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 [unnamed_addr] [AddrSpace] [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] [unnamed_addr] 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
701 the aliasee expression. ``unnamed_addr`` ones are only guaranteed to point
704 Since aliases are only a second name, some restrictions apply, of which
705 some can only be checked when producing an object file:
707 * The expression defining the aliasee must be computable at assembly
708 time. Since it is just a name, no relocations can be used.
710 * No alias in the expression can be weak as the possibility of the
711 intermediate alias being overridden cannot be represented in an
714 * No global value in the expression can be a declaration, since that
715 would require a relocation, which is not possible.
717 .. _namedmetadatastructure:
722 Named metadata is a collection of metadata. :ref:`Metadata
723 nodes <metadata>` (but not metadata strings) are the only valid
724 operands for a named metadata.
728 ; Some unnamed metadata nodes, which are referenced by the named metadata.
729 !0 = metadata !{metadata !"zero"}
730 !1 = metadata !{metadata !"one"}
731 !2 = metadata !{metadata !"two"}
733 !name = !{!0, !1, !2}
740 The return type and each parameter of a function type may have a set of
741 *parameter attributes* associated with them. Parameter attributes are
742 used to communicate additional information about the result or
743 parameters of a function. Parameter attributes are considered to be part
744 of the function, not of the function type, so functions with different
745 parameter attributes can have the same function type.
747 Parameter attributes are simple keywords that follow the type specified.
748 If multiple parameter attributes are needed, they are space separated.
753 declare i32 @printf(i8* noalias nocapture, ...)
754 declare i32 @atoi(i8 zeroext)
755 declare signext i8 @returns_signed_char()
757 Note that any attributes for the function result (``nounwind``,
758 ``readonly``) come immediately after the argument list.
760 Currently, only the following parameter attributes are defined:
763 This indicates to the code generator that the parameter or return
764 value should be zero-extended to the extent required by the target's
765 ABI (which is usually 32-bits, but is 8-bits for a i1 on x86-64) by
766 the caller (for a parameter) or the callee (for a return value).
768 This indicates to the code generator that the parameter or return
769 value should be sign-extended to the extent required by the target's
770 ABI (which is usually 32-bits) by the caller (for a parameter) or
771 the callee (for a return value).
773 This indicates that this parameter or return value should be treated
774 in a special target-dependent fashion during while emitting code for
775 a function call or return (usually, by putting it in a register as
776 opposed to memory, though some targets use it to distinguish between
777 two different kinds of registers). Use of this attribute is
780 This indicates that the pointer parameter should really be passed by
781 value to the function. The attribute implies that a hidden copy of
782 the pointee is made between the caller and the callee, so the callee
783 is unable to modify the value in the caller. This attribute is only
784 valid on LLVM pointer arguments. It is generally used to pass
785 structs and arrays by value, but is also valid on pointers to
786 scalars. The copy is considered to belong to the caller not the
787 callee (for example, ``readonly`` functions should not write to
788 ``byval`` parameters). This is not a valid attribute for return
791 The byval attribute also supports specifying an alignment with the
792 align attribute. It indicates the alignment of the stack slot to
793 form and the known alignment of the pointer specified to the call
794 site. If the alignment is not specified, then the code generator
795 makes a target-specific assumption.
801 The ``inalloca`` argument attribute allows the caller to take the
802 address of outgoing stack arguments. An ``inalloca`` argument must
803 be a pointer to stack memory produced by an ``alloca`` instruction.
804 The alloca, or argument allocation, must also be tagged with the
805 inalloca keyword. Only the past argument may have the ``inalloca``
806 attribute, and that argument is guaranteed to be passed in memory.
808 An argument allocation may be used by a call at most once because
809 the call may deallocate it. The ``inalloca`` attribute cannot be
810 used in conjunction with other attributes that affect argument
811 storage, like ``inreg``, ``nest``, ``sret``, or ``byval``. The
812 ``inalloca`` attribute also disables LLVM's implicit lowering of
813 large aggregate return values, which means that frontend authors
814 must lower them with ``sret`` pointers.
816 When the call site is reached, the argument allocation must have
817 been the most recent stack allocation that is still live, or the
818 results are undefined. It is possible to allocate additional stack
819 space after an argument allocation and before its call site, but it
820 must be cleared off with :ref:`llvm.stackrestore
823 See :doc:`InAlloca` for more information on how to use this
827 This indicates that the pointer parameter specifies the address of a
828 structure that is the return value of the function in the source
829 program. This pointer must be guaranteed by the caller to be valid:
830 loads and stores to the structure may be assumed by the callee
831 not to trap and to be properly aligned. This may only be applied to
832 the first parameter. This is not a valid attribute for return
838 This indicates that pointer values :ref:`based <pointeraliasing>` on
839 the argument or return value do not alias pointer values which are
840 not *based* on it, ignoring certain "irrelevant" dependencies. For a
841 call to the parent function, dependencies between memory references
842 from before or after the call and from those during the call are
843 "irrelevant" to the ``noalias`` keyword for the arguments and return
844 value used in that call. The caller shares the responsibility with
845 the callee for ensuring that these requirements are met. For further
846 details, please see the discussion of the NoAlias response in :ref:`alias
847 analysis <Must, May, or No>`.
849 Note that this definition of ``noalias`` is intentionally similar
850 to the definition of ``restrict`` in C99 for function arguments,
851 though it is slightly weaker.
853 For function return values, C99's ``restrict`` is not meaningful,
854 while LLVM's ``noalias`` is.
856 This indicates that the callee does not make any copies of the
857 pointer that outlive the callee itself. This is not a valid
858 attribute for return values.
863 This indicates that the pointer parameter can be excised using the
864 :ref:`trampoline intrinsics <int_trampoline>`. This is not a valid
865 attribute for return values and can only be applied to one parameter.
868 This indicates that the function always returns the argument as its return
869 value. This is an optimization hint to the code generator when generating
870 the caller, allowing tail call optimization and omission of register saves
871 and restores in some cases; it is not checked or enforced when generating
872 the callee. The parameter and the function return type must be valid
873 operands for the :ref:`bitcast instruction <i_bitcast>`. This is not a
874 valid attribute for return values and can only be applied to one parameter.
877 This indicates that the parameter or return pointer is not null. This
878 attribute may only be applied to pointer typed parameters. This is not
879 checked or enforced by LLVM, the caller must ensure that the pointer
880 passed in is non-null, or the callee must ensure that the returned pointer
885 Garbage Collector Names
886 -----------------------
888 Each function may specify a garbage collector name, which is simply a
893 define void @f() gc "name" { ... }
895 The compiler declares the supported values of *name*. Specifying a
896 collector which will cause the compiler to alter its output in order to
897 support the named garbage collection algorithm.
904 Prefix data is data associated with a function which the code generator
905 will emit immediately before the function body. The purpose of this feature
906 is to allow frontends to associate language-specific runtime metadata with
907 specific functions and make it available through the function pointer while
908 still allowing the function pointer to be called. To access the data for a
909 given function, a program may bitcast the function pointer to a pointer to
910 the constant's type. This implies that the IR symbol points to the start
913 To maintain the semantics of ordinary function calls, the prefix data must
914 have a particular format. Specifically, it must begin with a sequence of
915 bytes which decode to a sequence of machine instructions, valid for the
916 module's target, which transfer control to the point immediately succeeding
917 the prefix data, without performing any other visible action. This allows
918 the inliner and other passes to reason about the semantics of the function
919 definition without needing to reason about the prefix data. Obviously this
920 makes the format of the prefix data highly target dependent.
922 Prefix data is laid out as if it were an initializer for a global variable
923 of the prefix data's type. No padding is automatically placed between the
924 prefix data and the function body. If padding is required, it must be part
927 A trivial example of valid prefix data for the x86 architecture is ``i8 144``,
928 which encodes the ``nop`` instruction:
932 define void @f() prefix i8 144 { ... }
934 Generally prefix data can be formed by encoding a relative branch instruction
935 which skips the metadata, as in this example of valid prefix data for the
936 x86_64 architecture, where the first two bytes encode ``jmp .+10``:
940 %0 = type <{ i8, i8, i8* }>
942 define void @f() prefix %0 <{ i8 235, i8 8, i8* @md}> { ... }
944 A function may have prefix data but no body. This has similar semantics
945 to the ``available_externally`` linkage in that the data may be used by the
946 optimizers but will not be emitted in the object file.
953 Attribute groups are groups of attributes that are referenced by objects within
954 the IR. They are important for keeping ``.ll`` files readable, because a lot of
955 functions will use the same set of attributes. In the degenerative case of a
956 ``.ll`` file that corresponds to a single ``.c`` file, the single attribute
957 group will capture the important command line flags used to build that file.
959 An attribute group is a module-level object. To use an attribute group, an
960 object references the attribute group's ID (e.g. ``#37``). An object may refer
961 to more than one attribute group. In that situation, the attributes from the
962 different groups are merged.
964 Here is an example of attribute groups for a function that should always be
965 inlined, has a stack alignment of 4, and which shouldn't use SSE instructions:
969 ; Target-independent attributes:
970 attributes #0 = { alwaysinline alignstack=4 }
972 ; Target-dependent attributes:
973 attributes #1 = { "no-sse" }
975 ; Function @f has attributes: alwaysinline, alignstack=4, and "no-sse".
976 define void @f() #0 #1 { ... }
983 Function attributes are set to communicate additional information about
984 a function. Function attributes are considered to be part of the
985 function, not of the function type, so functions with different function
986 attributes can have the same function type.
988 Function attributes are simple keywords that follow the type specified.
989 If multiple attributes are needed, they are space separated. For
994 define void @f() noinline { ... }
995 define void @f() alwaysinline { ... }
996 define void @f() alwaysinline optsize { ... }
997 define void @f() optsize { ... }
1000 This attribute indicates that, when emitting the prologue and
1001 epilogue, the backend should forcibly align the stack pointer.
1002 Specify the desired alignment, which must be a power of two, in
1005 This attribute indicates that the inliner should attempt to inline
1006 this function into callers whenever possible, ignoring any active
1007 inlining size threshold for this caller.
1009 This indicates that the callee function at a call site should be
1010 recognized as a built-in function, even though the function's declaration
1011 uses the ``nobuiltin`` attribute. This is only valid at call sites for
1012 direct calls to functions which are declared with the ``nobuiltin``
1015 This attribute indicates that this function is rarely called. When
1016 computing edge weights, basic blocks post-dominated by a cold
1017 function call are also considered to be cold; and, thus, given low
1020 This attribute indicates that the source code contained a hint that
1021 inlining this function is desirable (such as the "inline" keyword in
1022 C/C++). It is just a hint; it imposes no requirements on the
1025 This attribute indicates that the function should be added to a
1026 jump-instruction table at code-generation time, and that all address-taken
1027 references to this function should be replaced with a reference to the
1028 appropriate jump-instruction-table function pointer. Note that this creates
1029 a new pointer for the original function, which means that code that depends
1030 on function-pointer identity can break. So, any function annotated with
1031 ``jumptable`` must also be ``unnamed_addr``.
1033 This attribute suggests that optimization passes and code generator
1034 passes make choices that keep the code size of this function as small
1035 as possible and perform optimizations that may sacrifice runtime
1036 performance in order to minimize the size of the generated code.
1038 This attribute disables prologue / epilogue emission for the
1039 function. This can have very system-specific consequences.
1041 This indicates that the callee function at a call site is not recognized as
1042 a built-in function. LLVM will retain the original call and not replace it
1043 with equivalent code based on the semantics of the built-in function, unless
1044 the call site uses the ``builtin`` attribute. This is valid at call sites
1045 and on function declarations and definitions.
1047 This attribute indicates that calls to the function cannot be
1048 duplicated. A call to a ``noduplicate`` function may be moved
1049 within its parent function, but may not be duplicated within
1050 its parent function.
1052 A function containing a ``noduplicate`` call may still
1053 be an inlining candidate, provided that the call is not
1054 duplicated by inlining. That implies that the function has
1055 internal linkage and only has one call site, so the original
1056 call is dead after inlining.
1058 This attributes disables implicit floating point instructions.
1060 This attribute indicates that the inliner should never inline this
1061 function in any situation. This attribute may not be used together
1062 with the ``alwaysinline`` attribute.
1064 This attribute suppresses lazy symbol binding for the function. This
1065 may make calls to the function faster, at the cost of extra program
1066 startup time if the function is not called during program startup.
1068 This attribute indicates that the code generator should not use a
1069 red zone, even if the target-specific ABI normally permits it.
1071 This function attribute indicates that the function never returns
1072 normally. This produces undefined behavior at runtime if the
1073 function ever does dynamically return.
1075 This function attribute indicates that the function never returns
1076 with an unwind or exceptional control flow. If the function does
1077 unwind, its runtime behavior is undefined.
1079 This function attribute indicates that the function is not optimized
1080 by any optimization or code generator passes with the
1081 exception of interprocedural optimization passes.
1082 This attribute cannot be used together with the ``alwaysinline``
1083 attribute; this attribute is also incompatible
1084 with the ``minsize`` attribute and the ``optsize`` attribute.
1086 This attribute requires the ``noinline`` attribute to be specified on
1087 the function as well, so the function is never inlined into any caller.
1088 Only functions with the ``alwaysinline`` attribute are valid
1089 candidates for inlining into the body of this function.
1091 This attribute suggests that optimization passes and code generator
1092 passes make choices that keep the code size of this function low,
1093 and otherwise do optimizations specifically to reduce code size as
1094 long as they do not significantly impact runtime performance.
1096 On a function, this attribute indicates that the function computes its
1097 result (or decides to unwind an exception) based strictly on its arguments,
1098 without dereferencing any pointer arguments or otherwise accessing
1099 any mutable state (e.g. memory, control registers, etc) visible to
1100 caller functions. It does not write through any pointer arguments
1101 (including ``byval`` arguments) and never changes any state visible
1102 to callers. This means that it cannot unwind exceptions by calling
1103 the ``C++`` exception throwing methods.
1105 On an argument, this attribute indicates that the function does not
1106 dereference that pointer argument, even though it may read or write the
1107 memory that the pointer points to if accessed through other pointers.
1109 On a function, this attribute indicates that the function does not write
1110 through any pointer arguments (including ``byval`` arguments) or otherwise
1111 modify any state (e.g. memory, control registers, etc) visible to
1112 caller functions. It may dereference pointer arguments and read
1113 state that may be set in the caller. A readonly function always
1114 returns the same value (or unwinds an exception identically) when
1115 called with the same set of arguments and global state. It cannot
1116 unwind an exception by calling the ``C++`` exception throwing
1119 On an argument, this attribute indicates that the function does not write
1120 through this pointer argument, even though it may write to the memory that
1121 the pointer points to.
1123 This attribute indicates that this function can return twice. The C
1124 ``setjmp`` is an example of such a function. The compiler disables
1125 some optimizations (like tail calls) in the caller of these
1127 ``sanitize_address``
1128 This attribute indicates that AddressSanitizer checks
1129 (dynamic address safety analysis) are enabled for this function.
1131 This attribute indicates that MemorySanitizer checks (dynamic detection
1132 of accesses to uninitialized memory) are enabled for this function.
1134 This attribute indicates that ThreadSanitizer checks
1135 (dynamic thread safety analysis) are enabled for this function.
1137 This attribute indicates that the function should emit a stack
1138 smashing protector. It is in the form of a "canary" --- a random value
1139 placed on the stack before the local variables that's checked upon
1140 return from the function to see if it has been overwritten. A
1141 heuristic is used to determine if a function needs stack protectors
1142 or not. The heuristic used will enable protectors for functions with:
1144 - Character arrays larger than ``ssp-buffer-size`` (default 8).
1145 - Aggregates containing character arrays larger than ``ssp-buffer-size``.
1146 - Calls to alloca() with variable sizes or constant sizes greater than
1147 ``ssp-buffer-size``.
1149 Variables that are identified as requiring a protector will be arranged
1150 on the stack such that they are adjacent to the stack protector guard.
1152 If a function that has an ``ssp`` attribute is inlined into a
1153 function that doesn't have an ``ssp`` attribute, then the resulting
1154 function will have an ``ssp`` attribute.
1156 This attribute indicates that the function should *always* emit a
1157 stack smashing protector. This overrides the ``ssp`` function
1160 Variables that are identified as requiring a protector will be arranged
1161 on the stack such that they are adjacent to the stack protector guard.
1162 The specific layout rules are:
1164 #. Large arrays and structures containing large arrays
1165 (``>= ssp-buffer-size``) are closest to the stack protector.
1166 #. Small arrays and structures containing small arrays
1167 (``< ssp-buffer-size``) are 2nd closest to the protector.
1168 #. Variables that have had their address taken are 3rd closest to the
1171 If a function that has an ``sspreq`` attribute is inlined into a
1172 function that doesn't have an ``sspreq`` attribute or which has an
1173 ``ssp`` or ``sspstrong`` attribute, then the resulting function will have
1174 an ``sspreq`` attribute.
1176 This attribute indicates that the function should emit a stack smashing
1177 protector. This attribute causes a strong heuristic to be used when
1178 determining if a function needs stack protectors. The strong heuristic
1179 will enable protectors for functions with:
1181 - Arrays of any size and type
1182 - Aggregates containing an array of any size and type.
1183 - Calls to alloca().
1184 - Local variables that have had their address taken.
1186 Variables that are identified as requiring a protector will be arranged
1187 on the stack such that they are adjacent to the stack protector guard.
1188 The specific layout rules are:
1190 #. Large arrays and structures containing large arrays
1191 (``>= ssp-buffer-size``) are closest to the stack protector.
1192 #. Small arrays and structures containing small arrays
1193 (``< ssp-buffer-size``) are 2nd closest to the protector.
1194 #. Variables that have had their address taken are 3rd closest to the
1197 This overrides the ``ssp`` function attribute.
1199 If a function that has an ``sspstrong`` attribute is inlined into a
1200 function that doesn't have an ``sspstrong`` attribute, then the
1201 resulting function will have an ``sspstrong`` attribute.
1203 This attribute indicates that the ABI being targeted requires that
1204 an unwind table entry be produce for this function even if we can
1205 show that no exceptions passes by it. This is normally the case for
1206 the ELF x86-64 abi, but it can be disabled for some compilation
1211 Module-Level Inline Assembly
1212 ----------------------------
1214 Modules may contain "module-level inline asm" blocks, which corresponds
1215 to the GCC "file scope inline asm" blocks. These blocks are internally
1216 concatenated by LLVM and treated as a single unit, but may be separated
1217 in the ``.ll`` file if desired. The syntax is very simple:
1219 .. code-block:: llvm
1221 module asm "inline asm code goes here"
1222 module asm "more can go here"
1224 The strings can contain any character by escaping non-printable
1225 characters. The escape sequence used is simply "\\xx" where "xx" is the
1226 two digit hex code for the number.
1228 The inline asm code is simply printed to the machine code .s file when
1229 assembly code is generated.
1231 .. _langref_datalayout:
1236 A module may specify a target specific data layout string that specifies
1237 how data is to be laid out in memory. The syntax for the data layout is
1240 .. code-block:: llvm
1242 target datalayout = "layout specification"
1244 The *layout specification* consists of a list of specifications
1245 separated by the minus sign character ('-'). Each specification starts
1246 with a letter and may include other information after the letter to
1247 define some aspect of the data layout. The specifications accepted are
1251 Specifies that the target lays out data in big-endian form. That is,
1252 the bits with the most significance have the lowest address
1255 Specifies that the target lays out data in little-endian form. That
1256 is, the bits with the least significance have the lowest address
1259 Specifies the natural alignment of the stack in bits. Alignment
1260 promotion of stack variables is limited to the natural stack
1261 alignment to avoid dynamic stack realignment. The stack alignment
1262 must be a multiple of 8-bits. If omitted, the natural stack
1263 alignment defaults to "unspecified", which does not prevent any
1264 alignment promotions.
