
2132 "``metadata !{ i32 0, metadata !"test" }``". Unlike other
2133 constants that are meant to be interpreted as part of the
2134 instruction stream, metadata is a place to attach additional
2135 information such as debug info.
2137 Global Variable and Function Addresses
2138 --------------------------------------
2140 The addresses of :ref:`global variables <globalvars>` and
2141 :ref:`functions <functionstructure>` are always implicitly valid
2142 (link-time) constants. These constants are explicitly referenced when
2143 the :ref:`identifier for the global <identifiers>` is used and always have
2144 :ref:`pointer <t_pointer>` type. For example, the following is a legal LLVM
2147 .. code-block:: llvm
2151 @Z = global [2 x i32*] [ i32* @X, i32* @Y ]
2158 The string '``undef``' can be used anywhere a constant is expected, and
2159 indicates that the user of the value may receive an unspecified
2160 bit-pattern. Undefined values may be of any type (other than '``label``'
2161 or '``void``') and be used anywhere a constant is permitted.
2163 Undefined values are useful because they indicate to the compiler that
2164 the program is well defined no matter what value is used. This gives the
2165 compiler more freedom to optimize. Here are some examples of
2166 (potentially surprising) transformations that are valid (in pseudo IR):
2168 .. code-block:: llvm
2178 This is safe because all of the output bits are affected by the undef
2179 bits. Any output bit can have a zero or one depending on the input bits.
2181 .. code-block:: llvm
2192 These logical operations have bits that are not always affected by the
2193 input. For example, if ``%X`` has a zero bit, then the output of the
2194 '``and``' operation will always be a zero for that bit, no matter what
2195 the corresponding bit from the '``undef``' is. As such, it is unsafe to
2196 optimize or assume that the result of the '``and``' is '``undef``'.
2197 However, it is safe to assume that all bits of the '``undef``' could be
2198 0, and optimize the '``and``' to 0. Likewise, it is safe to assume that
2199 all the bits of the '``undef``' operand to the '``or``' could be set,
2200 allowing the '``or``' to be folded to -1.
2202 .. code-block:: llvm
2204 %A = select undef, %X, %Y
2205 %B = select undef, 42, %Y
2206 %C = select %X, %Y, undef
2216 This set of examples shows that undefined '``select``' (and conditional
2217 branch) conditions can go *either way*, but they have to come from one
2218 of the two operands. In the ``%A`` example, if ``%X`` and ``%Y`` were
2219 both known to have a clear low bit, then ``%A`` would have to have a
2220 cleared low bit. However, in the ``%C`` example, the optimizer is
2221 allowed to assume that the '``undef``' operand could be the same as
2222 ``%Y``, allowing the whole '``select``' to be eliminated.
2224 .. code-block:: llvm
2226 %A = xor undef, undef
2243 This example points out that two '``undef``' operands are not
2244 necessarily the same. This can be surprising to people (and also matches
2245 C semantics) where they assume that "``X^X``" is always zero, even if
2246 ``X`` is undefined. This isn't true for a number of reasons, but the
2247 short answer is that an '``undef``' "variable" can arbitrarily change
2248 its value over its "live range". This is true because the variable
2249 doesn't actually *have a live range*. Instead, the value is logically
2250 read from arbitrary registers that happen to be around when needed, so
2251 the value is not necessarily consistent over time. In fact, ``%A`` and
2252 ``%C`` need to have the same semantics or the core LLVM "replace all
2253 uses with" concept would not hold.
2255 .. code-block:: llvm
2263 These examples show the crucial difference between an *undefined value*
2264 and *undefined behavior*. An undefined value (like '``undef``') is
2265 allowed to have an arbitrary bit-pattern. This means that the ``%A``
2266 operation can be constant folded to '``undef``', because the '``undef``'
2267 could be an SNaN, and ``fdiv`` is not (currently) defined on SNaN's.
2268 However, in the second example, we can make a more aggressive
2269 assumption: because the ``undef`` is allowed to be an arbitrary value,
2270 we are allowed to assume that it could be zero. Since a divide by zero
2271 has *undefined behavior*, we are allowed to assume that the operation
2272 does not execute at all. This allows us to delete the divide and all
2273 code after it. Because the undefined operation "can't happen", the
2274 optimizer can assume that it occurs in dead code.
2276 .. code-block:: llvm
2278 a: store undef -> %X
2279 b: store %X -> undef
2284 These examples reiterate the ``fdiv`` example: a store *of* an undefined
2285 value can be assumed to not have any effect; we can assume that the
2286 value is overwritten with bits that happen to match what was already
2287 there. However, a store *to* an undefined location could clobber
2288 arbitrary memory, therefore, it has undefined behavior.
2295 Poison values are similar to :ref:`undef values <undefvalues>`, however
2296 they also represent the fact that an instruction or constant expression
2297 which cannot evoke side effects has nevertheless detected a condition
2298 which results in undefined behavior.
2300 There is currently no way of representing a poison value in the IR; they
2301 only exist when produced by operations such as :ref:`add <i_add>` with
2304 Poison value behavior is defined in terms of value *dependence*:
2306 - Values other than :ref:`phi <i_phi>` nodes depend on their operands.
2307 - :ref:`Phi <i_phi>` nodes depend on the operand corresponding to
2308 their dynamic predecessor basic block.
2309 - Function arguments depend on the corresponding actual argument values
2310 in the dynamic callers of their functions.
2311 - :ref:`Call <i_call>` instructions depend on the :ref:`ret <i_ret>`
2312 instructions that dynamically transfer control back to them.
2313 - :ref:`Invoke <i_invoke>` instructions depend on the
2314 :ref:`ret <i_ret>`, :ref:`resume <i_resume>`, or exception-throwing
2315 call instructions that dynamically transfer control back to them.
2316 - Non-volatile loads and stores depend on the most recent stores to all
2317 of the referenced memory addresses, following the order in the IR
2318 (including loads and stores implied by intrinsics such as
2319 :ref:`@llvm.memcpy <int_memcpy>`.)
2320 - An instruction with externally visible side effects depends on the
2321 most recent preceding instruction with externally visible side
2322 effects, following the order in the IR. (This includes :ref:`volatile
2323 operations <volatile>`.)
2324 - An instruction *control-depends* on a :ref:`terminator
2325 instruction <terminators>` if the terminator instruction has
2326 multiple successors and the instruction is always executed when
2327 control transfers to one of the successors, and may not be executed
2328 when control is transferred to another.
2329 - Additionally, an instruction also *control-depends* on a terminator
2330 instruction if the set of instructions it otherwise depends on would
2331 be different if the terminator had transferred control to a different
2333 - Dependence is transitive.
2335 Poison Values have the same behavior as :ref:`undef values <undefvalues>`,
2336 with the additional affect that any instruction which has a *dependence*
2337 on a poison value has undefined behavior.
2339 Here are some examples:
2341 .. code-block:: llvm
2344 %poison = sub nuw i32 0, 1 ; Results in a poison value.
2345 %still_poison = and i32 %poison, 0 ; 0, but also poison.
2346 %poison_yet_again = getelementptr i32* @h, i32 %still_poison
2347 store i32 0, i32* %poison_yet_again ; memory at @h[0] is poisoned
2349 store i32 %poison, i32* @g ; Poison value stored to memory.
2350 %poison2 = load i32* @g ; Poison value loaded back from memory.
2352 store volatile i32 %poison, i32* @g ; External observation; undefined behavior.
2354 %narrowaddr = bitcast i32* @g to i16*
2355 %wideaddr = bitcast i32* @g to i64*
2356 %poison3 = load i16* %narrowaddr ; Returns a poison value.
2357 %poison4 = load i64* %wideaddr ; Returns a poison value.
2359 %cmp = icmp slt i32 %poison, 0 ; Returns a poison value.
2360 br i1 %cmp, label %true, label %end ; Branch to either destination.
2363 store volatile i32 0, i32* @g ; This is control-dependent on %cmp, so
2364 ; it has undefined behavior.
2368 %p = phi i32 [ 0, %entry ], [ 1, %true ]
2369 ; Both edges into this PHI are
2370 ; control-dependent on %cmp, so this
2371 ; always results in a poison value.
2373 store volatile i32 0, i32* @g ; This would depend on the store in %true
2374 ; if %cmp is true, or the store in %entry
2375 ; otherwise, so this is undefined behavior.
2377 br i1 %cmp, label %second_true, label %second_end
2378 ; The same branch again, but this time the
2379 ; true block doesn't have side effects.
2386 store volatile i32 0, i32* @g ; This time, the instruction always depends
2387 ; on the store in %end. Also, it is
2388 ; control-equivalent to %end, so this is
2389 ; well-defined (ignoring earlier undefined
2390 ; behavior in this example).
2394 Addresses of Basic Blocks
2395 -------------------------
2397 ``blockaddress(@function, %block)``
2399 The '``blockaddress``' constant computes the address of the specified
2400 basic block in the specified function, and always has an ``i8*`` type.
2401 Taking the address of the entry block is illegal.
2403 This value only has defined behavior when used as an operand to the
2404 ':ref:`indirectbr <i_indirectbr>`' instruction, or for comparisons
2405 against null. Pointer equality tests between labels addresses results in
2406 undefined behavior --- though, again, comparison against null is ok, and
2407 no label is equal to the null pointer. This may be passed around as an
2408 opaque pointer sized value as long as the bits are not inspected. This
2409 allows ``ptrtoint`` and arithmetic to be performed on these values so
2410 long as the original value is reconstituted before the ``indirectbr``
2413 Finally, some targets may provide defined semantics when using the value
2414 as the operand to an inline assembly, but that is target specific.
2418 Constant Expressions
2419 --------------------
2421 Constant expressions are used to allow expressions involving other
2422 constants to be used as constants. Constant expressions may be of any
2423 :ref:`first class <t_firstclass>` type and may involve any LLVM operation
2424 that does not have side effects (e.g. load and call are not supported).
2425 The following is the syntax for constant expressions:
2427 ``trunc (CST to TYPE)``
2428 Truncate a constant to another type. The bit size of CST must be
2429 larger than the bit size of TYPE. Both types must be integers.
2430 ``zext (CST to TYPE)``
2431 Zero extend a constant to another type. The bit size of CST must be
2432 smaller than the bit size of TYPE. Both types must be integers.
2433 ``sext (CST to TYPE)``
2434 Sign extend a constant to another type. The bit size of CST must be
2435 smaller than the bit size of TYPE. Both types must be integers.
2436 ``fptrunc (CST to TYPE)``
2437 Truncate a floating point constant to another floating point type.
2438 The size of CST must be larger than the size of TYPE. Both types
2439 must be floating point.
2440 ``fpext (CST to TYPE)``
2441 Floating point extend a constant to another type. The size of CST
2442 must be smaller or equal to the size of TYPE. Both types must be
2444 ``fptoui (CST to TYPE)``
2445 Convert a floating point constant to the corresponding unsigned
2446 integer constant. TYPE must be a scalar or vector integer type. CST
2447 must be of scalar or vector floating point type. Both CST and TYPE
2448 must be scalars, or vectors of the same number of elements. If the
2449 value won't fit in the integer type, the results are undefined.
2450 ``fptosi (CST to TYPE)``
2451 Convert a floating point constant to the corresponding signed
2452 integer constant. TYPE must be a scalar or vector integer type. CST
2453 must be of scalar or vector floating point type. Both CST and TYPE
2454 must be scalars, or vectors of the same number of elements. If the
2455 value won't fit in the integer type, the results are undefined.
2456 ``uitofp (CST to TYPE)``
2457 Convert an unsigned integer constant to the corresponding floating
2458 point constant. TYPE must be a scalar or vector floating point type.
2459 CST must be of scalar or vector integer type. Both CST and TYPE must
2460 be scalars, or vectors of the same number of elements. If the value
2461 won't fit in the floating point type, the results are undefined.
2462 ``sitofp (CST to TYPE)``
2463 Convert a signed integer constant to the corresponding floating
2464 point constant. TYPE must be a scalar or vector floating point type.
2465 CST must be of scalar or vector integer type. Both CST and TYPE must
2466 be scalars, or vectors of the same number of elements. If the value
2467 won't fit in the floating point type, the results are undefined.
2468 ``ptrtoint (CST to TYPE)``
2469 Convert a pointer typed constant to the corresponding integer
2470 constant. ``TYPE`` must be an integer type. ``CST`` must be of
2471 pointer type. The ``CST`` value is zero extended, truncated, or
2472 unchanged to make it fit in ``TYPE``.
2473 ``inttoptr (CST to TYPE)``
2474 Convert an integer constant to a pointer constant. TYPE must be a
2475 pointer type. CST must be of integer type. The CST value is zero
2476 extended, truncated, or unchanged to make it fit in a pointer size.
2477 This one is *really* dangerous!
2478 ``bitcast (CST to TYPE)``
2479 Convert a constant, CST, to another TYPE. The constraints of the
2480 operands are the same as those for the :ref:`bitcast
2481 instruction <i_bitcast>`.
2482 ``addrspacecast (CST to TYPE)``
2483 Convert a constant pointer or constant vector of pointer, CST, to another
2484 TYPE in a different address space. The constraints of the operands are the
2485 same as those for the :ref:`addrspacecast instruction <i_addrspacecast>`.
2486 ``getelementptr (CSTPTR, IDX0, IDX1, ...)``, ``getelementptr inbounds (CSTPTR, IDX0, IDX1, ...)``
2487 Perform the :ref:`getelementptr operation <i_getelementptr>` on
2488 constants. As with the :ref:`getelementptr <i_getelementptr>`
2489 instruction, the index list may have zero or more indexes, which are
2490 required to make sense for the type of "CSTPTR".
2491 ``select (COND, VAL1, VAL2)``
2492 Perform the :ref:`select operation <i_select>` on constants.
2493 ``icmp COND (VAL1, VAL2)``
2494 Performs the :ref:`icmp operation <i_icmp>` on constants.
2495 ``fcmp COND (VAL1, VAL2)``
2496 Performs the :ref:`fcmp operation <i_fcmp>` on constants.
2497 ``extractelement (VAL, IDX)``
2498 Perform the :ref:`extractelement operation <i_extractelement>` on
2500 ``insertelement (VAL, ELT, IDX)``
2501 Perform the :ref:`insertelement operation <i_insertelement>` on
2503 ``shufflevector (VEC1, VEC2, IDXMASK)``
2504 Perform the :ref:`shufflevector operation <i_shufflevector>` on
2506 ``extractvalue (VAL, IDX0, IDX1, ...)``
2507 Perform the :ref:`extractvalue operation <i_extractvalue>` on
2508 constants. The index list is interpreted in a similar manner as
2509 indices in a ':ref:`getelementptr <i_getelementptr>`' operation. At
2510 least one index value must be specified.
2511 ``insertvalue (VAL, ELT, IDX0, IDX1, ...)``
2512 Perform the :ref:`insertvalue operation <i_insertvalue>` on constants.
2513 The index list is interpreted in a similar manner as indices in a
2514 ':ref:`getelementptr <i_getelementptr>`' operation. At least one index
2515 value must be specified.
2516 ``OPCODE (LHS, RHS)``
2517 Perform the specified operation of the LHS and RHS constants. OPCODE
2518 may be any of the :ref:`binary <binaryops>` or :ref:`bitwise
2519 binary <bitwiseops>` operations. The constraints on operands are
2520 the same as those for the corresponding instruction (e.g. no bitwise
2521 operations on floating point values are allowed).
2528 Inline Assembler Expressions
2529 ----------------------------
2531 LLVM supports inline assembler expressions (as opposed to :ref:`Module-Level
2532 Inline Assembly <moduleasm>`) through the use of a special value. This
2533 value represents the inline assembler as a string (containing the
2534 instructions to emit), a list of operand constraints (stored as a
2535 string), a flag that indicates whether or not the inline asm expression
2536 has side effects, and a flag indicating whether the function containing
2537 the asm needs to align its stack conservatively. An example inline
2538 assembler expression is:
2540 .. code-block:: llvm
2542 i32 (i32) asm "bswap $0", "=r,r"
2544 Inline assembler expressions may **only** be used as the callee operand
2545 of a :ref:`call <i_call>` or an :ref:`invoke <i_invoke>` instruction.
2546 Thus, typically we have:
2548 .. code-block:: llvm
2550 %X = call i32 asm "bswap $0", "=r,r"(i32 %Y)
2552 Inline asms with side effects not visible in the constraint list must be
2553 marked as having side effects. This is done through the use of the
2554 '``sideeffect``' keyword, like so:
2556 .. code-block:: llvm
2558 call void asm sideeffect "eieio", ""()
2560 In some cases inline asms will contain code that will not work unless
2561 the stack is aligned in some way, such as calls or SSE instructions on
2562 x86, yet will not contain code that does that alignment within the asm.
2563 The compiler should make conservative assumptions about what the asm
2564 might contain and should generate its usual stack alignment code in the
2565 prologue if the '``alignstack``' keyword is present:
2567 .. code-block:: llvm
2569 call void asm alignstack "eieio", ""()
2571 Inline asms also support using non-standard assembly dialects. The
2572 assumed dialect is ATT. When the '``inteldialect``' keyword is present,
2573 the inline asm is using the Intel dialect. Currently, ATT and Intel are
2574 the only supported dialects. An example is:
2576 .. code-block:: llvm
2578 call void asm inteldialect "eieio", ""()
2580 If multiple keywords appear the '``sideeffect``' keyword must come
2581 first, the '``alignstack``' keyword second and the '``inteldialect``'
2587 The call instructions that wrap inline asm nodes may have a
2588 "``!srcloc``" MDNode attached to it that contains a list of constant
2589 integers. If present, the code generator will use the integer as the
2590 location cookie value when report errors through the ``LLVMContext``
2591 error reporting mechanisms. This allows a front-end to correlate backend
2592 errors that occur with inline asm back to the source code that produced
2595 .. code-block:: llvm
2597 call void asm sideeffect "something bad", ""(), !srcloc !42
2599 !42 = !{ i32 1234567 }
2601 It is up to the front-end to make sense of the magic numbers it places
2602 in the IR. If the MDNode contains multiple constants, the code generator
2603 will use the one that corresponds to the line of the asm that the error
2608 Metadata Nodes and Metadata Strings
2609 -----------------------------------
2611 LLVM IR allows metadata to be attached to instructions in the program
2612 that can convey extra information about the code to the optimizers and
2613 code generator. One example application of metadata is source-level
2614 debug information. There are two metadata primitives: strings and nodes.
2615 All metadata has the ``metadata`` type and is identified in syntax by a
2616 preceding exclamation point ('``!``').
2618 A metadata string is a string surrounded by double quotes. It can
2619 contain any character by escaping non-printable characters with
2620 "``\xx``" where "``xx``" is the two digit hex code. For example:
2623 Metadata nodes are represented with notation similar to structure
2624 constants (a comma separated list of elements, surrounded by braces and
2625 preceded by an exclamation point). Metadata nodes can have any values as
2626 their operand. For example:
2628 .. code-block:: llvm
2630 !{ metadata !"test\00", i32 10}
2632 A :ref:`named metadata <namedmetadatastructure>` is a collection of
2633 metadata nodes, which can be looked up in the module symbol table. For
2636 .. code-block:: llvm
2638 !foo = metadata !{!4, !3}
2640 Metadata can be used as function arguments. Here ``llvm.dbg.value``
2641 function is using two metadata arguments:
2643 .. code-block:: llvm
2645 call void @llvm.dbg.value(metadata !24, i64 0, metadata !25)
2647 Metadata can be attached with an instruction. Here metadata ``!21`` is
2648 attached to the ``add`` instruction using the ``!dbg`` identifier:
2650 .. code-block:: llvm
2652 %indvar.next = add i64 %indvar, 1, !dbg !21
2654 More information about specific metadata nodes recognized by the
2655 optimizers and code generator is found below.
2660 In LLVM IR, memory does not have types, so LLVM's own type system is not
2661 suitable for doing TBAA. Instead, metadata is added to the IR to
2662 describe a type system of a higher level language. This can be used to
2663 implement typical C/C++ TBAA, but it can also be used to implement
2664 custom alias analysis behavior for other languages.
2666 The current metadata format is very simple. TBAA metadata nodes have up
2667 to three fields, e.g.:
2669 .. code-block:: llvm
2671 !0 = metadata !{ metadata !"an example type tree" }
2672 !1 = metadata !{ metadata !"int", metadata !0 }
2673 !2 = metadata !{ metadata !"float", metadata !0 }
2674 !3 = metadata !{ metadata !"const float", metadata !2, i64 1 }
2676 The first field is an identity field. It can be any value, usually a
2677 metadata string, which uniquely identifies the type. The most important
2678 name in the tree is the name of the root node. Two trees with different
2679 root node names are entirely disjoint, even if they have leaves with
2682 The second field identifies the type's parent node in the tree, or is
2683 null or omitted for a root node. A type is considered to alias all of
2684 its descendants and all of its ancestors in the tree. Also, a type is
2685 considered to alias all types in other trees, so that bitcode produced
2686 from multiple front-ends is handled conservatively.
2688 If the third field is present, it's an integer which if equal to 1
2689 indicates that the type is "constant" (meaning
2690 ``pointsToConstantMemory`` should return true; see `other useful
2691 AliasAnalysis methods <AliasAnalysis.html#OtherItfs>`_).
2693 '``tbaa.struct``' Metadata
2694 ^^^^^^^^^^^^^^^^^^^^^^^^^^
2696 The :ref:`llvm.memcpy <int_memcpy>` is often used to implement
2697 aggregate assignment operations in C and similar languages, however it
2698 is defined to copy a contiguous region of memory, which is more than
2699 strictly necessary for aggregate types which contain holes due to
2700 padding. Also, it doesn't contain any TBAA information about the fields
2703 ``!tbaa.struct`` metadata can describe which memory subregions in a
2704 memcpy are padding and what the TBAA tags of the struct are.
2706 The current metadata format is very simple. ``!tbaa.struct`` metadata
2707 nodes are a list of operands which are in conceptual groups of three.
2708 For each group of three, the first operand gives the byte offset of a
2709 field in bytes, the second gives its size in bytes, and the third gives
2712 .. code-block:: llvm
2714 !4 = metadata !{ i64 0, i64 4, metadata !1, i64 8, i64 4, metadata !2 }
2716 This describes a struct with two fields. The first is at offset 0 bytes
2717 with size 4 bytes, and has tbaa tag !1. The second is at offset 8 bytes
2718 and has size 4 bytes and has tbaa tag !2.
2720 Note that the fields need not be contiguous. In this example, there is a
2721 4 byte gap between the two fields. This gap represents padding which
2722 does not carry useful data and need not be preserved.
2724 '``fpmath``' Metadata
2725 ^^^^^^^^^^^^^^^^^^^^^
2727 ``fpmath`` metadata may be attached to any instruction of floating point
2728 type. It can be used to express the maximum acceptable error in the
2729 result of that instruction, in ULPs, thus potentially allowing the
2730 compiler to use a more efficient but less accurate method of computing
2731 it. ULP is defined as follows:
2733 If ``x`` is a real number that lies between two finite consecutive
2734 floating-point numbers ``a`` and ``b``, without being equal to one
2735 of them, then ``ulp(x) = |b - a|``, otherwise ``ulp(x)`` is the
2736 distance between the two non-equal finite floating-point numbers
2737 nearest ``x``. Moreover, ``ulp(NaN)`` is ``NaN``.
2739 The metadata node shall consist of a single positive floating point
2740 number representing the maximum relative error, for example:
2742 .. code-block:: llvm
2744 !0 = metadata !{ float 2.5 } ; maximum acceptable inaccuracy is 2.5 ULPs
2746 '``range``' Metadata
2747 ^^^^^^^^^^^^^^^^^^^^
2749 ``range`` metadata may be attached only to loads of integer types. It
2750 expresses the possible ranges the loaded value is in. The ranges are
2751 represented with a flattened list of integers. The loaded value is known
2752 to be in the union of the ranges defined by each consecutive pair. Each
2753 pair has the following properties:
2755 - The type must match the type loaded by the instruction.
2756 - The pair ``a,b`` represents the range ``[a,b)``.
2757 - Both ``a`` and ``b`` are constants.
2758 - The range is allowed to wrap.
2759 - The range should not represent the full or empty set. That is,
2762 In addition, the pairs must be in signed order of the lower bound and
2763 they must be non-contiguous.
2767 .. code-block:: llvm
2769 %a = load i8* %x, align 1, !range !0 ; Can only be 0 or 1
2770 %b = load i8* %y, align 1, !range !1 ; Can only be 255 (-1), 0 or 1
2771 %c = load i8* %z, align 1, !range !2 ; Can only be 0, 1, 3, 4 or 5
2772 %d = load i8* %z, align 1, !range !3 ; Can only be -2, -1, 3, 4 or 5
2774 !0 = metadata !{ i8 0, i8 2 }
2775 !1 = metadata !{ i8 255, i8 2 }
2776 !2 = metadata !{ i8 0, i8 2, i8 3, i8 6 }
2777 !3 = metadata !{ i8 -2, i8 0, i8 3, i8 6 }
2782 It is sometimes useful to attach information to loop constructs. Currently,
2783 loop metadata is implemented as metadata attached to the branch instruction
2784 in the loop latch block. This type of metadata refer to a metadata node that is
2785 guaranteed to be separate for each loop. The loop identifier metadata is
2786 specified with the name ``llvm.loop``.