1265 ``p[n]:<size>:<abi>:<pref>``
1266 This specifies the *size* of a pointer and its ``<abi>`` and
1267 ``<pref>``\erred alignments for address space ``n``. All sizes are in
1268 bits. The address space, ``n`` is optional, and if not specified,
1269 denotes the default address space 0. The value of ``n`` must be
1270 in the range [1,2^23).
1271 ``i<size>:<abi>:<pref>``
1272 This specifies the alignment for an integer type of a given bit
1273 ``<size>``. The value of ``<size>`` must be in the range [1,2^23).
1274 ``v<size>:<abi>:<pref>``
1275 This specifies the alignment for a vector type of a given bit
1277 ``f<size>:<abi>:<pref>``
1278 This specifies the alignment for a floating point type of a given bit
1279 ``<size>``. Only values of ``<size>`` that are supported by the target
1280 will work. 32 (float) and 64 (double) are supported on all targets; 80
1281 or 128 (different flavors of long double) are also supported on some
1284 This specifies the alignment for an object of aggregate type.
1286 If present, specifies that llvm names are mangled in the output. The
1289 * ``e``: ELF mangling: Private symbols get a ``.L`` prefix.
1290 * ``m``: Mips mangling: Private symbols get a ``$`` prefix.
1291 * ``o``: Mach-O mangling: Private symbols get ``L`` prefix. Other
1292 symbols get a ``_`` prefix.
1293 * ``w``: Windows COFF prefix: Similar to Mach-O, but stdcall and fastcall
1294 functions also get a suffix based on the frame size.
1295 ``n<size1>:<size2>:<size3>...``
1296 This specifies a set of native integer widths for the target CPU in
1297 bits. For example, it might contain ``n32`` for 32-bit PowerPC,
1298 ``n32:64`` for PowerPC 64, or ``n8:16:32:64`` for X86-64. Elements of
1299 this set are considered to support most general arithmetic operations
1302 On every specification that takes a ``<abi>:<pref>``, specifying the
1303 ``<pref>`` alignment is optional. If omitted, the preceding ``:``
1304 should be omitted too and ``<pref>`` will be equal to ``<abi>``.
1306 When constructing the data layout for a given target, LLVM starts with a
1307 default set of specifications which are then (possibly) overridden by
1308 the specifications in the ``datalayout`` keyword. The default
1309 specifications are given in this list:
1311 - ``E`` - big endian
1312 - ``p:64:64:64`` - 64-bit pointers with 64-bit alignment.
1313 - ``p[n]:64:64:64`` - Other address spaces are assumed to be the
1314 same as the default address space.
1315 - ``S0`` - natural stack alignment is unspecified
1316 - ``i1:8:8`` - i1 is 8-bit (byte) aligned
1317 - ``i8:8:8`` - i8 is 8-bit (byte) aligned
1318 - ``i16:16:16`` - i16 is 16-bit aligned
1319 - ``i32:32:32`` - i32 is 32-bit aligned
1320 - ``i64:32:64`` - i64 has ABI alignment of 32-bits but preferred
1321 alignment of 64-bits
1322 - ``f16:16:16`` - half is 16-bit aligned
1323 - ``f32:32:32`` - float is 32-bit aligned
1324 - ``f64:64:64`` - double is 64-bit aligned
1325 - ``f128:128:128`` - quad is 128-bit aligned
1326 - ``v64:64:64`` - 64-bit vector is 64-bit aligned
1327 - ``v128:128:128`` - 128-bit vector is 128-bit aligned
1328 - ``a:0:64`` - aggregates are 64-bit aligned
1330 When LLVM is determining the alignment for a given type, it uses the
1333 #. If the type sought is an exact match for one of the specifications,
1334 that specification is used.
1335 #. If no match is found, and the type sought is an integer type, then
1336 the smallest integer type that is larger than the bitwidth of the
1337 sought type is used. If none of the specifications are larger than
1338 the bitwidth then the largest integer type is used. For example,
1339 given the default specifications above, the i7 type will use the
1340 alignment of i8 (next largest) while both i65 and i256 will use the
1341 alignment of i64 (largest specified).
1342 #. If no match is found, and the type sought is a vector type, then the
1343 largest vector type that is smaller than the sought vector type will
1344 be used as a fall back. This happens because <128 x double> can be
1345 implemented in terms of 64 <2 x double>, for example.
1347 The function of the data layout string may not be what you expect.
1348 Notably, this is not a specification from the frontend of what alignment
1349 the code generator should use.
1351 Instead, if specified, the target data layout is required to match what
1352 the ultimate *code generator* expects. This string is used by the
1353 mid-level optimizers to improve code, and this only works if it matches
1354 what the ultimate code generator uses. If you would like to generate IR
1355 that does not embed this target-specific detail into the IR, then you
1356 don't have to specify the string. This will disable some optimizations
1357 that require precise layout information, but this also prevents those
1358 optimizations from introducing target specificity into the IR.
1365 A module may specify a target triple string that describes the target
1366 host. The syntax for the target triple is simply:
1368 .. code-block:: llvm
1370 target triple = "x86_64-apple-macosx10.7.0"
1372 The *target triple* string consists of a series of identifiers delimited
1373 by the minus sign character ('-'). The canonical forms are:
1377 ARCHITECTURE-VENDOR-OPERATING_SYSTEM
1378 ARCHITECTURE-VENDOR-OPERATING_SYSTEM-ENVIRONMENT
1380 This information is passed along to the backend so that it generates
1381 code for the proper architecture. It's possible to override this on the
1382 command line with the ``-mtriple`` command line option.
1384 .. _pointeraliasing:
1386 Pointer Aliasing Rules
1387 ----------------------
1389 Any memory access must be done through a pointer value associated with
1390 an address range of the memory access, otherwise the behavior is
1391 undefined. Pointer values are associated with address ranges according
1392 to the following rules:
1394 - A pointer value is associated with the addresses associated with any
1395 value it is *based* on.
1396 - An address of a global variable is associated with the address range
1397 of the variable's storage.
1398 - The result value of an allocation instruction is associated with the
1399 address range of the allocated storage.
1400 - A null pointer in the default address-space is associated with no
1402 - An integer constant other than zero or a pointer value returned from
1403 a function not defined within LLVM may be associated with address
1404 ranges allocated through mechanisms other than those provided by
1405 LLVM. Such ranges shall not overlap with any ranges of addresses
1406 allocated by mechanisms provided by LLVM.
1408 A pointer value is *based* on another pointer value according to the
1411 - A pointer value formed from a ``getelementptr`` operation is *based*
1412 on the first operand of the ``getelementptr``.
1413 - The result value of a ``bitcast`` is *based* on the operand of the
1415 - A pointer value formed by an ``inttoptr`` is *based* on all pointer
1416 values that contribute (directly or indirectly) to the computation of
1417 the pointer's value.
1418 - The "*based* on" relationship is transitive.
1420 Note that this definition of *"based"* is intentionally similar to the
1421 definition of *"based"* in C99, though it is slightly weaker.
1423 LLVM IR does not associate types with memory. The result type of a
1424 ``load`` merely indicates the size and alignment of the memory from
1425 which to load, as well as the interpretation of the value. The first
1426 operand type of a ``store`` similarly only indicates the size and
1427 alignment of the store.
1429 Consequently, type-based alias analysis, aka TBAA, aka
1430 ``-fstrict-aliasing``, is not applicable to general unadorned LLVM IR.
1431 :ref:`Metadata <metadata>` may be used to encode additional information
1432 which specialized optimization passes may use to implement type-based
1437 Volatile Memory Accesses
1438 ------------------------
1440 Certain memory accesses, such as :ref:`load <i_load>`'s,
1441 :ref:`store <i_store>`'s, and :ref:`llvm.memcpy <int_memcpy>`'s may be
1442 marked ``volatile``. The optimizers must not change the number of
1443 volatile operations or change their order of execution relative to other
1444 volatile operations. The optimizers *may* change the order of volatile
1445 operations relative to non-volatile operations. This is not Java's
1446 "volatile" and has no cross-thread synchronization behavior.
1448 IR-level volatile loads and stores cannot safely be optimized into
1449 llvm.memcpy or llvm.memmove intrinsics even when those intrinsics are
1450 flagged volatile. Likewise, the backend should never split or merge
1451 target-legal volatile load/store instructions.
1453 .. admonition:: Rationale
1455 Platforms may rely on volatile loads and stores of natively supported
1456 data width to be executed as single instruction. For example, in C
1457 this holds for an l-value of volatile primitive type with native
1458 hardware support, but not necessarily for aggregate types. The
1459 frontend upholds these expectations, which are intentionally
1460 unspecified in the IR. The rules above ensure that IR transformation
1461 do not violate the frontend's contract with the language.
1465 Memory Model for Concurrent Operations
1466 --------------------------------------
1468 The LLVM IR does not define any way to start parallel threads of
1469 execution or to register signal handlers. Nonetheless, there are
1470 platform-specific ways to create them, and we define LLVM IR's behavior
1471 in their presence. This model is inspired by the C++0x memory model.
1473 For a more informal introduction to this model, see the :doc:`Atomics`.
1475 We define a *happens-before* partial order as the least partial order
1478 - Is a superset of single-thread program order, and
1479 - When a *synchronizes-with* ``b``, includes an edge from ``a`` to
1480 ``b``. *Synchronizes-with* pairs are introduced by platform-specific
1481 techniques, like pthread locks, thread creation, thread joining,
1482 etc., and by atomic instructions. (See also :ref:`Atomic Memory Ordering
1483 Constraints <ordering>`).
1485 Note that program order does not introduce *happens-before* edges
1486 between a thread and signals executing inside that thread.
1488 Every (defined) read operation (load instructions, memcpy, atomic
1489 loads/read-modify-writes, etc.) R reads a series of bytes written by
1490 (defined) write operations (store instructions, atomic
1491 stores/read-modify-writes, memcpy, etc.). For the purposes of this
1492 section, initialized globals are considered to have a write of the
1493 initializer which is atomic and happens before any other read or write
1494 of the memory in question. For each byte of a read R, R\ :sub:`byte`
1495 may see any write to the same byte, except:
1497 - If write\ :sub:`1` happens before write\ :sub:`2`, and
1498 write\ :sub:`2` happens before R\ :sub:`byte`, then
1499 R\ :sub:`byte` does not see write\ :sub:`1`.
1500 - If R\ :sub:`byte` happens before write\ :sub:`3`, then
1501 R\ :sub:`byte` does not see write\ :sub:`3`.
1503 Given that definition, R\ :sub:`byte` is defined as follows:
1505 - If R is volatile, the result is target-dependent. (Volatile is
1506 supposed to give guarantees which can support ``sig_atomic_t`` in
1507 C/C++, and may be used for accesses to addresses which do not behave
1508 like normal memory. It does not generally provide cross-thread
1510 - Otherwise, if there is no write to the same byte that happens before
1511 R\ :sub:`byte`, R\ :sub:`byte` returns ``undef`` for that byte.
1512 - Otherwise, if R\ :sub:`byte` may see exactly one write,
1513 R\ :sub:`byte` returns the value written by that write.
1514 - Otherwise, if R is atomic, and all the writes R\ :sub:`byte` may
1515 see are atomic, it chooses one of the values written. See the :ref:`Atomic
1516 Memory Ordering Constraints <ordering>` section for additional
1517 constraints on how the choice is made.
1518 - Otherwise R\ :sub:`byte` returns ``undef``.
1520 R returns the value composed of the series of bytes it read. This
1521 implies that some bytes within the value may be ``undef`` **without**
1522 the entire value being ``undef``. Note that this only defines the
1523 semantics of the operation; it doesn't mean that targets will emit more
1524 than one instruction to read the series of bytes.
1526 Note that in cases where none of the atomic intrinsics are used, this
1527 model places only one restriction on IR transformations on top of what
1528 is required for single-threaded execution: introducing a store to a byte
1529 which might not otherwise be stored is not allowed in general.
1530 (Specifically, in the case where another thread might write to and read
1531 from an address, introducing a store can change a load that may see
1532 exactly one write into a load that may see multiple writes.)
1536 Atomic Memory Ordering Constraints
1537 ----------------------------------
1539 Atomic instructions (:ref:`cmpxchg <i_cmpxchg>`,
1540 :ref:`atomicrmw <i_atomicrmw>`, :ref:`fence <i_fence>`,
1541 :ref:`atomic load <i_load>`, and :ref:`atomic store <i_store>`) take
1542 ordering parameters that determine which other atomic instructions on
1543 the same address they *synchronize with*. These semantics are borrowed
1544 from Java and C++0x, but are somewhat more colloquial. If these
1545 descriptions aren't precise enough, check those specs (see spec
1546 references in the :doc:`atomics guide <Atomics>`).
1547 :ref:`fence <i_fence>` instructions treat these orderings somewhat
1548 differently since they don't take an address. See that instruction's
1549 documentation for details.
1551 For a simpler introduction to the ordering constraints, see the
1555 The set of values that can be read is governed by the happens-before
1556 partial order. A value cannot be read unless some operation wrote
1557 it. This is intended to provide a guarantee strong enough to model
1558 Java's non-volatile shared variables. This ordering cannot be
1559 specified for read-modify-write operations; it is not strong enough
1560 to make them atomic in any interesting way.
1562 In addition to the guarantees of ``unordered``, there is a single
1563 total order for modifications by ``monotonic`` operations on each
1564 address. All modification orders must be compatible with the
1565 happens-before order. There is no guarantee that the modification
1566 orders can be combined to a global total order for the whole program
1567 (and this often will not be possible). The read in an atomic
1568 read-modify-write operation (:ref:`cmpxchg <i_cmpxchg>` and
1569 :ref:`atomicrmw <i_atomicrmw>`) reads the value in the modification
1570 order immediately before the value it writes. If one atomic read
1571 happens before another atomic read of the same address, the later
1572 read must see the same value or a later value in the address's
1573 modification order. This disallows reordering of ``monotonic`` (or
1574 stronger) operations on the same address. If an address is written
1575 ``monotonic``-ally by one thread, and other threads ``monotonic``-ally
1576 read that address repeatedly, the other threads must eventually see
1577 the write. This corresponds to the C++0x/C1x
1578 ``memory_order_relaxed``.
1580 In addition to the guarantees of ``monotonic``, a
1581 *synchronizes-with* edge may be formed with a ``release`` operation.
1582 This is intended to model C++'s ``memory_order_acquire``.
1584 In addition to the guarantees of ``monotonic``, if this operation
1585 writes a value which is subsequently read by an ``acquire``
1586 operation, it *synchronizes-with* that operation. (This isn't a
1587 complete description; see the C++0x definition of a release
1588 sequence.) This corresponds to the C++0x/C1x
1589 ``memory_order_release``.
1590 ``acq_rel`` (acquire+release)
1591 Acts as both an ``acquire`` and ``release`` operation on its
1592 address. This corresponds to the C++0x/C1x ``memory_order_acq_rel``.
1593 ``seq_cst`` (sequentially consistent)
1594 In addition to the guarantees of ``acq_rel`` (``acquire`` for an
1595 operation which only reads, ``release`` for an operation which only
1596 writes), there is a global total order on all
1597 sequentially-consistent operations on all addresses, which is
1598 consistent with the *happens-before* partial order and with the
1599 modification orders of all the affected addresses. Each
1600 sequentially-consistent read sees the last preceding write to the
1601 same address in this global order. This corresponds to the C++0x/C1x
1602 ``memory_order_seq_cst`` and Java volatile.
1606 If an atomic operation is marked ``singlethread``, it only *synchronizes
1607 with* or participates in modification and seq\_cst total orderings with
1608 other operations running in the same thread (for example, in signal
1616 LLVM IR floating-point binary ops (:ref:`fadd <i_fadd>`,
1617 :ref:`fsub <i_fsub>`, :ref:`fmul <i_fmul>`, :ref:`fdiv <i_fdiv>`,
1618 :ref:`frem <i_frem>`) have the following flags that can set to enable
1619 otherwise unsafe floating point operations
1622 No NaNs - Allow optimizations to assume the arguments and result are not
1623 NaN. Such optimizations are required to retain defined behavior over
1624 NaNs, but the value of the result is undefined.
1627 No Infs - Allow optimizations to assume the arguments and result are not
1628 +/-Inf. Such optimizations are required to retain defined behavior over
1629 +/-Inf, but the value of the result is undefined.
1632 No Signed Zeros - Allow optimizations to treat the sign of a zero
1633 argument or result as insignificant.
1636 Allow Reciprocal - Allow optimizations to use the reciprocal of an
1637 argument rather than perform division.
1640 Fast - Allow algebraically equivalent transformations that may
1641 dramatically change results in floating point (e.g. reassociate). This
1642 flag implies all the others.
1649 The LLVM type system is one of the most important features of the
1650 intermediate representation. Being typed enables a number of
1651 optimizations to be performed on the intermediate representation
1652 directly, without having to do extra analyses on the side before the
1653 transformation. A strong type system makes it easier to read the
1654 generated code and enables novel analyses and transformations that are
1655 not feasible to perform on normal three address code representations.
1665 The void type does not represent any value and has no size.
1683 The function type can be thought of as a function signature. It consists of a
1684 return type and a list of formal parameter types. The return type of a function
1685 type is a void type or first class type --- except for :ref:`label <t_label>`
1686 and :ref:`metadata <t_metadata>` types.
1692 <returntype> (<parameter list>)
1694 ...where '``<parameter list>``' is a comma-separated list of type
1695 specifiers. Optionally, the parameter list may include a type ``...``, which
1696 indicates that the function takes a variable number of arguments. Variable
1697 argument functions can access their arguments with the :ref:`variable argument
1698 handling intrinsic <int_varargs>` functions. '``<returntype>``' is any type
1699 except :ref:`label <t_label>` and :ref:`metadata <t_metadata>`.
1703 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1704 | ``i32 (i32)`` | function taking an ``i32``, returning an ``i32`` |
1705 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1706 | ``float (i16, i32 *) *`` | :ref:`Pointer <t_pointer>` to a function that takes an ``i16`` and a :ref:`pointer <t_pointer>` to ``i32``, returning ``float``. |
1707 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1708 | ``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. |
1709 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1710 | ``{i32, i32} (i32)`` | A function taking an ``i32``, returning a :ref:`structure <t_struct>` containing two ``i32`` values |
1711 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1718 The :ref:`first class <t_firstclass>` types are perhaps the most important.
1719 Values of these types are the only ones which can be produced by
1727 These are the types that are valid in registers from CodeGen's perspective.
1736 The integer type is a very simple type that simply specifies an
1737 arbitrary bit width for the integer type desired. Any bit width from 1
1738 bit to 2\ :sup:`23`\ -1 (about 8 million) can be specified.
1746 The number of bits the integer will occupy is specified by the ``N``
1752 +----------------+------------------------------------------------+
1753 | ``i1`` | a single-bit integer. |
1754 +----------------+------------------------------------------------+
1755 | ``i32`` | a 32-bit integer. |
1756 +----------------+------------------------------------------------+
1757 | ``i1942652`` | a really big integer of over 1 million bits. |
1758 +----------------+------------------------------------------------+
1762 Floating Point Types
1763 """"""""""""""""""""
1772 - 16-bit floating point value
1775 - 32-bit floating point value
1778 - 64-bit floating point value
1781 - 128-bit floating point value (112-bit mantissa)
1784 - 80-bit floating point value (X87)
1787 - 128-bit floating point value (two 64-bits)
1794 The x86_mmx type represents a value held in an MMX register on an x86
1795 machine. The operations allowed on it are quite limited: parameters and
1796 return values, load and store, and bitcast. User-specified MMX
1797 instructions are represented as intrinsic or asm calls with arguments
1798 and/or results of this type. There are no arrays, vectors or constants
1815 The pointer type is used to specify memory locations. Pointers are
1816 commonly used to reference objects in memory.