2788 The loop identifier metadata is implemented using a metadata that refers to
2789 itself to avoid merging it with any other identifier metadata, e.g.,
2790 during module linkage or function inlining. That is, each loop should refer
2791 to their own identification metadata even if they reside in separate functions.
2792 The following example contains loop identifier metadata for two separate loop
2795 .. code-block:: llvm
2797 !0 = metadata !{ metadata !0 }
2798 !1 = metadata !{ metadata !1 }
2800 The loop identifier metadata can be used to specify additional per-loop
2801 metadata. Any operands after the first operand can be treated as user-defined
2802 metadata. For example the ``llvm.vectorizer.unroll`` metadata is understood
2803 by the loop vectorizer to indicate how many times to unroll the loop:
2805 .. code-block:: llvm
2807 br i1 %exitcond, label %._crit_edge, label %.lr.ph, !llvm.loop !0
2809 !0 = metadata !{ metadata !0, metadata !1 }
2810 !1 = metadata !{ metadata !"llvm.vectorizer.unroll", i32 2 }
2815 Metadata types used to annotate memory accesses with information helpful
2816 for optimizations are prefixed with ``llvm.mem``.
2818 '``llvm.mem.parallel_loop_access``' Metadata
2819 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2821 The ``llvm.mem.parallel_loop_access`` metadata refers to a loop identifier,
2822 or metadata containing a list of loop identifiers for nested loops.
2823 The metadata is attached to memory accessing instructions and denotes that
2824 no loop carried memory dependence exist between it and other instructions denoted
2825 with the same loop identifier.
2827 Precisely, given two instructions ``m1`` and ``m2`` that both have the
2828 ``llvm.mem.parallel_loop_access`` metadata, with ``L1`` and ``L2`` being the
2829 set of loops associated with that metadata, respectively, then there is no loop
2830 carried dependence between ``m1`` and ``m2`` for loops ``L1`` or
2833 As a special case, if all memory accessing instructions in a loop have
2834 ``llvm.mem.parallel_loop_access`` metadata that refers to that loop, then the
2835 loop has no loop carried memory dependences and is considered to be a parallel
2838 Note that if not all memory access instructions have such metadata referring to
2839 the loop, then the loop is considered not being trivially parallel. Additional
2840 memory dependence analysis is required to make that determination. As a fail
2841 safe mechanism, this causes loops that were originally parallel to be considered
2842 sequential (if optimization passes that are unaware of the parallel semantics
2843 insert new memory instructions into the loop body).
2845 Example of a loop that is considered parallel due to its correct use of
2846 both ``llvm.loop`` and ``llvm.mem.parallel_loop_access``
2847 metadata types that refer to the same loop identifier metadata.
2849 .. code-block:: llvm
2853 %val0 = load i32* %arrayidx, !llvm.mem.parallel_loop_access !0
2855 store i32 %val0, i32* %arrayidx1, !llvm.mem.parallel_loop_access !0
2857 br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !0
2861 !0 = metadata !{ metadata !0 }
2863 It is also possible to have nested parallel loops. In that case the
2864 memory accesses refer to a list of loop identifier metadata nodes instead of
2865 the loop identifier metadata node directly:
2867 .. code-block:: llvm
2871 %val1 = load i32* %arrayidx3, !llvm.mem.parallel_loop_access !2
2873 br label %inner.for.body
2877 %val0 = load i32* %arrayidx1, !llvm.mem.parallel_loop_access !0
2879 store i32 %val0, i32* %arrayidx2, !llvm.mem.parallel_loop_access !0
2881 br i1 %exitcond, label %inner.for.end, label %inner.for.body, !llvm.loop !1
2885 store i32 %val1, i32* %arrayidx4, !llvm.mem.parallel_loop_access !2
2887 br i1 %exitcond, label %outer.for.end, label %outer.for.body, !llvm.loop !2
2889 outer.for.end: ; preds = %for.body
2891 !0 = metadata !{ metadata !1, metadata !2 } ; a list of loop identifiers
2892 !1 = metadata !{ metadata !1 } ; an identifier for the inner loop
2893 !2 = metadata !{ metadata !2 } ; an identifier for the outer loop
2895 '``llvm.vectorizer``'
2896 ^^^^^^^^^^^^^^^^^^^^^
2898 Metadata prefixed with ``llvm.vectorizer`` is used to control per-loop
2899 vectorization parameters such as vectorization factor and unroll factor.
2901 ``llvm.vectorizer`` metadata should be used in conjunction with ``llvm.loop``
2902 loop identification metadata.
2904 '``llvm.vectorizer.unroll``' Metadata
2905 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2907 This metadata instructs the loop vectorizer to unroll the specified
2908 loop exactly ``N`` times.
2910 The first operand is the string ``llvm.vectorizer.unroll`` and the second
2911 operand is an integer specifying the unroll factor. For example:
2913 .. code-block:: llvm
2915 !0 = metadata !{ metadata !"llvm.vectorizer.unroll", i32 4 }
2917 Note that setting ``llvm.vectorizer.unroll`` to 1 disables unrolling of the
2920 If ``llvm.vectorizer.unroll`` is set to 0 then the amount of unrolling will be
2921 determined automatically.
2923 '``llvm.vectorizer.width``' Metadata
2924 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2926 This metadata sets the target width of the vectorizer to ``N``. Without
2927 this metadata, the vectorizer will choose a width automatically.
2928 Regardless of this metadata, the vectorizer will only vectorize loops if
2929 it believes it is valid to do so.
2931 The first operand is the string ``llvm.vectorizer.width`` and the second
2932 operand is an integer specifying the width. For example:
2934 .. code-block:: llvm
2936 !0 = metadata !{ metadata !"llvm.vectorizer.width", i32 4 }
2938 Note that setting ``llvm.vectorizer.width`` to 1 disables vectorization of the
2941 If ``llvm.vectorizer.width`` is set to 0 then the width will be determined
2944 Module Flags Metadata
2945 =====================
2947 Information about the module as a whole is difficult to convey to LLVM's
2948 subsystems. The LLVM IR isn't sufficient to transmit this information.
2949 The ``llvm.module.flags`` named metadata exists in order to facilitate
2950 this. These flags are in the form of key / value pairs --- much like a
2951 dictionary --- making it easy for any subsystem who cares about a flag to
2954 The ``llvm.module.flags`` metadata contains a list of metadata triplets.
2955 Each triplet has the following form:
2957 - The first element is a *behavior* flag, which specifies the behavior
2958 when two (or more) modules are merged together, and it encounters two
2959 (or more) metadata with the same ID. The supported behaviors are
2961 - The second element is a metadata string that is a unique ID for the
2962 metadata. Each module may only have one flag entry for each unique ID (not
2963 including entries with the **Require** behavior).
2964 - The third element is the value of the flag.
2966 When two (or more) modules are merged together, the resulting
2967 ``llvm.module.flags`` metadata is the union of the modules' flags. That is, for
2968 each unique metadata ID string, there will be exactly one entry in the merged
2969 modules ``llvm.module.flags`` metadata table, and the value for that entry will
2970 be determined by the merge behavior flag, as described below. The only exception
2971 is that entries with the *Require* behavior are always preserved.
2973 The following behaviors are supported:
2984 Emits an error if two values disagree, otherwise the resulting value
2985 is that of the operands.
2989 Emits a warning if two values disagree. The result value will be the
2990 operand for the flag from the first module being linked.
2994 Adds a requirement that another module flag be present and have a
2995 specified value after linking is performed. The value must be a
2996 metadata pair, where the first element of the pair is the ID of the
2997 module flag to be restricted, and the second element of the pair is
2998 the value the module flag should be restricted to. This behavior can
2999 be used to restrict the allowable results (via triggering of an
3000 error) of linking IDs with the **Override** behavior.
3004 Uses the specified value, regardless of the behavior or value of the
3005 other module. If both modules specify **Override**, but the values
3006 differ, an error will be emitted.
3010 Appends the two values, which are required to be metadata nodes.
3014 Appends the two values, which are required to be metadata
3015 nodes. However, duplicate entries in the second list are dropped
3016 during the append operation.
3018 It is an error for a particular unique flag ID to have multiple behaviors,
3019 except in the case of **Require** (which adds restrictions on another metadata
3020 value) or **Override**.
3022 An example of module flags:
3024 .. code-block:: llvm
3026 !0 = metadata !{ i32 1, metadata !"foo", i32 1 }
3027 !1 = metadata !{ i32 4, metadata !"bar", i32 37 }
3028 !2 = metadata !{ i32 2, metadata !"qux", i32 42 }
3029 !3 = metadata !{ i32 3, metadata !"qux",
3031 metadata !"foo", i32 1
3034 !llvm.module.flags = !{ !0, !1, !2, !3 }
3036 - Metadata ``!0`` has the ID ``!"foo"`` and the value '1'. The behavior
3037 if two or more ``!"foo"`` flags are seen is to emit an error if their
3038 values are not equal.
3040 - Metadata ``!1`` has the ID ``!"bar"`` and the value '37'. The
3041 behavior if two or more ``!"bar"`` flags are seen is to use the value
3044 - Metadata ``!2`` has the ID ``!"qux"`` and the value '42'. The
3045 behavior if two or more ``!"qux"`` flags are seen is to emit a
3046 warning if their values are not equal.
3048 - Metadata ``!3`` has the ID ``!"qux"`` and the value:
3052 metadata !{ metadata !"foo", i32 1 }
3054 The behavior is to emit an error if the ``llvm.module.flags`` does not
3055 contain a flag with the ID ``!"foo"`` that has the value '1' after linking is
3058 Objective-C Garbage Collection Module Flags Metadata
3059 ----------------------------------------------------
3061 On the Mach-O platform, Objective-C stores metadata about garbage
3062 collection in a special section called "image info". The metadata
3063 consists of a version number and a bitmask specifying what types of
3064 garbage collection are supported (if any) by the file. If two or more
3065 modules are linked together their garbage collection metadata needs to
3066 be merged rather than appended together.
3068 The Objective-C garbage collection module flags metadata consists of the
3069 following key-value pairs:
3078 * - ``Objective-C Version``
3079 - **[Required]** --- The Objective-C ABI version. Valid values are 1 and 2.
3081 * - ``Objective-C Image Info Version``
3082 - **[Required]** --- The version of the image info section. Currently
3085 * - ``Objective-C Image Info Section``
3086 - **[Required]** --- The section to place the metadata. Valid values are
3087 ``"__OBJC, __image_info, regular"`` for Objective-C ABI version 1, and
3088 ``"__DATA,__objc_imageinfo, regular, no_dead_strip"`` for
3089 Objective-C ABI version 2.
3091 * - ``Objective-C Garbage Collection``
3092 - **[Required]** --- Specifies whether garbage collection is supported or
3093 not. Valid values are 0, for no garbage collection, and 2, for garbage
3094 collection supported.
3096 * - ``Objective-C GC Only``
3097 - **[Optional]** --- Specifies that only garbage collection is supported.
3098 If present, its value must be 6. This flag requires that the
3099 ``Objective-C Garbage Collection`` flag have the value 2.
3101 Some important flag interactions:
3103 - If a module with ``Objective-C Garbage Collection`` set to 0 is
3104 merged with a module with ``Objective-C Garbage Collection`` set to
3105 2, then the resulting module has the
3106 ``Objective-C Garbage Collection`` flag set to 0.
3107 - A module with ``Objective-C Garbage Collection`` set to 0 cannot be
3108 merged with a module with ``Objective-C GC Only`` set to 6.
3110 Automatic Linker Flags Module Flags Metadata
3111 --------------------------------------------
3113 Some targets support embedding flags to the linker inside individual object
3114 files. Typically this is used in conjunction with language extensions which
3115 allow source files to explicitly declare the libraries they depend on, and have
3116 these automatically be transmitted to the linker via object files.
3118 These flags are encoded in the IR using metadata in the module flags section,
3119 using the ``Linker Options`` key. The merge behavior for this flag is required
3120 to be ``AppendUnique``, and the value for the key is expected to be a metadata
3121 node which should be a list of other metadata nodes, each of which should be a
3122 list of metadata strings defining linker options.
3124 For example, the following metadata section specifies two separate sets of
3125 linker options, presumably to link against ``libz`` and the ``Cocoa``
3128 !0 = metadata !{ i32 6, metadata !"Linker Options",
3130 metadata !{ metadata !"-lz" },
3131 metadata !{ metadata !"-framework", metadata !"Cocoa" } } }
3132 !llvm.module.flags = !{ !0 }
3134 The metadata encoding as lists of lists of options, as opposed to a collapsed
3135 list of options, is chosen so that the IR encoding can use multiple option
3136 strings to specify e.g., a single library, while still having that specifier be
3137 preserved as an atomic element that can be recognized by a target specific
3138 assembly writer or object file emitter.
3140 Each individual option is required to be either a valid option for the target's
3141 linker, or an option that is reserved by the target specific assembly writer or
3142 object file emitter. No other aspect of these options is defined by the IR.
3144 .. _intrinsicglobalvariables:
3146 Intrinsic Global Variables
3147 ==========================
3149 LLVM has a number of "magic" global variables that contain data that
3150 affect code generation or other IR semantics. These are documented here.
3151 All globals of this sort should have a section specified as
3152 "``llvm.metadata``". This section and all globals that start with
3153 "``llvm.``" are reserved for use by LLVM.
3157 The '``llvm.used``' Global Variable
3158 -----------------------------------
3160 The ``@llvm.used`` global is an array which has
3161 :ref:`appending linkage <linkage_appending>`. This array contains a list of
3162 pointers to named global variables, functions and aliases which may optionally
3163 have a pointer cast formed of bitcast or getelementptr. For example, a legal
3166 .. code-block:: llvm
3171 @llvm.used = appending global [2 x i8*] [
3173 i8* bitcast (i32* @Y to i8*)
3174 ], section "llvm.metadata"
3176 If a symbol appears in the ``@llvm.used`` list, then the compiler, assembler,
3177 and linker are required to treat the symbol as if there is a reference to the
3178 symbol that it cannot see (which is why they have to be named). For example, if
3179 a variable has internal linkage and no references other than that from the
3180 ``@llvm.used`` list, it cannot be deleted. This is commonly used to represent
3181 references from inline asms and other things the compiler cannot "see", and
3182 corresponds to "``attribute((used))``" in GNU C.
3184 On some targets, the code generator must emit a directive to the
3185 assembler or object file to prevent the assembler and linker from
3186 molesting the symbol.
3188 .. _gv_llvmcompilerused:
3190 The '``llvm.compiler.used``' Global Variable
3191 --------------------------------------------
3193 The ``@llvm.compiler.used`` directive is the same as the ``@llvm.used``
3194 directive, except that it only prevents the compiler from touching the
3195 symbol. On targets that support it, this allows an intelligent linker to
3196 optimize references to the symbol without being impeded as it would be
3199 This is a rare construct that should only be used in rare circumstances,
3200 and should not be exposed to source languages.
3202 .. _gv_llvmglobalctors:
3204 The '``llvm.global_ctors``' Global Variable
3205 -------------------------------------------
3207 .. code-block:: llvm
3209 %0 = type { i32, void ()*, i8* }
3210 @llvm.global_ctors = appending global [1 x %0] [%0 { i32 65535, void ()* @ctor, i8* @data }]
3212 The ``@llvm.global_ctors`` array contains a list of constructor
3213 functions, priorities, and an optional associated global or function.
3214 The functions referenced by this array will be called in ascending order
3215 of priority (i.e. lowest first) when the module is loaded. The order of
3216 functions with the same priority is not defined.
3218 If the third field is present, non-null, and points to a global variable
3219 or function, the initializer function will only run if the associated
3220 data from the current module is not discarded.
3222 .. _llvmglobaldtors:
3224 The '``llvm.global_dtors``' Global Variable
3225 -------------------------------------------
3227 .. code-block:: llvm
3229 %0 = type { i32, void ()*, i8* }
3230 @llvm.global_dtors = appending global [1 x %0] [%0 { i32 65535, void ()* @dtor, i8* @data }]
3232 The ``@llvm.global_dtors`` array contains a list of destructor
3233 functions, priorities, and an optional associated global or function.
3234 The functions referenced by this array will be called in descending
3235 order of priority (i.e. highest first) when the module is unloaded. The
3236 order of functions with the same priority is not defined.
3238 If the third field is present, non-null, and points to a global variable
3239 or function, the destructor function will only run if the associated
3240 data from the current module is not discarded.
3242 Instruction Reference
3243 =====================
3245 The LLVM instruction set consists of several different classifications
3246 of instructions: :ref:`terminator instructions <terminators>`, :ref:`binary
3247 instructions <binaryops>`, :ref:`bitwise binary
3248 instructions <bitwiseops>`, :ref:`memory instructions <memoryops>`, and
3249 :ref:`other instructions <otherops>`.
3253 Terminator Instructions
3254 -----------------------
3256 As mentioned :ref:`previously <functionstructure>`, every basic block in a
3257 program ends with a "Terminator" instruction, which indicates which
3258 block should be executed after the current block is finished. These
3259 terminator instructions typically yield a '``void``' value: they produce
3260 control flow, not values (the one exception being the
3261 ':ref:`invoke <i_invoke>`' instruction).
3263 The terminator instructions are: ':ref:`ret <i_ret>`',
3264 ':ref:`br <i_br>`', ':ref:`switch <i_switch>`',
3265 ':ref:`indirectbr <i_indirectbr>`', ':ref:`invoke <i_invoke>`',
3266 ':ref:`resume <i_resume>`', and ':ref:`unreachable <i_unreachable>`'.
3270 '``ret``' Instruction
3271 ^^^^^^^^^^^^^^^^^^^^^
3278 ret <type> <value> ; Return a value from a non-void function
3279 ret void ; Return from void function
3284 The '``ret``' instruction is used to return control flow (and optionally
3285 a value) from a function back to the caller.
3287 There are two forms of the '``ret``' instruction: one that returns a
3288 value and then causes control flow, and one that just causes control
3294 The '``ret``' instruction optionally accepts a single argument, the
3295 return value. The type of the return value must be a ':ref:`first
3296 class <t_firstclass>`' type.
3298 A function is not :ref:`well formed <wellformed>` if it it has a non-void
3299 return type and contains a '``ret``' instruction with no return value or
3300 a return value with a type that does not match its type, or if it has a
3301 void return type and contains a '``ret``' instruction with a return
3307 When the '``ret``' instruction is executed, control flow returns back to
3308 the calling function's context. If the caller is a
3309 ":ref:`call <i_call>`" instruction, execution continues at the
3310 instruction after the call. If the caller was an
3311 ":ref:`invoke <i_invoke>`" instruction, execution continues at the
3312 beginning of the "normal" destination block. If the instruction returns
3313 a value, that value shall set the call or invoke instruction's return
3319 .. code-block:: llvm
3321 ret i32 5 ; Return an integer value of 5
3322 ret void ; Return from a void function
3323 ret { i32, i8 } { i32 4, i8 2 } ; Return a struct of values 4 and 2
3327 '``br``' Instruction
3328 ^^^^^^^^^^^^^^^^^^^^
3335 br i1 <cond>, label <iftrue>, label <iffalse>
3336 br label <dest> ; Unconditional branch
3341 The '``br``' instruction is used to cause control flow to transfer to a
3342 different basic block in the current function. There are two forms of
3343 this instruction, corresponding to a conditional branch and an
3344 unconditional branch.
3349 The conditional branch form of the '``br``' instruction takes a single
3350 '``i1``' value and two '``label``' values. The unconditional form of the
3351 '``br``' instruction takes a single '``label``' value as a target.
3356 Upon execution of a conditional '``br``' instruction, the '``i1``'
3357 argument is evaluated. If the value is ``true``, control flows to the
3358 '``iftrue``' ``label`` argument. If "cond" is ``false``, control flows
3359 to the '``iffalse``' ``label`` argument.
3364 .. code-block:: llvm
3367 %cond = icmp eq i32 %a, %b
3368 br i1 %cond, label %IfEqual, label %IfUnequal
3376 '``switch``' Instruction
3377 ^^^^^^^^^^^^^^^^^^^^^^^^
3384 switch <intty> <value>, label <defaultdest> [ <intty> <val>, label <dest> ... ]
3389 The '``switch``' instruction is used to transfer control flow to one of
3390 several different places. It is a generalization of the '``br``'
3391 instruction, allowing a branch to occur to one of many possible
3397 The '``switch``' instruction uses three parameters: an integer
3398 comparison value '``value``', a default '``label``' destination, and an
3399 array of pairs of comparison value constants and '``label``'s. The table
3400 is not allowed to contain duplicate constant entries.
3405 The ``switch`` instruction specifies a table of values and destinations.
3406 When the '``switch``' instruction is executed, this table is searched
3407 for the given value. If the value is found, control flow is transferred
3408 to the corresponding destination; otherwise, control flow is transferred
3409 to the default destination.
3414 Depending on properties of the target machine and the particular
3415 ``switch`` instruction, this instruction may be code generated in
3416 different ways. For example, it could be generated as a series of
3417 chained conditional branches or with a lookup table.
3422 .. code-block:: llvm
3424 ; Emulate a conditional br instruction
3425 %Val = zext i1 %value to i32
3426 switch i32 %Val, label %truedest [ i32 0, label %falsedest ]
3428 ; Emulate an unconditional br instruction
3429 switch i32 0, label %dest [ ]
3431 ; Implement a jump table:
3432 switch i32 %val, label %otherwise [ i32 0, label %onzero
3434 i32 2, label %ontwo ]
3438 '``indirectbr``' Instruction
3439 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3446 indirectbr <somety>* <address>, [ label <dest1>, label <dest2>, ... ]
3451 The '``indirectbr``' instruction implements an indirect branch to a
3452 label within the current function, whose address is specified by
3453 "``address``". Address must be derived from a
3454 :ref:`blockaddress <blockaddress>` constant.
3459 The '``address``' argument is the address of the label to jump to. The
3460 rest of the arguments indicate the full set of possible destinations
3461 that the address may point to. Blocks are allowed to occur multiple
3462 times in the destination list, though this isn't particularly useful.
3464 This destination list is required so that dataflow analysis has an
3465 accurate understanding of the CFG.
3470 Control transfers to the block specified in the address argument. All
3471 possible destination blocks must be listed in the label list, otherwise
3472 this instruction has undefined behavior. This implies that jumps to
3473 labels defined in other functions have undefined behavior as well.
3478 This is typically implemented with a jump through a register.
3483 .. code-block:: llvm
3485 indirectbr i8* %Addr, [ label %bb1, label %bb2, label %bb3 ]
3489 '``invoke``' Instruction
3490 ^^^^^^^^^^^^^^^^^^^^^^^^
3497 <result> = invoke [cconv] [ret attrs] <ptr to function ty> <function ptr val>(<function args>) [fn attrs]
3498 to label <normal label> unwind label <exception label>
3503 The '``invoke``' instruction causes control to transfer to a specified
3504 function, with the possibility of control flow transfer to either the
3505 '``normal``' label or the '``exception``' label. If the callee function
3506 returns with the "``ret``" instruction, control flow will return to the
3507 "normal" label. If the callee (or any indirect callees) returns via the
3508 ":ref:`resume <i_resume>`" instruction or other exception handling
3509 mechanism, control is interrupted and continued at the dynamically
3510 nearest "exception" label.
3512 The '``exception``' label is a `landing
3513 pad <ExceptionHandling.html#overview>`_ for the exception. As such,
3514 '``exception``' label is required to have the
3515 ":ref:`landingpad <i_landingpad>`" instruction, which contains the
3516 information about the behavior of the program after unwinding happens,
3517 as its first non-PHI instruction. The restrictions on the
3518 "``landingpad``" instruction's tightly couples it to the "``invoke``"
3519 instruction, so that the important information contained within the
3520 "``landingpad``" instruction can't be lost through normal code motion.
3525 This instruction requires several arguments:
3527 #. The optional "cconv" marker indicates which :ref:`calling
3528 convention <callingconv>` the call should use. If none is
3529 specified, the call defaults to using C calling conventions.
3530 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
3531 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
3533 #. '``ptr to function ty``': shall be the signature of the pointer to
3534 function value being invoked. In most cases, this is a direct
3535 function invocation, but indirect ``invoke``'s are just as possible,
3536 branching off an arbitrary pointer to function value.
3537 #. '``function ptr val``': An LLVM value containing a pointer to a
3538 function to be invoked.
3539 #. '``function args``': argument list whose types match the function
3540 signature argument types and parameter attributes. All arguments must
3541 be of :ref:`first class <t_firstclass>` type. If the function signature
3542 indicates the function accepts a variable number of arguments, the
3543 extra arguments can be specified.