1818 Pointer types may have an optional address space attribute defining the
1819 numbered address space where the pointed-to object resides. The default
1820 address space is number zero. The semantics of non-zero address spaces
1821 are target-specific.
1823 Note that LLVM does not permit pointers to void (``void*``) nor does it
1824 permit pointers to labels (``label*``). Use ``i8*`` instead.
1834 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1835 | ``[4 x i32]*`` | A :ref:`pointer <t_pointer>` to :ref:`array <t_array>` of four ``i32`` values. |
1836 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1837 | ``i32 (i32*) *`` | A :ref:`pointer <t_pointer>` to a :ref:`function <t_function>` that takes an ``i32*``, returning an ``i32``. |
1838 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1839 | ``i32 addrspace(5)*`` | A :ref:`pointer <t_pointer>` to an ``i32`` value that resides in address space #5. |
1840 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1849 A vector type is a simple derived type that represents a vector of
1850 elements. Vector types are used when multiple primitive data are
1851 operated in parallel using a single instruction (SIMD). A vector type
1852 requires a size (number of elements) and an underlying primitive data
1853 type. Vector types are considered :ref:`first class <t_firstclass>`.
1859 < <# elements> x <elementtype> >
1861 The number of elements is a constant integer value larger than 0;
1862 elementtype may be any integer or floating point type, or a pointer to
1863 these types. Vectors of size zero are not allowed.
1867 +-------------------+--------------------------------------------------+
1868 | ``<4 x i32>`` | Vector of 4 32-bit integer values. |
1869 +-------------------+--------------------------------------------------+
1870 | ``<8 x float>`` | Vector of 8 32-bit floating-point values. |
1871 +-------------------+--------------------------------------------------+
1872 | ``<2 x i64>`` | Vector of 2 64-bit integer values. |
1873 +-------------------+--------------------------------------------------+
1874 | ``<4 x i64*>`` | Vector of 4 pointers to 64-bit integer values. |
1875 +-------------------+--------------------------------------------------+
1884 The label type represents code labels.
1899 The metadata type represents embedded metadata. No derived types may be
1900 created from metadata except for :ref:`function <t_function>` arguments.
1913 Aggregate Types are a subset of derived types that can contain multiple
1914 member types. :ref:`Arrays <t_array>` and :ref:`structs <t_struct>` are
1915 aggregate types. :ref:`Vectors <t_vector>` are not considered to be
1925 The array type is a very simple derived type that arranges elements
1926 sequentially in memory. The array type requires a size (number of
1927 elements) and an underlying data type.
1933 [<# elements> x <elementtype>]
1935 The number of elements is a constant integer value; ``elementtype`` may
1936 be any type with a size.
1940 +------------------+--------------------------------------+
1941 | ``[40 x i32]`` | Array of 40 32-bit integer values. |
1942 +------------------+--------------------------------------+
1943 | ``[41 x i32]`` | Array of 41 32-bit integer values. |
1944 +------------------+--------------------------------------+
1945 | ``[4 x i8]`` | Array of 4 8-bit integer values. |
1946 +------------------+--------------------------------------+
1948 Here are some examples of multidimensional arrays:
1950 +-----------------------------+----------------------------------------------------------+
1951 | ``[3 x [4 x i32]]`` | 3x4 array of 32-bit integer values. |
1952 +-----------------------------+----------------------------------------------------------+
1953 | ``[12 x [10 x float]]`` | 12x10 array of single precision floating point values. |
1954 +-----------------------------+----------------------------------------------------------+
1955 | ``[2 x [3 x [4 x i16]]]`` | 2x3x4 array of 16-bit integer values. |
1956 +-----------------------------+----------------------------------------------------------+
1958 There is no restriction on indexing beyond the end of the array implied
1959 by a static type (though there are restrictions on indexing beyond the
1960 bounds of an allocated object in some cases). This means that
1961 single-dimension 'variable sized array' addressing can be implemented in
1962 LLVM with a zero length array type. An implementation of 'pascal style
1963 arrays' in LLVM could use the type "``{ i32, [0 x float]}``", for
1973 The structure type is used to represent a collection of data members
1974 together in memory. The elements of a structure may be any type that has
1977 Structures in memory are accessed using '``load``' and '``store``' by
1978 getting a pointer to a field with the '``getelementptr``' instruction.
1979 Structures in registers are accessed using the '``extractvalue``' and
1980 '``insertvalue``' instructions.
1982 Structures may optionally be "packed" structures, which indicate that
1983 the alignment of the struct is one byte, and that there is no padding
1984 between the elements. In non-packed structs, padding between field types
1985 is inserted as defined by the DataLayout string in the module, which is
1986 required to match what the underlying code generator expects.
1988 Structures can either be "literal" or "identified". A literal structure
1989 is defined inline with other types (e.g. ``{i32, i32}*``) whereas
1990 identified types are always defined at the top level with a name.
1991 Literal types are uniqued by their contents and can never be recursive
1992 or opaque since there is no way to write one. Identified types can be
1993 recursive, can be opaqued, and are never uniqued.
1999 %T1 = type { <type list> } ; Identified normal struct type
2000 %T2 = type <{ <type list> }> ; Identified packed struct type
2004 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2005 | ``{ i32, i32, i32 }`` | A triple of three ``i32`` values |
2006 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2007 | ``{ 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``. |
2008 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2009 | ``<{ i8, i32 }>`` | A packed struct known to be 5 bytes in size. |
2010 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2014 Opaque Structure Types
2015 """"""""""""""""""""""
2019 Opaque structure types are used to represent named structure types that
2020 do not have a body specified. This corresponds (for example) to the C
2021 notion of a forward declared structure.
2032 +--------------+-------------------+
2033 | ``opaque`` | An opaque type. |
2034 +--------------+-------------------+
2041 LLVM has several different basic types of constants. This section
2042 describes them all and their syntax.
2047 **Boolean constants**
2048 The two strings '``true``' and '``false``' are both valid constants
2050 **Integer constants**
2051 Standard integers (such as '4') are constants of the
2052 :ref:`integer <t_integer>` type. Negative numbers may be used with
2054 **Floating point constants**
2055 Floating point constants use standard decimal notation (e.g.
2056 123.421), exponential notation (e.g. 1.23421e+2), or a more precise
2057 hexadecimal notation (see below). The assembler requires the exact
2058 decimal value of a floating-point constant. For example, the
2059 assembler accepts 1.25 but rejects 1.3 because 1.3 is a repeating
2060 decimal in binary. Floating point constants must have a :ref:`floating
2061 point <t_floating>` type.
2062 **Null pointer constants**
2063 The identifier '``null``' is recognized as a null pointer constant
2064 and must be of :ref:`pointer type <t_pointer>`.
2066 The one non-intuitive notation for constants is the hexadecimal form of
2067 floating point constants. For example, the form
2068 '``double 0x432ff973cafa8000``' is equivalent to (but harder to read
2069 than) '``double 4.5e+15``'. The only time hexadecimal floating point
2070 constants are required (and the only time that they are generated by the
2071 disassembler) is when a floating point constant must be emitted but it
2072 cannot be represented as a decimal floating point number in a reasonable
2073 number of digits. For example, NaN's, infinities, and other special
2074 values are represented in their IEEE hexadecimal format so that assembly
2075 and disassembly do not cause any bits to change in the constants.
2077 When using the hexadecimal form, constants of types half, float, and
2078 double are represented using the 16-digit form shown above (which
2079 matches the IEEE754 representation for double); half and float values
2080 must, however, be exactly representable as IEEE 754 half and single
2081 precision, respectively. Hexadecimal format is always used for long
2082 double, and there are three forms of long double. The 80-bit format used
2083 by x86 is represented as ``0xK`` followed by 20 hexadecimal digits. The
2084 128-bit format used by PowerPC (two adjacent doubles) is represented by
2085 ``0xM`` followed by 32 hexadecimal digits. The IEEE 128-bit format is
2086 represented by ``0xL`` followed by 32 hexadecimal digits. Long doubles
2087 will only work if they match the long double format on your target.
2088 The IEEE 16-bit format (half precision) is represented by ``0xH``
2089 followed by 4 hexadecimal digits. All hexadecimal formats are big-endian
2090 (sign bit at the left).
2092 There are no constants of type x86_mmx.
2094 .. _complexconstants:
2099 Complex constants are a (potentially recursive) combination of simple
2100 constants and smaller complex constants.
2102 **Structure constants**
2103 Structure constants are represented with notation similar to
2104 structure type definitions (a comma separated list of elements,
2105 surrounded by braces (``{}``)). For example:
2106 "``{ i32 4, float 17.0, i32* @G }``", where "``@G``" is declared as
2107 "``@G = external global i32``". Structure constants must have
2108 :ref:`structure type <t_struct>`, and the number and types of elements
2109 must match those specified by the type.
2111 Array constants are represented with notation similar to array type
2112 definitions (a comma separated list of elements, surrounded by
2113 square brackets (``[]``)). For example:
2114 "``[ i32 42, i32 11, i32 74 ]``". Array constants must have
2115 :ref:`array type <t_array>`, and the number and types of elements must
2116 match those specified by the type.
2117 **Vector constants**
2118 Vector constants are represented with notation similar to vector
2119 type definitions (a comma separated list of elements, surrounded by
2120 less-than/greater-than's (``<>``)). For example:
2121 "``< i32 42, i32 11, i32 74, i32 100 >``". Vector constants
2122 must have :ref:`vector type <t_vector>`, and the number and types of
2123 elements must match those specified by the type.
2124 **Zero initialization**
2125 The string '``zeroinitializer``' can be used to zero initialize a
2126 value to zero of *any* type, including scalar and
2127 :ref:`aggregate <t_aggregate>` types. This is often used to avoid
2128 having to print large zero initializers (e.g. for large arrays) and
2129 is always exactly equivalent to using explicit zero initializers.
2131 A metadata node is a structure-like constant with :ref:`metadata
2132 type <t_metadata>`. For example:
2133 "``metadata !{ i32 0, metadata !"test" }``". Unlike other
2134 constants that are meant to be interpreted as part of the
2135 instruction stream, metadata is a place to attach additional
2136 information such as debug info.
2138 Global Variable and Function Addresses
2139 --------------------------------------
2141 The addresses of :ref:`global variables <globalvars>` and
2142 :ref:`functions <functionstructure>` are always implicitly valid
2143 (link-time) constants. These constants are explicitly referenced when
2144 the :ref:`identifier for the global <identifiers>` is used and always have
2145 :ref:`pointer <t_pointer>` type. For example, the following is a legal LLVM
2148 .. code-block:: llvm
2152 @Z = global [2 x i32*] [ i32* @X, i32* @Y ]
2159 The string '``undef``' can be used anywhere a constant is expected, and
2160 indicates that the user of the value may receive an unspecified
2161 bit-pattern. Undefined values may be of any type (other than '``label``'
2162 or '``void``') and be used anywhere a constant is permitted.
2164 Undefined values are useful because they indicate to the compiler that
2165 the program is well defined no matter what value is used. This gives the
2166 compiler more freedom to optimize. Here are some examples of
2167 (potentially surprising) transformations that are valid (in pseudo IR):
2169 .. code-block:: llvm
2179 This is safe because all of the output bits are affected by the undef
2180 bits. Any output bit can have a zero or one depending on the input bits.
2182 .. code-block:: llvm
2193 These logical operations have bits that are not always affected by the
2194 input. For example, if ``%X`` has a zero bit, then the output of the
2195 '``and``' operation will always be a zero for that bit, no matter what
2196 the corresponding bit from the '``undef``' is. As such, it is unsafe to
2197 optimize or assume that the result of the '``and``' is '``undef``'.
2198 However, it is safe to assume that all bits of the '``undef``' could be
2199 0, and optimize the '``and``' to 0. Likewise, it is safe to assume that
2200 all the bits of the '``undef``' operand to the '``or``' could be set,
2201 allowing the '``or``' to be folded to -1.
2203 .. code-block:: llvm
2205 %A = select undef, %X, %Y
2206 %B = select undef, 42, %Y
2207 %C = select %X, %Y, undef
2217 This set of examples shows that undefined '``select``' (and conditional
2218 branch) conditions can go *either way*, but they have to come from one
2219 of the two operands. In the ``%A`` example, if ``%X`` and ``%Y`` were
2220 both known to have a clear low bit, then ``%A`` would have to have a
2221 cleared low bit. However, in the ``%C`` example, the optimizer is
2222 allowed to assume that the '``undef``' operand could be the same as
2223 ``%Y``, allowing the whole '``select``' to be eliminated.
2225 .. code-block:: llvm
2227 %A = xor undef, undef
2244 This example points out that two '``undef``' operands are not
2245 necessarily the same. This can be surprising to people (and also matches
2246 C semantics) where they assume that "``X^X``" is always zero, even if
2247 ``X`` is undefined. This isn't true for a number of reasons, but the
2248 short answer is that an '``undef``' "variable" can arbitrarily change
2249 its value over its "live range". This is true because the variable
2250 doesn't actually *have a live range*. Instead, the value is logically
2251 read from arbitrary registers that happen to be around when needed, so
2252 the value is not necessarily consistent over time. In fact, ``%A`` and
2253 ``%C`` need to have the same semantics or the core LLVM "replace all
2254 uses with" concept would not hold.
2256 .. code-block:: llvm
2264 These examples show the crucial difference between an *undefined value*
2265 and *undefined behavior*. An undefined value (like '``undef``') is
2266 allowed to have an arbitrary bit-pattern. This means that the ``%A``
2267 operation can be constant folded to '``undef``', because the '``undef``'
2268 could be an SNaN, and ``fdiv`` is not (currently) defined on SNaN's.
2269 However, in the second example, we can make a more aggressive
2270 assumption: because the ``undef`` is allowed to be an arbitrary value,
2271 we are allowed to assume that it could be zero. Since a divide by zero
2272 has *undefined behavior*, we are allowed to assume that the operation
2273 does not execute at all. This allows us to delete the divide and all
2274 code after it. Because the undefined operation "can't happen", the
2275 optimizer can assume that it occurs in dead code.
2277 .. code-block:: llvm
2279 a: store undef -> %X
2280 b: store %X -> undef
2285 These examples reiterate the ``fdiv`` example: a store *of* an undefined
2286 value can be assumed to not have any effect; we can assume that the
2287 value is overwritten with bits that happen to match what was already
2288 there. However, a store *to* an undefined location could clobber
2289 arbitrary memory, therefore, it has undefined behavior.
2296 Poison values are similar to :ref:`undef values <undefvalues>`, however
2297 they also represent the fact that an instruction or constant expression
2298 which cannot evoke side effects has nevertheless detected a condition
2299 which results in undefined behavior.
2301 There is currently no way of representing a poison value in the IR; they
2302 only exist when produced by operations such as :ref:`add <i_add>` with
2305 Poison value behavior is defined in terms of value *dependence*:
2307 - Values other than :ref:`phi <i_phi>` nodes depend on their operands.
2308 - :ref:`Phi <i_phi>` nodes depend on the operand corresponding to
2309 their dynamic predecessor basic block.
2310 - Function arguments depend on the corresponding actual argument values
2311 in the dynamic callers of their functions.
2312 - :ref:`Call <i_call>` instructions depend on the :ref:`ret <i_ret>`
2313 instructions that dynamically transfer control back to them.
2314 - :ref:`Invoke <i_invoke>` instructions depend on the
2315 :ref:`ret <i_ret>`, :ref:`resume <i_resume>`, or exception-throwing
2316 call instructions that dynamically transfer control back to them.
2317 - Non-volatile loads and stores depend on the most recent stores to all
2318 of the referenced memory addresses, following the order in the IR
2319 (including loads and stores implied by intrinsics such as
2320 :ref:`@llvm.memcpy <int_memcpy>`.)
2321 - An instruction with externally visible side effects depends on the
2322 most recent preceding instruction with externally visible side
2323 effects, following the order in the IR. (This includes :ref:`volatile
2324 operations <volatile>`.)
2325 - An instruction *control-depends* on a :ref:`terminator
2326 instruction <terminators>` if the terminator instruction has
2327 multiple successors and the instruction is always executed when
2328 control transfers to one of the successors, and may not be executed
2329 when control is transferred to another.
2330 - Additionally, an instruction also *control-depends* on a terminator
2331 instruction if the set of instructions it otherwise depends on would
2332 be different if the terminator had transferred control to a different
2334 - Dependence is transitive.
2336 Poison Values have the same behavior as :ref:`undef values <undefvalues>`,
2337 with the additional affect that any instruction which has a *dependence*
2338 on a poison value has undefined behavior.
2340 Here are some examples:
2342 .. code-block:: llvm
2345 %poison = sub nuw i32 0, 1 ; Results in a poison value.
2346 %still_poison = and i32 %poison, 0 ; 0, but also poison.
2347 %poison_yet_again = getelementptr i32* @h, i32 %still_poison
2348 store i32 0, i32* %poison_yet_again ; memory at @h[0] is poisoned
2350 store i32 %poison, i32* @g ; Poison value stored to memory.
2351 %poison2 = load i32* @g ; Poison value loaded back from memory.
2353 store volatile i32 %poison, i32* @g ; External observation; undefined behavior.
2355 %narrowaddr = bitcast i32* @g to i16*
2356 %wideaddr = bitcast i32* @g to i64*
2357 %poison3 = load i16* %narrowaddr ; Returns a poison value.
2358 %poison4 = load i64* %wideaddr ; Returns a poison value.
2360 %cmp = icmp slt i32 %poison, 0 ; Returns a poison value.
2361 br i1 %cmp, label %true, label %end ; Branch to either destination.
2364 store volatile i32 0, i32* @g ; This is control-dependent on %cmp, so
2365 ; it has undefined behavior.
2369 %p = phi i32 [ 0, %entry ], [ 1, %true ]
2370 ; Both edges into this PHI are
2371 ; control-dependent on %cmp, so this
2372 ; always results in a poison value.
2374 store volatile i32 0, i32* @g ; This would depend on the store in %true
2375 ; if %cmp is true, or the store in %entry
2376 ; otherwise, so this is undefined behavior.
2378 br i1 %cmp, label %second_true, label %second_end
2379 ; The same branch again, but this time the
2380 ; true block doesn't have side effects.
2387 store volatile i32 0, i32* @g ; This time, the instruction always depends
2388 ; on the store in %end. Also, it is
2389 ; control-equivalent to %end, so this is
2390 ; well-defined (ignoring earlier undefined
2391 ; behavior in this example).
2395 Addresses of Basic Blocks
2396 -------------------------
2398 ``blockaddress(@function, %block)``
2400 The '``blockaddress``' constant computes the address of the specified
2401 basic block in the specified function, and always has an ``i8*`` type.
2402 Taking the address of the entry block is illegal.
2404 This value only has defined behavior when used as an operand to the
2405 ':ref:`indirectbr <i_indirectbr>`' instruction, or for comparisons
2406 against null. Pointer equality tests between labels addresses results in
2407 undefined behavior --- though, again, comparison against null is ok, and
2408 no label is equal to the null pointer. This may be passed around as an
2409 opaque pointer sized value as long as the bits are not inspected. This
2410 allows ``ptrtoint`` and arithmetic to be performed on these values so
2411 long as the original value is reconstituted before the ``indirectbr``
2414 Finally, some targets may provide defined semantics when using the value
2415 as the operand to an inline assembly, but that is target specific.