3544 #. '``normal label``': the label reached when the called function
3545 executes a '``ret``' instruction.
3546 #. '``exception label``': the label reached when a callee returns via
3547 the :ref:`resume <i_resume>` instruction or other exception handling
3549 #. The optional :ref:`function attributes <fnattrs>` list. Only
3550 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
3551 attributes are valid here.
3556 This instruction is designed to operate as a standard '``call``'
3557 instruction in most regards. The primary difference is that it
3558 establishes an association with a label, which is used by the runtime
3559 library to unwind the stack.
3561 This instruction is used in languages with destructors to ensure that
3562 proper cleanup is performed in the case of either a ``longjmp`` or a
3563 thrown exception. Additionally, this is important for implementation of
3564 '``catch``' clauses in high-level languages that support them.
3566 For the purposes of the SSA form, the definition of the value returned
3567 by the '``invoke``' instruction is deemed to occur on the edge from the
3568 current block to the "normal" label. If the callee unwinds then no
3569 return value is available.
3574 .. code-block:: llvm
3576 %retval = invoke i32 @Test(i32 15) to label %Continue
3577 unwind label %TestCleanup ; {i32}:retval set
3578 %retval = invoke coldcc i32 %Testfnptr(i32 15) to label %Continue
3579 unwind label %TestCleanup ; {i32}:retval set
3583 '``resume``' Instruction
3584 ^^^^^^^^^^^^^^^^^^^^^^^^
3591 resume <type> <value>
3596 The '``resume``' instruction is a terminator instruction that has no
3602 The '``resume``' instruction requires one argument, which must have the
3603 same type as the result of any '``landingpad``' instruction in the same
3609 The '``resume``' instruction resumes propagation of an existing
3610 (in-flight) exception whose unwinding was interrupted with a
3611 :ref:`landingpad <i_landingpad>` instruction.
3616 .. code-block:: llvm
3618 resume { i8*, i32 } %exn
3622 '``unreachable``' Instruction
3623 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3635 The '``unreachable``' instruction has no defined semantics. This
3636 instruction is used to inform the optimizer that a particular portion of
3637 the code is not reachable. This can be used to indicate that the code
3638 after a no-return function cannot be reached, and other facts.
3643 The '``unreachable``' instruction has no defined semantics.
3650 Binary operators are used to do most of the computation in a program.
3651 They require two operands of the same type, execute an operation on
3652 them, and produce a single value. The operands might represent multiple
3653 data, as is the case with the :ref:`vector <t_vector>` data type. The
3654 result value has the same type as its operands.
3656 There are several different binary operators:
3660 '``add``' Instruction
3661 ^^^^^^^^^^^^^^^^^^^^^
3668 <result> = add <ty> <op1>, <op2> ; yields {ty}:result
3669 <result> = add nuw <ty> <op1>, <op2> ; yields {ty}:result
3670 <result> = add nsw <ty> <op1>, <op2> ; yields {ty}:result
3671 <result> = add nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3676 The '``add``' instruction returns the sum of its two operands.
3681 The two arguments to the '``add``' instruction must be
3682 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3683 arguments must have identical types.
3688 The value produced is the integer sum of the two operands.
3690 If the sum has unsigned overflow, the result returned is the
3691 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
3694 Because LLVM integers use a two's complement representation, this
3695 instruction is appropriate for both signed and unsigned integers.
3697 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3698 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3699 result value of the ``add`` is a :ref:`poison value <poisonvalues>` if
3700 unsigned and/or signed overflow, respectively, occurs.
3705 .. code-block:: llvm
3707 <result> = add i32 4, %var ; yields {i32}:result = 4 + %var
3711 '``fadd``' Instruction
3712 ^^^^^^^^^^^^^^^^^^^^^^
3719 <result> = fadd [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3724 The '``fadd``' instruction returns the sum of its two operands.
3729 The two arguments to the '``fadd``' instruction must be :ref:`floating
3730 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3731 Both arguments must have identical types.
3736 The value produced is the floating point sum of the two operands. This
3737 instruction can also take any number of :ref:`fast-math flags <fastmath>`,
3738 which are optimization hints to enable otherwise unsafe floating point
3744 .. code-block:: llvm
3746 <result> = fadd float 4.0, %var ; yields {float}:result = 4.0 + %var
3748 '``sub``' Instruction
3749 ^^^^^^^^^^^^^^^^^^^^^
3756 <result> = sub <ty> <op1>, <op2> ; yields {ty}:result
3757 <result> = sub nuw <ty> <op1>, <op2> ; yields {ty}:result
3758 <result> = sub nsw <ty> <op1>, <op2> ; yields {ty}:result
3759 <result> = sub nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3764 The '``sub``' instruction returns the difference of its two operands.
3766 Note that the '``sub``' instruction is used to represent the '``neg``'
3767 instruction present in most other intermediate representations.
3772 The two arguments to the '``sub``' instruction must be
3773 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3774 arguments must have identical types.
3779 The value produced is the integer difference of the two operands.
3781 If the difference has unsigned overflow, the result returned is the
3782 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
3785 Because LLVM integers use a two's complement representation, this
3786 instruction is appropriate for both signed and unsigned integers.
3788 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3789 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3790 result value of the ``sub`` is a :ref:`poison value <poisonvalues>` if
3791 unsigned and/or signed overflow, respectively, occurs.
3796 .. code-block:: llvm
3798 <result> = sub i32 4, %var ; yields {i32}:result = 4 - %var
3799 <result> = sub i32 0, %val ; yields {i32}:result = -%var
3803 '``fsub``' Instruction
3804 ^^^^^^^^^^^^^^^^^^^^^^
3811 <result> = fsub [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3816 The '``fsub``' instruction returns the difference of its two operands.
3818 Note that the '``fsub``' instruction is used to represent the '``fneg``'
3819 instruction present in most other intermediate representations.
3824 The two arguments to the '``fsub``' instruction must be :ref:`floating
3825 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3826 Both arguments must have identical types.
3831 The value produced is the floating point difference of the two operands.
3832 This instruction can also take any number of :ref:`fast-math
3833 flags <fastmath>`, which are optimization hints to enable otherwise
3834 unsafe floating point optimizations:
3839 .. code-block:: llvm
3841 <result> = fsub float 4.0, %var ; yields {float}:result = 4.0 - %var
3842 <result> = fsub float -0.0, %val ; yields {float}:result = -%var
3844 '``mul``' Instruction
3845 ^^^^^^^^^^^^^^^^^^^^^
3852 <result> = mul <ty> <op1>, <op2> ; yields {ty}:result
3853 <result> = mul nuw <ty> <op1>, <op2> ; yields {ty}:result
3854 <result> = mul nsw <ty> <op1>, <op2> ; yields {ty}:result
3855 <result> = mul nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3860 The '``mul``' instruction returns the product of its two operands.
3865 The two arguments to the '``mul``' instruction must be
3866 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3867 arguments must have identical types.
3872 The value produced is the integer product of the two operands.
3874 If the result of the multiplication has unsigned overflow, the result
3875 returned is the mathematical result modulo 2\ :sup:`n`\ , where n is the
3876 bit width of the result.
3878 Because LLVM integers use a two's complement representation, and the
3879 result is the same width as the operands, this instruction returns the
3880 correct result for both signed and unsigned integers. If a full product
3881 (e.g. ``i32`` * ``i32`` -> ``i64``) is needed, the operands should be
3882 sign-extended or zero-extended as appropriate to the width of the full
3885 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3886 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3887 result value of the ``mul`` is a :ref:`poison value <poisonvalues>` if
3888 unsigned and/or signed overflow, respectively, occurs.
3893 .. code-block:: llvm
3895 <result> = mul i32 4, %var ; yields {i32}:result = 4 * %var
3899 '``fmul``' Instruction
3900 ^^^^^^^^^^^^^^^^^^^^^^
3907 <result> = fmul [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3912 The '``fmul``' instruction returns the product of its two operands.
3917 The two arguments to the '``fmul``' instruction must be :ref:`floating
3918 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3919 Both arguments must have identical types.
3924 The value produced is the floating point product of the two operands.
3925 This instruction can also take any number of :ref:`fast-math
3926 flags <fastmath>`, which are optimization hints to enable otherwise
3927 unsafe floating point optimizations:
3932 .. code-block:: llvm
3934 <result> = fmul float 4.0, %var ; yields {float}:result = 4.0 * %var
3936 '``udiv``' Instruction
3937 ^^^^^^^^^^^^^^^^^^^^^^
3944 <result> = udiv <ty> <op1>, <op2> ; yields {ty}:result
3945 <result> = udiv exact <ty> <op1>, <op2> ; yields {ty}:result
3950 The '``udiv``' instruction returns the quotient of its two operands.
3955 The two arguments to the '``udiv``' instruction must be
3956 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3957 arguments must have identical types.
3962 The value produced is the unsigned integer quotient of the two operands.
3964 Note that unsigned integer division and signed integer division are
3965 distinct operations; for signed integer division, use '``sdiv``'.
3967 Division by zero leads to undefined behavior.
3969 If the ``exact`` keyword is present, the result value of the ``udiv`` is
3970 a :ref:`poison value <poisonvalues>` if %op1 is not a multiple of %op2 (as
3971 such, "((a udiv exact b) mul b) == a").
3976 .. code-block:: llvm
3978 <result> = udiv i32 4, %var ; yields {i32}:result = 4 / %var
3980 '``sdiv``' Instruction
3981 ^^^^^^^^^^^^^^^^^^^^^^
3988 <result> = sdiv <ty> <op1>, <op2> ; yields {ty}:result
3989 <result> = sdiv exact <ty> <op1>, <op2> ; yields {ty}:result
3994 The '``sdiv``' instruction returns the quotient of its two operands.
3999 The two arguments to the '``sdiv``' instruction must be
4000 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4001 arguments must have identical types.
4006 The value produced is the signed integer quotient of the two operands
4007 rounded towards zero.
4009 Note that signed integer division and unsigned integer division are
4010 distinct operations; for unsigned integer division, use '``udiv``'.
4012 Division by zero leads to undefined behavior. Overflow also leads to
4013 undefined behavior; this is a rare case, but can occur, for example, by
4014 doing a 32-bit division of -2147483648 by -1.
4016 If the ``exact`` keyword is present, the result value of the ``sdiv`` is
4017 a :ref:`poison value <poisonvalues>` if the result would be rounded.
4022 .. code-block:: llvm
4024 <result> = sdiv i32 4, %var ; yields {i32}:result = 4 / %var
4028 '``fdiv``' Instruction
4029 ^^^^^^^^^^^^^^^^^^^^^^
4036 <result> = fdiv [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
4041 The '``fdiv``' instruction returns the quotient of its two operands.
4046 The two arguments to the '``fdiv``' instruction must be :ref:`floating
4047 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
4048 Both arguments must have identical types.
4053 The value produced is the floating point quotient of the two operands.
4054 This instruction can also take any number of :ref:`fast-math
4055 flags <fastmath>`, which are optimization hints to enable otherwise
4056 unsafe floating point optimizations:
4061 .. code-block:: llvm
4063 <result> = fdiv float 4.0, %var ; yields {float}:result = 4.0 / %var
4065 '``urem``' Instruction
4066 ^^^^^^^^^^^^^^^^^^^^^^
4073 <result> = urem <ty> <op1>, <op2> ; yields {ty}:result
4078 The '``urem``' instruction returns the remainder from the unsigned
4079 division of its two arguments.
4084 The two arguments to the '``urem``' instruction must be
4085 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4086 arguments must have identical types.
4091 This instruction returns the unsigned integer *remainder* of a division.
4092 This instruction always performs an unsigned division to get the
4095 Note that unsigned integer remainder and signed integer remainder are
4096 distinct operations; for signed integer remainder, use '``srem``'.
4098 Taking the remainder of a division by zero leads to undefined behavior.
4103 .. code-block:: llvm
4105 <result> = urem i32 4, %var ; yields {i32}:result = 4 % %var
4107 '``srem``' Instruction
4108 ^^^^^^^^^^^^^^^^^^^^^^
4115 <result> = srem <ty> <op1>, <op2> ; yields {ty}:result
4120 The '``srem``' instruction returns the remainder from the signed
4121 division of its two operands. This instruction can also take
4122 :ref:`vector <t_vector>` versions of the values in which case the elements
4128 The two arguments to the '``srem``' instruction must be
4129 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4130 arguments must have identical types.
4135 This instruction returns the *remainder* of a division (where the result
4136 is either zero or has the same sign as the dividend, ``op1``), not the
4137 *modulo* operator (where the result is either zero or has the same sign
4138 as the divisor, ``op2``) of a value. For more information about the
4139 difference, see `The Math
4140 Forum <http://mathforum.org/dr.math/problems/anne.4.28.99.html>`_. For a
4141 table of how this is implemented in various languages, please see
4143 operation <http://en.wikipedia.org/wiki/Modulo_operation>`_.
4145 Note that signed integer remainder and unsigned integer remainder are
4146 distinct operations; for unsigned integer remainder, use '``urem``'.
4148 Taking the remainder of a division by zero leads to undefined behavior.
4149 Overflow also leads to undefined behavior; this is a rare case, but can
4150 occur, for example, by taking the remainder of a 32-bit division of
4151 -2147483648 by -1. (The remainder doesn't actually overflow, but this
4152 rule lets srem be implemented using instructions that return both the
4153 result of the division and the remainder.)
4158 .. code-block:: llvm
4160 <result> = srem i32 4, %var ; yields {i32}:result = 4 % %var
4164 '``frem``' Instruction
4165 ^^^^^^^^^^^^^^^^^^^^^^
4172 <result> = frem [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
4177 The '``frem``' instruction returns the remainder from the division of
4183 The two arguments to the '``frem``' instruction must be :ref:`floating
4184 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
4185 Both arguments must have identical types.
4190 This instruction returns the *remainder* of a division. The remainder
4191 has the same sign as the dividend. This instruction can also take any
4192 number of :ref:`fast-math flags <fastmath>`, which are optimization hints
4193 to enable otherwise unsafe floating point optimizations:
4198 .. code-block:: llvm
4200 <result> = frem float 4.0, %var ; yields {float}:result = 4.0 % %var
4204 Bitwise Binary Operations
4205 -------------------------
4207 Bitwise binary operators are used to do various forms of bit-twiddling
4208 in a program. They are generally very efficient instructions and can
4209 commonly be strength reduced from other instructions. They require two
4210 operands of the same type, execute an operation on them, and produce a
4211 single value. The resulting value is the same type as its operands.
4213 '``shl``' Instruction
4214 ^^^^^^^^^^^^^^^^^^^^^
4221 <result> = shl <ty> <op1>, <op2> ; yields {ty}:result
4222 <result> = shl nuw <ty> <op1>, <op2> ; yields {ty}:result
4223 <result> = shl nsw <ty> <op1>, <op2> ; yields {ty}:result
4224 <result> = shl nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
4229 The '``shl``' instruction returns the first operand shifted to the left
4230 a specified number of bits.
4235 Both arguments to the '``shl``' instruction must be the same
4236 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4237 '``op2``' is treated as an unsigned value.
4242 The value produced is ``op1`` \* 2\ :sup:`op2` mod 2\ :sup:`n`,
4243 where ``n`` is the width of the result. If ``op2`` is (statically or
4244 dynamically) negative or equal to or larger than the number of bits in
4245 ``op1``, the result is undefined. If the arguments are vectors, each
4246 vector element of ``op1`` is shifted by the corresponding shift amount
4249 If the ``nuw`` keyword is present, then the shift produces a :ref:`poison
4250 value <poisonvalues>` if it shifts out any non-zero bits. If the
4251 ``nsw`` keyword is present, then the shift produces a :ref:`poison
4252 value <poisonvalues>` if it shifts out any bits that disagree with the
4253 resultant sign bit. As such, NUW/NSW have the same semantics as they
4254 would if the shift were expressed as a mul instruction with the same
4255 nsw/nuw bits in (mul %op1, (shl 1, %op2)).
4260 .. code-block:: llvm
4262 <result> = shl i32 4, %var ; yields {i32}: 4 << %var
4263 <result> = shl i32 4, 2 ; yields {i32}: 16
4264 <result> = shl i32 1, 10 ; yields {i32}: 1024
4265 <result> = shl i32 1, 32 ; undefined
4266 <result> = shl <2 x i32> < i32 1, i32 1>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 2, i32 4>
4268 '``lshr``' Instruction
4269 ^^^^^^^^^^^^^^^^^^^^^^
4276 <result> = lshr <ty> <op1>, <op2> ; yields {ty}:result
4277 <result> = lshr exact <ty> <op1>, <op2> ; yields {ty}:result
4282 The '``lshr``' instruction (logical shift right) returns the first
4283 operand shifted to the right a specified number of bits with zero fill.
4288 Both arguments to the '``lshr``' instruction must be the same
4289 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4290 '``op2``' is treated as an unsigned value.
4295 This instruction always performs a logical shift right operation. The
4296 most significant bits of the result will be filled with zero bits after
4297 the shift. If ``op2`` is (statically or dynamically) equal to or larger
4298 than the number of bits in ``op1``, the result is undefined. If the
4299 arguments are vectors, each vector element of ``op1`` is shifted by the
4300 corresponding shift amount in ``op2``.
4302 If the ``exact`` keyword is present, the result value of the ``lshr`` is
4303 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
4309 .. code-block:: llvm
4311 <result> = lshr i32 4, 1 ; yields {i32}:result = 2
4312 <result> = lshr i32 4, 2 ; yields {i32}:result = 1
4313 <result> = lshr i8 4, 3 ; yields {i8}:result = 0
4314 <result> = lshr i8 -2, 1 ; yields {i8}:result = 0x7F
4315 <result> = lshr i32 1, 32 ; undefined
4316 <result> = lshr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 0x7FFFFFFF, i32 1>
4318 '``ashr``' Instruction
4319 ^^^^^^^^^^^^^^^^^^^^^^
4326 <result> = ashr <ty> <op1>, <op2> ; yields {ty}:result
4327 <result> = ashr exact <ty> <op1>, <op2> ; yields {ty}:result
4332 The '``ashr``' instruction (arithmetic shift right) returns the first
4333 operand shifted to the right a specified number of bits with sign
4339 Both arguments to the '``ashr``' instruction must be the same
4340 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4341 '``op2``' is treated as an unsigned value.
4346 This instruction always performs an arithmetic shift right operation,
4347 The most significant bits of the result will be filled with the sign bit
4348 of ``op1``. If ``op2`` is (statically or dynamically) equal to or larger
4349 than the number of bits in ``op1``, the result is undefined. If the
4350 arguments are vectors, each vector element of ``op1`` is shifted by the
4351 corresponding shift amount in ``op2``.
4353 If the ``exact`` keyword is present, the result value of the ``ashr`` is
4354 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
4360 .. code-block:: llvm
4362 <result> = ashr i32 4, 1 ; yields {i32}:result = 2
4363 <result> = ashr i32 4, 2 ; yields {i32}:result = 1
4364 <result> = ashr i8 4, 3 ; yields {i8}:result = 0
4365 <result> = ashr i8 -2, 1 ; yields {i8}:result = -1
4366 <result> = ashr i32 1, 32 ; undefined
4367 <result> = ashr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 3> ; yields: result=<2 x i32> < i32 -1, i32 0>
4369 '``and``' Instruction
4370 ^^^^^^^^^^^^^^^^^^^^^
4377 <result> = and <ty> <op1>, <op2> ; yields {ty}:result
4382 The '``and``' instruction returns the bitwise logical and of its two
4388 The two arguments to the '``and``' instruction must be
4389 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4390 arguments must have identical types.
4395 The truth table used for the '``and``' instruction is:
4412 .. code-block:: llvm
4414 <result> = and i32 4, %var ; yields {i32}:result = 4 & %var
4415 <result> = and i32 15, 40 ; yields {i32}:result = 8
4416 <result> = and i32 4, 8 ; yields {i32}:result = 0
4418 '``or``' Instruction
4419 ^^^^^^^^^^^^^^^^^^^^
4426 <result> = or <ty> <op1>, <op2> ; yields {ty}:result
4431 The '``or``' instruction returns the bitwise logical inclusive or of its
4437 The two arguments to the '``or``' instruction must be
4438 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4439 arguments must have identical types.
4444 The truth table used for the '``or``' instruction is:
4463 <result> = or i32 4, %var ; yields {i32}:result = 4 | %var
4464 <result> = or i32 15, 40 ; yields {i32}:result = 47
4465 <result> = or i32 4, 8 ; yields {i32}:result = 12
4467 '``xor``' Instruction
4468 ^^^^^^^^^^^^^^^^^^^^^
4475 <result> = xor <ty> <op1>, <op2> ; yields {ty}:result
4480 The '``xor``' instruction returns the bitwise logical exclusive or of
4481 its two operands. The ``xor`` is used to implement the "one's
4482 complement" operation, which is the "~" operator in C.
4487 The two arguments to the '``xor``' instruction must be
4488 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4489 arguments must have identical types.
4494 The truth table used for the '``xor``' instruction is:
4511 .. code-block:: llvm
4513 <result> = xor i32 4, %var ; yields {i32}:result = 4 ^ %var
4514 <result> = xor i32 15, 40 ; yields {i32}:result = 39
4515 <result> = xor i32 4, 8 ; yields {i32}:result = 12
4516 <result> = xor i32 %V, -1 ; yields {i32}:result = ~%V
4521 LLVM supports several instructions to represent vector operations in a
4522 target-independent manner. These instructions cover the element-access
4523 and vector-specific operations needed to process vectors effectively.
4524 While LLVM does directly support these vector operations, many
4525 sophisticated algorithms will want to use target-specific intrinsics to
4526 take full advantage of a specific target.
4528 .. _i_extractelement:
4530 '``extractelement``' Instruction
4531 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4538 <result> = extractelement <n x <ty>> <val>, <ty2> <idx> ; yields <ty>
4543 The '``extractelement``' instruction extracts a single scalar element
4544 from a vector at a specified index.
4549 The first operand of an '``extractelement``' instruction is a value of
4550 :ref:`vector <t_vector>` type. The second operand is an index indicating
4551 the position from which to extract the element. The index may be a
4552 variable of any integer type.
4557 The result is a scalar of the same type as the element type of ``val``.
4558 Its value is the value at position ``idx`` of ``val``. If ``idx``
4559 exceeds the length of ``val``, the results are undefined.
4564 .. code-block:: llvm
4566 <result> = extractelement <4 x i32> %vec, i32 0 ; yields i32
4568 .. _i_insertelement:
4570 '``insertelement``' Instruction
4571 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4578 <result> = insertelement <n x <ty>> <val>, <ty> <elt>, <ty2> <idx> ; yields <n x <ty>>
4583 The '``insertelement``' instruction inserts a scalar element into a
4584 vector at a specified index.
4589 The first operand of an '``insertelement``' instruction is a value of
4590 :ref:`vector <t_vector>` type. The second operand is a scalar value whose
4591 type must equal the element type of the first operand. The third operand
4592 is an index indicating the position at which to insert the value. The
4593 index may be a variable of any integer type.
4598 The result is a vector of the same type as ``val``. Its element values
4599 are those of ``val`` except at position ``idx``, where it gets the value
4600 ``elt``. If ``idx`` exceeds the length of ``val``, the results are
4606 .. code-block:: llvm
4608 <result> = insertelement <4 x i32> %vec, i32 1, i32 0 ; yields <4 x i32>
4610 .. _i_shufflevector:
4612 '``shufflevector``' Instruction
4613 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4620 <result> = shufflevector <n x <ty>> <v1>, <n x <ty>> <v2>, <m x i32> <mask> ; yields <m x <ty>>
4625 The '``shufflevector``' instruction constructs a permutation of elements
4626 from two input vectors, returning a vector with the same element type as
4627 the input and length that is the same as the shuffle mask.
4632 The first two operands of a '``shufflevector``' instruction are vectors
4633 with the same type. The third argument is a shuffle mask whose element
4634 type is always 'i32'. The result of the instruction is a vector whose
4635 length is the same as the shuffle mask and whose element type is the
4636 same as the element type of the first two operands.
4638 The shuffle mask operand is required to be a constant vector with either
4639 constant integer or undef values.
4644 The elements of the two input vectors are numbered from left to right
4645 across both of the vectors. The shuffle mask operand specifies, for each
4646 element of the result vector, which element of the two input vectors the
4647 result element gets. The element selector may be undef (meaning "don't
4648 care") and the second operand may be undef if performing a shuffle from
4654 .. code-block:: llvm
4656 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
4657 <4 x i32> <i32 0, i32 4, i32 1, i32 5> ; yields <4 x i32>
4658 <result> = shufflevector <4 x i32> %v1, <4 x i32> undef,
4659 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32> - Identity shuffle.
4660 <result> = shufflevector <8 x i32> %v1, <8 x i32> undef,
4661 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32>
4662 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
4663 <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7 > ; yields <8 x i32>
4665 Aggregate Operations
4666 --------------------
4668 LLVM supports several instructions for working with
4669 :ref:`aggregate <t_aggregate>` values.
4673 '``extractvalue``' Instruction
4674 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4681 <result> = extractvalue <aggregate type> <val>, <idx>{, <idx>}*
4686 The '``extractvalue``' instruction extracts the value of a member field
4687 from an :ref:`aggregate <t_aggregate>` value.
4692 The first operand of an '``extractvalue``' instruction is a value of
4693 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The operands are
4694 constant indices to specify which value to extract in a similar manner
4695 as indices in a '``getelementptr``' instruction.
4697 The major differences to ``getelementptr`` indexing are:
4699 - Since the value being indexed is not a pointer, the first index is
4700 omitted and assumed to be zero.
4701 - At least one index must be specified.
4702 - Not only struct indices but also array indices must be in bounds.
4707 The result is the value at the position in the aggregate specified by
4713 .. code-block:: llvm
4715 <result> = extractvalue {i32, float} %agg, 0 ; yields i32
4719 '``insertvalue``' Instruction
4720 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4727 <result> = insertvalue <aggregate type> <val>, <ty> <elt>, <idx>{, <idx>}* ; yields <aggregate type>
4732 The '``insertvalue``' instruction inserts a value into a member field in
4733 an :ref:`aggregate <t_aggregate>` value.
4738 The first operand of an '``insertvalue``' instruction is a value of
4739 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The second operand is
4740 a first-class value to insert. The following operands are constant
4741 indices indicating the position at which to insert the value in a
4742 similar manner as indices in a '``extractvalue``' instruction. The value
4743 to insert must have the same type as the value identified by the
4749 The result is an aggregate of the same type as ``val``. Its value is
4750 that of ``val`` except that the value at the position specified by the
4751 indices is that of ``elt``.
4756 .. code-block:: llvm
4758 %agg1 = insertvalue {i32, float} undef, i32 1, 0 ; yields {i32 1, float undef}
4759 %agg2 = insertvalue {i32, float} %agg1, float %val, 1 ; yields {i32 1, float %val}
4760 %agg3 = insertvalue {i32, {float}} %agg1, float %val, 1, 0 ; yields {i32 1, float %val}
4764 Memory Access and Addressing Operations
4765 ---------------------------------------
4767 A key design point of an SSA-based representation is how it represents
4768 memory. In LLVM, no memory locations are in SSA form, which makes things
4769 very simple. This section describes how to read, write, and allocate
4774 '``alloca``' Instruction
4775 ^^^^^^^^^^^^^^^^^^^^^^^^
4782 <result> = alloca [inalloca] <type> [, <ty> <NumElements>] [, align <alignment>] ; yields {type*}:result
4787 The '``alloca``' instruction allocates memory on the stack frame of the
4788 currently executing function, to be automatically released when this
4789 function returns to its caller. The object is always allocated in the
4790 generic address space (address space zero).
4795 The '``alloca``' instruction allocates ``sizeof(<type>)*NumElements``
4796 bytes of memory on the runtime stack, returning a pointer of the
4797 appropriate type to the program. If "NumElements" is specified, it is
4798 the number of elements allocated, otherwise "NumElements" is defaulted
4799 to be one. If a constant alignment is specified, the value result of the
4800 allocation is guaranteed to be aligned to at least that boundary. If not
4801 specified, or if zero, the target can choose to align the allocation on
4802 any convenient boundary compatible with the type.
4804 '``type``' may be any sized type.
4809 Memory is allocated; a pointer is returned. The operation is undefined
4810 if there is insufficient stack space for the allocation. '``alloca``'d
4811 memory is automatically released when the function returns. The
4812 '``alloca``' instruction is commonly used to represent automatic
4813 variables that must have an address available. When the function returns
4814 (either with the ``ret`` or ``resume`` instructions), the memory is
4815 reclaimed. Allocating zero bytes is legal, but the result is undefined.
4816 The order in which memory is allocated (ie., which way the stack grows)
4822 .. code-block:: llvm
4824 %ptr = alloca i32 ; yields {i32*}:ptr
4825 %ptr = alloca i32, i32 4 ; yields {i32*}:ptr
4826 %ptr = alloca i32, i32 4, align 1024 ; yields {i32*}:ptr
4827 %ptr = alloca i32, align 1024 ; yields {i32*}:ptr
4831 '``load``' Instruction
4832 ^^^^^^^^^^^^^^^^^^^^^^
4839 <result> = load [volatile] <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>][, !invariant.load !<index>]
4840 <result> = load atomic [volatile] <ty>* <pointer> [singlethread] <ordering>, align <alignment>
4841 !<index> = !{ i32 1 }
4846 The '``load``' instruction is used to read from memory.
4851 The argument to the ``load`` instruction specifies the memory address
4852 from which to load. The pointer must point to a :ref:`first
4853 class <t_firstclass>` type. If the ``load`` is marked as ``volatile``,
4854 then the optimizer is not allowed to modify the number or order of
4855 execution of this ``load`` with other :ref:`volatile
4856 operations <volatile>`.
4858 If the ``load`` is marked as ``atomic``, it takes an extra
4859 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
4860 ``release`` and ``acq_rel`` orderings are not valid on ``load``
4861 instructions. Atomic loads produce :ref:`defined <memmodel>` results
4862 when they may see multiple atomic stores. The type of the pointee must
4863 be an integer type whose bit width is a power of two greater than or
4864 equal to eight and less than or equal to a target-specific size limit.
4865 ``align`` must be explicitly specified on atomic loads, and the load has
4866 undefined behavior if the alignment is not set to a value which is at
4867 least the size in bytes of the pointee. ``!nontemporal`` does not have
4868 any defined semantics for atomic loads.
4870 The optional constant ``align`` argument specifies the alignment of the
4871 operation (that is, the alignment of the memory address). A value of 0
4872 or an omitted ``align`` argument means that the operation has the ABI
4873 alignment for the target. It is the responsibility of the code emitter
4874 to ensure that the alignment information is correct. Overestimating the
4875 alignment results in undefined behavior. Underestimating the alignment
4876 may produce less efficient code. An alignment of 1 is always safe.
4878 The optional ``!nontemporal`` metadata must reference a single
4879 metadata name ``<index>`` corresponding to a metadata node with one
4880 ``i32`` entry of value 1. The existence of the ``!nontemporal``
4881 metadata on the instruction tells the optimizer and code generator
4882 that this load is not expected to be reused in the cache. The code
4883 generator may select special instructions to save cache bandwidth, such
4884 as the ``MOVNT`` instruction on x86.
4886 The optional ``!invariant.load`` metadata must reference a single
4887 metadata name ``<index>`` corresponding to a metadata node with no
4888 entries. The existence of the ``!invariant.load`` metadata on the
4889 instruction tells the optimizer and code generator that this load
4890 address points to memory which does not change value during program
4891 execution. The optimizer may then move this load around, for example, by
4892 hoisting it out of loops using loop invariant code motion.
4897 The location of memory pointed to is loaded. If the value being loaded
4898 is of scalar type then the number of bytes read does not exceed the
4899 minimum number of bytes needed to hold all bits of the type. For
4900 example, loading an ``i24`` reads at most three bytes. When loading a
4901 value of a type like ``i20`` with a size that is not an integral number
4902 of bytes, the result is undefined if the value was not originally
4903 written using a store of the same type.
4908 .. code-block:: llvm
4910 %ptr = alloca i32 ; yields {i32*}:ptr
4911 store i32 3, i32* %ptr ; yields {void}
4912 %val = load i32* %ptr ; yields {i32}:val = i32 3
4916 '``store``' Instruction
4917 ^^^^^^^^^^^^^^^^^^^^^^^
4924 store [volatile] <ty> <value>, <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>] ; yields {void}
4925 store atomic [volatile] <ty> <value>, <ty>* <pointer> [singlethread] <ordering>, align <alignment> ; yields {void}
4930 The '``store``' instruction is used to write to memory.
4935 There are two arguments to the ``store`` instruction: a value to store
4936 and an address at which to store it. The type of the ``<pointer>``
4937 operand must be a pointer to the :ref:`first class <t_firstclass>` type of
4938 the ``<value>`` operand. If the ``store`` is marked as ``volatile``,
4939 then the optimizer is not allowed to modify the number or order of
4940 execution of this ``store`` with other :ref:`volatile
4941 operations <volatile>`.
4943 If the ``store`` is marked as ``atomic``, it takes an extra
4944 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
4945 ``acquire`` and ``acq_rel`` orderings aren't valid on ``store``
4946 instructions. Atomic loads produce :ref:`defined <memmodel>` results
4947 when they may see multiple atomic stores. The type of the pointee must
4948 be an integer type whose bit width is a power of two greater than or
4949 equal to eight and less than or equal to a target-specific size limit.
4950 ``align`` must be explicitly specified on atomic stores, and the store
4951 has undefined behavior if the alignment is not set to a value which is
4952 at least the size in bytes of the pointee. ``!nontemporal`` does not
4953 have any defined semantics for atomic stores.
4955 The optional constant ``align`` argument specifies the alignment of the
4956 operation (that is, the alignment of the memory address). A value of 0
4957 or an omitted ``align`` argument means that the operation has the ABI
4958 alignment for the target. It is the responsibility of the code emitter
4959 to ensure that the alignment information is correct. Overestimating the
4960 alignment results in undefined behavior. Underestimating the
4961 alignment may produce less efficient code. An alignment of 1 is always
4964 The optional ``!nontemporal`` metadata must reference a single metadata
4965 name ``<index>`` corresponding to a metadata node with one ``i32`` entry of
4966 value 1. The existence of the ``!nontemporal`` metadata on the instruction
4967 tells the optimizer and code generator that this load is not expected to
4968 be reused in the cache. The code generator may select special
4969 instructions to save cache bandwidth, such as the MOVNT instruction on
4975 The contents of memory are updated to contain ``<value>`` at the
4976 location specified by the ``<pointer>`` operand. If ``<value>`` is
4977 of scalar type then the number of bytes written does not exceed the
4978 minimum number of bytes needed to hold all bits of the type. For
4979 example, storing an ``i24`` writes at most three bytes. When writing a
4980 value of a type like ``i20`` with a size that is not an integral number
4981 of bytes, it is unspecified what happens to the extra bits that do not
4982 belong to the type, but they will typically be overwritten.
4987 .. code-block:: llvm
4989 %ptr = alloca i32 ; yields {i32*}:ptr
4990 store i32 3, i32* %ptr ; yields {void}
4991 %val = load i32* %ptr ; yields {i32}:val = i32 3
4995 '``fence``' Instruction
4996 ^^^^^^^^^^^^^^^^^^^^^^^
5003 fence [singlethread] <ordering> ; yields {void}
5008 The '``fence``' instruction is used to introduce happens-before edges
5014 '``fence``' instructions take an :ref:`ordering <ordering>` argument which
5015 defines what *synchronizes-with* edges they add. They can only be given
5016 ``acquire``, ``release``, ``acq_rel``, and ``seq_cst`` orderings.
5021 A fence A which has (at least) ``release`` ordering semantics
5022 *synchronizes with* a fence B with (at least) ``acquire`` ordering
5023 semantics if and only if there exist atomic operations X and Y, both
5024 operating on some atomic object M, such that A is sequenced before X, X
5025 modifies M (either directly or through some side effect of a sequence
5026 headed by X), Y is sequenced before B, and Y observes M. This provides a
5027 *happens-before* dependency between A and B. Rather than an explicit
5028 ``fence``, one (but not both) of the atomic operations X or Y might
5029 provide a ``release`` or ``acquire`` (resp.) ordering constraint and
5030 still *synchronize-with* the explicit ``fence`` and establish the
5031 *happens-before* edge.
5033 A ``fence`` which has ``seq_cst`` ordering, in addition to having both
5034 ``acquire`` and ``release`` semantics specified above, participates in
5035 the global program order of other ``seq_cst`` operations and/or fences.
5037 The optional ":ref:`singlethread <singlethread>`" argument specifies
5038 that the fence only synchronizes with other fences in the same thread.
5039 (This is useful for interacting with signal handlers.)
5044 .. code-block:: llvm
5046 fence acquire ; yields {void}
5047 fence singlethread seq_cst ; yields {void}
5051 '``cmpxchg``' Instruction
5052 ^^^^^^^^^^^^^^^^^^^^^^^^^
5059 cmpxchg [volatile] <ty>* <pointer>, <ty> <cmp>, <ty> <new> [singlethread] <success ordering> <failure ordering> ; yields {ty}
5064 The '``cmpxchg``' instruction is used to atomically modify memory. It
5065 loads a value in memory and compares it to a given value. If they are
5066 equal, it stores a new value into the memory.
5071 There are three arguments to the '``cmpxchg``' instruction: an address
5072 to operate on, a value to compare to the value currently be at that
5073 address, and a new value to place at that address if the compared values
5074 are equal. The type of '<cmp>' must be an integer type whose bit width
5075 is a power of two greater than or equal to eight and less than or equal
5076 to a target-specific size limit. '<cmp>' and '<new>' must have the same
5077 type, and the type of '<pointer>' must be a pointer to that type. If the
5078 ``cmpxchg`` is marked as ``volatile``, then the optimizer is not allowed
5079 to modify the number or order of execution of this ``cmpxchg`` with
5080 other :ref:`volatile operations <volatile>`.
5082 The success and failure :ref:`ordering <ordering>` arguments specify how this
5083 ``cmpxchg`` synchronizes with other atomic operations. The both ordering
5084 parameters must be at least ``monotonic``, the ordering constraint on failure
5085 must be no stronger than that on success, and the failure ordering cannot be
5086 either ``release`` or ``acq_rel``.
5088 The optional "``singlethread``" argument declares that the ``cmpxchg``
5089 is only atomic with respect to code (usually signal handlers) running in
5090 the same thread as the ``cmpxchg``. Otherwise the cmpxchg is atomic with
5091 respect to all other code in the system.
5093 The pointer passed into cmpxchg must have alignment greater than or
5094 equal to the size in memory of the operand.
5099 The contents of memory at the location specified by the '``<pointer>``'
5100 operand is read and compared to '``<cmp>``'; if the read value is the
5101 equal, '``<new>``' is written. The original value at the location is
5104 A successful ``cmpxchg`` is a read-modify-write instruction for the purpose of
5105 identifying release sequences. A failed ``cmpxchg`` is equivalent to an atomic
5106 load with an ordering parameter determined the second ordering parameter.
5111 .. code-block:: llvm
5114 %orig = atomic load i32* %ptr unordered ; yields {i32}
5118 %cmp = phi i32 [ %orig, %entry ], [%old, %loop]
5119 %squared = mul i32 %cmp, %cmp
5120 %old = cmpxchg i32* %ptr, i32 %cmp, i32 %squared acq_rel monotonic ; yields {i32}
5121 %success = icmp eq i32 %cmp, %old
5122 br i1 %success, label %done, label %loop
5129 '``atomicrmw``' Instruction
5130 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
5137 atomicrmw [volatile] <operation> <ty>* <pointer>, <ty> <value> [singlethread] <ordering> ; yields {ty}
5142 The '``atomicrmw``' instruction is used to atomically modify memory.
5147 There are three arguments to the '``atomicrmw``' instruction: an
5148 operation to apply, an address whose value to modify, an argument to the
5149 operation. The operation must be one of the following keywords:
5163 The type of '<value>' must be an integer type whose bit width is a power
5164 of two greater than or equal to eight and less than or equal to a
5165 target-specific size limit. The type of the '``<pointer>``' operand must
5166 be a pointer to that type. If the ``atomicrmw`` is marked as
5167 ``volatile``, then the optimizer is not allowed to modify the number or
5168 order of execution of this ``atomicrmw`` with other :ref:`volatile
5169 operations <volatile>`.
5174 The contents of memory at the location specified by the '``<pointer>``'
5175 operand are atomically read, modified, and written back. The original
5176 value at the location is returned. The modification is specified by the
5179 - xchg: ``*ptr = val``
5180 - add: ``*ptr = *ptr + val``
5181 - sub: ``*ptr = *ptr - val``
5182 - and: ``*ptr = *ptr & val``
5183 - nand: ``*ptr = ~(*ptr & val)``
5184 - or: ``*ptr = *ptr | val``
5185 - xor: ``*ptr = *ptr ^ val``
5186 - max: ``*ptr = *ptr > val ? *ptr : val`` (using a signed comparison)
5187 - min: ``*ptr = *ptr < val ? *ptr : val`` (using a signed comparison)
5188 - umax: ``*ptr = *ptr > val ? *ptr : val`` (using an unsigned
5190 - umin: ``*ptr = *ptr < val ? *ptr : val`` (using an unsigned
5196 .. code-block:: llvm
5198 %old = atomicrmw add i32* %ptr, i32 1 acquire ; yields {i32}
5200 .. _i_getelementptr:
5202 '``getelementptr``' Instruction
5203 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5210 <result> = getelementptr <pty>* <ptrval>{, <ty> <idx>}*
5211 <result> = getelementptr inbounds <pty>* <ptrval>{, <ty> <idx>}*
5212 <result> = getelementptr <ptr vector> ptrval, <vector index type> idx
5217 The '``getelementptr``' instruction is used to get the address of a
5218 subelement of an :ref:`aggregate <t_aggregate>` data structure. It performs
5219 address calculation only and does not access memory.
5224 The first argument is always a pointer or a vector of pointers, and
5225 forms the basis of the calculation. The remaining arguments are indices
5226 that indicate which of the elements of the aggregate object are indexed.
5227 The interpretation of each index is dependent on the type being indexed
5228 into. The first index always indexes the pointer value given as the
5229 first argument, the second index indexes a value of the type pointed to
5230 (not necessarily the value directly pointed to, since the first index
5231 can be non-zero), etc. The first type indexed into must be a pointer
5232 value, subsequent types can be arrays, vectors, and structs. Note that
5233 subsequent types being indexed into can never be pointers, since that
5234 would require loading the pointer before continuing calculation.
5236 The type of each index argument depends on the type it is indexing into.
5237 When indexing into a (optionally packed) structure, only ``i32`` integer
5238 **constants** are allowed (when using a vector of indices they must all
5239 be the **same** ``i32`` integer constant). When indexing into an array,
5240 pointer or vector, integers of any width are allowed, and they are not
5241 required to be constant. These integers are treated as signed values
5244 For example, let's consider a C code fragment and how it gets compiled
5260 int *foo(struct ST *s) {
5261 return &s[1].Z.B[5][13];
5264 The LLVM code generated by Clang is:
5266 .. code-block:: llvm
5268 %struct.RT = type { i8, [10 x [20 x i32]], i8 }
5269 %struct.ST = type { i32, double, %struct.RT }
5271 define i32* @foo(%struct.ST* %s) nounwind uwtable readnone optsize ssp {
5273 %arrayidx = getelementptr inbounds %struct.ST* %s, i64 1, i32 2, i32 1, i64 5, i64 13
5280 In the example above, the first index is indexing into the
5281 '``%struct.ST*``' type, which is a pointer, yielding a '``%struct.ST``'
5282 = '``{ i32, double, %struct.RT }``' type, a structure. The second index
5283 indexes into the third element of the structure, yielding a
5284 '``%struct.RT``' = '``{ i8 , [10 x [20 x i32]], i8 }``' type, another
5285 structure. The third index indexes into the second element of the
5286 structure, yielding a '``[10 x [20 x i32]]``' type, an array. The two
5287 dimensions of the array are subscripted into, yielding an '``i32``'
5288 type. The '``getelementptr``' instruction returns a pointer to this
5289 element, thus computing a value of '``i32*``' type.
5291 Note that it is perfectly legal to index partially through a structure,
5292 returning a pointer to an inner element. Because of this, the LLVM code
5293 for the given testcase is equivalent to:
5295 .. code-block:: llvm
5297 define i32* @foo(%struct.ST* %s) {
5298 %t1 = getelementptr %struct.ST* %s, i32 1 ; yields %struct.ST*:%t1
5299 %t2 = getelementptr %struct.ST* %t1, i32 0, i32 2 ; yields %struct.RT*:%t2
5300 %t3 = getelementptr %struct.RT* %t2, i32 0, i32 1 ; yields [10 x [20 x i32]]*:%t3
5301 %t4 = getelementptr [10 x [20 x i32]]* %t3, i32 0, i32 5 ; yields [20 x i32]*:%t4
5302 %t5 = getelementptr [20 x i32]* %t4, i32 0, i32 13 ; yields i32*:%t5
5306 If the ``inbounds`` keyword is present, the result value of the
5307 ``getelementptr`` is a :ref:`poison value <poisonvalues>` if the base
5308 pointer is not an *in bounds* address of an allocated object, or if any
5309 of the addresses that would be formed by successive addition of the
5310 offsets implied by the indices to the base address with infinitely
5311 precise signed arithmetic are not an *in bounds* address of that
5312 allocated object. The *in bounds* addresses for an allocated object are
5313 all the addresses that point into the object, plus the address one byte
5314 past the end. In cases where the base is a vector of pointers the
5315 ``inbounds`` keyword applies to each of the computations element-wise.
5317 If the ``inbounds`` keyword is not present, the offsets are added to the
5318 base address with silently-wrapping two's complement arithmetic. If the
5319 offsets have a different width from the pointer, they are sign-extended
5320 or truncated to the width of the pointer. The result value of the
5321 ``getelementptr`` may be outside the object pointed to by the base
5322 pointer. The result value may not necessarily be used to access memory
5323 though, even if it happens to point into allocated storage. See the
5324 :ref:`Pointer Aliasing Rules <pointeraliasing>` section for more
5327 The getelementptr instruction is often confusing. For some more insight
5328 into how it works, see :doc:`the getelementptr FAQ <GetElementPtr>`.
5333 .. code-block:: llvm
5335 ; yields [12 x i8]*:aptr
5336 %aptr = getelementptr {i32, [12 x i8]}* %saptr, i64 0, i32 1
5338 %vptr = getelementptr {i32, <2 x i8>}* %svptr, i64 0, i32 1, i32 1
5340 %eptr = getelementptr [12 x i8]* %aptr, i64 0, i32 1
5342 %iptr = getelementptr [10 x i32]* @arr, i16 0, i16 0
5344 In cases where the pointer argument is a vector of pointers, each index
5345 must be a vector with the same number of elements. For example:
5347 .. code-block:: llvm
5349 %A = getelementptr <4 x i8*> %ptrs, <4 x i64> %offsets,
5351 Conversion Operations
5352 ---------------------
5354 The instructions in this category are the conversion instructions
5355 (casting) which all take a single operand and a type. They perform
5356 various bit conversions on the operand.
5358 '``trunc .. to``' Instruction
5359 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5366 <result> = trunc <ty> <value> to <ty2> ; yields ty2
5371 The '``trunc``' instruction truncates its operand to the type ``ty2``.
5376 The '``trunc``' instruction takes a value to trunc, and a type to trunc
5377 it to. Both types must be of :ref:`integer <t_integer>` types, or vectors
5378 of the same number of integers. The bit size of the ``value`` must be
5379 larger than the bit size of the destination type, ``ty2``. Equal sized
5380 types are not allowed.
5385 The '``trunc``' instruction truncates the high order bits in ``value``
5386 and converts the remaining bits to ``ty2``. Since the source size must
5387 be larger than the destination size, ``trunc`` cannot be a *no-op cast*.
5388 It will always truncate bits.
5393 .. code-block:: llvm
5395 %X = trunc i32 257 to i8 ; yields i8:1
5396 %Y = trunc i32 123 to i1 ; yields i1:true
5397 %Z = trunc i32 122 to i1 ; yields i1:false
5398 %W = trunc <2 x i16> <i16 8, i16 7> to <2 x i8> ; yields <i8 8, i8 7>
5400 '``zext .. to``' Instruction
5401 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5408 <result> = zext <ty> <value> to <ty2> ; yields ty2
5413 The '``zext``' instruction zero extends its operand to type ``ty2``.
5418 The '``zext``' instruction takes a value to cast, and a type to cast it
5419 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
5420 the same number of integers. The bit size of the ``value`` must be
5421 smaller than the bit size of the destination type, ``ty2``.
5426 The ``zext`` fills the high order bits of the ``value`` with zero bits
5427 until it reaches the size of the destination type, ``ty2``.
5429 When zero extending from i1, the result will always be either 0 or 1.
5434 .. code-block:: llvm
5436 %X = zext i32 257 to i64 ; yields i64:257
5437 %Y = zext i1 true to i32 ; yields i32:1
5438 %Z = zext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
5440 '``sext .. to``' Instruction
5441 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5448 <result> = sext <ty> <value> to <ty2> ; yields ty2
5453 The '``sext``' sign extends ``value`` to the type ``ty2``.
5458 The '``sext``' instruction takes a value to cast, and a type to cast it
5459 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
5460 the same number of integers. The bit size of the ``value`` must be
5461 smaller than the bit size of the destination type, ``ty2``.