2419 Constant Expressions
2420 --------------------
2422 Constant expressions are used to allow expressions involving other
2423 constants to be used as constants. Constant expressions may be of any
2424 :ref:`first class <t_firstclass>` type and may involve any LLVM operation
2425 that does not have side effects (e.g. load and call are not supported).
2426 The following is the syntax for constant expressions:
2428 ``trunc (CST to TYPE)``
2429 Truncate a constant to another type. The bit size of CST must be
2430 larger than the bit size of TYPE. Both types must be integers.
2431 ``zext (CST to TYPE)``
2432 Zero extend a constant to another type. The bit size of CST must be
2433 smaller than the bit size of TYPE. Both types must be integers.
2434 ``sext (CST to TYPE)``
2435 Sign extend a constant to another type. The bit size of CST must be
2436 smaller than the bit size of TYPE. Both types must be integers.
2437 ``fptrunc (CST to TYPE)``
2438 Truncate a floating point constant to another floating point type.
2439 The size of CST must be larger than the size of TYPE. Both types
2440 must be floating point.
2441 ``fpext (CST to TYPE)``
2442 Floating point extend a constant to another type. The size of CST
2443 must be smaller or equal to the size of TYPE. Both types must be
2445 ``fptoui (CST to TYPE)``
2446 Convert a floating point constant to the corresponding unsigned
2447 integer constant. TYPE must be a scalar or vector integer type. CST
2448 must be of scalar or vector floating point type. Both CST and TYPE
2449 must be scalars, or vectors of the same number of elements. If the
2450 value won't fit in the integer type, the results are undefined.
2451 ``fptosi (CST to TYPE)``
2452 Convert a floating point constant to the corresponding signed
2453 integer constant. TYPE must be a scalar or vector integer type. CST
2454 must be of scalar or vector floating point type. Both CST and TYPE
2455 must be scalars, or vectors of the same number of elements. If the
2456 value won't fit in the integer type, the results are undefined.
2457 ``uitofp (CST to TYPE)``
2458 Convert an unsigned integer constant to the corresponding floating
2459 point constant. TYPE must be a scalar or vector floating point type.
2460 CST must be of scalar or vector integer type. Both CST and TYPE must
2461 be scalars, or vectors of the same number of elements. If the value
2462 won't fit in the floating point type, the results are undefined.
2463 ``sitofp (CST to TYPE)``
2464 Convert a signed integer constant to the corresponding floating
2465 point constant. TYPE must be a scalar or vector floating point type.
2466 CST must be of scalar or vector integer type. Both CST and TYPE must
2467 be scalars, or vectors of the same number of elements. If the value
2468 won't fit in the floating point type, the results are undefined.
2469 ``ptrtoint (CST to TYPE)``
2470 Convert a pointer typed constant to the corresponding integer
2471 constant. ``TYPE`` must be an integer type. ``CST`` must be of
2472 pointer type. The ``CST`` value is zero extended, truncated, or
2473 unchanged to make it fit in ``TYPE``.
2474 ``inttoptr (CST to TYPE)``
2475 Convert an integer constant to a pointer constant. TYPE must be a
2476 pointer type. CST must be of integer type. The CST value is zero
2477 extended, truncated, or unchanged to make it fit in a pointer size.
2478 This one is *really* dangerous!
2479 ``bitcast (CST to TYPE)``
2480 Convert a constant, CST, to another TYPE. The constraints of the
2481 operands are the same as those for the :ref:`bitcast
2482 instruction <i_bitcast>`.
2483 ``addrspacecast (CST to TYPE)``
2484 Convert a constant pointer or constant vector of pointer, CST, to another
2485 TYPE in a different address space. The constraints of the operands are the
2486 same as those for the :ref:`addrspacecast instruction <i_addrspacecast>`.
2487 ``getelementptr (CSTPTR, IDX0, IDX1, ...)``, ``getelementptr inbounds (CSTPTR, IDX0, IDX1, ...)``
2488 Perform the :ref:`getelementptr operation <i_getelementptr>` on
2489 constants. As with the :ref:`getelementptr <i_getelementptr>`
2490 instruction, the index list may have zero or more indexes, which are
2491 required to make sense for the type of "CSTPTR".
2492 ``select (COND, VAL1, VAL2)``
2493 Perform the :ref:`select operation <i_select>` on constants.
2494 ``icmp COND (VAL1, VAL2)``
2495 Performs the :ref:`icmp operation <i_icmp>` on constants.
2496 ``fcmp COND (VAL1, VAL2)``
2497 Performs the :ref:`fcmp operation <i_fcmp>` on constants.
2498 ``extractelement (VAL, IDX)``
2499 Perform the :ref:`extractelement operation <i_extractelement>` on
2501 ``insertelement (VAL, ELT, IDX)``
2502 Perform the :ref:`insertelement operation <i_insertelement>` on
2504 ``shufflevector (VEC1, VEC2, IDXMASK)``
2505 Perform the :ref:`shufflevector operation <i_shufflevector>` on
2507 ``extractvalue (VAL, IDX0, IDX1, ...)``
2508 Perform the :ref:`extractvalue operation <i_extractvalue>` on
2509 constants. The index list is interpreted in a similar manner as
2510 indices in a ':ref:`getelementptr <i_getelementptr>`' operation. At
2511 least one index value must be specified.
2512 ``insertvalue (VAL, ELT, IDX0, IDX1, ...)``
2513 Perform the :ref:`insertvalue operation <i_insertvalue>` on constants.
2514 The index list is interpreted in a similar manner as indices in a
2515 ':ref:`getelementptr <i_getelementptr>`' operation. At least one index
2516 value must be specified.
2517 ``OPCODE (LHS, RHS)``
2518 Perform the specified operation of the LHS and RHS constants. OPCODE
2519 may be any of the :ref:`binary <binaryops>` or :ref:`bitwise
2520 binary <bitwiseops>` operations. The constraints on operands are
2521 the same as those for the corresponding instruction (e.g. no bitwise
2522 operations on floating point values are allowed).
2529 Inline Assembler Expressions
2530 ----------------------------
2532 LLVM supports inline assembler expressions (as opposed to :ref:`Module-Level
2533 Inline Assembly <moduleasm>`) through the use of a special value. This
2534 value represents the inline assembler as a string (containing the
2535 instructions to emit), a list of operand constraints (stored as a
2536 string), a flag that indicates whether or not the inline asm expression
2537 has side effects, and a flag indicating whether the function containing
2538 the asm needs to align its stack conservatively. An example inline
2539 assembler expression is:
2541 .. code-block:: llvm
2543 i32 (i32) asm "bswap $0", "=r,r"
2545 Inline assembler expressions may **only** be used as the callee operand
2546 of a :ref:`call <i_call>` or an :ref:`invoke <i_invoke>` instruction.
2547 Thus, typically we have:
2549 .. code-block:: llvm
2551 %X = call i32 asm "bswap $0", "=r,r"(i32 %Y)
2553 Inline asms with side effects not visible in the constraint list must be
2554 marked as having side effects. This is done through the use of the
2555 '``sideeffect``' keyword, like so:
2557 .. code-block:: llvm
2559 call void asm sideeffect "eieio", ""()
2561 In some cases inline asms will contain code that will not work unless
2562 the stack is aligned in some way, such as calls or SSE instructions on
2563 x86, yet will not contain code that does that alignment within the asm.
2564 The compiler should make conservative assumptions about what the asm
2565 might contain and should generate its usual stack alignment code in the
2566 prologue if the '``alignstack``' keyword is present:
2568 .. code-block:: llvm
2570 call void asm alignstack "eieio", ""()
2572 Inline asms also support using non-standard assembly dialects. The
2573 assumed dialect is ATT. When the '``inteldialect``' keyword is present,
2574 the inline asm is using the Intel dialect. Currently, ATT and Intel are
2575 the only supported dialects. An example is:
2577 .. code-block:: llvm
2579 call void asm inteldialect "eieio", ""()
2581 If multiple keywords appear the '``sideeffect``' keyword must come
2582 first, the '``alignstack``' keyword second and the '``inteldialect``'
2588 The call instructions that wrap inline asm nodes may have a
2589 "``!srcloc``" MDNode attached to it that contains a list of constant
2590 integers. If present, the code generator will use the integer as the
2591 location cookie value when report errors through the ``LLVMContext``
2592 error reporting mechanisms. This allows a front-end to correlate backend
2593 errors that occur with inline asm back to the source code that produced
2596 .. code-block:: llvm
2598 call void asm sideeffect "something bad", ""(), !srcloc !42
2600 !42 = !{ i32 1234567 }
2602 It is up to the front-end to make sense of the magic numbers it places
2603 in the IR. If the MDNode contains multiple constants, the code generator
2604 will use the one that corresponds to the line of the asm that the error
2609 Metadata Nodes and Metadata Strings
2610 -----------------------------------
2612 LLVM IR allows metadata to be attached to instructions in the program
2613 that can convey extra information about the code to the optimizers and
2614 code generator. One example application of metadata is source-level
2615 debug information. There are two metadata primitives: strings and nodes.
2616 All metadata has the ``metadata`` type and is identified in syntax by a
2617 preceding exclamation point ('``!``').
2619 A metadata string is a string surrounded by double quotes. It can
2620 contain any character by escaping non-printable characters with
2621 "``\xx``" where "``xx``" is the two digit hex code. For example:
2624 Metadata nodes are represented with notation similar to structure
2625 constants (a comma separated list of elements, surrounded by braces and
2626 preceded by an exclamation point). Metadata nodes can have any values as
2627 their operand. For example:
2629 .. code-block:: llvm
2631 !{ metadata !"test\00", i32 10}
2633 A :ref:`named metadata <namedmetadatastructure>` is a collection of
2634 metadata nodes, which can be looked up in the module symbol table. For
2637 .. code-block:: llvm
2639 !foo = metadata !{!4, !3}
2641 Metadata can be used as function arguments. Here ``llvm.dbg.value``
2642 function is using two metadata arguments:
2644 .. code-block:: llvm
2646 call void @llvm.dbg.value(metadata !24, i64 0, metadata !25)
2648 Metadata can be attached with an instruction. Here metadata ``!21`` is
2649 attached to the ``add`` instruction using the ``!dbg`` identifier:
2651 .. code-block:: llvm
2653 %indvar.next = add i64 %indvar, 1, !dbg !21
2655 More information about specific metadata nodes recognized by the
2656 optimizers and code generator is found below.
2661 In LLVM IR, memory does not have types, so LLVM's own type system is not
2662 suitable for doing TBAA. Instead, metadata is added to the IR to
2663 describe a type system of a higher level language. This can be used to
2664 implement typical C/C++ TBAA, but it can also be used to implement
2665 custom alias analysis behavior for other languages.
2667 The current metadata format is very simple. TBAA metadata nodes have up
2668 to three fields, e.g.:
2670 .. code-block:: llvm
2672 !0 = metadata !{ metadata !"an example type tree" }
2673 !1 = metadata !{ metadata !"int", metadata !0 }
2674 !2 = metadata !{ metadata !"float", metadata !0 }
2675 !3 = metadata !{ metadata !"const float", metadata !2, i64 1 }
2677 The first field is an identity field. It can be any value, usually a
2678 metadata string, which uniquely identifies the type. The most important
2679 name in the tree is the name of the root node. Two trees with different
2680 root node names are entirely disjoint, even if they have leaves with
2683 The second field identifies the type's parent node in the tree, or is
2684 null or omitted for a root node. A type is considered to alias all of
2685 its descendants and all of its ancestors in the tree. Also, a type is
2686 considered to alias all types in other trees, so that bitcode produced
2687 from multiple front-ends is handled conservatively.
2689 If the third field is present, it's an integer which if equal to 1
2690 indicates that the type is "constant" (meaning
2691 ``pointsToConstantMemory`` should return true; see `other useful
2692 AliasAnalysis methods <AliasAnalysis.html#OtherItfs>`_).
2694 '``tbaa.struct``' Metadata
2695 ^^^^^^^^^^^^^^^^^^^^^^^^^^
2697 The :ref:`llvm.memcpy <int_memcpy>` is often used to implement
2698 aggregate assignment operations in C and similar languages, however it
2699 is defined to copy a contiguous region of memory, which is more than
2700 strictly necessary for aggregate types which contain holes due to
2701 padding. Also, it doesn't contain any TBAA information about the fields
2704 ``!tbaa.struct`` metadata can describe which memory subregions in a
2705 memcpy are padding and what the TBAA tags of the struct are.
2707 The current metadata format is very simple. ``!tbaa.struct`` metadata
2708 nodes are a list of operands which are in conceptual groups of three.
2709 For each group of three, the first operand gives the byte offset of a
2710 field in bytes, the second gives its size in bytes, and the third gives
2713 .. code-block:: llvm
2715 !4 = metadata !{ i64 0, i64 4, metadata !1, i64 8, i64 4, metadata !2 }
2717 This describes a struct with two fields. The first is at offset 0 bytes
2718 with size 4 bytes, and has tbaa tag !1. The second is at offset 8 bytes
2719 and has size 4 bytes and has tbaa tag !2.
2721 Note that the fields need not be contiguous. In this example, there is a
2722 4 byte gap between the two fields. This gap represents padding which
2723 does not carry useful data and need not be preserved.
2725 '``fpmath``' Metadata
2726 ^^^^^^^^^^^^^^^^^^^^^
2728 ``fpmath`` metadata may be attached to any instruction of floating point
2729 type. It can be used to express the maximum acceptable error in the
2730 result of that instruction, in ULPs, thus potentially allowing the
2731 compiler to use a more efficient but less accurate method of computing
2732 it. ULP is defined as follows:
2734 If ``x`` is a real number that lies between two finite consecutive
2735 floating-point numbers ``a`` and ``b``, without being equal to one
2736 of them, then ``ulp(x) = |b - a|``, otherwise ``ulp(x)`` is the
2737 distance between the two non-equal finite floating-point numbers
2738 nearest ``x``. Moreover, ``ulp(NaN)`` is ``NaN``.
2740 The metadata node shall consist of a single positive floating point
2741 number representing the maximum relative error, for example:
2743 .. code-block:: llvm
2745 !0 = metadata !{ float 2.5 } ; maximum acceptable inaccuracy is 2.5 ULPs
2747 '``range``' Metadata
2748 ^^^^^^^^^^^^^^^^^^^^
2750 ``range`` metadata may be attached only to loads of integer types. It
2751 expresses the possible ranges the loaded value is in. The ranges are
2752 represented with a flattened list of integers. The loaded value is known
2753 to be in the union of the ranges defined by each consecutive pair. Each
2754 pair has the following properties:
2756 - The type must match the type loaded by the instruction.
2757 - The pair ``a,b`` represents the range ``[a,b)``.
2758 - Both ``a`` and ``b`` are constants.
2759 - The range is allowed to wrap.
2760 - The range should not represent the full or empty set. That is,
2763 In addition, the pairs must be in signed order of the lower bound and
2764 they must be non-contiguous.
2768 .. code-block:: llvm
2770 %a = load i8* %x, align 1, !range !0 ; Can only be 0 or 1
2771 %b = load i8* %y, align 1, !range !1 ; Can only be 255 (-1), 0 or 1
2772 %c = load i8* %z, align 1, !range !2 ; Can only be 0, 1, 3, 4 or 5
2773 %d = load i8* %z, align 1, !range !3 ; Can only be -2, -1, 3, 4 or 5
2775 !0 = metadata !{ i8 0, i8 2 }
2776 !1 = metadata !{ i8 255, i8 2 }
2777 !2 = metadata !{ i8 0, i8 2, i8 3, i8 6 }
2778 !3 = metadata !{ i8 -2, i8 0, i8 3, i8 6 }
2783 It is sometimes useful to attach information to loop constructs. Currently,
2784 loop metadata is implemented as metadata attached to the branch instruction
2785 in the loop latch block. This type of metadata refer to a metadata node that is
2786 guaranteed to be separate for each loop. The loop identifier metadata is
2787 specified with the name ``llvm.loop``.
2789 The loop identifier metadata is implemented using a metadata that refers to
2790 itself to avoid merging it with any other identifier metadata, e.g.,
2791 during module linkage or function inlining. That is, each loop should refer
2792 to their own identification metadata even if they reside in separate functions.
2793 The following example contains loop identifier metadata for two separate loop
2796 .. code-block:: llvm
2798 !0 = metadata !{ metadata !0 }
2799 !1 = metadata !{ metadata !1 }
2801 The loop identifier metadata can be used to specify additional per-loop
2802 metadata. Any operands after the first operand can be treated as user-defined
2803 metadata. For example the ``llvm.vectorizer.unroll`` metadata is understood
2804 by the loop vectorizer to indicate how many times to unroll the loop:
2806 .. code-block:: llvm
2808 br i1 %exitcond, label %._crit_edge, label %.lr.ph, !llvm.loop !0
2810 !0 = metadata !{ metadata !0, metadata !1 }
2811 !1 = metadata !{ metadata !"llvm.vectorizer.unroll", i32 2 }
2816 Metadata types used to annotate memory accesses with information helpful
2817 for optimizations are prefixed with ``llvm.mem``.
2819 '``llvm.mem.parallel_loop_access``' Metadata
2820 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2822 The ``llvm.mem.parallel_loop_access`` metadata refers to a loop identifier,
2823 or metadata containing a list of loop identifiers for nested loops.
2824 The metadata is attached to memory accessing instructions and denotes that
2825 no loop carried memory dependence exist between it and other instructions denoted
2826 with the same loop identifier.
2828 Precisely, given two instructions ``m1`` and ``m2`` that both have the
2829 ``llvm.mem.parallel_loop_access`` metadata, with ``L1`` and ``L2`` being the
2830 set of loops associated with that metadata, respectively, then there is no loop
2831 carried dependence between ``m1`` and ``m2`` for loops in both ``L1`` and
2834 As a special case, if all memory accessing instructions in a loop have
2835 ``llvm.mem.parallel_loop_access`` metadata that refers to that loop, then the
2836 loop has no loop carried memory dependences and is considered to be a parallel
2839 Note that if not all memory access instructions have such metadata referring to
2840 the loop, then the loop is considered not being trivially parallel. Additional
2841 memory dependence analysis is required to make that determination. As a fail
2842 safe mechanism, this causes loops that were originally parallel to be considered
2843 sequential (if optimization passes that are unaware of the parallel semantics
2844 insert new memory instructions into the loop body).
2846 Example of a loop that is considered parallel due to its correct use of
2847 both ``llvm.loop`` and ``llvm.mem.parallel_loop_access``
2848 metadata types that refer to the same loop identifier metadata.
2850 .. code-block:: llvm
2854 %val0 = load i32* %arrayidx, !llvm.mem.parallel_loop_access !0
2856 store i32 %val0, i32* %arrayidx1, !llvm.mem.parallel_loop_access !0
2858 br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !0
2862 !0 = metadata !{ metadata !0 }
2864 It is also possible to have nested parallel loops. In that case the
2865 memory accesses refer to a list of loop identifier metadata nodes instead of
2866 the loop identifier metadata node directly:
2868 .. code-block:: llvm
2872 %val1 = load i32* %arrayidx3, !llvm.mem.parallel_loop_access !2
2874 br label %inner.for.body
2878 %val0 = load i32* %arrayidx1, !llvm.mem.parallel_loop_access !0
2880 store i32 %val0, i32* %arrayidx2, !llvm.mem.parallel_loop_access !0
2882 br i1 %exitcond, label %inner.for.end, label %inner.for.body, !llvm.loop !1
2886 store i32 %val1, i32* %arrayidx4, !llvm.mem.parallel_loop_access !2
2888 br i1 %exitcond, label %outer.for.end, label %outer.for.body, !llvm.loop !2
2890 outer.for.end: ; preds = %for.body
2892 !0 = metadata !{ metadata !1, metadata !2 } ; a list of loop identifiers
2893 !1 = metadata !{ metadata !1 } ; an identifier for the inner loop
2894 !2 = metadata !{ metadata !2 } ; an identifier for the outer loop
2896 '``llvm.vectorizer``'
2897 ^^^^^^^^^^^^^^^^^^^^^
2899 Metadata prefixed with ``llvm.vectorizer`` is used to control per-loop
2900 vectorization parameters such as vectorization factor and unroll factor.