5466 The '``sext``' instruction performs a sign extension by copying the sign
5467 bit (highest order bit) of the ``value`` until it reaches the bit size
5468 of the type ``ty2``.
5470 When sign extending from i1, the extension always results in -1 or 0.
5475 .. code-block:: llvm
5477 %X = sext i8 -1 to i16 ; yields i16 :65535
5478 %Y = sext i1 true to i32 ; yields i32:-1
5479 %Z = sext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
5481 '``fptrunc .. to``' Instruction
5482 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5489 <result> = fptrunc <ty> <value> to <ty2> ; yields ty2
5494 The '``fptrunc``' instruction truncates ``value`` to type ``ty2``.
5499 The '``fptrunc``' instruction takes a :ref:`floating point <t_floating>`
5500 value to cast and a :ref:`floating point <t_floating>` type to cast it to.
5501 The size of ``value`` must be larger than the size of ``ty2``. This
5502 implies that ``fptrunc`` cannot be used to make a *no-op cast*.
5507 The '``fptrunc``' instruction truncates a ``value`` from a larger
5508 :ref:`floating point <t_floating>` type to a smaller :ref:`floating
5509 point <t_floating>` type. If the value cannot fit within the
5510 destination type, ``ty2``, then the results are undefined.
5515 .. code-block:: llvm
5517 %X = fptrunc double 123.0 to float ; yields float:123.0
5518 %Y = fptrunc double 1.0E+300 to float ; yields undefined
5520 '``fpext .. to``' Instruction
5521 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5528 <result> = fpext <ty> <value> to <ty2> ; yields ty2
5533 The '``fpext``' extends a floating point ``value`` to a larger floating
5539 The '``fpext``' instruction takes a :ref:`floating point <t_floating>`
5540 ``value`` to cast, and a :ref:`floating point <t_floating>` type to cast it
5541 to. The source type must be smaller than the destination type.
5546 The '``fpext``' instruction extends the ``value`` from a smaller
5547 :ref:`floating point <t_floating>` type to a larger :ref:`floating
5548 point <t_floating>` type. The ``fpext`` cannot be used to make a
5549 *no-op cast* because it always changes bits. Use ``bitcast`` to make a
5550 *no-op cast* for a floating point cast.
5555 .. code-block:: llvm
5557 %X = fpext float 3.125 to double ; yields double:3.125000e+00
5558 %Y = fpext double %X to fp128 ; yields fp128:0xL00000000000000004000900000000000
5560 '``fptoui .. to``' Instruction
5561 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5568 <result> = fptoui <ty> <value> to <ty2> ; yields ty2
5573 The '``fptoui``' converts a floating point ``value`` to its unsigned
5574 integer equivalent of type ``ty2``.
5579 The '``fptoui``' instruction takes a value to cast, which must be a
5580 scalar or vector :ref:`floating point <t_floating>` value, and a type to
5581 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
5582 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
5583 type with the same number of elements as ``ty``
5588 The '``fptoui``' instruction converts its :ref:`floating
5589 point <t_floating>` operand into the nearest (rounding towards zero)
5590 unsigned integer value. If the value cannot fit in ``ty2``, the results
5596 .. code-block:: llvm
5598 %X = fptoui double 123.0 to i32 ; yields i32:123
5599 %Y = fptoui float 1.0E+300 to i1 ; yields undefined:1
5600 %Z = fptoui float 1.04E+17 to i8 ; yields undefined:1
5602 '``fptosi .. to``' Instruction
5603 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5610 <result> = fptosi <ty> <value> to <ty2> ; yields ty2
5615 The '``fptosi``' instruction converts :ref:`floating point <t_floating>`
5616 ``value`` to type ``ty2``.
5621 The '``fptosi``' instruction takes a value to cast, which must be a
5622 scalar or vector :ref:`floating point <t_floating>` value, and a type to
5623 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
5624 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
5625 type with the same number of elements as ``ty``
5630 The '``fptosi``' instruction converts its :ref:`floating
5631 point <t_floating>` operand into the nearest (rounding towards zero)
5632 signed integer value. If the value cannot fit in ``ty2``, the results
5638 .. code-block:: llvm
5640 %X = fptosi double -123.0 to i32 ; yields i32:-123
5641 %Y = fptosi float 1.0E-247 to i1 ; yields undefined:1
5642 %Z = fptosi float 1.04E+17 to i8 ; yields undefined:1
5644 '``uitofp .. to``' Instruction
5645 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5652 <result> = uitofp <ty> <value> to <ty2> ; yields ty2
5657 The '``uitofp``' instruction regards ``value`` as an unsigned integer
5658 and converts that value to the ``ty2`` type.
5663 The '``uitofp``' instruction takes a value to cast, which must be a
5664 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
5665 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
5666 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
5667 type with the same number of elements as ``ty``
5672 The '``uitofp``' instruction interprets its operand as an unsigned
5673 integer quantity and converts it to the corresponding floating point
5674 value. If the value cannot fit in the floating point value, the results
5680 .. code-block:: llvm
5682 %X = uitofp i32 257 to float ; yields float:257.0
5683 %Y = uitofp i8 -1 to double ; yields double:255.0
5685 '``sitofp .. to``' Instruction
5686 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5693 <result> = sitofp <ty> <value> to <ty2> ; yields ty2
5698 The '``sitofp``' instruction regards ``value`` as a signed integer and
5699 converts that value to the ``ty2`` type.
5704 The '``sitofp``' instruction takes a value to cast, which must be a
5705 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
5706 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
5707 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
5708 type with the same number of elements as ``ty``
5713 The '``sitofp``' instruction interprets its operand as a signed integer
5714 quantity and converts it to the corresponding floating point value. If
5715 the value cannot fit in the floating point value, the results are
5721 .. code-block:: llvm
5723 %X = sitofp i32 257 to float ; yields float:257.0
5724 %Y = sitofp i8 -1 to double ; yields double:-1.0
5728 '``ptrtoint .. to``' Instruction
5729 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5736 <result> = ptrtoint <ty> <value> to <ty2> ; yields ty2
5741 The '``ptrtoint``' instruction converts the pointer or a vector of
5742 pointers ``value`` to the integer (or vector of integers) type ``ty2``.
5747 The '``ptrtoint``' instruction takes a ``value`` to cast, which must be
5748 a a value of type :ref:`pointer <t_pointer>` or a vector of pointers, and a
5749 type to cast it to ``ty2``, which must be an :ref:`integer <t_integer>` or
5750 a vector of integers type.
5755 The '``ptrtoint``' instruction converts ``value`` to integer type
5756 ``ty2`` by interpreting the pointer value as an integer and either
5757 truncating or zero extending that value to the size of the integer type.
5758 If ``value`` is smaller than ``ty2`` then a zero extension is done. If
5759 ``value`` is larger than ``ty2`` then a truncation is done. If they are
5760 the same size, then nothing is done (*no-op cast*) other than a type
5766 .. code-block:: llvm
5768 %X = ptrtoint i32* %P to i8 ; yields truncation on 32-bit architecture
5769 %Y = ptrtoint i32* %P to i64 ; yields zero extension on 32-bit architecture
5770 %Z = ptrtoint <4 x i32*> %P to <4 x i64>; yields vector zero extension for a vector of addresses on 32-bit architecture
5774 '``inttoptr .. to``' Instruction
5775 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5782 <result> = inttoptr <ty> <value> to <ty2> ; yields ty2
5787 The '``inttoptr``' instruction converts an integer ``value`` to a
5788 pointer type, ``ty2``.
5793 The '``inttoptr``' instruction takes an :ref:`integer <t_integer>` value to
5794 cast, and a type to cast it to, which must be a :ref:`pointer <t_pointer>`
5800 The '``inttoptr``' instruction converts ``value`` to type ``ty2`` by
5801 applying either a zero extension or a truncation depending on the size
5802 of the integer ``value``. If ``value`` is larger than the size of a
5803 pointer then a truncation is done. If ``value`` is smaller than the size
5804 of a pointer then a zero extension is done. If they are the same size,
5805 nothing is done (*no-op cast*).
5810 .. code-block:: llvm
5812 %X = inttoptr i32 255 to i32* ; yields zero extension on 64-bit architecture
5813 %Y = inttoptr i32 255 to i32* ; yields no-op on 32-bit architecture
5814 %Z = inttoptr i64 0 to i32* ; yields truncation on 32-bit architecture
5815 %Z = inttoptr <4 x i32> %G to <4 x i8*>; yields truncation of vector G to four pointers
5819 '``bitcast .. to``' Instruction
5820 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5827 <result> = bitcast <ty> <value> to <ty2> ; yields ty2
5832 The '``bitcast``' instruction converts ``value`` to type ``ty2`` without
5838 The '``bitcast``' instruction takes a value to cast, which must be a
5839 non-aggregate first class value, and a type to cast it to, which must
5840 also be a non-aggregate :ref:`first class <t_firstclass>` type. The
5841 bit sizes of ``value`` and the destination type, ``ty2``, must be
5842 identical. If the source type is a pointer, the destination type must
5843 also be a pointer of the same size. This instruction supports bitwise
5844 conversion of vectors to integers and to vectors of other types (as
5845 long as they have the same size).
5850 The '``bitcast``' instruction converts ``value`` to type ``ty2``. It
5851 is always a *no-op cast* because no bits change with this
5852 conversion. The conversion is done as if the ``value`` had been stored
5853 to memory and read back as type ``ty2``. Pointer (or vector of
5854 pointers) types may only be converted to other pointer (or vector of
5855 pointers) types with the same address space through this instruction.
5856 To convert pointers to other types, use the :ref:`inttoptr <i_inttoptr>`
5857 or :ref:`ptrtoint <i_ptrtoint>` instructions first.
5862 .. code-block:: llvm
5864 %X = bitcast i8 255 to i8 ; yields i8 :-1
5865 %Y = bitcast i32* %x to sint* ; yields sint*:%x
5866 %Z = bitcast <2 x int> %V to i64; ; yields i64: %V
5867 %Z = bitcast <2 x i32*> %V to <2 x i64*> ; yields <2 x i64*>
5869 .. _i_addrspacecast:
5871 '``addrspacecast .. to``' Instruction
5872 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5879 <result> = addrspacecast <pty> <ptrval> to <pty2> ; yields pty2
5884 The '``addrspacecast``' instruction converts ``ptrval`` from ``pty`` in
5885 address space ``n`` to type ``pty2`` in address space ``m``.
5890 The '``addrspacecast``' instruction takes a pointer or vector of pointer value
5891 to cast and a pointer type to cast it to, which must have a different
5897 The '``addrspacecast``' instruction converts the pointer value
5898 ``ptrval`` to type ``pty2``. It can be a *no-op cast* or a complex
5899 value modification, depending on the target and the address space
5900 pair. Pointer conversions within the same address space must be
5901 performed with the ``bitcast`` instruction. Note that if the address space
5902 conversion is legal then both result and operand refer to the same memory
5908 .. code-block:: llvm
5910 %X = addrspacecast i32* %x to i32 addrspace(1)* ; yields i32 addrspace(1)*:%x
5911 %Y = addrspacecast i32 addrspace(1)* %y to i64 addrspace(2)* ; yields i64 addrspace(2)*:%y
5912 %Z = addrspacecast <4 x i32*> %z to <4 x float addrspace(3)*> ; yields <4 x float addrspace(3)*>:%z
5919 The instructions in this category are the "miscellaneous" instructions,
5920 which defy better classification.
5924 '``icmp``' Instruction
5925 ^^^^^^^^^^^^^^^^^^^^^^
5932 <result> = icmp <cond> <ty> <op1>, <op2> ; yields {i1} or {<N x i1>}:result
5937 The '``icmp``' instruction returns a boolean value or a vector of
5938 boolean values based on comparison of its two integer, integer vector,
5939 pointer, or pointer vector operands.
5944 The '``icmp``' instruction takes three operands. The first operand is
5945 the condition code indicating the kind of comparison to perform. It is
5946 not a value, just a keyword. The possible condition code are:
5949 #. ``ne``: not equal
5950 #. ``ugt``: unsigned greater than
5951 #. ``uge``: unsigned greater or equal
5952 #. ``ult``: unsigned less than
5953 #. ``ule``: unsigned less or equal
5954 #. ``sgt``: signed greater than
5955 #. ``sge``: signed greater or equal
5956 #. ``slt``: signed less than
5957 #. ``sle``: signed less or equal
5959 The remaining two arguments must be :ref:`integer <t_integer>` or
5960 :ref:`pointer <t_pointer>` or integer :ref:`vector <t_vector>` typed. They
5961 must also be identical types.
5966 The '``icmp``' compares ``op1`` and ``op2`` according to the condition
5967 code given as ``cond``. The comparison performed always yields either an
5968 :ref:`i1 <t_integer>` or vector of ``i1`` result, as follows:
5970 #. ``eq``: yields ``true`` if the operands are equal, ``false``
5971 otherwise. No sign interpretation is necessary or performed.
5972 #. ``ne``: yields ``true`` if the operands are unequal, ``false``
5973 otherwise. No sign interpretation is necessary or performed.
5974 #. ``ugt``: interprets the operands as unsigned values and yields
5975 ``true`` if ``op1`` is greater than ``op2``.
5976 #. ``uge``: interprets the operands as unsigned values and yields
5977 ``true`` if ``op1`` is greater than or equal to ``op2``.
5978 #. ``ult``: interprets the operands as unsigned values and yields
5979 ``true`` if ``op1`` is less than ``op2``.
5980 #. ``ule``: interprets the operands as unsigned values and yields
5981 ``true`` if ``op1`` is less than or equal to ``op2``.
5982 #. ``sgt``: interprets the operands as signed values and yields ``true``
5983 if ``op1`` is greater than ``op2``.
5984 #. ``sge``: interprets the operands as signed values and yields ``true``
5985 if ``op1`` is greater than or equal to ``op2``.
5986 #. ``slt``: interprets the operands as signed values and yields ``true``
5987 if ``op1`` is less than ``op2``.
5988 #. ``sle``: interprets the operands as signed values and yields ``true``
5989 if ``op1`` is less than or equal to ``op2``.
5991 If the operands are :ref:`pointer <t_pointer>` typed, the pointer values
5992 are compared as if they were integers.
5994 If the operands are integer vectors, then they are compared element by
5995 element. The result is an ``i1`` vector with the same number of elements
5996 as the values being compared. Otherwise, the result is an ``i1``.
6001 .. code-block:: llvm
6003 <result> = icmp eq i32 4, 5 ; yields: result=false
6004 <result> = icmp ne float* %X, %X ; yields: result=false
6005 <result> = icmp ult i16 4, 5 ; yields: result=true
6006 <result> = icmp sgt i16 4, 5 ; yields: result=false
6007 <result> = icmp ule i16 -4, 5 ; yields: result=false
6008 <result> = icmp sge i16 4, 5 ; yields: result=false
6010 Note that the code generator does not yet support vector types with the
6011 ``icmp`` instruction.
6015 '``fcmp``' Instruction
6016 ^^^^^^^^^^^^^^^^^^^^^^
6023 <result> = fcmp <cond> <ty> <op1>, <op2> ; yields {i1} or {<N x i1>}:result
6028 The '``fcmp``' instruction returns a boolean value or vector of boolean
6029 values based on comparison of its operands.
6031 If the operands are floating point scalars, then the result type is a
6032 boolean (:ref:`i1 <t_integer>`).
6034 If the operands are floating point vectors, then the result type is a
6035 vector of boolean with the same number of elements as the operands being
6041 The '``fcmp``' instruction takes three operands. The first operand is
6042 the condition code indicating the kind of comparison to perform. It is
6043 not a value, just a keyword. The possible condition code are:
6045 #. ``false``: no comparison, always returns false
6046 #. ``oeq``: ordered and equal
6047 #. ``ogt``: ordered and greater than
6048 #. ``oge``: ordered and greater than or equal
6049 #. ``olt``: ordered and less than
6050 #. ``ole``: ordered and less than or equal
6051 #. ``one``: ordered and not equal
6052 #. ``ord``: ordered (no nans)
6053 #. ``ueq``: unordered or equal
6054 #. ``ugt``: unordered or greater than
6055 #. ``uge``: unordered or greater than or equal
6056 #. ``ult``: unordered or less than
6057 #. ``ule``: unordered or less than or equal
6058 #. ``une``: unordered or not equal
6059 #. ``uno``: unordered (either nans)
6060 #. ``true``: no comparison, always returns true
6062 *Ordered* means that neither operand is a QNAN while *unordered* means
6063 that either operand may be a QNAN.
6065 Each of ``val1`` and ``val2`` arguments must be either a :ref:`floating
6066 point <t_floating>` type or a :ref:`vector <t_vector>` of floating point
6067 type. They must have identical types.
6072 The '``fcmp``' instruction compares ``op1`` and ``op2`` according to the
6073 condition code given as ``cond``. If the operands are vectors, then the
6074 vectors are compared element by element. Each comparison performed
6075 always yields an :ref:`i1 <t_integer>` result, as follows:
6077 #. ``false``: always yields ``false``, regardless of operands.
6078 #. ``oeq``: yields ``true`` if both operands are not a QNAN and ``op1``
6079 is equal to ``op2``.
6080 #. ``ogt``: yields ``true`` if both operands are not a QNAN and ``op1``
6081 is greater than ``op2``.
6082 #. ``oge``: yields ``true`` if both operands are not a QNAN and ``op1``
6083 is greater than or equal to ``op2``.
6084 #. ``olt``: yields ``true`` if both operands are not a QNAN and ``op1``
6085 is less than ``op2``.
6086 #. ``ole``: yields ``true`` if both operands are not a QNAN and ``op1``
6087 is less than or equal to ``op2``.
6088 #. ``one``: yields ``true`` if both operands are not a QNAN and ``op1``
6089 is not equal to ``op2``.
6090 #. ``ord``: yields ``true`` if both operands are not a QNAN.
6091 #. ``ueq``: yields ``true`` if either operand is a QNAN or ``op1`` is
6093 #. ``ugt``: yields ``true`` if either operand is a QNAN or ``op1`` is
6094 greater than ``op2``.
6095 #. ``uge``: yields ``true`` if either operand is a QNAN or ``op1`` is
6096 greater than or equal to ``op2``.
6097 #. ``ult``: yields ``true`` if either operand is a QNAN or ``op1`` is
6099 #. ``ule``: yields ``true`` if either operand is a QNAN or ``op1`` is
6100 less than or equal to ``op2``.
6101 #. ``une``: yields ``true`` if either operand is a QNAN or ``op1`` is
6102 not equal to ``op2``.
6103 #. ``uno``: yields ``true`` if either operand is a QNAN.
6104 #. ``true``: always yields ``true``, regardless of operands.
6109 .. code-block:: llvm
6111 <result> = fcmp oeq float 4.0, 5.0 ; yields: result=false
6112 <result> = fcmp one float 4.0, 5.0 ; yields: result=true
6113 <result> = fcmp olt float 4.0, 5.0 ; yields: result=true
6114 <result> = fcmp ueq double 1.0, 2.0 ; yields: result=false
6116 Note that the code generator does not yet support vector types with the
6117 ``fcmp`` instruction.
6121 '``phi``' Instruction
6122 ^^^^^^^^^^^^^^^^^^^^^
6129 <result> = phi <ty> [ <val0>, <label0>], ...
6134 The '``phi``' instruction is used to implement the φ node in the SSA
6135 graph representing the function.
6140 The type of the incoming values is specified with the first type field.
6141 After this, the '``phi``' instruction takes a list of pairs as
6142 arguments, with one pair for each predecessor basic block of the current
6143 block. Only values of :ref:`first class <t_firstclass>` type may be used as
6144 the value arguments to the PHI node. Only labels may be used as the
6147 There must be no non-phi instructions between the start of a basic block
6148 and the PHI instructions: i.e. PHI instructions must be first in a basic
6151 For the purposes of the SSA form, the use of each incoming value is
6152 deemed to occur on the edge from the corresponding predecessor block to
6153 the current block (but after any definition of an '``invoke``'
6154 instruction's return value on the same edge).
6159 At runtime, the '``phi``' instruction logically takes on the value
6160 specified by the pair corresponding to the predecessor basic block that
6161 executed just prior to the current block.
6166 .. code-block:: llvm
6168 Loop: ; Infinite loop that counts from 0 on up...
6169 %indvar = phi i32 [ 0, %LoopHeader ], [ %nextindvar, %Loop ]
6170 %nextindvar = add i32 %indvar, 1
6175 '``select``' Instruction
6176 ^^^^^^^^^^^^^^^^^^^^^^^^
6183 <result> = select selty <cond>, <ty> <val1>, <ty> <val2> ; yields ty
6185 selty is either i1 or {<N x i1>}
6190 The '``select``' instruction is used to choose one value based on a
6191 condition, without IR-level branching.
6196 The '``select``' instruction requires an 'i1' value or a vector of 'i1'
6197 values indicating the condition, and two values of the same :ref:`first
6198 class <t_firstclass>` type. If the val1/val2 are vectors and the
6199 condition is a scalar, then entire vectors are selected, not individual
6205 If the condition is an i1 and it evaluates to 1, the instruction returns
6206 the first value argument; otherwise, it returns the second value
6209 If the condition is a vector of i1, then the value arguments must be
6210 vectors of the same size, and the selection is done element by element.
6215 .. code-block:: llvm
6217 %X = select i1 true, i8 17, i8 42 ; yields i8:17
6221 '``call``' Instruction
6222 ^^^^^^^^^^^^^^^^^^^^^^
6229 <result> = [tail | musttail] call [cconv] [ret attrs] <ty> [<fnty>*] <fnptrval>(<function args>) [fn attrs]
6234 The '``call``' instruction represents a simple function call.
6239 This instruction requires several arguments:
6241 #. The optional ``tail`` and ``musttail`` markers indicate that the optimizers
6242 should perform tail call optimization. The ``tail`` marker is a hint that
6243 `can be ignored <CodeGenerator.html#sibcallopt>`_. The ``musttail`` marker
6244 means that the call must be tail call optimized in order for the program to
6245 be correct. The ``musttail`` marker provides these guarantees:
6247 #. The call will not cause unbounded stack growth if it is part of a
6248 recursive cycle in the call graph.
6249 #. Arguments with the :ref:`inalloca <attr_inalloca>` attribute are
6252 Both markers imply that the callee does not access allocas or varargs from
6253 the caller. Calls marked ``musttail`` must obey the following additional
6256 - The call must immediately precede a :ref:`ret <i_ret>` instruction,
6257 or a pointer bitcast followed by a ret instruction.
6258 - The ret instruction must return the (possibly bitcasted) value
6259 produced by the call or void.
6260 - The caller and callee prototypes must match. Pointer types of
6261 parameters or return types may differ in pointee type, but not
6263 - The calling conventions of the caller and callee must match.
6264 - All ABI-impacting function attributes, such as sret, byval, inreg,
6265 returned, and inalloca, must match.
6267 Tail call optimization for calls marked ``tail`` is guaranteed to occur if
6268 the following conditions are met:
6270 - Caller and callee both have the calling convention ``fastcc``.
6271 - The call is in tail position (ret immediately follows call and ret
6272 uses value of call or is void).
6273 - Option ``-tailcallopt`` is enabled, or
6274 ``llvm::GuaranteedTailCallOpt`` is ``true``.
6275 - `Platform specific constraints are
6276 met. <CodeGenerator.html#tailcallopt>`_
6278 #. The optional "cconv" marker indicates which :ref:`calling
6279 convention <callingconv>` the call should use. If none is
6280 specified, the call defaults to using C calling conventions. The
6281 calling convention of the call must match the calling convention of
6282 the target function, or else the behavior is undefined.
6283 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
6284 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
6286 #. '``ty``': the type of the call instruction itself which is also the
6287 type of the return value. Functions that return no value are marked
6289 #. '``fnty``': shall be the signature of the pointer to function value
6290 being invoked. The argument types must match the types implied by
6291 this signature. This type can be omitted if the function is not
6292 varargs and if the function type does not return a pointer to a
6294 #. '``fnptrval``': An LLVM value containing a pointer to a function to
6295 be invoked. In most cases, this is a direct function invocation, but
6296 indirect ``call``'s are just as possible, calling an arbitrary pointer
6298 #. '``function args``': argument list whose types match the function
6299 signature argument types and parameter attributes. All arguments must
6300 be of :ref:`first class <t_firstclass>` type. If the function signature
6301 indicates the function accepts a variable number of arguments, the
6302 extra arguments can be specified.
6303 #. The optional :ref:`function attributes <fnattrs>` list. Only
6304 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
6305 attributes are valid here.
6310 The '``call``' instruction is used to cause control flow to transfer to
6311 a specified function, with its incoming arguments bound to the specified
6312 values. Upon a '``ret``' instruction in the called function, control
6313 flow continues with the instruction after the function call, and the
6314 return value of the function is bound to the result argument.
6319 .. code-block:: llvm
6321 %retval = call i32 @test(i32 %argc)
6322 call i32 (i8*, ...)* @printf(i8* %msg, i32 12, i8 42) ; yields i32
6323 %X = tail call i32 @foo() ; yields i32
6324 %Y = tail call fastcc i32 @foo() ; yields i32
6325 call void %foo(i8 97 signext)
6327 %struct.A = type { i32, i8 }
6328 %r = call %struct.A @foo() ; yields { 32, i8 }
6329 %gr = extractvalue %struct.A %r, 0 ; yields i32
6330 %gr1 = extractvalue %struct.A %r, 1 ; yields i8
6331 %Z = call void @foo() noreturn ; indicates that %foo never returns normally
6332 %ZZ = call zeroext i32 @bar() ; Return value is %zero extended
6334 llvm treats calls to some functions with names and arguments that match
6335 the standard C99 library as being the C99 library functions, and may
6336 perform optimizations or generate code for them under that assumption.
6337 This is something we'd like to change in the future to provide better
6338 support for freestanding environments and non-C-based languages.
6342 '``va_arg``' Instruction
6343 ^^^^^^^^^^^^^^^^^^^^^^^^
6350 <resultval> = va_arg <va_list*> <arglist>, <argty>
6355 The '``va_arg``' instruction is used to access arguments passed through
6356 the "variable argument" area of a function call. It is used to implement
6357 the ``va_arg`` macro in C.
6362 This instruction takes a ``va_list*`` value and the type of the
6363 argument. It returns a value of the specified argument type and
6364 increments the ``va_list`` to point to the next argument. The actual
6365 type of ``va_list`` is target specific.
6370 The '``va_arg``' instruction loads an argument of the specified type
6371 from the specified ``va_list`` and causes the ``va_list`` to point to
6372 the next argument. For more information, see the variable argument
6373 handling :ref:`Intrinsic Functions <int_varargs>`.
6375 It is legal for this instruction to be called in a function which does
6376 not take a variable number of arguments, for example, the ``vfprintf``
6379 ``va_arg`` is an LLVM instruction instead of an :ref:`intrinsic
6380 function <intrinsics>` because it takes a type as an argument.
6385 See the :ref:`variable argument processing <int_varargs>` section.
6387 Note that the code generator does not yet fully support va\_arg on many
6388 targets. Also, it does not currently support va\_arg with aggregate
6389 types on any target.
6393 '``landingpad``' Instruction
6394 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6401 <resultval> = landingpad <resultty> personality <type> <pers_fn> <clause>+
6402 <resultval> = landingpad <resultty> personality <type> <pers_fn> cleanup <clause>*
6404 <clause> := catch <type> <value>
6405 <clause> := filter <array constant type> <array constant>
6410 The '``landingpad``' instruction is used by `LLVM's exception handling
6411 system <ExceptionHandling.html#overview>`_ to specify that a basic block
6412 is a landing pad --- one where the exception lands, and corresponds to the
6413 code found in the ``catch`` portion of a ``try``/``catch`` sequence. It
6414 defines values supplied by the personality function (``pers_fn``) upon
6415 re-entry to the function. The ``resultval`` has the type ``resultty``.
6420 This instruction takes a ``pers_fn`` value. This is the personality
6421 function associated with the unwinding mechanism. The optional
6422 ``cleanup`` flag indicates that the landing pad block is a cleanup.
6424 A ``clause`` begins with the clause type --- ``catch`` or ``filter`` --- and
6425 contains the global variable representing the "type" that may be caught
6426 or filtered respectively. Unlike the ``catch`` clause, the ``filter``
6427 clause takes an array constant as its argument. Use
6428 "``[0 x i8**] undef``" for a filter which cannot throw. The
6429 '``landingpad``' instruction must contain *at least* one ``clause`` or
6430 the ``cleanup`` flag.
6435 The '``landingpad``' instruction defines the values which are set by the
6436 personality function (``pers_fn``) upon re-entry to the function, and
6437 therefore the "result type" of the ``landingpad`` instruction. As with
6438 calling conventions, how the personality function results are
6439 represented in LLVM IR is target specific.
6441 The clauses are applied in order from top to bottom. If two
6442 ``landingpad`` instructions are merged together through inlining, the
6443 clauses from the calling function are appended to the list of clauses.
6444 When the call stack is being unwound due to an exception being thrown,
6445 the exception is compared against each ``clause`` in turn. If it doesn't
6446 match any of the clauses, and the ``cleanup`` flag is not set, then
6447 unwinding continues further up the call stack.
6449 The ``landingpad`` instruction has several restrictions:
6451 - A landing pad block is a basic block which is the unwind destination
6452 of an '``invoke``' instruction.
6453 - A landing pad block must have a '``landingpad``' instruction as its
6454 first non-PHI instruction.
6455 - There can be only one '``landingpad``' instruction within the landing
6457 - A basic block that is not a landing pad block may not include a
6458 '``landingpad``' instruction.
6459 - All '``landingpad``' instructions in a function must have the same
6460 personality function.
6465 .. code-block:: llvm
6467 ;; A landing pad which can catch an integer.
6468 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6470 ;; A landing pad that is a cleanup.
6471 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6473 ;; A landing pad which can catch an integer and can only throw a double.
6474 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6476 filter [1 x i8**] [@_ZTId]
6483 LLVM supports the notion of an "intrinsic function". These functions
6484 have well known names and semantics and are required to follow certain
6485 restrictions. Overall, these intrinsics represent an extension mechanism
6486 for the LLVM language that does not require changing all of the
6487 transformations in LLVM when adding to the language (or the bitcode
6488 reader/writer, the parser, etc...).
6490 Intrinsic function names must all start with an "``llvm.``" prefix. This
6491 prefix is reserved in LLVM for intrinsic names; thus, function names may
6492 not begin with this prefix. Intrinsic functions must always be external
6493 functions: you cannot define the body of intrinsic functions. Intrinsic
6494 functions may only be used in call or invoke instructions: it is illegal
6495 to take the address of an intrinsic function. Additionally, because
6496 intrinsic functions are part of the LLVM language, it is required if any
6497 are added that they be documented here.
6499 Some intrinsic functions can be overloaded, i.e., the intrinsic
6500 represents a family of functions that perform the same operation but on
6501 different data types. Because LLVM can represent over 8 million
6502 different integer types, overloading is used commonly to allow an
6503 intrinsic function to operate on any integer type. One or more of the
6504 argument types or the result type can be overloaded to accept any
6505 integer type. Argument types may also be defined as exactly matching a
6506 previous argument's type or the result type. This allows an intrinsic
6507 function which accepts multiple arguments, but needs all of them to be
6508 of the same type, to only be overloaded with respect to a single
6509 argument or the result.
6511 Overloaded intrinsics will have the names of its overloaded argument
6512 types encoded into its function name, each preceded by a period. Only
6513 those types which are overloaded result in a name suffix. Arguments
6514 whose type is matched against another type do not. For example, the
6515 ``llvm.ctpop`` function can take an integer of any width and returns an
6516 integer of exactly the same integer width. This leads to a family of
6517 functions such as ``i8 @llvm.ctpop.i8(i8 %val)`` and
6518 ``i29 @llvm.ctpop.i29(i29 %val)``. Only one type, the return type, is
6519 overloaded, and only one type suffix is required. Because the argument's
6520 type is matched against the return type, it does not require its own
6523 To learn how to add an intrinsic function, please see the `Extending
6524 LLVM Guide <ExtendingLLVM.html>`_.
6528 Variable Argument Handling Intrinsics
6529 -------------------------------------
6531 Variable argument support is defined in LLVM with the
6532 :ref:`va_arg <i_va_arg>` instruction and these three intrinsic
6533 functions. These functions are related to the similarly named macros
6534 defined in the ``<stdarg.h>`` header file.
6536 All of these functions operate on arguments that use a target-specific
6537 value type "``va_list``". The LLVM assembly language reference manual
6538 does not define what this type is, so all transformations should be
6539 prepared to handle these functions regardless of the type used.
6541 This example shows how the :ref:`va_arg <i_va_arg>` instruction and the
6542 variable argument handling intrinsic functions are used.
6544 .. code-block:: llvm
6546 define i32 @test(i32 %X, ...) {
6547 ; Initialize variable argument processing
6549 %ap2 = bitcast i8** %ap to i8*
6550 call void @llvm.va_start(i8* %ap2)
6552 ; Read a single integer argument
6553 %tmp = va_arg i8** %ap, i32
6555 ; Demonstrate usage of llvm.va_copy and llvm.va_end
6557 %aq2 = bitcast i8** %aq to i8*
6558 call void @llvm.va_copy(i8* %aq2, i8* %ap2)
6559 call void @llvm.va_end(i8* %aq2)
6561 ; Stop processing of arguments.
6562 call void @llvm.va_end(i8* %ap2)
6566 declare void @llvm.va_start(i8*)
6567 declare void @llvm.va_copy(i8*, i8*)
6568 declare void @llvm.va_end(i8*)
6572 '``llvm.va_start``' Intrinsic
6573 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6580 declare void @llvm.va_start(i8* <arglist>)
6585 The '``llvm.va_start``' intrinsic initializes ``*<arglist>`` for
6586 subsequent use by ``va_arg``.
6591 The argument is a pointer to a ``va_list`` element to initialize.
6596 The '``llvm.va_start``' intrinsic works just like the ``va_start`` macro
6597 available in C. In a target-dependent way, it initializes the
6598 ``va_list`` element to which the argument points, so that the next call
6599 to ``va_arg`` will produce the first variable argument passed to the
6600 function. Unlike the C ``va_start`` macro, this intrinsic does not need
6601 to know the last argument of the function as the compiler can figure
6604 '``llvm.va_end``' Intrinsic
6605 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6612 declare void @llvm.va_end(i8* <arglist>)
6617 The '``llvm.va_end``' intrinsic destroys ``*<arglist>``, which has been
6618 initialized previously with ``llvm.va_start`` or ``llvm.va_copy``.
6623 The argument is a pointer to a ``va_list`` to destroy.
6628 The '``llvm.va_end``' intrinsic works just like the ``va_end`` macro
6629 available in C. In a target-dependent way, it destroys the ``va_list``
6630 element to which the argument points. Calls to
6631 :ref:`llvm.va_start <int_va_start>` and
6632 :ref:`llvm.va_copy <int_va_copy>` must be matched exactly with calls to
6637 '``llvm.va_copy``' Intrinsic
6638 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6645 declare void @llvm.va_copy(i8* <destarglist>, i8* <srcarglist>)
6650 The '``llvm.va_copy``' intrinsic copies the current argument position
6651 from the source argument list to the destination argument list.
6656 The first argument is a pointer to a ``va_list`` element to initialize.
6657 The second argument is a pointer to a ``va_list`` element to copy from.
6662 The '``llvm.va_copy``' intrinsic works just like the ``va_copy`` macro
6663 available in C. In a target-dependent way, it copies the source
6664 ``va_list`` element into the destination ``va_list`` element. This
6665 intrinsic is necessary because the `` llvm.va_start`` intrinsic may be
6666 arbitrarily complex and require, for example, memory allocation.
6668 Accurate Garbage Collection Intrinsics
6669 --------------------------------------
6671 LLVM support for `Accurate Garbage Collection <GarbageCollection.html>`_
6672 (GC) requires the implementation and generation of these intrinsics.
6673 These intrinsics allow identification of :ref:`GC roots on the
6674 stack <int_gcroot>`, as well as garbage collector implementations that
6675 require :ref:`read <int_gcread>` and :ref:`write <int_gcwrite>` barriers.
6676 Front-ends for type-safe garbage collected languages should generate
6677 these intrinsics to make use of the LLVM garbage collectors. For more
6678 details, see `Accurate Garbage Collection with
6679 LLVM <GarbageCollection.html>`_.
6681 The garbage collection intrinsics only operate on objects in the generic
6682 address space (address space zero).
6686 '``llvm.gcroot``' Intrinsic
6687 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6694 declare void @llvm.gcroot(i8** %ptrloc, i8* %metadata)
6699 The '``llvm.gcroot``' intrinsic declares the existence of a GC root to
6700 the code generator, and allows some metadata to be associated with it.
6705 The first argument specifies the address of a stack object that contains
6706 the root pointer. The second pointer (which must be either a constant or
6707 a global value address) contains the meta-data to be associated with the
6713 At runtime, a call to this intrinsic stores a null pointer into the
6714 "ptrloc" location. At compile-time, the code generator generates
6715 information to allow the runtime to find the pointer at GC safe points.
6716 The '``llvm.gcroot``' intrinsic may only be used in a function which
6717 :ref:`specifies a GC algorithm <gc>`.
6721 '``llvm.gcread``' Intrinsic
6722 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6729 declare i8* @llvm.gcread(i8* %ObjPtr, i8** %Ptr)
6734 The '``llvm.gcread``' intrinsic identifies reads of references from heap
6735 locations, allowing garbage collector implementations that require read
6741 The second argument is the address to read from, which should be an
6742 address allocated from the garbage collector. The first object is a
6743 pointer to the start of the referenced object, if needed by the language
6744 runtime (otherwise null).
6749 The '``llvm.gcread``' intrinsic has the same semantics as a load
6750 instruction, but may be replaced with substantially more complex code by
6751 the garbage collector runtime, as needed. The '``llvm.gcread``'
6752 intrinsic may only be used in a function which :ref:`specifies a GC
6757 '``llvm.gcwrite``' Intrinsic
6758 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6765 declare void @llvm.gcwrite(i8* %P1, i8* %Obj, i8** %P2)
6770 The '``llvm.gcwrite``' intrinsic identifies writes of references to heap
6771 locations, allowing garbage collector implementations that require write
6772 barriers (such as generational or reference counting collectors).
6777 The first argument is the reference to store, the second is the start of
6778 the object to store it to, and the third is the address of the field of
6779 Obj to store to. If the runtime does not require a pointer to the
6780 object, Obj may be null.
6785 The '``llvm.gcwrite``' intrinsic has the same semantics as a store
6786 instruction, but may be replaced with substantially more complex code by
6787 the garbage collector runtime, as needed. The '``llvm.gcwrite``'
6788 intrinsic may only be used in a function which :ref:`specifies a GC
6791 Code Generator Intrinsics
6792 -------------------------
6794 These intrinsics are provided by LLVM to expose special features that
6795 may only be implemented with code generator support.
6797 '``llvm.returnaddress``' Intrinsic
6798 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6805 declare i8 *@llvm.returnaddress(i32 <level>)
6810 The '``llvm.returnaddress``' intrinsic attempts to compute a
6811 target-specific value indicating the return address of the current
6812 function or one of its callers.
6817 The argument to this intrinsic indicates which function to return the
6818 address for. Zero indicates the calling function, one indicates its
6819 caller, etc. The argument is **required** to be a constant integer
6825 The '``llvm.returnaddress``' intrinsic either returns a pointer
6826 indicating the return address of the specified call frame, or zero if it
6827 cannot be identified. The value returned by this intrinsic is likely to
6828 be incorrect or 0 for arguments other than zero, so it should only be
6829 used for debugging purposes.
6831 Note that calling this intrinsic does not prevent function inlining or
6832 other aggressive transformations, so the value returned may not be that
6833 of the obvious source-language caller.
6835 '``llvm.frameaddress``' Intrinsic
6836 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6843 declare i8* @llvm.frameaddress(i32 <level>)
6848 The '``llvm.frameaddress``' intrinsic attempts to return the
6849 target-specific frame pointer value for the specified stack frame.
6854 The argument to this intrinsic indicates which function to return the
6855 frame pointer for. Zero indicates the calling function, one indicates
6856 its caller, etc. The argument is **required** to be a constant integer
6862 The '``llvm.frameaddress``' intrinsic either returns a pointer
6863 indicating the frame address of the specified call frame, or zero if it
6864 cannot be identified. The value returned by this intrinsic is likely to
6865 be incorrect or 0 for arguments other than zero, so it should only be
6866 used for debugging purposes.
6868 Note that calling this intrinsic does not prevent function inlining or
6869 other aggressive transformations, so the value returned may not be that
6870 of the obvious source-language caller.
6872 .. _int_read_register:
6873 .. _int_write_register:
6875 '``llvm.read_register``' and '``llvm.write_register``' Intrinsics
6876 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6883 declare i32 @llvm.read_register.i32(metadata)
6884 declare i64 @llvm.read_register.i64(metadata)
6885 declare void @llvm.write_register.i32(metadata, i32 @value)
6886 declare void @llvm.write_register.i64(metadata, i64 @value)
6887 !0 = metadata !{metadata !"sp\00"}
6892 The '``llvm.read_register``' and '``llvm.write_register``' intrinsics
6893 provides access to the named register. The register must be valid on
6894 the architecture being compiled to. The type needs to be compatible
6895 with the register being read.
6900 The '``llvm.read_register``' intrinsic returns the current value of the
6901 register, where possible. The '``llvm.write_register``' intrinsic sets
6902 the current value of the register, where possible.
6904 This is useful to implement named register global variables that need
6905 to always be mapped to a specific register, as is common practice on
6906 bare-metal programs including OS kernels.
6908 The compiler doesn't check for register availability or use of the used
6909 register in surrounding code, including inline assembly. Because of that,
6910 allocatable registers are not supported.
6912 Warning: So far it only works with the stack pointer on selected
6913 architectures (ARM, AArch64, PowerPC and x86_64). Significant amount of
6914 work is needed to support other registers and even more so, allocatable
6919 '``llvm.stacksave``' Intrinsic
6920 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6927 declare i8* @llvm.stacksave()
6932 The '``llvm.stacksave``' intrinsic is used to remember the current state
6933 of the function stack, for use with
6934 :ref:`llvm.stackrestore <int_stackrestore>`. This is useful for
6935 implementing language features like scoped automatic variable sized
6941 This intrinsic returns a opaque pointer value that can be passed to
6942 :ref:`llvm.stackrestore <int_stackrestore>`. When an
6943 ``llvm.stackrestore`` intrinsic is executed with a value saved from
6944 ``llvm.stacksave``, it effectively restores the state of the stack to
6945 the state it was in when the ``llvm.stacksave`` intrinsic executed. In
6946 practice, this pops any :ref:`alloca <i_alloca>` blocks from the stack that
6947 were allocated after the ``llvm.stacksave`` was executed.
6949 .. _int_stackrestore:
6951 '``llvm.stackrestore``' Intrinsic
6952 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6959 declare void @llvm.stackrestore(i8* %ptr)
6964 The '``llvm.stackrestore``' intrinsic is used to restore the state of
6965 the function stack to the state it was in when the corresponding
6966 :ref:`llvm.stacksave <int_stacksave>` intrinsic executed. This is
6967 useful for implementing language features like scoped automatic variable
6968 sized arrays in C99.
6973 See the description for :ref:`llvm.stacksave <int_stacksave>`.
6975 '``llvm.prefetch``' Intrinsic
6976 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6983 declare void @llvm.prefetch(i8* <address>, i32 <rw>, i32 <locality>, i32 <cache type>)
6988 The '``llvm.prefetch``' intrinsic is a hint to the code generator to
6989 insert a prefetch instruction if supported; otherwise, it is a noop.
6990 Prefetches have no effect on the behavior of the program but can change
6991 its performance characteristics.
6996 ``address`` is the address to be prefetched, ``rw`` is the specifier
6997 determining if the fetch should be for a read (0) or write (1), and
6998 ``locality`` is a temporal locality specifier ranging from (0) - no
6999 locality, to (3) - extremely local keep in cache. The ``cache type``
7000 specifies whether the prefetch is performed on the data (1) or
7001 instruction (0) cache. The ``rw``, ``locality`` and ``cache type``
7002 arguments must be constant integers.
7007 This intrinsic does not modify the behavior of the program. In
7008 particular, prefetches cannot trap and do not produce a value. On
7009 targets that support this intrinsic, the prefetch can provide hints to
7010 the processor cache for better performance.
7012 '``llvm.pcmarker``' Intrinsic
7013 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7020 declare void @llvm.pcmarker(i32 <id>)
7025 The '``llvm.pcmarker``' intrinsic is a method to export a Program
7026 Counter (PC) in a region of code to simulators and other tools. The
7027 method is target specific, but it is expected that the marker will use
7028 exported symbols to transmit the PC of the marker. The marker makes no
7029 guarantees that it will remain with any specific instruction after
7030 optimizations. It is possible that the presence of a marker will inhibit
7031 optimizations. The intended use is to be inserted after optimizations to
7032 allow correlations of simulation runs.
7037 ``id`` is a numerical id identifying the marker.
7042 This intrinsic does not modify the behavior of the program. Backends
7043 that do not support this intrinsic may ignore it.
7045 '``llvm.readcyclecounter``' Intrinsic
7046 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7053 declare i64 @llvm.readcyclecounter()
7058 The '``llvm.readcyclecounter``' intrinsic provides access to the cycle
7059 counter register (or similar low latency, high accuracy clocks) on those
7060 targets that support it. On X86, it should map to RDTSC. On Alpha, it
7061 should map to RPCC. As the backing counters overflow quickly (on the
7062 order of 9 seconds on alpha), this should only be used for small
7068 When directly supported, reading the cycle counter should not modify any
7069 memory. Implementations are allowed to either return a application
7070 specific value or a system wide value. On backends without support, this
7071 is lowered to a constant 0.
7073 Note that runtime support may be conditional on the privilege-level code is
7074 running at and the host platform.
7076 '``llvm.clear_cache``' Intrinsic
7077 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7084 declare void @llvm.clear_cache(i8*, i8*)
7089 The '``llvm.clear_cache``' intrinsic ensures visibility of modifications
7090 in the specified range to the execution unit of the processor. On
7091 targets with non-unified instruction and data cache, the implementation
7092 flushes the instruction cache.
7097 On platforms with coherent instruction and data caches (e.g. x86), this
7098 intrinsic is a nop. On platforms with non-coherent instruction and data
7099 cache (e.g. ARM, MIPS), the intrinsic is lowered either to appropriate
7100 instructions or a system call, if cache flushing requires special
7103 The default behavior is to emit a call to ``__clear_cache`` from the run
7106 This instrinsic does *not* empty the instruction pipeline. Modifications
7107 of the current function are outside the scope of the intrinsic.
7109 Standard C Library Intrinsics
7110 -----------------------------
7112 LLVM provides intrinsics for a few important standard C library
7113 functions. These intrinsics allow source-language front-ends to pass
7114 information about the alignment of the pointer arguments to the code
7115 generator, providing opportunity for more efficient code generation.
7119 '``llvm.memcpy``' Intrinsic
7120 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7125 This is an overloaded intrinsic. You can use ``llvm.memcpy`` on any
7126 integer bit width and for different address spaces. Not all targets
7127 support all bit widths however.
7131 declare void @llvm.memcpy.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
7132 i32 <len>, i32 <align>, i1 <isvolatile>)
7133 declare void @llvm.memcpy.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
7134 i64 <len>, i32 <align>, i1 <isvolatile>)
7139 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
7140 source location to the destination location.
7142 Note that, unlike the standard libc function, the ``llvm.memcpy.*``
7143 intrinsics do not return a value, takes extra alignment/isvolatile
7144 arguments and the pointers can be in specified address spaces.
7149 The first argument is a pointer to the destination, the second is a
7150 pointer to the source. The third argument is an integer argument
7151 specifying the number of bytes to copy, the fourth argument is the
7152 alignment of the source and destination locations, and the fifth is a
7153 boolean indicating a volatile access.
7155 If the call to this intrinsic has an alignment value that is not 0 or 1,
7156 then the caller guarantees that both the source and destination pointers
7157 are aligned to that boundary.
7159 If the ``isvolatile`` parameter is ``true``, the ``llvm.memcpy`` call is
7160 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
7161 very cleanly specified and it is unwise to depend on it.
7166 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
7167 source location to the destination location, which are not allowed to
7168 overlap. It copies "len" bytes of memory over. If the argument is known
7169 to be aligned to some boundary, this can be specified as the fourth
7170 argument, otherwise it should be set to 0 or 1 (both meaning no alignment).
7172 '``llvm.memmove``' Intrinsic
7173 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7178 This is an overloaded intrinsic. You can use llvm.memmove on any integer
7179 bit width and for different address space. Not all targets support all
7184 declare void @llvm.memmove.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
7185 i32 <len>, i32 <align>, i1 <isvolatile>)
7186 declare void @llvm.memmove.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
7187 i64 <len>, i32 <align>, i1 <isvolatile>)
7192 The '``llvm.memmove.*``' intrinsics move a block of memory from the
7193 source location to the destination location. It is similar to the
7194 '``llvm.memcpy``' intrinsic but allows the two memory locations to
7197 Note that, unlike the standard libc function, the ``llvm.memmove.*``
7198 intrinsics do not return a value, takes extra alignment/isvolatile
7199 arguments and the pointers can be in specified address spaces.
7204 The first argument is a pointer to the destination, the second is a
7205 pointer to the source. The third argument is an integer argument
7206 specifying the number of bytes to copy, the fourth argument is the
7207 alignment of the source and destination locations, and the fifth is a
7208 boolean indicating a volatile access.