2902 ``llvm.vectorizer`` metadata should be used in conjunction with ``llvm.loop``
2903 loop identification metadata.
2905 '``llvm.vectorizer.unroll``' Metadata
2906 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2908 This metadata instructs the loop vectorizer to unroll the specified
2909 loop exactly ``N`` times.
2911 The first operand is the string ``llvm.vectorizer.unroll`` and the second
2912 operand is an integer specifying the unroll factor. For example:
2914 .. code-block:: llvm
2916 !0 = metadata !{ metadata !"llvm.vectorizer.unroll", i32 4 }
2918 Note that setting ``llvm.vectorizer.unroll`` to 1 disables unrolling of the
2921 If ``llvm.vectorizer.unroll`` is set to 0 then the amount of unrolling will be
2922 determined automatically.
2924 '``llvm.vectorizer.width``' Metadata
2925 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2927 This metadata sets the target width of the vectorizer to ``N``. Without
2928 this metadata, the vectorizer will choose a width automatically.
2929 Regardless of this metadata, the vectorizer will only vectorize loops if
2930 it believes it is valid to do so.
2932 The first operand is the string ``llvm.vectorizer.width`` and the second
2933 operand is an integer specifying the width. For example:
2935 .. code-block:: llvm
2937 !0 = metadata !{ metadata !"llvm.vectorizer.width", i32 4 }
2939 Note that setting ``llvm.vectorizer.width`` to 1 disables vectorization of the
2942 If ``llvm.vectorizer.width`` is set to 0 then the width will be determined
2945 Module Flags Metadata
2946 =====================
2948 Information about the module as a whole is difficult to convey to LLVM's
2949 subsystems. The LLVM IR isn't sufficient to transmit this information.
2950 The ``llvm.module.flags`` named metadata exists in order to facilitate
2951 this. These flags are in the form of key / value pairs --- much like a
2952 dictionary --- making it easy for any subsystem who cares about a flag to
2955 The ``llvm.module.flags`` metadata contains a list of metadata triplets.
2956 Each triplet has the following form:
2958 - The first element is a *behavior* flag, which specifies the behavior
2959 when two (or more) modules are merged together, and it encounters two
2960 (or more) metadata with the same ID. The supported behaviors are
2962 - The second element is a metadata string that is a unique ID for the
2963 metadata. Each module may only have one flag entry for each unique ID (not
2964 including entries with the **Require** behavior).
2965 - The third element is the value of the flag.
2967 When two (or more) modules are merged together, the resulting
2968 ``llvm.module.flags`` metadata is the union of the modules' flags. That is, for
2969 each unique metadata ID string, there will be exactly one entry in the merged
2970 modules ``llvm.module.flags`` metadata table, and the value for that entry will
2971 be determined by the merge behavior flag, as described below. The only exception
2972 is that entries with the *Require* behavior are always preserved.
2974 The following behaviors are supported:
2985 Emits an error if two values disagree, otherwise the resulting value
2986 is that of the operands.
2990 Emits a warning if two values disagree. The result value will be the
2991 operand for the flag from the first module being linked.
2995 Adds a requirement that another module flag be present and have a
2996 specified value after linking is performed. The value must be a
2997 metadata pair, where the first element of the pair is the ID of the
2998 module flag to be restricted, and the second element of the pair is
2999 the value the module flag should be restricted to. This behavior can
3000 be used to restrict the allowable results (via triggering of an
3001 error) of linking IDs with the **Override** behavior.
3005 Uses the specified value, regardless of the behavior or value of the
3006 other module. If both modules specify **Override**, but the values
3007 differ, an error will be emitted.
3011 Appends the two values, which are required to be metadata nodes.
3015 Appends the two values, which are required to be metadata
3016 nodes. However, duplicate entries in the second list are dropped
3017 during the append operation.
3019 It is an error for a particular unique flag ID to have multiple behaviors,
3020 except in the case of **Require** (which adds restrictions on another metadata
3021 value) or **Override**.
3023 An example of module flags:
3025 .. code-block:: llvm
3027 !0 = metadata !{ i32 1, metadata !"foo", i32 1 }
3028 !1 = metadata !{ i32 4, metadata !"bar", i32 37 }
3029 !2 = metadata !{ i32 2, metadata !"qux", i32 42 }
3030 !3 = metadata !{ i32 3, metadata !"qux",
3032 metadata !"foo", i32 1
3035 !llvm.module.flags = !{ !0, !1, !2, !3 }
3037 - Metadata ``!0`` has the ID ``!"foo"`` and the value '1'. The behavior
3038 if two or more ``!"foo"`` flags are seen is to emit an error if their
3039 values are not equal.
3041 - Metadata ``!1`` has the ID ``!"bar"`` and the value '37'. The
3042 behavior if two or more ``!"bar"`` flags are seen is to use the value
3045 - Metadata ``!2`` has the ID ``!"qux"`` and the value '42'. The
3046 behavior if two or more ``!"qux"`` flags are seen is to emit a
3047 warning if their values are not equal.
3049 - Metadata ``!3`` has the ID ``!"qux"`` and the value:
3053 metadata !{ metadata !"foo", i32 1 }
3055 The behavior is to emit an error if the ``llvm.module.flags`` does not
3056 contain a flag with the ID ``!"foo"`` that has the value '1' after linking is
3059 Objective-C Garbage Collection Module Flags Metadata
3060 ----------------------------------------------------
3062 On the Mach-O platform, Objective-C stores metadata about garbage
3063 collection in a special section called "image info". The metadata
3064 consists of a version number and a bitmask specifying what types of
3065 garbage collection are supported (if any) by the file. If two or more
3066 modules are linked together their garbage collection metadata needs to
3067 be merged rather than appended together.
3069 The Objective-C garbage collection module flags metadata consists of the
3070 following key-value pairs:
3079 * - ``Objective-C Version``
3080 - **[Required]** --- The Objective-C ABI version. Valid values are 1 and 2.
3082 * - ``Objective-C Image Info Version``
3083 - **[Required]** --- The version of the image info section. Currently
3086 * - ``Objective-C Image Info Section``
3087 - **[Required]** --- The section to place the metadata. Valid values are
3088 ``"__OBJC, __image_info, regular"`` for Objective-C ABI version 1, and
3089 ``"__DATA,__objc_imageinfo, regular, no_dead_strip"`` for
3090 Objective-C ABI version 2.
3092 * - ``Objective-C Garbage Collection``
3093 - **[Required]** --- Specifies whether garbage collection is supported or
3094 not. Valid values are 0, for no garbage collection, and 2, for garbage
3095 collection supported.
3097 * - ``Objective-C GC Only``
3098 - **[Optional]** --- Specifies that only garbage collection is supported.
3099 If present, its value must be 6. This flag requires that the
3100 ``Objective-C Garbage Collection`` flag have the value 2.
3102 Some important flag interactions:
3104 - If a module with ``Objective-C Garbage Collection`` set to 0 is
3105 merged with a module with ``Objective-C Garbage Collection`` set to
3106 2, then the resulting module has the
3107 ``Objective-C Garbage Collection`` flag set to 0.
3108 - A module with ``Objective-C Garbage Collection`` set to 0 cannot be
3109 merged with a module with ``Objective-C GC Only`` set to 6.
3111 Automatic Linker Flags Module Flags Metadata
3112 --------------------------------------------
3114 Some targets support embedding flags to the linker inside individual object
3115 files. Typically this is used in conjunction with language extensions which
3116 allow source files to explicitly declare the libraries they depend on, and have
3117 these automatically be transmitted to the linker via object files.
3119 These flags are encoded in the IR using metadata in the module flags section,
3120 using the ``Linker Options`` key. The merge behavior for this flag is required
3121 to be ``AppendUnique``, and the value for the key is expected to be a metadata
3122 node which should be a list of other metadata nodes, each of which should be a
3123 list of metadata strings defining linker options.
3125 For example, the following metadata section specifies two separate sets of
3126 linker options, presumably to link against ``libz`` and the ``Cocoa``
3129 !0 = metadata !{ i32 6, metadata !"Linker Options",
3131 metadata !{ metadata !"-lz" },
3132 metadata !{ metadata !"-framework", metadata !"Cocoa" } } }
3133 !llvm.module.flags = !{ !0 }
3135 The metadata encoding as lists of lists of options, as opposed to a collapsed
3136 list of options, is chosen so that the IR encoding can use multiple option
3137 strings to specify e.g., a single library, while still having that specifier be
3138 preserved as an atomic element that can be recognized by a target specific
3139 assembly writer or object file emitter.
3141 Each individual option is required to be either a valid option for the target's
3142 linker, or an option that is reserved by the target specific assembly writer or
3143 object file emitter. No other aspect of these options is defined by the IR.
3145 .. _intrinsicglobalvariables:
3147 Intrinsic Global Variables
3148 ==========================
3150 LLVM has a number of "magic" global variables that contain data that
3151 affect code generation or other IR semantics. These are documented here.
3152 All globals of this sort should have a section specified as
3153 "``llvm.metadata``". This section and all globals that start with
3154 "``llvm.``" are reserved for use by LLVM.
3158 The '``llvm.used``' Global Variable
3159 -----------------------------------
3161 The ``@llvm.used`` global is an array which has
3162 :ref:`appending linkage <linkage_appending>`. This array contains a list of
3163 pointers to named global variables, functions and aliases which may optionally
3164 have a pointer cast formed of bitcast or getelementptr. For example, a legal
3167 .. code-block:: llvm
3172 @llvm.used = appending global [2 x i8*] [
3174 i8* bitcast (i32* @Y to i8*)
3175 ], section "llvm.metadata"
3177 If a symbol appears in the ``@llvm.used`` list, then the compiler, assembler,
3178 and linker are required to treat the symbol as if there is a reference to the
3179 symbol that it cannot see (which is why they have to be named). For example, if
3180 a variable has internal linkage and no references other than that from the
3181 ``@llvm.used`` list, it cannot be deleted. This is commonly used to represent
3182 references from inline asms and other things the compiler cannot "see", and
3183 corresponds to "``attribute((used))``" in GNU C.
3185 On some targets, the code generator must emit a directive to the
3186 assembler or object file to prevent the assembler and linker from
3187 molesting the symbol.
3189 .. _gv_llvmcompilerused:
3191 The '``llvm.compiler.used``' Global Variable
3192 --------------------------------------------
3194 The ``@llvm.compiler.used`` directive is the same as the ``@llvm.used``
3195 directive, except that it only prevents the compiler from touching the
3196 symbol. On targets that support it, this allows an intelligent linker to
3197 optimize references to the symbol without being impeded as it would be
3200 This is a rare construct that should only be used in rare circumstances,
3201 and should not be exposed to source languages.
3203 .. _gv_llvmglobalctors:
3205 The '``llvm.global_ctors``' Global Variable
3206 -------------------------------------------
3208 .. code-block:: llvm
3210 %0 = type { i32, void ()*, i8* }
3211 @llvm.global_ctors = appending global [1 x %0] [%0 { i32 65535, void ()* @ctor, i8* @data }]
3213 The ``@llvm.global_ctors`` array contains a list of constructor
3214 functions, priorities, and an optional associated global or function.
3215 The functions referenced by this array will be called in ascending order
3216 of priority (i.e. lowest first) when the module is loaded. The order of
3217 functions with the same priority is not defined.
3219 If the third field is present, non-null, and points to a global variable
3220 or function, the initializer function will only run if the associated
3221 data from the current module is not discarded.
3223 .. _llvmglobaldtors:
3225 The '``llvm.global_dtors``' Global Variable
3226 -------------------------------------------
3228 .. code-block:: llvm
3230 %0 = type { i32, void ()*, i8* }
3231 @llvm.global_dtors = appending global [1 x %0] [%0 { i32 65535, void ()* @dtor, i8* @data }]
3233 The ``@llvm.global_dtors`` array contains a list of destructor
3234 functions, priorities, and an optional associated global or function.
3235 The functions referenced by this array will be called in descending
3236 order of priority (i.e. highest first) when the module is unloaded. The
3237 order of functions with the same priority is not defined.
3239 If the third field is present, non-null, and points to a global variable
3240 or function, the destructor function will only run if the associated
3241 data from the current module is not discarded.
3243 Instruction Reference
3244 =====================
3246 The LLVM instruction set consists of several different classifications
3247 of instructions: :ref:`terminator instructions <terminators>`, :ref:`binary
3248 instructions <binaryops>`, :ref:`bitwise binary
3249 instructions <bitwiseops>`, :ref:`memory instructions <memoryops>`, and
3250 :ref:`other instructions <otherops>`.
3254 Terminator Instructions
3255 -----------------------
3257 As mentioned :ref:`previously <functionstructure>`, every basic block in a
3258 program ends with a "Terminator" instruction, which indicates which
3259 block should be executed after the current block is finished. These
3260 terminator instructions typically yield a '``void``' value: they produce
3261 control flow, not values (the one exception being the
3262 ':ref:`invoke <i_invoke>`' instruction).
3264 The terminator instructions are: ':ref:`ret <i_ret>`',
3265 ':ref:`br <i_br>`', ':ref:`switch <i_switch>`',
3266 ':ref:`indirectbr <i_indirectbr>`', ':ref:`invoke <i_invoke>`',
3267 ':ref:`resume <i_resume>`', and ':ref:`unreachable <i_unreachable>`'.
3271 '``ret``' Instruction
3272 ^^^^^^^^^^^^^^^^^^^^^
3279 ret <type> <value> ; Return a value from a non-void function
3280 ret void ; Return from void function
3285 The '``ret``' instruction is used to return control flow (and optionally
3286 a value) from a function back to the caller.
3288 There are two forms of the '``ret``' instruction: one that returns a
3289 value and then causes control flow, and one that just causes control
3295 The '``ret``' instruction optionally accepts a single argument, the
3296 return value. The type of the return value must be a ':ref:`first
3297 class <t_firstclass>`' type.
3299 A function is not :ref:`well formed <wellformed>` if it it has a non-void
3300 return type and contains a '``ret``' instruction with no return value or
3301 a return value with a type that does not match its type, or if it has a
3302 void return type and contains a '``ret``' instruction with a return
3308 When the '``ret``' instruction is executed, control flow returns back to
3309 the calling function's context. If the caller is a
3310 ":ref:`call <i_call>`" instruction, execution continues at the
3311 instruction after the call. If the caller was an
3312 ":ref:`invoke <i_invoke>`" instruction, execution continues at the
3313 beginning of the "normal" destination block. If the instruction returns
3314 a value, that value shall set the call or invoke instruction's return
3320 .. code-block:: llvm
3322 ret i32 5 ; Return an integer value of 5
3323 ret void ; Return from a void function
3324 ret { i32, i8 } { i32 4, i8 2 } ; Return a struct of values 4 and 2
3328 '``br``' Instruction
3329 ^^^^^^^^^^^^^^^^^^^^
3336 br i1 <cond>, label <iftrue>, label <iffalse>
3337 br label <dest> ; Unconditional branch
3342 The '``br``' instruction is used to cause control flow to transfer to a
3343 different basic block in the current function. There are two forms of
3344 this instruction, corresponding to a conditional branch and an
3345 unconditional branch.
3350 The conditional branch form of the '``br``' instruction takes a single
3351 '``i1``' value and two '``label``' values. The unconditional form of the
3352 '``br``' instruction takes a single '``label``' value as a target.
3357 Upon execution of a conditional '``br``' instruction, the '``i1``'
3358 argument is evaluated. If the value is ``true``, control flows to the
3359 '``iftrue``' ``label`` argument. If "cond" is ``false``, control flows
3360 to the '``iffalse``' ``label`` argument.
3365 .. code-block:: llvm
3368 %cond = icmp eq i32 %a, %b
3369 br i1 %cond, label %IfEqual, label %IfUnequal
3377 '``switch``' Instruction
3378 ^^^^^^^^^^^^^^^^^^^^^^^^
3385 switch <intty> <value>, label <defaultdest> [ <intty> <val>, label <dest> ... ]
3390 The '``switch``' instruction is used to transfer control flow to one of
3391 several different places. It is a generalization of the '``br``'
3392 instruction, allowing a branch to occur to one of many possible
3398 The '``switch``' instruction uses three parameters: an integer
3399 comparison value '``value``', a default '``label``' destination, and an
3400 array of pairs of comparison value constants and '``label``'s. The table
3401 is not allowed to contain duplicate constant entries.
3406 The ``switch`` instruction specifies a table of values and destinations.
3407 When the '``switch``' instruction is executed, this table is searched
3408 for the given value. If the value is found, control flow is transferred
3409 to the corresponding destination; otherwise, control flow is transferred
3410 to the default destination.
3415 Depending on properties of the target machine and the particular
3416 ``switch`` instruction, this instruction may be code generated in
3417 different ways. For example, it could be generated as a series of
3418 chained conditional branches or with a lookup table.
3423 .. code-block:: llvm
3425 ; Emulate a conditional br instruction
3426 %Val = zext i1 %value to i32
3427 switch i32 %Val, label %truedest [ i32 0, label %falsedest ]
3429 ; Emulate an unconditional br instruction
3430 switch i32 0, label %dest [ ]
3432 ; Implement a jump table:
3433 switch i32 %val, label %otherwise [ i32 0, label %onzero
3435 i32 2, label %ontwo ]
3439 '``indirectbr``' Instruction
3440 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3447 indirectbr <somety>* <address>, [ label <dest1>, label <dest2>, ... ]
3452 The '``indirectbr``' instruction implements an indirect branch to a
3453 label within the current function, whose address is specified by
3454 "``address``". Address must be derived from a
3455 :ref:`blockaddress <blockaddress>` constant.
3460 The '``address``' argument is the address of the label to jump to. The
3461 rest of the arguments indicate the full set of possible destinations
3462 that the address may point to. Blocks are allowed to occur multiple
3463 times in the destination list, though this isn't particularly useful.
3465 This destination list is required so that dataflow analysis has an
3466 accurate understanding of the CFG.
3471 Control transfers to the block specified in the address argument. All
3472 possible destination blocks must be listed in the label list, otherwise
3473 this instruction has undefined behavior. This implies that jumps to
3474 labels defined in other functions have undefined behavior as well.
3479 This is typically implemented with a jump through a register.
3484 .. code-block:: llvm
3486 indirectbr i8* %Addr, [ label %bb1, label %bb2, label %bb3 ]
3490 '``invoke``' Instruction
3491 ^^^^^^^^^^^^^^^^^^^^^^^^
3498 <result> = invoke [cconv] [ret attrs] <ptr to function ty> <function ptr val>(<function args>) [fn attrs]
3499 to label <normal label> unwind label <exception label>
3504 The '``invoke``' instruction causes control to transfer to a specified
3505 function, with the possibility of control flow transfer to either the
3506 '``normal``' label or the '``exception``' label. If the callee function
3507 returns with the "``ret``" instruction, control flow will return to the
3508 "normal" label. If the callee (or any indirect callees) returns via the
3509 ":ref:`resume <i_resume>`" instruction or other exception handling
3510 mechanism, control is interrupted and continued at the dynamically
3511 nearest "exception" label.