7210 If the call to this intrinsic has an alignment value that is not 0 or 1,
7211 then the caller guarantees that the source and destination pointers are
7212 aligned to that boundary.
7214 If the ``isvolatile`` parameter is ``true``, the ``llvm.memmove`` call
7215 is a :ref:`volatile operation <volatile>`. The detailed access behavior is
7216 not very cleanly specified and it is unwise to depend on it.
7221 The '``llvm.memmove.*``' intrinsics copy a block of memory from the
7222 source location to the destination location, which may overlap. It
7223 copies "len" bytes of memory over. If the argument is known to be
7224 aligned to some boundary, this can be specified as the fourth argument,
7225 otherwise it should be set to 0 or 1 (both meaning no alignment).
7227 '``llvm.memset.*``' Intrinsics
7228 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7233 This is an overloaded intrinsic. You can use llvm.memset on any integer
7234 bit width and for different address spaces. However, not all targets
7235 support all bit widths.
7239 declare void @llvm.memset.p0i8.i32(i8* <dest>, i8 <val>,
7240 i32 <len>, i32 <align>, i1 <isvolatile>)
7241 declare void @llvm.memset.p0i8.i64(i8* <dest>, i8 <val>,
7242 i64 <len>, i32 <align>, i1 <isvolatile>)
7247 The '``llvm.memset.*``' intrinsics fill a block of memory with a
7248 particular byte value.
7250 Note that, unlike the standard libc function, the ``llvm.memset``
7251 intrinsic does not return a value and takes extra alignment/volatile
7252 arguments. Also, the destination can be in an arbitrary address space.
7257 The first argument is a pointer to the destination to fill, the second
7258 is the byte value with which to fill it, the third argument is an
7259 integer argument specifying the number of bytes to fill, and the fourth
7260 argument is the known alignment of the destination location.
7262 If the call to this intrinsic has an alignment value that is not 0 or 1,
7263 then the caller guarantees that the destination pointer is aligned to
7266 If the ``isvolatile`` parameter is ``true``, the ``llvm.memset`` call is
7267 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
7268 very cleanly specified and it is unwise to depend on it.
7273 The '``llvm.memset.*``' intrinsics fill "len" bytes of memory starting
7274 at the destination location. If the argument is known to be aligned to
7275 some boundary, this can be specified as the fourth argument, otherwise
7276 it should be set to 0 or 1 (both meaning no alignment).
7278 '``llvm.sqrt.*``' Intrinsic
7279 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7284 This is an overloaded intrinsic. You can use ``llvm.sqrt`` on any
7285 floating point or vector of floating point type. Not all targets support
7290 declare float @llvm.sqrt.f32(float %Val)
7291 declare double @llvm.sqrt.f64(double %Val)
7292 declare x86_fp80 @llvm.sqrt.f80(x86_fp80 %Val)
7293 declare fp128 @llvm.sqrt.f128(fp128 %Val)
7294 declare ppc_fp128 @llvm.sqrt.ppcf128(ppc_fp128 %Val)
7299 The '``llvm.sqrt``' intrinsics return the sqrt of the specified operand,
7300 returning the same value as the libm '``sqrt``' functions would. Unlike
7301 ``sqrt`` in libm, however, ``llvm.sqrt`` has undefined behavior for
7302 negative numbers other than -0.0 (which allows for better optimization,
7303 because there is no need to worry about errno being set).
7304 ``llvm.sqrt(-0.0)`` is defined to return -0.0 like IEEE sqrt.
7309 The argument and return value are floating point numbers of the same
7315 This function returns the sqrt of the specified operand if it is a
7316 nonnegative floating point number.
7318 '``llvm.powi.*``' Intrinsic
7319 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7324 This is an overloaded intrinsic. You can use ``llvm.powi`` on any
7325 floating point or vector of floating point type. Not all targets support
7330 declare float @llvm.powi.f32(float %Val, i32 %power)
7331 declare double @llvm.powi.f64(double %Val, i32 %power)
7332 declare x86_fp80 @llvm.powi.f80(x86_fp80 %Val, i32 %power)
7333 declare fp128 @llvm.powi.f128(fp128 %Val, i32 %power)
7334 declare ppc_fp128 @llvm.powi.ppcf128(ppc_fp128 %Val, i32 %power)
7339 The '``llvm.powi.*``' intrinsics return the first operand raised to the
7340 specified (positive or negative) power. The order of evaluation of
7341 multiplications is not defined. When a vector of floating point type is
7342 used, the second argument remains a scalar integer value.
7347 The second argument is an integer power, and the first is a value to
7348 raise to that power.
7353 This function returns the first value raised to the second power with an
7354 unspecified sequence of rounding operations.
7356 '``llvm.sin.*``' Intrinsic
7357 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7362 This is an overloaded intrinsic. You can use ``llvm.sin`` on any
7363 floating point or vector of floating point type. Not all targets support
7368 declare float @llvm.sin.f32(float %Val)
7369 declare double @llvm.sin.f64(double %Val)
7370 declare x86_fp80 @llvm.sin.f80(x86_fp80 %Val)
7371 declare fp128 @llvm.sin.f128(fp128 %Val)
7372 declare ppc_fp128 @llvm.sin.ppcf128(ppc_fp128 %Val)
7377 The '``llvm.sin.*``' intrinsics return the sine of the operand.
7382 The argument and return value are floating point numbers of the same
7388 This function returns the sine of the specified operand, returning the
7389 same values as the libm ``sin`` functions would, and handles error
7390 conditions in the same way.
7392 '``llvm.cos.*``' Intrinsic
7393 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7398 This is an overloaded intrinsic. You can use ``llvm.cos`` on any
7399 floating point or vector of floating point type. Not all targets support
7404 declare float @llvm.cos.f32(float %Val)
7405 declare double @llvm.cos.f64(double %Val)
7406 declare x86_fp80 @llvm.cos.f80(x86_fp80 %Val)
7407 declare fp128 @llvm.cos.f128(fp128 %Val)
7408 declare ppc_fp128 @llvm.cos.ppcf128(ppc_fp128 %Val)
7413 The '``llvm.cos.*``' intrinsics return the cosine of the operand.
7418 The argument and return value are floating point numbers of the same
7424 This function returns the cosine of the specified operand, returning the
7425 same values as the libm ``cos`` functions would, and handles error
7426 conditions in the same way.
7428 '``llvm.pow.*``' Intrinsic
7429 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7434 This is an overloaded intrinsic. You can use ``llvm.pow`` on any
7435 floating point or vector of floating point type. Not all targets support
7440 declare float @llvm.pow.f32(float %Val, float %Power)
7441 declare double @llvm.pow.f64(double %Val, double %Power)
7442 declare x86_fp80 @llvm.pow.f80(x86_fp80 %Val, x86_fp80 %Power)
7443 declare fp128 @llvm.pow.f128(fp128 %Val, fp128 %Power)
7444 declare ppc_fp128 @llvm.pow.ppcf128(ppc_fp128 %Val, ppc_fp128 Power)
7449 The '``llvm.pow.*``' intrinsics return the first operand raised to the
7450 specified (positive or negative) power.
7455 The second argument is a floating point power, and the first is a value
7456 to raise to that power.
7461 This function returns the first value raised to the second power,
7462 returning the same values as the libm ``pow`` functions would, and
7463 handles error conditions in the same way.
7465 '``llvm.exp.*``' Intrinsic
7466 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7471 This is an overloaded intrinsic. You can use ``llvm.exp`` on any
7472 floating point or vector of floating point type. Not all targets support
7477 declare float @llvm.exp.f32(float %Val)
7478 declare double @llvm.exp.f64(double %Val)
7479 declare x86_fp80 @llvm.exp.f80(x86_fp80 %Val)
7480 declare fp128 @llvm.exp.f128(fp128 %Val)
7481 declare ppc_fp128 @llvm.exp.ppcf128(ppc_fp128 %Val)
7486 The '``llvm.exp.*``' intrinsics perform the exp function.
7491 The argument and return value are floating point numbers of the same
7497 This function returns the same values as the libm ``exp`` functions
7498 would, and handles error conditions in the same way.
7500 '``llvm.exp2.*``' Intrinsic
7501 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7506 This is an overloaded intrinsic. You can use ``llvm.exp2`` on any
7507 floating point or vector of floating point type. Not all targets support
7512 declare float @llvm.exp2.f32(float %Val)
7513 declare double @llvm.exp2.f64(double %Val)
7514 declare x86_fp80 @llvm.exp2.f80(x86_fp80 %Val)
7515 declare fp128 @llvm.exp2.f128(fp128 %Val)
7516 declare ppc_fp128 @llvm.exp2.ppcf128(ppc_fp128 %Val)
7521 The '``llvm.exp2.*``' intrinsics perform the exp2 function.
7526 The argument and return value are floating point numbers of the same
7532 This function returns the same values as the libm ``exp2`` functions
7533 would, and handles error conditions in the same way.
7535 '``llvm.log.*``' Intrinsic
7536 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7541 This is an overloaded intrinsic. You can use ``llvm.log`` on any
7542 floating point or vector of floating point type. Not all targets support
7547 declare float @llvm.log.f32(float %Val)
7548 declare double @llvm.log.f64(double %Val)
7549 declare x86_fp80 @llvm.log.f80(x86_fp80 %Val)
7550 declare fp128 @llvm.log.f128(fp128 %Val)
7551 declare ppc_fp128 @llvm.log.ppcf128(ppc_fp128 %Val)
7556 The '``llvm.log.*``' intrinsics perform the log function.
7561 The argument and return value are floating point numbers of the same
7567 This function returns the same values as the libm ``log`` functions
7568 would, and handles error conditions in the same way.
7570 '``llvm.log10.*``' Intrinsic
7571 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7576 This is an overloaded intrinsic. You can use ``llvm.log10`` on any
7577 floating point or vector of floating point type. Not all targets support
7582 declare float @llvm.log10.f32(float %Val)
7583 declare double @llvm.log10.f64(double %Val)
7584 declare x86_fp80 @llvm.log10.f80(x86_fp80 %Val)
7585 declare fp128 @llvm.log10.f128(fp128 %Val)
7586 declare ppc_fp128 @llvm.log10.ppcf128(ppc_fp128 %Val)
7591 The '``llvm.log10.*``' intrinsics perform the log10 function.
7596 The argument and return value are floating point numbers of the same
7602 This function returns the same values as the libm ``log10`` functions
7603 would, and handles error conditions in the same way.
7605 '``llvm.log2.*``' Intrinsic
7606 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7611 This is an overloaded intrinsic. You can use ``llvm.log2`` on any
7612 floating point or vector of floating point type. Not all targets support
7617 declare float @llvm.log2.f32(float %Val)
7618 declare double @llvm.log2.f64(double %Val)
7619 declare x86_fp80 @llvm.log2.f80(x86_fp80 %Val)
7620 declare fp128 @llvm.log2.f128(fp128 %Val)
7621 declare ppc_fp128 @llvm.log2.ppcf128(ppc_fp128 %Val)
7626 The '``llvm.log2.*``' intrinsics perform the log2 function.
7631 The argument and return value are floating point numbers of the same
7637 This function returns the same values as the libm ``log2`` functions
7638 would, and handles error conditions in the same way.
7640 '``llvm.fma.*``' Intrinsic
7641 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7646 This is an overloaded intrinsic. You can use ``llvm.fma`` on any
7647 floating point or vector of floating point type. Not all targets support
7652 declare float @llvm.fma.f32(float %a, float %b, float %c)
7653 declare double @llvm.fma.f64(double %a, double %b, double %c)
7654 declare x86_fp80 @llvm.fma.f80(x86_fp80 %a, x86_fp80 %b, x86_fp80 %c)
7655 declare fp128 @llvm.fma.f128(fp128 %a, fp128 %b, fp128 %c)
7656 declare ppc_fp128 @llvm.fma.ppcf128(ppc_fp128 %a, ppc_fp128 %b, ppc_fp128 %c)
7661 The '``llvm.fma.*``' intrinsics perform the fused multiply-add
7667 The argument and return value are floating point numbers of the same
7673 This function returns the same values as the libm ``fma`` functions
7674 would, and does not set errno.
7676 '``llvm.fabs.*``' Intrinsic
7677 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7682 This is an overloaded intrinsic. You can use ``llvm.fabs`` on any
7683 floating point or vector of floating point type. Not all targets support
7688 declare float @llvm.fabs.f32(float %Val)
7689 declare double @llvm.fabs.f64(double %Val)
7690 declare x86_fp80 @llvm.fabs.f80(x86_fp80 %Val)
7691 declare fp128 @llvm.fabs.f128(fp128 %Val)
7692 declare ppc_fp128 @llvm.fabs.ppcf128(ppc_fp128 %Val)
7697 The '``llvm.fabs.*``' intrinsics return the absolute value of the
7703 The argument and return value are floating point numbers of the same
7709 This function returns the same values as the libm ``fabs`` functions
7710 would, and handles error conditions in the same way.
7712 '``llvm.copysign.*``' Intrinsic
7713 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7718 This is an overloaded intrinsic. You can use ``llvm.copysign`` on any
7719 floating point or vector of floating point type. Not all targets support
7724 declare float @llvm.copysign.f32(float %Mag, float %Sgn)
7725 declare double @llvm.copysign.f64(double %Mag, double %Sgn)
7726 declare x86_fp80 @llvm.copysign.f80(x86_fp80 %Mag, x86_fp80 %Sgn)
7727 declare fp128 @llvm.copysign.f128(fp128 %Mag, fp128 %Sgn)
7728 declare ppc_fp128 @llvm.copysign.ppcf128(ppc_fp128 %Mag, ppc_fp128 %Sgn)
7733 The '``llvm.copysign.*``' intrinsics return a value with the magnitude of the
7734 first operand and the sign of the second operand.
7739 The arguments and return value are floating point numbers of the same
7745 This function returns the same values as the libm ``copysign``
7746 functions would, and handles error conditions in the same way.
7748 '``llvm.floor.*``' Intrinsic
7749 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7754 This is an overloaded intrinsic. You can use ``llvm.floor`` on any
7755 floating point or vector of floating point type. Not all targets support
7760 declare float @llvm.floor.f32(float %Val)
7761 declare double @llvm.floor.f64(double %Val)
7762 declare x86_fp80 @llvm.floor.f80(x86_fp80 %Val)
7763 declare fp128 @llvm.floor.f128(fp128 %Val)
7764 declare ppc_fp128 @llvm.floor.ppcf128(ppc_fp128 %Val)
7769 The '``llvm.floor.*``' intrinsics return the floor of the operand.
7774 The argument and return value are floating point numbers of the same
7780 This function returns the same values as the libm ``floor`` functions
7781 would, and handles error conditions in the same way.
7783 '``llvm.ceil.*``' Intrinsic
7784 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7789 This is an overloaded intrinsic. You can use ``llvm.ceil`` on any
7790 floating point or vector of floating point type. Not all targets support
7795 declare float @llvm.ceil.f32(float %Val)
7796 declare double @llvm.ceil.f64(double %Val)
7797 declare x86_fp80 @llvm.ceil.f80(x86_fp80 %Val)
7798 declare fp128 @llvm.ceil.f128(fp128 %Val)
7799 declare ppc_fp128 @llvm.ceil.ppcf128(ppc_fp128 %Val)
7804 The '``llvm.ceil.*``' intrinsics return the ceiling of the operand.
7809 The argument and return value are floating point numbers of the same
7815 This function returns the same values as the libm ``ceil`` functions
7816 would, and handles error conditions in the same way.
7818 '``llvm.trunc.*``' Intrinsic
7819 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7824 This is an overloaded intrinsic. You can use ``llvm.trunc`` on any
7825 floating point or vector of floating point type. Not all targets support
7830 declare float @llvm.trunc.f32(float %Val)
7831 declare double @llvm.trunc.f64(double %Val)
7832 declare x86_fp80 @llvm.trunc.f80(x86_fp80 %Val)
7833 declare fp128 @llvm.trunc.f128(fp128 %Val)
7834 declare ppc_fp128 @llvm.trunc.ppcf128(ppc_fp128 %Val)
7839 The '``llvm.trunc.*``' intrinsics returns the operand rounded to the
7840 nearest integer not larger in magnitude than the operand.
7845 The argument and return value are floating point numbers of the same
7851 This function returns the same values as the libm ``trunc`` functions
7852 would, and handles error conditions in the same way.
7854 '``llvm.rint.*``' Intrinsic
7855 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7860 This is an overloaded intrinsic. You can use ``llvm.rint`` on any
7861 floating point or vector of floating point type. Not all targets support
7866 declare float @llvm.rint.f32(float %Val)
7867 declare double @llvm.rint.f64(double %Val)
7868 declare x86_fp80 @llvm.rint.f80(x86_fp80 %Val)
7869 declare fp128 @llvm.rint.f128(fp128 %Val)
7870 declare ppc_fp128 @llvm.rint.ppcf128(ppc_fp128 %Val)
7875 The '``llvm.rint.*``' intrinsics returns the operand rounded to the
7876 nearest integer. It may raise an inexact floating-point exception if the
7877 operand isn't an integer.
7882 The argument and return value are floating point numbers of the same
7888 This function returns the same values as the libm ``rint`` functions
7889 would, and handles error conditions in the same way.
7891 '``llvm.nearbyint.*``' Intrinsic
7892 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7897 This is an overloaded intrinsic. You can use ``llvm.nearbyint`` on any
7898 floating point or vector of floating point type. Not all targets support
7903 declare float @llvm.nearbyint.f32(float %Val)
7904 declare double @llvm.nearbyint.f64(double %Val)
7905 declare x86_fp80 @llvm.nearbyint.f80(x86_fp80 %Val)
7906 declare fp128 @llvm.nearbyint.f128(fp128 %Val)
7907 declare ppc_fp128 @llvm.nearbyint.ppcf128(ppc_fp128 %Val)
7912 The '``llvm.nearbyint.*``' intrinsics returns the operand rounded to the
7918 The argument and return value are floating point numbers of the same
7924 This function returns the same values as the libm ``nearbyint``
7925 functions would, and handles error conditions in the same way.
7927 '``llvm.round.*``' Intrinsic
7928 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7933 This is an overloaded intrinsic. You can use ``llvm.round`` on any
7934 floating point or vector of floating point type. Not all targets support
7939 declare float @llvm.round.f32(float %Val)
7940 declare double @llvm.round.f64(double %Val)
7941 declare x86_fp80 @llvm.round.f80(x86_fp80 %Val)
7942 declare fp128 @llvm.round.f128(fp128 %Val)
7943 declare ppc_fp128 @llvm.round.ppcf128(ppc_fp128 %Val)
7948 The '``llvm.round.*``' intrinsics returns the operand rounded to the
7954 The argument and return value are floating point numbers of the same
7960 This function returns the same values as the libm ``round``
7961 functions would, and handles error conditions in the same way.
7963 Bit Manipulation Intrinsics
7964 ---------------------------
7966 LLVM provides intrinsics for a few important bit manipulation
7967 operations. These allow efficient code generation for some algorithms.
7969 '``llvm.bswap.*``' Intrinsics
7970 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7975 This is an overloaded intrinsic function. You can use bswap on any
7976 integer type that is an even number of bytes (i.e. BitWidth % 16 == 0).
7980 declare i16 @llvm.bswap.i16(i16 <id>)
7981 declare i32 @llvm.bswap.i32(i32 <id>)
7982 declare i64 @llvm.bswap.i64(i64 <id>)
7987 The '``llvm.bswap``' family of intrinsics is used to byte swap integer
7988 values with an even number of bytes (positive multiple of 16 bits).
7989 These are useful for performing operations on data that is not in the
7990 target's native byte order.
7995 The ``llvm.bswap.i16`` intrinsic returns an i16 value that has the high
7996 and low byte of the input i16 swapped. Similarly, the ``llvm.bswap.i32``
7997 intrinsic returns an i32 value that has the four bytes of the input i32
7998 swapped, so that if the input bytes are numbered 0, 1, 2, 3 then the
7999 returned i32 will have its bytes in 3, 2, 1, 0 order. The
8000 ``llvm.bswap.i48``, ``llvm.bswap.i64`` and other intrinsics extend this
8001 concept to additional even-byte lengths (6 bytes, 8 bytes and more,
8004 '``llvm.ctpop.*``' Intrinsic
8005 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8010 This is an overloaded intrinsic. You can use llvm.ctpop on any integer
8011 bit width, or on any vector with integer elements. Not all targets
8012 support all bit widths or vector types, however.
8016 declare i8 @llvm.ctpop.i8(i8 <src>)
8017 declare i16 @llvm.ctpop.i16(i16 <src>)
8018 declare i32 @llvm.ctpop.i32(i32 <src>)
8019 declare i64 @llvm.ctpop.i64(i64 <src>)
8020 declare i256 @llvm.ctpop.i256(i256 <src>)
8021 declare <2 x i32> @llvm.ctpop.v2i32(<2 x i32> <src>)
8026 The '``llvm.ctpop``' family of intrinsics counts the number of bits set
8032 The only argument is the value to be counted. The argument may be of any
8033 integer type, or a vector with integer elements. The return type must
8034 match the argument type.
8039 The '``llvm.ctpop``' intrinsic counts the 1's in a variable, or within
8040 each element of a vector.
8042 '``llvm.ctlz.*``' Intrinsic
8043 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8048 This is an overloaded intrinsic. You can use ``llvm.ctlz`` on any
8049 integer bit width, or any vector whose elements are integers. Not all
8050 targets support all bit widths or vector types, however.
8054 declare i8 @llvm.ctlz.i8 (i8 <src>, i1 <is_zero_undef>)
8055 declare i16 @llvm.ctlz.i16 (i16 <src>, i1 <is_zero_undef>)
8056 declare i32 @llvm.ctlz.i32 (i32 <src>, i1 <is_zero_undef>)
8057 declare i64 @llvm.ctlz.i64 (i64 <src>, i1 <is_zero_undef>)
8058 declare i256 @llvm.ctlz.i256(i256 <src>, i1 <is_zero_undef>)
8059 declase <2 x i32> @llvm.ctlz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
8064 The '``llvm.ctlz``' family of intrinsic functions counts the number of
8065 leading zeros in a variable.
8070 The first argument is the value to be counted. This argument may be of
8071 any integer type, or a vectory with integer element type. The return
8072 type must match the first argument type.
8074 The second argument must be a constant and is a flag to indicate whether
8075 the intrinsic should ensure that a zero as the first argument produces a
8076 defined result. Historically some architectures did not provide a
8077 defined result for zero values as efficiently, and many algorithms are
8078 now predicated on avoiding zero-value inputs.
8083 The '``llvm.ctlz``' intrinsic counts the leading (most significant)
8084 zeros in a variable, or within each element of the vector. If
8085 ``src == 0`` then the result is the size in bits of the type of ``src``
8086 if ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
8087 ``llvm.ctlz(i32 2) = 30``.
8089 '``llvm.cttz.*``' Intrinsic
8090 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8095 This is an overloaded intrinsic. You can use ``llvm.cttz`` on any
8096 integer bit width, or any vector of integer elements. Not all targets
8097 support all bit widths or vector types, however.
8101 declare i8 @llvm.cttz.i8 (i8 <src>, i1 <is_zero_undef>)
8102 declare i16 @llvm.cttz.i16 (i16 <src>, i1 <is_zero_undef>)
8103 declare i32 @llvm.cttz.i32 (i32 <src>, i1 <is_zero_undef>)
8104 declare i64 @llvm.cttz.i64 (i64 <src>, i1 <is_zero_undef>)
8105 declare i256 @llvm.cttz.i256(i256 <src>, i1 <is_zero_undef>)
8106 declase <2 x i32> @llvm.cttz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
8111 The '``llvm.cttz``' family of intrinsic functions counts the number of
8117 The first argument is the value to be counted. This argument may be of
8118 any integer type, or a vectory with integer element type. The return
8119 type must match the first argument type.
8121 The second argument must be a constant and is a flag to indicate whether
8122 the intrinsic should ensure that a zero as the first argument produces a
8123 defined result. Historically some architectures did not provide a
8124 defined result for zero values as efficiently, and many algorithms are
8125 now predicated on avoiding zero-value inputs.
8130 The '``llvm.cttz``' intrinsic counts the trailing (least significant)
8131 zeros in a variable, or within each element of a vector. If ``src == 0``
8132 then the result is the size in bits of the type of ``src`` if
8133 ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
8134 ``llvm.cttz(2) = 1``.
8136 Arithmetic with Overflow Intrinsics
8137 -----------------------------------
8139 LLVM provides intrinsics for some arithmetic with overflow operations.
8141 '``llvm.sadd.with.overflow.*``' Intrinsics
8142 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8147 This is an overloaded intrinsic. You can use ``llvm.sadd.with.overflow``
8148 on any integer bit width.
8152 declare {i16, i1} @llvm.sadd.with.overflow.i16(i16 %a, i16 %b)
8153 declare {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
8154 declare {i64, i1} @llvm.sadd.with.overflow.i64(i64 %a, i64 %b)
8159 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
8160 a signed addition of the two arguments, and indicate whether an overflow
8161 occurred during the signed summation.