3513 The '``exception``' label is a `landing
3514 pad <ExceptionHandling.html#overview>`_ for the exception. As such,
3515 '``exception``' label is required to have the
3516 ":ref:`landingpad <i_landingpad>`" instruction, which contains the
3517 information about the behavior of the program after unwinding happens,
3518 as its first non-PHI instruction. The restrictions on the
3519 "``landingpad``" instruction's tightly couples it to the "``invoke``"
3520 instruction, so that the important information contained within the
3521 "``landingpad``" instruction can't be lost through normal code motion.
3526 This instruction requires several arguments:
3528 #. The optional "cconv" marker indicates which :ref:`calling
3529 convention <callingconv>` the call should use. If none is
3530 specified, the call defaults to using C calling conventions.
3531 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
3532 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
3534 #. '``ptr to function ty``': shall be the signature of the pointer to
3535 function value being invoked. In most cases, this is a direct
3536 function invocation, but indirect ``invoke``'s are just as possible,
3537 branching off an arbitrary pointer to function value.
3538 #. '``function ptr val``': An LLVM value containing a pointer to a
3539 function to be invoked.
3540 #. '``function args``': argument list whose types match the function
3541 signature argument types and parameter attributes. All arguments must
3542 be of :ref:`first class <t_firstclass>` type. If the function signature
3543 indicates the function accepts a variable number of arguments, the
3544 extra arguments can be specified.
3545 #. '``normal label``': the label reached when the called function
3546 executes a '``ret``' instruction.
3547 #. '``exception label``': the label reached when a callee returns via
3548 the :ref:`resume <i_resume>` instruction or other exception handling
3550 #. The optional :ref:`function attributes <fnattrs>` list. Only
3551 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
3552 attributes are valid here.
3557 This instruction is designed to operate as a standard '``call``'
3558 instruction in most regards. The primary difference is that it
3559 establishes an association with a label, which is used by the runtime
3560 library to unwind the stack.
3562 This instruction is used in languages with destructors to ensure that
3563 proper cleanup is performed in the case of either a ``longjmp`` or a
3564 thrown exception. Additionally, this is important for implementation of
3565 '``catch``' clauses in high-level languages that support them.
3567 For the purposes of the SSA form, the definition of the value returned
3568 by the '``invoke``' instruction is deemed to occur on the edge from the
3569 current block to the "normal" label. If the callee unwinds then no
3570 return value is available.
3575 .. code-block:: llvm
3577 %retval = invoke i32 @Test(i32 15) to label %Continue
3578 unwind label %TestCleanup ; {i32}:retval set
3579 %retval = invoke coldcc i32 %Testfnptr(i32 15) to label %Continue
3580 unwind label %TestCleanup ; {i32}:retval set
3584 '``resume``' Instruction
3585 ^^^^^^^^^^^^^^^^^^^^^^^^
3592 resume <type> <value>
3597 The '``resume``' instruction is a terminator instruction that has no
3603 The '``resume``' instruction requires one argument, which must have the
3604 same type as the result of any '``landingpad``' instruction in the same
3610 The '``resume``' instruction resumes propagation of an existing
3611 (in-flight) exception whose unwinding was interrupted with a
3612 :ref:`landingpad <i_landingpad>` instruction.
3617 .. code-block:: llvm
3619 resume { i8*, i32 } %exn
3623 '``unreachable``' Instruction
3624 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3636 The '``unreachable``' instruction has no defined semantics. This
3637 instruction is used to inform the optimizer that a particular portion of
3638 the code is not reachable. This can be used to indicate that the code
3639 after a no-return function cannot be reached, and other facts.
3644 The '``unreachable``' instruction has no defined semantics.
3651 Binary operators are used to do most of the computation in a program.
3652 They require two operands of the same type, execute an operation on
3653 them, and produce a single value. The operands might represent multiple
3654 data, as is the case with the :ref:`vector <t_vector>` data type. The
3655 result value has the same type as its operands.
3657 There are several different binary operators:
3661 '``add``' Instruction
3662 ^^^^^^^^^^^^^^^^^^^^^
3669 <result> = add <ty> <op1>, <op2> ; yields {ty}:result
3670 <result> = add nuw <ty> <op1>, <op2> ; yields {ty}:result
3671 <result> = add nsw <ty> <op1>, <op2> ; yields {ty}:result
3672 <result> = add nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3677 The '``add``' instruction returns the sum of its two operands.
3682 The two arguments to the '``add``' instruction must be
3683 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3684 arguments must have identical types.
3689 The value produced is the integer sum of the two operands.
3691 If the sum has unsigned overflow, the result returned is the
3692 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
3695 Because LLVM integers use a two's complement representation, this
3696 instruction is appropriate for both signed and unsigned integers.
3698 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3699 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3700 result value of the ``add`` is a :ref:`poison value <poisonvalues>` if
3701 unsigned and/or signed overflow, respectively, occurs.
3706 .. code-block:: llvm
3708 <result> = add i32 4, %var ; yields {i32}:result = 4 + %var
3712 '``fadd``' Instruction
3713 ^^^^^^^^^^^^^^^^^^^^^^
3720 <result> = fadd [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3725 The '``fadd``' instruction returns the sum of its two operands.
3730 The two arguments to the '``fadd``' instruction must be :ref:`floating
3731 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3732 Both arguments must have identical types.
3737 The value produced is the floating point sum of the two operands. This
3738 instruction can also take any number of :ref:`fast-math flags <fastmath>`,
3739 which are optimization hints to enable otherwise unsafe floating point
3745 .. code-block:: llvm
3747 <result> = fadd float 4.0, %var ; yields {float}:result = 4.0 + %var
3749 '``sub``' Instruction
3750 ^^^^^^^^^^^^^^^^^^^^^
3757 <result> = sub <ty> <op1>, <op2> ; yields {ty}:result
3758 <result> = sub nuw <ty> <op1>, <op2> ; yields {ty}:result
3759 <result> = sub nsw <ty> <op1>, <op2> ; yields {ty}:result
3760 <result> = sub nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3765 The '``sub``' instruction returns the difference of its two operands.
3767 Note that the '``sub``' instruction is used to represent the '``neg``'
3768 instruction present in most other intermediate representations.
3773 The two arguments to the '``sub``' instruction must be
3774 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3775 arguments must have identical types.
3780 The value produced is the integer difference of the two operands.
3782 If the difference has unsigned overflow, the result returned is the
3783 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
3786 Because LLVM integers use a two's complement representation, this
3787 instruction is appropriate for both signed and unsigned integers.
3789 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3790 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3791 result value of the ``sub`` is a :ref:`poison value <poisonvalues>` if
3792 unsigned and/or signed overflow, respectively, occurs.
3797 .. code-block:: llvm
3799 <result> = sub i32 4, %var ; yields {i32}:result = 4 - %var
3800 <result> = sub i32 0, %val ; yields {i32}:result = -%var
3804 '``fsub``' Instruction
3805 ^^^^^^^^^^^^^^^^^^^^^^
3812 <result> = fsub [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3817 The '``fsub``' instruction returns the difference of its two operands.
3819 Note that the '``fsub``' instruction is used to represent the '``fneg``'
3820 instruction present in most other intermediate representations.
3825 The two arguments to the '``fsub``' instruction must be :ref:`floating
3826 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3827 Both arguments must have identical types.
3832 The value produced is the floating point difference of the two operands.
3833 This instruction can also take any number of :ref:`fast-math
3834 flags <fastmath>`, which are optimization hints to enable otherwise
3835 unsafe floating point optimizations:
3840 .. code-block:: llvm
3842 <result> = fsub float 4.0, %var ; yields {float}:result = 4.0 - %var
3843 <result> = fsub float -0.0, %val ; yields {float}:result = -%var
3845 '``mul``' Instruction
3846 ^^^^^^^^^^^^^^^^^^^^^
3853 <result> = mul <ty> <op1>, <op2> ; yields {ty}:result
3854 <result> = mul nuw <ty> <op1>, <op2> ; yields {ty}:result
3855 <result> = mul nsw <ty> <op1>, <op2> ; yields {ty}:result
3856 <result> = mul nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3861 The '``mul``' instruction returns the product of its two operands.
3866 The two arguments to the '``mul``' instruction must be
3867 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3868 arguments must have identical types.
3873 The value produced is the integer product of the two operands.
3875 If the result of the multiplication has unsigned overflow, the result
3876 returned is the mathematical result modulo 2\ :sup:`n`\ , where n is the
3877 bit width of the result.
3879 Because LLVM integers use a two's complement representation, and the
3880 result is the same width as the operands, this instruction returns the
3881 correct result for both signed and unsigned integers. If a full product
3882 (e.g. ``i32`` * ``i32`` -> ``i64``) is needed, the operands should be
3883 sign-extended or zero-extended as appropriate to the width of the full
3886 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3887 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3888 result value of the ``mul`` is a :ref:`poison value <poisonvalues>` if
3889 unsigned and/or signed overflow, respectively, occurs.
3894 .. code-block:: llvm
3896 <result> = mul i32 4, %var ; yields {i32}:result = 4 * %var
3900 '``fmul``' Instruction
3901 ^^^^^^^^^^^^^^^^^^^^^^
3908 <result> = fmul [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3913 The '``fmul``' instruction returns the product of its two operands.
3918 The two arguments to the '``fmul``' instruction must be :ref:`floating
3919 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3920 Both arguments must have identical types.
3925 The value produced is the floating point product of the two operands.
3926 This instruction can also take any number of :ref:`fast-math
3927 flags <fastmath>`, which are optimization hints to enable otherwise
3928 unsafe floating point optimizations:
3933 .. code-block:: llvm
3935 <result> = fmul float 4.0, %var ; yields {float}:result = 4.0 * %var
3937 '``udiv``' Instruction
3938 ^^^^^^^^^^^^^^^^^^^^^^
3945 <result> = udiv <ty> <op1>, <op2> ; yields {ty}:result
3946 <result> = udiv exact <ty> <op1>, <op2> ; yields {ty}:result
3951 The '``udiv``' instruction returns the quotient of its two operands.
3956 The two arguments to the '``udiv``' instruction must be
3957 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3958 arguments must have identical types.
3963 The value produced is the unsigned integer quotient of the two operands.
3965 Note that unsigned integer division and signed integer division are
3966 distinct operations; for signed integer division, use '``sdiv``'.
3968 Division by zero leads to undefined behavior.
3970 If the ``exact`` keyword is present, the result value of the ``udiv`` is
3971 a :ref:`poison value <poisonvalues>` if %op1 is not a multiple of %op2 (as
3972 such, "((a udiv exact b) mul b) == a").
3977 .. code-block:: llvm
3979 <result> = udiv i32 4, %var ; yields {i32}:result = 4 / %var
3981 '``sdiv``' Instruction
3982 ^^^^^^^^^^^^^^^^^^^^^^
3989 <result> = sdiv <ty> <op1>, <op2> ; yields {ty}:result
3990 <result> = sdiv exact <ty> <op1>, <op2> ; yields {ty}:result
3995 The '``sdiv``' instruction returns the quotient of its two operands.
4000 The two arguments to the '``sdiv``' instruction must be
4001 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4002 arguments must have identical types.
4007 The value produced is the signed integer quotient of the two operands
4008 rounded towards zero.
4010 Note that signed integer division and unsigned integer division are
4011 distinct operations; for unsigned integer division, use '``udiv``'.
4013 Division by zero leads to undefined behavior. Overflow also leads to
4014 undefined behavior; this is a rare case, but can occur, for example, by
4015 doing a 32-bit division of -2147483648 by -1.
4017 If the ``exact`` keyword is present, the result value of the ``sdiv`` is
4018 a :ref:`poison value <poisonvalues>` if the result would be rounded.
4023 .. code-block:: llvm
4025 <result> = sdiv i32 4, %var ; yields {i32}:result = 4 / %var
4029 '``fdiv``' Instruction
4030 ^^^^^^^^^^^^^^^^^^^^^^
4037 <result> = fdiv [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
4042 The '``fdiv``' instruction returns the quotient of its two operands.
4047 The two arguments to the '``fdiv``' instruction must be :ref:`floating
4048 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
4049 Both arguments must have identical types.
4054 The value produced is the floating point quotient of the two operands.
4055 This instruction can also take any number of :ref:`fast-math
4056 flags <fastmath>`, which are optimization hints to enable otherwise
4057 unsafe floating point optimizations:
4062 .. code-block:: llvm
4064 <result> = fdiv float 4.0, %var ; yields {float}:result = 4.0 / %var
4066 '``urem``' Instruction
4067 ^^^^^^^^^^^^^^^^^^^^^^
4074 <result> = urem <ty> <op1>, <op2> ; yields {ty}:result
4079 The '``urem``' instruction returns the remainder from the unsigned
4080 division of its two arguments.
4085 The two arguments to the '``urem``' instruction must be
4086 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4087 arguments must have identical types.
4092 This instruction returns the unsigned integer *remainder* of a division.
4093 This instruction always performs an unsigned division to get the
4096 Note that unsigned integer remainder and signed integer remainder are
4097 distinct operations; for signed integer remainder, use '``srem``'.
4099 Taking the remainder of a division by zero leads to undefined behavior.
4104 .. code-block:: llvm
4106 <result> = urem i32 4, %var ; yields {i32}:result = 4 % %var
4108 '``srem``' Instruction
4109 ^^^^^^^^^^^^^^^^^^^^^^
4116 <result> = srem <ty> <op1>, <op2> ; yields {ty}:result
4121 The '``srem``' instruction returns the remainder from the signed
4122 division of its two operands. This instruction can also take
4123 :ref:`vector <t_vector>` versions of the values in which case the elements
4129 The two arguments to the '``srem``' instruction must be
4130 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4131 arguments must have identical types.
4136 This instruction returns the *remainder* of a division (where the result
4137 is either zero or has the same sign as the dividend, ``op1``), not the
4138 *modulo* operator (where the result is either zero or has the same sign
4139 as the divisor, ``op2``) of a value. For more information about the
4140 difference, see `The Math
4141 Forum <http://mathforum.org/dr.math/problems/anne.4.28.99.html>`_. For a
4142 table of how this is implemented in various languages, please see
4144 operation <http://en.wikipedia.org/wiki/Modulo_operation>`_.
4146 Note that signed integer remainder and unsigned integer remainder are
4147 distinct operations; for unsigned integer remainder, use '``urem``'.
4149 Taking the remainder of a division by zero leads to undefined behavior.
4150 Overflow also leads to undefined behavior; this is a rare case, but can
4151 occur, for example, by taking the remainder of a 32-bit division of
4152 -2147483648 by -1. (The remainder doesn't actually overflow, but this
4153 rule lets srem be implemented using instructions that return both the
4154 result of the division and the remainder.)
4159 .. code-block:: llvm
4161 <result> = srem i32 4, %var ; yields {i32}:result = 4 % %var
4165 '``frem``' Instruction
4166 ^^^^^^^^^^^^^^^^^^^^^^
4173 <result> = frem [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
4178 The '``frem``' instruction returns the remainder from the division of
4184 The two arguments to the '``frem``' instruction must be :ref:`floating
4185 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
4186 Both arguments must have identical types.
4191 This instruction returns the *remainder* of a division. The remainder
4192 has the same sign as the dividend. This instruction can also take any
4193 number of :ref:`fast-math flags <fastmath>`, which are optimization hints
4194 to enable otherwise unsafe floating point optimizations:
4199 .. code-block:: llvm
4201 <result> = frem float 4.0, %var ; yields {float}:result = 4.0 % %var
4205 Bitwise Binary Operations
4206 -------------------------
4208 Bitwise binary operators are used to do various forms of bit-twiddling
4209 in a program. They are generally very efficient instructions and can
4210 commonly be strength reduced from other instructions. They require two
4211 operands of the same type, execute an operation on them, and produce a
4212 single value. The resulting value is the same type as its operands.
4214 '``shl``' Instruction
4215 ^^^^^^^^^^^^^^^^^^^^^
4222 <result> = shl <ty> <op1>, <op2> ; yields {ty}:result
4223 <result> = shl nuw <ty> <op1>, <op2> ; yields {ty}:result
4224 <result> = shl nsw <ty> <op1>, <op2> ; yields {ty}:result
4225 <result> = shl nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
4230 The '``shl``' instruction returns the first operand shifted to the left
4231 a specified number of bits.
4236 Both arguments to the '``shl``' instruction must be the same
4237 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4238 '``op2``' is treated as an unsigned value.
4243 The value produced is ``op1`` \* 2\ :sup:`op2` mod 2\ :sup:`n`,
4244 where ``n`` is the width of the result. If ``op2`` is (statically or
4245 dynamically) negative or equal to or larger than the number of bits in
4246 ``op1``, the result is undefined. If the arguments are vectors, each
4247 vector element of ``op1`` is shifted by the corresponding shift amount
4250 If the ``nuw`` keyword is present, then the shift produces a :ref:`poison
4251 value <poisonvalues>` if it shifts out any non-zero bits. If the
4252 ``nsw`` keyword is present, then the shift produces a :ref:`poison
4253 value <poisonvalues>` if it shifts out any bits that disagree with the
4254 resultant sign bit. As such, NUW/NSW have the same semantics as they
4255 would if the shift were expressed as a mul instruction with the same
4256 nsw/nuw bits in (mul %op1, (shl 1, %op2)).
4261 .. code-block:: llvm
4263 <result> = shl i32 4, %var ; yields {i32}: 4 << %var
4264 <result> = shl i32 4, 2 ; yields {i32}: 16
4265 <result> = shl i32 1, 10 ; yields {i32}: 1024
4266 <result> = shl i32 1, 32 ; undefined
4267 <result> = shl <2 x i32> < i32 1, i32 1>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 2, i32 4>
4269 '``lshr``' Instruction
4270 ^^^^^^^^^^^^^^^^^^^^^^
4277 <result> = lshr <ty> <op1>, <op2> ; yields {ty}:result
4278 <result> = lshr exact <ty> <op1>, <op2> ; yields {ty}:result
4283 The '``lshr``' instruction (logical shift right) returns the first
4284 operand shifted to the right a specified number of bits with zero fill.
4289 Both arguments to the '``lshr``' instruction must be the same
4290 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4291 '``op2``' is treated as an unsigned value.
4296 This instruction always performs a logical shift right operation. The
4297 most significant bits of the result will be filled with zero bits after
4298 the shift. If ``op2`` is (statically or dynamically) equal to or larger
4299 than the number of bits in ``op1``, the result is undefined. If the
4300 arguments are vectors, each vector element of ``op1`` is shifted by the
4301 corresponding shift amount in ``op2``.
4303 If the ``exact`` keyword is present, the result value of the ``lshr`` is
4304 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
4310 .. code-block:: llvm
4312 <result> = lshr i32 4, 1 ; yields {i32}:result = 2
4313 <result> = lshr i32 4, 2 ; yields {i32}:result = 1
4314 <result> = lshr i8 4, 3 ; yields {i8}:result = 0
4315 <result> = lshr i8 -2, 1 ; yields {i8}:result = 0x7F
4316 <result> = lshr i32 1, 32 ; undefined
4317 <result> = lshr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 0x7FFFFFFF, i32 1>
4319 '``ashr``' Instruction
4320 ^^^^^^^^^^^^^^^^^^^^^^
4327 <result> = ashr <ty> <op1>, <op2> ; yields {ty}:result
4328 <result> = ashr exact <ty> <op1>, <op2> ; yields {ty}:result
4333 The '``ashr``' instruction (arithmetic shift right) returns the first
4334 operand shifted to the right a specified number of bits with sign
4340 Both arguments to the '``ashr``' instruction must be the same
4341 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4342 '``op2``' is treated as an unsigned value.