8166 The arguments (%a and %b) and the first element of the result structure
8167 may be of integer types of any bit width, but they must have the same
8168 bit width. The second element of the result structure must be of type
8169 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
8175 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
8176 a signed addition of the two variables. They return a structure --- the
8177 first element of which is the signed summation, and the second element
8178 of which is a bit specifying if the signed summation resulted in an
8184 .. code-block:: llvm
8186 %res = call {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
8187 %sum = extractvalue {i32, i1} %res, 0
8188 %obit = extractvalue {i32, i1} %res, 1
8189 br i1 %obit, label %overflow, label %normal
8191 '``llvm.uadd.with.overflow.*``' Intrinsics
8192 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8197 This is an overloaded intrinsic. You can use ``llvm.uadd.with.overflow``
8198 on any integer bit width.
8202 declare {i16, i1} @llvm.uadd.with.overflow.i16(i16 %a, i16 %b)
8203 declare {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
8204 declare {i64, i1} @llvm.uadd.with.overflow.i64(i64 %a, i64 %b)
8209 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
8210 an unsigned addition of the two arguments, and indicate whether a carry
8211 occurred during the unsigned summation.
8216 The arguments (%a and %b) and the first element of the result structure
8217 may be of integer types of any bit width, but they must have the same
8218 bit width. The second element of the result structure must be of type
8219 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
8225 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
8226 an unsigned addition of the two arguments. They return a structure --- the
8227 first element of which is the sum, and the second element of which is a
8228 bit specifying if the unsigned summation resulted in a carry.
8233 .. code-block:: llvm
8235 %res = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
8236 %sum = extractvalue {i32, i1} %res, 0
8237 %obit = extractvalue {i32, i1} %res, 1
8238 br i1 %obit, label %carry, label %normal
8240 '``llvm.ssub.with.overflow.*``' Intrinsics
8241 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8246 This is an overloaded intrinsic. You can use ``llvm.ssub.with.overflow``
8247 on any integer bit width.
8251 declare {i16, i1} @llvm.ssub.with.overflow.i16(i16 %a, i16 %b)
8252 declare {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
8253 declare {i64, i1} @llvm.ssub.with.overflow.i64(i64 %a, i64 %b)
8258 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
8259 a signed subtraction of the two arguments, and indicate whether an
8260 overflow occurred during the signed subtraction.
8265 The arguments (%a and %b) and the first element of the result structure
8266 may be of integer types of any bit width, but they must have the same
8267 bit width. The second element of the result structure must be of type
8268 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
8274 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
8275 a signed subtraction of the two arguments. They return a structure --- the
8276 first element of which is the subtraction, and the second element of
8277 which is a bit specifying if the signed subtraction resulted in an
8283 .. code-block:: llvm
8285 %res = call {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
8286 %sum = extractvalue {i32, i1} %res, 0
8287 %obit = extractvalue {i32, i1} %res, 1
8288 br i1 %obit, label %overflow, label %normal
8290 '``llvm.usub.with.overflow.*``' Intrinsics
8291 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8296 This is an overloaded intrinsic. You can use ``llvm.usub.with.overflow``
8297 on any integer bit width.
8301 declare {i16, i1} @llvm.usub.with.overflow.i16(i16 %a, i16 %b)
8302 declare {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
8303 declare {i64, i1} @llvm.usub.with.overflow.i64(i64 %a, i64 %b)
8308 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
8309 an unsigned subtraction of the two arguments, and indicate whether an
8310 overflow occurred during the unsigned subtraction.
8315 The arguments (%a and %b) and the first element of the result structure
8316 may be of integer types of any bit width, but they must have the same
8317 bit width. The second element of the result structure must be of type
8318 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
8324 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
8325 an unsigned subtraction of the two arguments. They return a structure ---
8326 the first element of which is the subtraction, and the second element of
8327 which is a bit specifying if the unsigned subtraction resulted in an
8333 .. code-block:: llvm
8335 %res = call {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
8336 %sum = extractvalue {i32, i1} %res, 0
8337 %obit = extractvalue {i32, i1} %res, 1
8338 br i1 %obit, label %overflow, label %normal
8340 '``llvm.smul.with.overflow.*``' Intrinsics
8341 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8346 This is an overloaded intrinsic. You can use ``llvm.smul.with.overflow``
8347 on any integer bit width.
8351 declare {i16, i1} @llvm.smul.with.overflow.i16(i16 %a, i16 %b)
8352 declare {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
8353 declare {i64, i1} @llvm.smul.with.overflow.i64(i64 %a, i64 %b)
8358 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
8359 a signed multiplication of the two arguments, and indicate whether an
8360 overflow occurred during the signed multiplication.
8365 The arguments (%a and %b) and the first element of the result structure
8366 may be of integer types of any bit width, but they must have the same
8367 bit width. The second element of the result structure must be of type
8368 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
8374 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
8375 a signed multiplication of the two arguments. They return a structure ---
8376 the first element of which is the multiplication, and the second element
8377 of which is a bit specifying if the signed multiplication resulted in an
8383 .. code-block:: llvm
8385 %res = call {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
8386 %sum = extractvalue {i32, i1} %res, 0
8387 %obit = extractvalue {i32, i1} %res, 1
8388 br i1 %obit, label %overflow, label %normal
8390 '``llvm.umul.with.overflow.*``' Intrinsics
8391 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8396 This is an overloaded intrinsic. You can use ``llvm.umul.with.overflow``
8397 on any integer bit width.
8401 declare {i16, i1} @llvm.umul.with.overflow.i16(i16 %a, i16 %b)
8402 declare {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
8403 declare {i64, i1} @llvm.umul.with.overflow.i64(i64 %a, i64 %b)
8408 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
8409 a unsigned multiplication of the two arguments, and indicate whether an
8410 overflow occurred during the unsigned multiplication.
8415 The arguments (%a and %b) and the first element of the result structure
8416 may be of integer types of any bit width, but they must have the same
8417 bit width. The second element of the result structure must be of type
8418 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
8424 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
8425 an unsigned multiplication of the two arguments. They return a structure ---
8426 the first element of which is the multiplication, and the second
8427 element of which is a bit specifying if the unsigned multiplication
8428 resulted in an overflow.
8433 .. code-block:: llvm
8435 %res = call {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
8436 %sum = extractvalue {i32, i1} %res, 0
8437 %obit = extractvalue {i32, i1} %res, 1
8438 br i1 %obit, label %overflow, label %normal
8440 Specialised Arithmetic Intrinsics
8441 ---------------------------------
8443 '``llvm.fmuladd.*``' Intrinsic
8444 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8451 declare float @llvm.fmuladd.f32(float %a, float %b, float %c)
8452 declare double @llvm.fmuladd.f64(double %a, double %b, double %c)
8457 The '``llvm.fmuladd.*``' intrinsic functions represent multiply-add
8458 expressions that can be fused if the code generator determines that (a) the
8459 target instruction set has support for a fused operation, and (b) that the
8460 fused operation is more efficient than the equivalent, separate pair of mul
8461 and add instructions.
8466 The '``llvm.fmuladd.*``' intrinsics each take three arguments: two
8467 multiplicands, a and b, and an addend c.
8476 %0 = call float @llvm.fmuladd.f32(%a, %b, %c)
8478 is equivalent to the expression a \* b + c, except that rounding will
8479 not be performed between the multiplication and addition steps if the
8480 code generator fuses the operations. Fusion is not guaranteed, even if
8481 the target platform supports it. If a fused multiply-add is required the
8482 corresponding llvm.fma.\* intrinsic function should be used
8483 instead. This never sets errno, just as '``llvm.fma.*``'.
8488 .. code-block:: llvm
8490 %r2 = call float @llvm.fmuladd.f32(float %a, float %b, float %c) ; yields {float}:r2 = (a * b) + c
8492 Half Precision Floating Point Intrinsics
8493 ----------------------------------------
8495 For most target platforms, half precision floating point is a
8496 storage-only format. This means that it is a dense encoding (in memory)
8497 but does not support computation in the format.
8499 This means that code must first load the half-precision floating point
8500 value as an i16, then convert it to float with
8501 :ref:`llvm.convert.from.fp16 <int_convert_from_fp16>`. Computation can
8502 then be performed on the float value (including extending to double
8503 etc). To store the value back to memory, it is first converted to float
8504 if needed, then converted to i16 with
8505 :ref:`llvm.convert.to.fp16 <int_convert_to_fp16>`, then storing as an
8508 .. _int_convert_to_fp16:
8510 '``llvm.convert.to.fp16``' Intrinsic
8511 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8518 declare i16 @llvm.convert.to.fp16(f32 %a)
8523 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion
8524 from single precision floating point format to half precision floating
8530 The intrinsic function contains single argument - the value to be
8536 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion
8537 from single precision floating point format to half precision floating
8538 point format. The return value is an ``i16`` which contains the
8544 .. code-block:: llvm
8546 %res = call i16 @llvm.convert.to.fp16(f32 %a)
8547 store i16 %res, i16* @x, align 2
8549 .. _int_convert_from_fp16:
8551 '``llvm.convert.from.fp16``' Intrinsic
8552 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8559 declare f32 @llvm.convert.from.fp16(i16 %a)
8564 The '``llvm.convert.from.fp16``' intrinsic function performs a
8565 conversion from half precision floating point format to single precision
8566 floating point format.
8571 The intrinsic function contains single argument - the value to be
8577 The '``llvm.convert.from.fp16``' intrinsic function performs a
8578 conversion from half single precision floating point format to single
8579 precision floating point format. The input half-float value is
8580 represented by an ``i16`` value.
8585 .. code-block:: llvm
8587 %a = load i16* @x, align 2
8588 %res = call f32 @llvm.convert.from.fp16(i16 %a)
8593 The LLVM debugger intrinsics (which all start with ``llvm.dbg.``
8594 prefix), are described in the `LLVM Source Level
8595 Debugging <SourceLevelDebugging.html#format_common_intrinsics>`_
8598 Exception Handling Intrinsics
8599 -----------------------------
8601 The LLVM exception handling intrinsics (which all start with
8602 ``llvm.eh.`` prefix), are described in the `LLVM Exception
8603 Handling <ExceptionHandling.html#format_common_intrinsics>`_ document.
8607 Trampoline Intrinsics
8608 ---------------------
8610 These intrinsics make it possible to excise one parameter, marked with
8611 the :ref:`nest <nest>` attribute, from a function. The result is a
8612 callable function pointer lacking the nest parameter - the caller does
8613 not need to provide a value for it. Instead, the value to use is stored
8614 in advance in a "trampoline", a block of memory usually allocated on the
8615 stack, which also contains code to splice the nest value into the
8616 argument list. This is used to implement the GCC nested function address
8619 For example, if the function is ``i32 f(i8* nest %c, i32 %x, i32 %y)``
8620 then the resulting function pointer has signature ``i32 (i32, i32)*``.
8621 It can be created as follows:
8623 .. code-block:: llvm
8625 %tramp = alloca [10 x i8], align 4 ; size and alignment only correct for X86
8626 %tramp1 = getelementptr [10 x i8]* %tramp, i32 0, i32 0
8627 call i8* @llvm.init.trampoline(i8* %tramp1, i8* bitcast (i32 (i8*, i32, i32)* @f to i8*), i8* %nval)
8628 %p = call i8* @llvm.adjust.trampoline(i8* %tramp1)
8629 %fp = bitcast i8* %p to i32 (i32, i32)*
8631 The call ``%val = call i32 %fp(i32 %x, i32 %y)`` is then equivalent to
8632 ``%val = call i32 %f(i8* %nval, i32 %x, i32 %y)``.
8636 '``llvm.init.trampoline``' Intrinsic
8637 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8644 declare void @llvm.init.trampoline(i8* <tramp>, i8* <func>, i8* <nval>)
8649 This fills the memory pointed to by ``tramp`` with executable code,
8650 turning it into a trampoline.
8655 The ``llvm.init.trampoline`` intrinsic takes three arguments, all
8656 pointers. The ``tramp`` argument must point to a sufficiently large and
8657 sufficiently aligned block of memory; this memory is written to by the
8658 intrinsic. Note that the size and the alignment are target-specific -
8659 LLVM currently provides no portable way of determining them, so a
8660 front-end that generates this intrinsic needs to have some
8661 target-specific knowledge. The ``func`` argument must hold a function
8662 bitcast to an ``i8*``.
8667 The block of memory pointed to by ``tramp`` is filled with target
8668 dependent code, turning it into a function. Then ``tramp`` needs to be
8669 passed to :ref:`llvm.adjust.trampoline <int_at>` to get a pointer which can
8670 be :ref:`bitcast (to a new function) and called <int_trampoline>`. The new
8671 function's signature is the same as that of ``func`` with any arguments
8672 marked with the ``nest`` attribute removed. At most one such ``nest``
8673 argument is allowed, and it must be of pointer type. Calling the new
8674 function is equivalent to calling ``func`` with the same argument list,
8675 but with ``nval`` used for the missing ``nest`` argument. If, after
8676 calling ``llvm.init.trampoline``, the memory pointed to by ``tramp`` is
8677 modified, then the effect of any later call to the returned function
8678 pointer is undefined.
8682 '``llvm.adjust.trampoline``' Intrinsic
8683 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8690 declare i8* @llvm.adjust.trampoline(i8* <tramp>)
8695 This performs any required machine-specific adjustment to the address of
8696 a trampoline (passed as ``tramp``).
8701 ``tramp`` must point to a block of memory which already has trampoline
8702 code filled in by a previous call to
8703 :ref:`llvm.init.trampoline <int_it>`.
8708 On some architectures the address of the code to be executed needs to be
8709 different to the address where the trampoline is actually stored. This
8710 intrinsic returns the executable address corresponding to ``tramp``
8711 after performing the required machine specific adjustments. The pointer
8712 returned can then be :ref:`bitcast and executed <int_trampoline>`.
8717 This class of intrinsics exists to information about the lifetime of
8718 memory objects and ranges where variables are immutable.
8722 '``llvm.lifetime.start``' Intrinsic
8723 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8730 declare void @llvm.lifetime.start(i64 <size>, i8* nocapture <ptr>)
8735 The '``llvm.lifetime.start``' intrinsic specifies the start of a memory
8741 The first argument is a constant integer representing the size of the
8742 object, or -1 if it is variable sized. The second argument is a pointer
8748 This intrinsic indicates that before this point in the code, the value
8749 of the memory pointed to by ``ptr`` is dead. This means that it is known
8750 to never be used and has an undefined value. A load from the pointer
8751 that precedes this intrinsic can be replaced with ``'undef'``.
8755 '``llvm.lifetime.end``' Intrinsic
8756 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8763 declare void @llvm.lifetime.end(i64 <size>, i8* nocapture <ptr>)
8768 The '``llvm.lifetime.end``' intrinsic specifies the end of a memory
8774 The first argument is a constant integer representing the size of the
8775 object, or -1 if it is variable sized. The second argument is a pointer
8781 This intrinsic indicates that after this point in the code, the value of
8782 the memory pointed to by ``ptr`` is dead. This means that it is known to
8783 never be used and has an undefined value. Any stores into the memory
8784 object following this intrinsic may be removed as dead.
8786 '``llvm.invariant.start``' Intrinsic
8787 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8794 declare {}* @llvm.invariant.start(i64 <size>, i8* nocapture <ptr>)
8799 The '``llvm.invariant.start``' intrinsic specifies that the contents of
8800 a memory object will not change.
8805 The first argument is a constant integer representing the size of the
8806 object, or -1 if it is variable sized. The second argument is a pointer
8812 This intrinsic indicates that until an ``llvm.invariant.end`` that uses
8813 the return value, the referenced memory location is constant and
8816 '``llvm.invariant.end``' Intrinsic
8817 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8824 declare void @llvm.invariant.end({}* <start>, i64 <size>, i8* nocapture <ptr>)
8829 The '``llvm.invariant.end``' intrinsic specifies that the contents of a
8830 memory object are mutable.
8835 The first argument is the matching ``llvm.invariant.start`` intrinsic.
8836 The second argument is a constant integer representing the size of the
8837 object, or -1 if it is variable sized and the third argument is a
8838 pointer to the object.
8843 This intrinsic indicates that the memory is mutable again.
8848 This class of intrinsics is designed to be generic and has no specific
8851 '``llvm.var.annotation``' Intrinsic
8852 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8859 declare void @llvm.var.annotation(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
8864 The '``llvm.var.annotation``' intrinsic.
8869 The first argument is a pointer to a value, the second is a pointer to a
8870 global string, the third is a pointer to a global string which is the
8871 source file name, and the last argument is the line number.
8876 This intrinsic allows annotation of local variables with arbitrary
8877 strings. This can be useful for special purpose optimizations that want
8878 to look for these annotations. These have no other defined use; they are
8879 ignored by code generation and optimization.
8881 '``llvm.ptr.annotation.*``' Intrinsic
8882 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8887 This is an overloaded intrinsic. You can use '``llvm.ptr.annotation``' on a
8888 pointer to an integer of any width. *NOTE* you must specify an address space for
8889 the pointer. The identifier for the default address space is the integer
8894 declare i8* @llvm.ptr.annotation.p<address space>i8(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
8895 declare i16* @llvm.ptr.annotation.p<address space>i16(i16* <val>, i8* <str>, i8* <str>, i32 <int>)
8896 declare i32* @llvm.ptr.annotation.p<address space>i32(i32* <val>, i8* <str>, i8* <str>, i32 <int>)
8897 declare i64* @llvm.ptr.annotation.p<address space>i64(i64* <val>, i8* <str>, i8* <str>, i32 <int>)
8898 declare i256* @llvm.ptr.annotation.p<address space>i256(i256* <val>, i8* <str>, i8* <str>, i32 <int>)
8903 The '``llvm.ptr.annotation``' intrinsic.
8908 The first argument is a pointer to an integer value of arbitrary bitwidth
8909 (result of some expression), the second is a pointer to a global string, the
8910 third is a pointer to a global string which is the source file name, and the
8911 last argument is the line number. It returns the value of the first argument.
8916 This intrinsic allows annotation of a pointer to an integer with arbitrary
8917 strings. This can be useful for special purpose optimizations that want to look
8918 for these annotations. These have no other defined use; they are ignored by code
8919 generation and optimization.
8921 '``llvm.annotation.*``' Intrinsic
8922 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8927 This is an overloaded intrinsic. You can use '``llvm.annotation``' on
8928 any integer bit width.
8932 declare i8 @llvm.annotation.i8(i8 <val>, i8* <str>, i8* <str>, i32 <int>)
8933 declare i16 @llvm.annotation.i16(i16 <val>, i8* <str>, i8* <str>, i32 <int>)
8934 declare i32 @llvm.annotation.i32(i32 <val>, i8* <str>, i8* <str>, i32 <int>)
8935 declare i64 @llvm.annotation.i64(i64 <val>, i8* <str>, i8* <str>, i32 <int>)
8936 declare i256 @llvm.annotation.i256(i256 <val>, i8* <str>, i8* <str>, i32 <int>)
8941 The '``llvm.annotation``' intrinsic.
8946 The first argument is an integer value (result of some expression), the
8947 second is a pointer to a global string, the third is a pointer to a
8948 global string which is the source file name, and the last argument is
8949 the line number. It returns the value of the first argument.
8954 This intrinsic allows annotations to be put on arbitrary expressions
8955 with arbitrary strings. This can be useful for special purpose
8956 optimizations that want to look for these annotations. These have no
8957 other defined use; they are ignored by code generation and optimization.
8959 '``llvm.trap``' Intrinsic
8960 ^^^^^^^^^^^^^^^^^^^^^^^^^
8967 declare void @llvm.trap() noreturn nounwind
8972 The '``llvm.trap``' intrinsic.
8982 This intrinsic is lowered to the target dependent trap instruction. If
8983 the target does not have a trap instruction, this intrinsic will be
8984 lowered to a call of the ``abort()`` function.
8986 '``llvm.debugtrap``' Intrinsic
8987 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8994 declare void @llvm.debugtrap() nounwind
8999 The '``llvm.debugtrap``' intrinsic.
9009 This intrinsic is lowered to code which is intended to cause an
9010 execution trap with the intention of requesting the attention of a
9013 '``llvm.stackprotector``' Intrinsic
9014 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9021 declare void @llvm.stackprotector(i8* <guard>, i8** <slot>)
9026 The ``llvm.stackprotector`` intrinsic takes the ``guard`` and stores it
9027 onto the stack at ``slot``. The stack slot is adjusted to ensure that it
9028 is placed on the stack before local variables.
9033 The ``llvm.stackprotector`` intrinsic requires two pointer arguments.
9034 The first argument is the value loaded from the stack guard
9035 ``@__stack_chk_guard``. The second variable is an ``alloca`` that has
9036 enough space to hold the value of the guard.
9041 This intrinsic causes the prologue/epilogue inserter to force the position of
9042 the ``AllocaInst`` stack slot to be before local variables on the stack. This is
9043 to ensure that if a local variable on the stack is overwritten, it will destroy
9044 the value of the guard. When the function exits, the guard on the stack is
9045 checked against the original guard by ``llvm.stackprotectorcheck``. If they are
9046 different, then ``llvm.stackprotectorcheck`` causes the program to abort by
9047 calling the ``__stack_chk_fail()`` function.
9049 '``llvm.stackprotectorcheck``' Intrinsic
9050 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9057 declare void @llvm.stackprotectorcheck(i8** <guard>)
9062 The ``llvm.stackprotectorcheck`` intrinsic compares ``guard`` against an already
9063 created stack protector and if they are not equal calls the
9064 ``__stack_chk_fail()`` function.
9069 The ``llvm.stackprotectorcheck`` intrinsic requires one pointer argument, the
9070 the variable ``@__stack_chk_guard``.
9075 This intrinsic is provided to perform the stack protector check by comparing
9076 ``guard`` with the stack slot created by ``llvm.stackprotector`` and if the
9077 values do not match call the ``__stack_chk_fail()`` function.
9079 The reason to provide this as an IR level intrinsic instead of implementing it
9080 via other IR operations is that in order to perform this operation at the IR
9081 level without an intrinsic, one would need to create additional basic blocks to
9082 handle the success/failure cases. This makes it difficult to stop the stack
9083 protector check from disrupting sibling tail calls in Codegen. With this
9084 intrinsic, we are able to generate the stack protector basic blocks late in
9085 codegen after the tail call decision has occurred.
9087 '``llvm.objectsize``' Intrinsic
9088 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9095 declare i32 @llvm.objectsize.i32(i8* <object>, i1 <min>)
9096 declare i64 @llvm.objectsize.i64(i8* <object>, i1 <min>)
9101 The ``llvm.objectsize`` intrinsic is designed to provide information to
9102 the optimizers to determine at compile time whether a) an operation
9103 (like memcpy) will overflow a buffer that corresponds to an object, or
9104 b) that a runtime check for overflow isn't necessary. An object in this
9105 context means an allocation of a specific class, structure, array, or
9111 The ``llvm.objectsize`` intrinsic takes two arguments. The first
9112 argument is a pointer to or into the ``object``. The second argument is
9113 a boolean and determines whether ``llvm.objectsize`` returns 0 (if true)
9114 or -1 (if false) when the object size is unknown. The second argument
9115 only accepts constants.
9120 The ``llvm.objectsize`` intrinsic is lowered to a constant representing
9121 the size of the object concerned. If the size cannot be determined at
9122 compile time, ``llvm.objectsize`` returns ``i32/i64 -1 or 0`` (depending
9123 on the ``min`` argument).
9125 '``llvm.expect``' Intrinsic
9126 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9131 This is an overloaded intrinsic. You can use ``llvm.expect`` on any
9136 declare i1 @llvm.expect.i1(i1 <val>, i1 <expected_val>)
9137 declare i32 @llvm.expect.i32(i32 <val>, i32 <expected_val>)
9138 declare i64 @llvm.expect.i64(i64 <val>, i64 <expected_val>)
9143 The ``llvm.expect`` intrinsic provides information about expected (the
9144 most probable) value of ``val``, which can be used by optimizers.
9149 The ``llvm.expect`` intrinsic takes two arguments. The first argument is
9150 a value. The second argument is an expected value, this needs to be a
9151 constant value, variables are not allowed.
9156 This intrinsic is lowered to the ``val``.
9158 '``llvm.donothing``' Intrinsic
9159 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9166 declare void @llvm.donothing() nounwind readnone
9171 The ``llvm.donothing`` intrinsic doesn't perform any operation. It's the
9172 only intrinsic that can be called with an invoke instruction.
9182 This intrinsic does nothing, and it's removed by optimizers and ignored
9185 Stack Map Intrinsics
9186 --------------------
9188 LLVM provides experimental intrinsics to support runtime patching
9189 mechanisms commonly desired in dynamic language JITs. These intrinsics
9190 are described in :doc:`StackMaps`.