4347 This instruction always performs an arithmetic shift right operation,
4348 The most significant bits of the result will be filled with the sign bit
4349 of ``op1``. If ``op2`` is (statically or dynamically) equal to or larger
4350 than the number of bits in ``op1``, the result is undefined. If the
4351 arguments are vectors, each vector element of ``op1`` is shifted by the
4352 corresponding shift amount in ``op2``.
4354 If the ``exact`` keyword is present, the result value of the ``ashr`` is
4355 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
4361 .. code-block:: llvm
4363 <result> = ashr i32 4, 1 ; yields {i32}:result = 2
4364 <result> = ashr i32 4, 2 ; yields {i32}:result = 1
4365 <result> = ashr i8 4, 3 ; yields {i8}:result = 0
4366 <result> = ashr i8 -2, 1 ; yields {i8}:result = -1
4367 <result> = ashr i32 1, 32 ; undefined
4368 <result> = ashr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 3> ; yields: result=<2 x i32> < i32 -1, i32 0>
4370 '``and``' Instruction
4371 ^^^^^^^^^^^^^^^^^^^^^
4378 <result> = and <ty> <op1>, <op2> ; yields {ty}:result
4383 The '``and``' instruction returns the bitwise logical and of its two
4389 The two arguments to the '``and``' instruction must be
4390 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4391 arguments must have identical types.
4396 The truth table used for the '``and``' instruction is:
4413 .. code-block:: llvm
4415 <result> = and i32 4, %var ; yields {i32}:result = 4 & %var
4416 <result> = and i32 15, 40 ; yields {i32}:result = 8
4417 <result> = and i32 4, 8 ; yields {i32}:result = 0
4419 '``or``' Instruction
4420 ^^^^^^^^^^^^^^^^^^^^
4427 <result> = or <ty> <op1>, <op2> ; yields {ty}:result
4432 The '``or``' instruction returns the bitwise logical inclusive or of its
4438 The two arguments to the '``or``' instruction must be
4439 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4440 arguments must have identical types.
4445 The truth table used for the '``or``' instruction is:
4464 <result> = or i32 4, %var ; yields {i32}:result = 4 | %var
4465 <result> = or i32 15, 40 ; yields {i32}:result = 47
4466 <result> = or i32 4, 8 ; yields {i32}:result = 12
4468 '``xor``' Instruction
4469 ^^^^^^^^^^^^^^^^^^^^^
4476 <result> = xor <ty> <op1>, <op2> ; yields {ty}:result
4481 The '``xor``' instruction returns the bitwise logical exclusive or of
4482 its two operands. The ``xor`` is used to implement the "one's
4483 complement" operation, which is the "~" operator in C.
4488 The two arguments to the '``xor``' instruction must be
4489 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4490 arguments must have identical types.
4495 The truth table used for the '``xor``' instruction is:
4512 .. code-block:: llvm
4514 <result> = xor i32 4, %var ; yields {i32}:result = 4 ^ %var
4515 <result> = xor i32 15, 40 ; yields {i32}:result = 39
4516 <result> = xor i32 4, 8 ; yields {i32}:result = 12
4517 <result> = xor i32 %V, -1 ; yields {i32}:result = ~%V
4522 LLVM supports several instructions to represent vector operations in a
4523 target-independent manner. These instructions cover the element-access
4524 and vector-specific operations needed to process vectors effectively.
4525 While LLVM does directly support these vector operations, many
4526 sophisticated algorithms will want to use target-specific intrinsics to
4527 take full advantage of a specific target.
4529 .. _i_extractelement:
4531 '``extractelement``' Instruction
4532 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4539 <result> = extractelement <n x <ty>> <val>, <ty2> <idx> ; yields <ty>
4544 The '``extractelement``' instruction extracts a single scalar element
4545 from a vector at a specified index.
4550 The first operand of an '``extractelement``' instruction is a value of
4551 :ref:`vector <t_vector>` type. The second operand is an index indicating
4552 the position from which to extract the element. The index may be a
4553 variable of any integer type.
4558 The result is a scalar of the same type as the element type of ``val``.
4559 Its value is the value at position ``idx`` of ``val``. If ``idx``
4560 exceeds the length of ``val``, the results are undefined.
4565 .. code-block:: llvm
4567 <result> = extractelement <4 x i32> %vec, i32 0 ; yields i32
4569 .. _i_insertelement:
4571 '``insertelement``' Instruction
4572 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4579 <result> = insertelement <n x <ty>> <val>, <ty> <elt>, <ty2> <idx> ; yields <n x <ty>>
4584 The '``insertelement``' instruction inserts a scalar element into a
4585 vector at a specified index.
4590 The first operand of an '``insertelement``' instruction is a value of
4591 :ref:`vector <t_vector>` type. The second operand is a scalar value whose
4592 type must equal the element type of the first operand. The third operand
4593 is an index indicating the position at which to insert the value. The
4594 index may be a variable of any integer type.
4599 The result is a vector of the same type as ``val``. Its element values
4600 are those of ``val`` except at position ``idx``, where it gets the value
4601 ``elt``. If ``idx`` exceeds the length of ``val``, the results are
4607 .. code-block:: llvm
4609 <result> = insertelement <4 x i32> %vec, i32 1, i32 0 ; yields <4 x i32>
4611 .. _i_shufflevector:
4613 '``shufflevector``' Instruction
4614 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4621 <result> = shufflevector <n x <ty>> <v1>, <n x <ty>> <v2>, <m x i32> <mask> ; yields <m x <ty>>
4626 The '``shufflevector``' instruction constructs a permutation of elements
4627 from two input vectors, returning a vector with the same element type as
4628 the input and length that is the same as the shuffle mask.
4633 The first two operands of a '``shufflevector``' instruction are vectors
4634 with the same type. The third argument is a shuffle mask whose element
4635 type is always 'i32'. The result of the instruction is a vector whose
4636 length is the same as the shuffle mask and whose element type is the
4637 same as the element type of the first two operands.
4639 The shuffle mask operand is required to be a constant vector with either
4640 constant integer or undef values.
4645 The elements of the two input vectors are numbered from left to right
4646 across both of the vectors. The shuffle mask operand specifies, for each
4647 element of the result vector, which element of the two input vectors the
4648 result element gets. The element selector may be undef (meaning "don't
4649 care") and the second operand may be undef if performing a shuffle from
4655 .. code-block:: llvm
4657 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
4658 <4 x i32> <i32 0, i32 4, i32 1, i32 5> ; yields <4 x i32>
4659 <result> = shufflevector <4 x i32> %v1, <4 x i32> undef,
4660 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32> - Identity shuffle.
4661 <result> = shufflevector <8 x i32> %v1, <8 x i32> undef,
4662 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32>
4663 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
4664 <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7 > ; yields <8 x i32>
4666 Aggregate Operations
4667 --------------------
4669 LLVM supports several instructions for working with
4670 :ref:`aggregate <t_aggregate>` values.
4674 '``extractvalue``' Instruction
4675 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4682 <result> = extractvalue <aggregate type> <val>, <idx>{, <idx>}*
4687 The '``extractvalue``' instruction extracts the value of a member field
4688 from an :ref:`aggregate <t_aggregate>` value.
4693 The first operand of an '``extractvalue``' instruction is a value of
4694 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The operands are
4695 constant indices to specify which value to extract in a similar manner
4696 as indices in a '``getelementptr``' instruction.
4698 The major differences to ``getelementptr`` indexing are:
4700 - Since the value being indexed is not a pointer, the first index is
4701 omitted and assumed to be zero.
4702 - At least one index must be specified.
4703 - Not only struct indices but also array indices must be in bounds.
4708 The result is the value at the position in the aggregate specified by
4714 .. code-block:: llvm
4716 <result> = extractvalue {i32, float} %agg, 0 ; yields i32
4720 '``insertvalue``' Instruction
4721 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4728 <result> = insertvalue <aggregate type> <val>, <ty> <elt>, <idx>{, <idx>}* ; yields <aggregate type>
4733 The '``insertvalue``' instruction inserts a value into a member field in
4734 an :ref:`aggregate <t_aggregate>` value.
4739 The first operand of an '``insertvalue``' instruction is a value of
4740 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The second operand is
4741 a first-class value to insert. The following operands are constant
4742 indices indicating the position at which to insert the value in a
4743 similar manner as indices in a '``extractvalue``' instruction. The value
4744 to insert must have the same type as the value identified by the
4750 The result is an aggregate of the same type as ``val``. Its value is
4751 that of ``val`` except that the value at the position specified by the
4752 indices is that of ``elt``.
4757 .. code-block:: llvm
4759 %agg1 = insertvalue {i32, float} undef, i32 1, 0 ; yields {i32 1, float undef}
4760 %agg2 = insertvalue {i32, float} %agg1, float %val, 1 ; yields {i32 1, float %val}
4761 %agg3 = insertvalue {i32, {float}} %agg1, float %val, 1, 0 ; yields {i32 1, float %val}
4765 Memory Access and Addressing Operations
4766 ---------------------------------------
4768 A key design point of an SSA-based representation is how it represents
4769 memory. In LLVM, no memory locations are in SSA form, which makes things
4770 very simple. This section describes how to read, write, and allocate
4775 '``alloca``' Instruction
4776 ^^^^^^^^^^^^^^^^^^^^^^^^
4783 <result> = alloca [inalloca] <type> [, <ty> <NumElements>] [, align <alignment>] ; yields {type*}:result
4788 The '``alloca``' instruction allocates memory on the stack frame of the
4789 currently executing function, to be automatically released when this
4790 function returns to its caller. The object is always allocated in the
4791 generic address space (address space zero).
4796 The '``alloca``' instruction allocates ``sizeof(<type>)*NumElements``
4797 bytes of memory on the runtime stack, returning a pointer of the
4798 appropriate type to the program. If "NumElements" is specified, it is
4799 the number of elements allocated, otherwise "NumElements" is defaulted
4800 to be one. If a constant alignment is specified, the value result of the
4801 allocation is guaranteed to be aligned to at least that boundary. If not
4802 specified, or if zero, the target can choose to align the allocation on
4803 any convenient boundary compatible with the type.
4805 '``type``' may be any sized type.
4810 Memory is allocated; a pointer is returned. The operation is undefined
4811 if there is insufficient stack space for the allocation. '``alloca``'d
4812 memory is automatically released when the function returns. The
4813 '``alloca``' instruction is commonly used to represent automatic
4814 variables that must have an address available. When the function returns
4815 (either with the ``ret`` or ``resume`` instructions), the memory is
4816 reclaimed. Allocating zero bytes is legal, but the result is undefined.
4817 The order in which memory is allocated (ie., which way the stack grows)
4823 .. code-block:: llvm
4825 %ptr = alloca i32 ; yields {i32*}:ptr
4826 %ptr = alloca i32, i32 4 ; yields {i32*}:ptr
4827 %ptr = alloca i32, i32 4, align 1024 ; yields {i32*}:ptr
4828 %ptr = alloca i32, align 1024 ; yields {i32*}:ptr
4832 '``load``' Instruction
4833 ^^^^^^^^^^^^^^^^^^^^^^
4840 <result> = load [volatile] <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>][, !invariant.load !<index>]
4841 <result> = load atomic [volatile] <ty>* <pointer> [singlethread] <ordering>, align <alignment>
4842 !<index> = !{ i32 1 }
4847 The '``load``' instruction is used to read from memory.
4852 The argument to the ``load`` instruction specifies the memory address
4853 from which to load. The pointer must point to a :ref:`first
4854 class <t_firstclass>` type. If the ``load`` is marked as ``volatile``,
4855 then the optimizer is not allowed to modify the number or order of
4856 execution of this ``load`` with other :ref:`volatile
4857 operations <volatile>`.
4859 If the ``load`` is marked as ``atomic``, it takes an extra
4860 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
4861 ``release`` and ``acq_rel`` orderings are not valid on ``load``
4862 instructions. Atomic loads produce :ref:`defined <memmodel>` results
4863 when they may see multiple atomic stores. The type of the pointee must
4864 be an integer type whose bit width is a power of two greater than or
4865 equal to eight and less than or equal to a target-specific size limit.
4866 ``align`` must be explicitly specified on atomic loads, and the load has
4867 undefined behavior if the alignment is not set to a value which is at
4868 least the size in bytes of the pointee. ``!nontemporal`` does not have
4869 any defined semantics for atomic loads.
4871 The optional constant ``align`` argument specifies the alignment of the
4872 operation (that is, the alignment of the memory address). A value of 0
4873 or an omitted ``align`` argument means that the operation has the ABI
4874 alignment for the target. It is the responsibility of the code emitter
4875 to ensure that the alignment information is correct. Overestimating the
4876 alignment results in undefined behavior. Underestimating the alignment
4877 may produce less efficient code. An alignment of 1 is always safe.
4879 The optional ``!nontemporal`` metadata must reference a single
4880 metadata name ``<index>`` corresponding to a metadata node with one
4881 ``i32`` entry of value 1. The existence of the ``!nontemporal``
4882 metadata on the instruction tells the optimizer and code generator
4883 that this load is not expected to be reused in the cache. The code
4884 generator may select special instructions to save cache bandwidth, such
4885 as the ``MOVNT`` instruction on x86.
4887 The optional ``!invariant.load`` metadata must reference a single
4888 metadata name ``<index>`` corresponding to a metadata node with no
4889 entries. The existence of the ``!invariant.load`` metadata on the
4890 instruction tells the optimizer and code generator that this load
4891 address points to memory which does not change value during program
4892 execution. The optimizer may then move this load around, for example, by
4893 hoisting it out of loops using loop invariant code motion.
4898 The location of memory pointed to is loaded. If the value being loaded
4899 is of scalar type then the number of bytes read does not exceed the
4900 minimum number of bytes needed to hold all bits of the type. For
4901 example, loading an ``i24`` reads at most three bytes. When loading a
4902 value of a type like ``i20`` with a size that is not an integral number
4903 of bytes, the result is undefined if the value was not originally
4904 written using a store of the same type.
4909 .. code-block:: llvm
4911 %ptr = alloca i32 ; yields {i32*}:ptr
4912 store i32 3, i32* %ptr ; yields {void}
4913 %val = load i32* %ptr ; yields {i32}:val = i32 3
4917 '``store``' Instruction
4918 ^^^^^^^^^^^^^^^^^^^^^^^
4925 store [volatile] <ty> <value>, <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>] ; yields {void}
4926 store atomic [volatile] <ty> <value>, <ty>* <pointer> [singlethread] <ordering>, align <alignment> ; yields {void}
4931 The '``store``' instruction is used to write to memory.
4936 There are two arguments to the ``store`` instruction: a value to store
4937 and an address at which to store it. The type of the ``<pointer>``
4938 operand must be a pointer to the :ref:`first class <t_firstclass>` type of
4939 the ``<value>`` operand. If the ``store`` is marked as ``volatile``,
4940 then the optimizer is not allowed to modify the number or order of
4941 execution of this ``store`` with other :ref:`volatile
4942 operations <volatile>`.
4944 If the ``store`` is marked as ``atomic``, it takes an extra
4945 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
4946 ``acquire`` and ``acq_rel`` orderings aren't valid on ``store``
4947 instructions. Atomic loads produce :ref:`defined <memmodel>` results
4948 when they may see multiple atomic stores. The type of the pointee must
4949 be an integer type whose bit width is a power of two greater than or
4950 equal to eight and less than or equal to a target-specific size limit.
4951 ``align`` must be explicitly specified on atomic stores, and the store
4952 has undefined behavior if the alignment is not set to a value which is
4953 at least the size in bytes of the pointee. ``!nontemporal`` does not
4954 have any defined semantics for atomic stores.
4956 The optional constant ``align`` argument specifies the alignment of the
4957 operation (that is, the alignment of the memory address). A value of 0
4958 or an omitted ``align`` argument means that the operation has the ABI
4959 alignment for the target. It is the responsibility of the code emitter
4960 to ensure that the alignment information is correct. Overestimating the
4961 alignment results in undefined behavior. Underestimating the
4962 alignment may produce less efficient code. An alignment of 1 is always
4965 The optional ``!nontemporal`` metadata must reference a single metadata
4966 name ``<index>`` corresponding to a metadata node with one ``i32`` entry of
4967 value 1. The existence of the ``!nontemporal`` metadata on the instruction
4968 tells the optimizer and code generator that this load is not expected to
4969 be reused in the cache. The code generator may select special
4970 instructions to save cache bandwidth, such as the MOVNT instruction on
4976 The contents of memory are updated to contain ``<value>`` at the
4977 location specified by the ``<pointer>`` operand. If ``<value>`` is
4978 of scalar type then the number of bytes written does not exceed the
4979 minimum number of bytes needed to hold all bits of the type. For
4980 example, storing an ``i24`` writes at most three bytes. When writing a
4981 value of a type like ``i20`` with a size that is not an integral number
4982 of bytes, it is unspecified what happens to the extra bits that do not
4983 belong to the type, but they will typically be overwritten.
4988 .. code-block:: llvm
4990 %ptr = alloca i32 ; yields {i32*}:ptr
4991 store i32 3, i32* %ptr ; yields {void}
4992 %val = load i32* %ptr ; yields {i32}:val = i32 3
4996 '``fence``' Instruction
4997 ^^^^^^^^^^^^^^^^^^^^^^^
5004 fence [singlethread] <ordering> ; yields {void}
5009 The '``fence``' instruction is used to introduce happens-before edges
5015 '``fence``' instructions take an :ref:`ordering <ordering>` argument which
5016 defines what *synchronizes-with* edges they add. They can only be given
5017 ``acquire``, ``release``, ``acq_rel``, and ``seq_cst`` orderings.
5022 A fence A which has (at least) ``release`` ordering semantics
5023 *synchronizes with* a fence B with (at least) ``acquire`` ordering
5024 semantics if and only if there exist atomic operations X and Y, both
5025 operating on some atomic object M, such that A is sequenced before X, X
5026 modifies M (either directly or through some side effect of a sequence
5027 headed by X), Y is sequenced before B, and Y observes M. This provides a
5028 *happens-before* dependency between A and B. Rather than an explicit
5029 ``fence``, one (but not both) of the atomic operations X or Y might
5030 provide a ``release`` or ``acquire`` (resp.) ordering constraint and
5031 still *synchronize-with* the explicit ``fence`` and establish the
5032 *happens-before* edge.
5034 A ``fence`` which has ``seq_cst`` ordering, in addition to having both
5035 ``acquire`` and ``release`` semantics specified above, participates in
5036 the global program order of other ``seq_cst`` operations and/or fences.
5038 The optional ":ref:`singlethread <singlethread>`" argument specifies
5039 that the fence only synchronizes with other fences in the same thread.
5040 (This is useful for interacting with signal handlers.)
5045 .. code-block:: llvm
5047 fence acquire ; yields {void}
5048 fence singlethread seq_cst ; yields {void}
5052 '``cmpxchg``' Instruction
5053 ^^^^^^^^^^^^^^^^^^^^^^^^^
5060 cmpxchg [volatile] <ty>* <pointer>, <ty> <cmp>, <ty> <new> [singlethread] <success ordering> <failure ordering> ; yields {ty}
5065 The '``cmpxchg``' instruction is used to atomically modify memory. It
5066 loads a value in memory and compares it to a given value. If they are
5067 equal, it stores a new value into the memory.
5072 There are three arguments to the '``cmpxchg``' instruction: an address
5073 to operate on, a value to compare to the value currently be at that
5074 address, and a new value to place at that address if the compared values
5075 are equal. The type of '<cmp>' must be an integer type whose bit width
5076 is a power of two greater than or equal to eight and less than or equal
5077 to a target-specific size limit. '<cmp>' and '<new>' must have the same
5078 type, and the type of '<pointer>' must be a pointer to that type. If the
5079 ``cmpxchg`` is marked as ``volatile``, then the optimizer is not allowed
5080 to modify the number or order of execution of this ``cmpxchg`` with
5081 other :ref:`volatile operations <volatile>`.
5083 The success and failure :ref:`ordering <ordering>` arguments specify how this
5084 ``cmpxchg`` synchronizes with other atomic operations. The both ordering
5085 parameters must be at least ``monotonic``, the ordering constraint on failure
5086 must be no stronger than that on success, and the failure ordering cannot be
5087 either ``release`` or ``acq_rel``.
5089 The optional "``singlethread``" argument declares that the ``cmpxchg``
5090 is only atomic with respect to code (usually signal handlers) running in
5091 the same thread as the ``cmpxchg``. Otherwise the cmpxchg is atomic with
5092 respect to all other code in the system.
5094 The pointer passed into cmpxchg must have alignment greater than or
5095 equal to the size in memory of the operand.
5100 The contents of memory at the location specified by the '``<pointer>``'
5101 operand is read and compared to '``<cmp>``'; if the read value is the
5102 equal, '``<new>``' is written. The original value at the location is
5105 A successful ``cmpxchg`` is a read-modify-write instruction for the purpose of
5106 identifying release sequences. A failed ``cmpxchg`` is equivalent to an atomic
5107 load with an ordering parameter determined the second ordering parameter.
5112 .. code-block:: llvm
5115 %orig = atomic load i32* %ptr unordered ; yields {i32}
5119 %cmp = phi i32 [ %orig, %entry ], [%old, %loop]
5120 %squared = mul i32 %cmp, %cmp
5121 %old = cmpxchg i32* %ptr, i32 %cmp, i32 %squared acq_rel monotonic ; yields {i32}
5122 %success = icmp eq i32 %cmp, %old
5123 br i1 %success, label %done, label %loop
5130 '``atomicrmw``' Instruction
5131 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
5138 atomicrmw [volatile] <operation> <ty>* <pointer>, <ty> <value> [singlethread] <ordering> ; yields {ty}
5143 The '``atomicrmw``' instruction is used to atomically modify memory.
5148 There are three arguments to the '``atomicrmw``' instruction: an
5149 operation to apply, an address whose value to modify, an argument to the
5150 operation. The operation must be one of the following keywords:
5164 The type of '<value>' must be an integer type whose bit width is a power
5165 of two greater than or equal to eight and less than or equal to a
5166 target-specific size limit. The type of the '``<pointer>``' operand must
5167 be a pointer to that type. If the ``atomicrmw`` is marked as
5168 ``volatile``, then the optimizer is not allowed to modify the number or
5169 order of execution of this ``atomicrmw`` with other :ref:`volatile
5170 operations <volatile>`.
5175 The contents of memory at the location specified by the '``<pointer>``'
5176 operand are atomically read, modified, and written back. The original
5177 value at the location is returned. The modification is specified by the
5180 - xchg: ``*ptr = val``
5181 - add: ``*ptr = *ptr + val``
5182 - sub: ``*ptr = *ptr - val``
5183 - and: ``*ptr = *ptr & val``
5184 - nand: ``*ptr = ~(*ptr & val)``
5185 - or: ``*ptr = *ptr | val``
5186 - xor: ``*ptr = *ptr ^ val``
5187 - max: ``*ptr = *ptr > val ? *ptr : val`` (using a signed comparison)
5188 - min: ``*ptr = *ptr < val ? *ptr : val`` (using a signed comparison)
5189 - umax: ``*ptr = *ptr > val ? *ptr : val`` (using an unsigned
5191 - umin: ``*ptr = *ptr < val ? *ptr : val`` (using an unsigned
5197 .. code-block:: llvm
5199 %old = atomicrmw add i32* %ptr, i32 1 acquire ; yields {i32}
5201 .. _i_getelementptr:
5203 '``getelementptr``' Instruction
5204 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5211 <result> = getelementptr <pty>* <ptrval>{, <ty> <idx>}*
5212 <result> = getelementptr inbounds <pty>* <ptrval>{, <ty> <idx>}*
5213 <result> = getelementptr <ptr vector> ptrval, <vector index type> idx
5218 The '``getelementptr``' instruction is used to get the address of a
5219 subelement of an :ref:`aggregate <t_aggregate>` data structure. It performs
5220 address calculation only and does not access memory.
5225 The first argument is always a pointer or a vector of pointers, and
5226 forms the basis of the calculation. The remaining arguments are indices
5227 that indicate which of the elements of the aggregate object are indexed.
5228 The interpretation of each index is dependent on the type being indexed
5229 into. The first index always indexes the pointer value given as the
5230 first argument, the second index indexes a value of the type pointed to
5231 (not necessarily the value directly pointed to, since the first index
5232 can be non-zero), etc. The first type indexed into must be a pointer
5233 value, subsequent types can be arrays, vectors, and structs. Note that
5234 subsequent types being indexed into can never be pointers, since that
5235 would require loading the pointer before continuing calculation.
5237 The type of each index argument depends on the type it is indexing into.
5238 When indexing into a (optionally packed) structure, only ``i32`` integer
5239 **constants** are allowed (when using a vector of indices they must all
5240 be the **same** ``i32`` integer constant). When indexing into an array,
5241 pointer or vector, integers of any width are allowed, and they are not
5242 required to be constant. These integers are treated as signed values
5245 For example, let's consider a C code fragment and how it gets compiled
5261 int *foo(struct ST *s) {
5262 return &s[1].Z.B[5][13];
5265 The LLVM code generated by Clang is:
5267 .. code-block:: llvm
5269 %struct.RT = type { i8, [10 x [20 x i32]], i8 }
5270 %struct.ST = type { i32, double, %struct.RT }
5272 define i32* @foo(%struct.ST* %s) nounwind uwtable readnone optsize ssp {
5274 %arrayidx = getelementptr inbounds %struct.ST* %s, i64 1, i32 2, i32 1, i64 5, i64 13
5281 In the example above, the first index is indexing into the
5282 '``%struct.ST*``' type, which is a pointer, yielding a '``%struct.ST``'
5283 = '``{ i32, double, %struct.RT }``' type, a structure. The second index
5284 indexes into the third element of the structure, yielding a
5285 '``%struct.RT``' = '``{ i8 , [10 x [20 x i32]], i8 }``' type, another
5286 structure. The third index indexes into the second element of the
5287 structure, yielding a '``[10 x [20 x i32]]``' type, an array. The two
5288 dimensions of the array are subscripted into, yielding an '``i32``'
5289 type. The '``getelementptr``' instruction returns a pointer to this
5290 element, thus computing a value of '``i32*``' type.
5292 Note that it is perfectly legal to index partially through a structure,
5293 returning a pointer to an inner element. Because of this, the LLVM code
5294 for the given testcase is equivalent to:
5296 .. code-block:: llvm
5298 define i32* @foo(%struct.ST* %s) {
5299 %t1 = getelementptr %struct.ST* %s, i32 1 ; yields %struct.ST*:%t1
5300 %t2 = getelementptr %struct.ST* %t1, i32 0, i32 2 ; yields %struct.RT*:%t2
5301 %t3 = getelementptr %struct.RT* %t2, i32 0, i32 1 ; yields [10 x [20 x i32]]*:%t3
5302 %t4 = getelementptr [10 x [20 x i32]]* %t3, i32 0, i32 5 ; yields [20 x i32]*:%t4
5303 %t5 = getelementptr [20 x i32]* %t4, i32 0, i32 13 ; yields i32*:%t5
5307 If the ``inbounds`` keyword is present, the result value of the
5308 ``getelementptr`` is a :ref:`poison value <poisonvalues>` if the base
5309 pointer is not an *in bounds* address of an allocated object, or if any
5310 of the addresses that would be formed by successive addition of the
5311 offsets implied by the indices to the base address with infinitely
5312 precise signed arithmetic are not an *in bounds* address of that
5313 allocated object. The *in bounds* addresses for an allocated object are
5314 all the addresses that point into the object, plus the address one byte
5315 past the end. In cases where the base is a vector of pointers the
5316 ``inbounds`` keyword applies to each of the computations element-wise.
5318 If the ``inbounds`` keyword is not present, the offsets are added to the
5319 base address with silently-wrapping two's complement arithmetic. If the
5320 offsets have a different width from the pointer, they are sign-extended
5321 or truncated to the width of the pointer. The result value of the
5322 ``getelementptr`` may be outside the object pointed to by the base
5323 pointer. The result value may not necessarily be used to access memory
5324 though, even if it happens to point into allocated storage. See the
5325 :ref:`Pointer Aliasing Rules <pointeraliasing>` section for more
5328 The getelementptr instruction is often confusing. For some more insight
5329 into how it works, see :doc:`the getelementptr FAQ <GetElementPtr>`.
5334 .. code-block:: llvm
5336 ; yields [12 x i8]*:aptr
5337 %aptr = getelementptr {i32, [12 x i8]}* %saptr, i64 0, i32 1
5339 %vptr = getelementptr {i32, <2 x i8>}* %svptr, i64 0, i32 1, i32 1
5341 %eptr = getelementptr [12 x i8]* %aptr, i64 0, i32 1
5343 %iptr = getelementptr [10 x i32]* @arr, i16 0, i16 0
5345 In cases where the pointer argument is a vector of pointers, each index
5346 must be a vector with the same number of elements. For example:
5348 .. code-block:: llvm
5350 %A = getelementptr <4 x i8*> %ptrs, <4 x i64> %offsets,
5352 Conversion Operations
5353 ---------------------
5355 The instructions in this category are the conversion instructions
5356 (casting) which all take a single operand and a type. They perform
5357 various bit conversions on the operand.
5359 '``trunc .. to``' Instruction
5360 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5367 <result> = trunc <ty> <value> to <ty2> ; yields ty2
5372 The '``trunc``' instruction truncates its operand to the type ``ty2``.
5377 The '``trunc``' instruction takes a value to trunc, and a type to trunc
5378 it to. Both types must be of :ref:`integer <t_integer>` types, or vectors
5379 of the same number of integers. The bit size of the ``value`` must be
5380 larger than the bit size of the destination type, ``ty2``. Equal sized
5381 types are not allowed.
5386 The '``trunc``' instruction truncates the high order bits in ``value``
5387 and converts the remaining bits to ``ty2``. Since the source size must
5388 be larger than the destination size, ``trunc`` cannot be a *no-op cast*.
5389 It will always truncate bits.
5394 .. code-block:: llvm
5396 %X = trunc i32 257 to i8 ; yields i8:1
5397 %Y = trunc i32 123 to i1 ; yields i1:true
5398 %Z = trunc i32 122 to i1 ; yields i1:false
5399 %W = trunc <2 x i16> <i16 8, i16 7> to <2 x i8> ; yields <i8 8, i8 7>
5401 '``zext .. to``' Instruction
5402 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5409 <result> = zext <ty> <value> to <ty2> ; yields ty2
5414 The '``zext``' instruction zero extends its operand to type ``ty2``.
5419 The '``zext``' instruction takes a value to cast, and a type to cast it
5420 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
5421 the same number of integers. The bit size of the ``value`` must be
5422 smaller than the bit size of the destination type, ``ty2``.
5427 The ``zext`` fills the high order bits of the ``value`` with zero bits
5428 until it reaches the size of the destination type, ``ty2``.
5430 When zero extending from i1, the result will always be either 0 or 1.
5435 .. code-block:: llvm
5437 %X = zext i32 257 to i64 ; yields i64:257
5438 %Y = zext i1 true to i32 ; yields i32:1
5439 %Z = zext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
5441 '``sext .. to``' Instruction
5442 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5449 <result> = sext <ty> <value> to <ty2> ; yields ty2
5454 The '``sext``' sign extends ``value`` to the type ``ty2``.
5459 The '``sext``' instruction takes a value to cast, and a type to cast it
5460 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
5461 the same number of integers. The bit size of the ``value`` must be
5462 smaller than the bit size of the destination type, ``ty2``.
5467 The '``sext``' instruction performs a sign extension by copying the sign
5468 bit (highest order bit) of the ``value`` until it reaches the bit size
5469 of the type ``ty2``.
5471 When sign extending from i1, the extension always results in -1 or 0.
5476 .. code-block:: llvm
5478 %X = sext i8 -1 to i16 ; yields i16 :65535
5479 %Y = sext i1 true to i32 ; yields i32:-1
5480 %Z = sext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
5482 '``fptrunc .. to``' Instruction
5483 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5490 <result> = fptrunc <ty> <value> to <ty2> ; yields ty2
5495 The '``fptrunc``' instruction truncates ``value`` to type ``ty2``.
5500 The '``fptrunc``' instruction takes a :ref:`floating point <t_floating>`
5501 value to cast and a :ref:`floating point <t_floating>` type to cast it to.
5502 The size of ``value`` must be larger than the size of ``ty2``. This
5503 implies that ``fptrunc`` cannot be used to make a *no-op cast*.
5508 The '``fptrunc``' instruction truncates a ``value`` from a larger
5509 :ref:`floating point <t_floating>` type to a smaller :ref:`floating
5510 point <t_floating>` type. If the value cannot fit within the
5511 destination type, ``ty2``, then the results are undefined.
5516 .. code-block:: llvm
5518 %X = fptrunc double 123.0 to float ; yields float:123.0
5519 %Y = fptrunc double 1.0E+300 to float ; yields undefined
5521 '``fpext .. to``' Instruction
5522 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5529 <result> = fpext <ty> <value> to <ty2> ; yields ty2
5534 The '``fpext``' extends a floating point ``value`` to a larger floating
5540 The '``fpext``' instruction takes a :ref:`floating point <t_floating>`
5541 ``value`` to cast, and a :ref:`floating point <t_floating>` type to cast it
5542 to. The source type must be smaller than the destination type.
5547 The '``fpext``' instruction extends the ``value`` from a smaller
5548 :ref:`floating point <t_floating>` type to a larger :ref:`floating
5549 point <t_floating>` type. The ``fpext`` cannot be used to make a
5550 *no-op cast* because it always changes bits. Use ``bitcast`` to make a
5551 *no-op cast* for a floating point cast.
5556 .. code-block:: llvm
5558 %X = fpext float 3.125 to double ; yields double:3.125000e+00
5559 %Y = fpext double %X to fp128 ; yields fp128:0xL00000000000000004000900000000000
5561 '``fptoui .. to``' Instruction
5562 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5569 <result> = fptoui <ty> <value> to <ty2> ; yields ty2
5574 The '``fptoui``' converts a floating point ``value`` to its unsigned
5575 integer equivalent of type ``ty2``.
5580 The '``fptoui``' instruction takes a value to cast, which must be a
5581 scalar or vector :ref:`floating point <t_floating>` value, and a type to
5582 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
5583 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
5584 type with the same number of elements as ``ty``
5589 The '``fptoui``' instruction converts its :ref:`floating
5590 point <t_floating>` operand into the nearest (rounding towards zero)
5591 unsigned integer value. If the value cannot fit in ``ty2``, the results
5597 .. code-block:: llvm
5599 %X = fptoui double 123.0 to i32 ; yields i32:123
5600 %Y = fptoui float 1.0E+300 to i1 ; yields undefined:1
5601 %Z = fptoui float 1.04E+17 to i8 ; yields undefined:1
5603 '``fptosi .. to``' Instruction
5604 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5611 <result> = fptosi <ty> <value> to <ty2> ; yields ty2
5616 The '``fptosi``' instruction converts :ref:`floating point <t_floating>`
5617 ``value`` to type ``ty2``.
5622 The '``fptosi``' instruction takes a value to cast, which must be a
5623 scalar or vector :ref:`floating point <t_floating>` value, and a type to
5624 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
5625 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
5626 type with the same number of elements as ``ty``
5631 The '``fptosi``' instruction converts its :ref:`floating
5632 point <t_floating>` operand into the nearest (rounding towards zero)
5633 signed integer value. If the value cannot fit in ``ty2``, the results
5639 .. code-block:: llvm
5641 %X = fptosi double -123.0 to i32 ; yields i32:-123
5642 %Y = fptosi float 1.0E-247 to i1 ; yields undefined:1
5643 %Z = fptosi float 1.04E+17 to i8 ; yields undefined:1
5645 '``uitofp .. to``' Instruction
5646 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5653 <result> = uitofp <ty> <value> to <ty2> ; yields ty2
5658 The '``uitofp``' instruction regards ``value`` as an unsigned integer
5659 and converts that value to the ``ty2`` type.
5664 The '``uitofp``' instruction takes a value to cast, which must be a
5665 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
5666 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
5667 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
5668 type with the same number of elements as ``ty``
5673 The '``uitofp``' instruction interprets its operand as an unsigned
5674 integer quantity and converts it to the corresponding floating point
5675 value. If the value cannot fit in the floating point value, the results
5681 .. code-block:: llvm
5683 %X = uitofp i32 257 to float ; yields float:257.0
5684 %Y = uitofp i8 -1 to double ; yields double:255.0
5686 '``sitofp .. to``' Instruction
5687 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5694 <result> = sitofp <ty> <value> to <ty2> ; yields ty2
5699 The '``sitofp``' instruction regards ``value`` as a signed integer and
5700 converts that value to the ``ty2`` type.
5705 The '``sitofp``' instruction takes a value to cast, which must be a
5706 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
5707 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
5708 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
5709 type with the same number of elements as ``ty``
5714 The '``sitofp``' instruction interprets its operand as a signed integer
5715 quantity and converts it to the corresponding floating point value. If
5716 the value cannot fit in the floating point value, the results are
5722 .. code-block:: llvm
5724 %X = sitofp i32 257 to float ; yields float:257.0
5725 %Y = sitofp i8 -1 to double ; yields double:-1.0
5729 '``ptrtoint .. to``' Instruction
5730 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5737 <result> = ptrtoint <ty> <value> to <ty2> ; yields ty2
5742 The '``ptrtoint``' instruction converts the pointer or a vector of
5743 pointers ``value`` to the integer (or vector of integers) type ``ty2``.
5748 The '``ptrtoint``' instruction takes a ``value`` to cast, which must be
5749 a a value of type :ref:`pointer <t_pointer>` or a vector of pointers, and a
5750 type to cast it to ``ty2``, which must be an :ref:`integer <t_integer>` or
5751 a vector of integers type.
5756 The '``ptrtoint``' instruction converts ``value`` to integer type
5757 ``ty2`` by interpreting the pointer value as an integer and either
5758 truncating or zero extending that value to the size of the integer type.
5759 If ``value`` is smaller than ``ty2`` then a zero extension is done. If
5760 ``value`` is larger than ``ty2`` then a truncation is done. If they are
5761 the same size, then nothing is done (*no-op cast*) other than a type
5767 .. code-block:: llvm
5769 %X = ptrtoint i32* %P to i8 ; yields truncation on 32-bit architecture
5770 %Y = ptrtoint i32* %P to i64 ; yields zero extension on 32-bit architecture
5771 %Z = ptrtoint <4 x i32*> %P to <4 x i64>; yields vector zero extension for a vector of addresses on 32-bit architecture
5775 '``inttoptr .. to``' Instruction
5776 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5783 <result> = inttoptr <ty> <value> to <ty2> ; yields ty2
5788 The '``inttoptr``' instruction converts an integer ``value`` to a
5789 pointer type, ``ty2``.
5794 The '``inttoptr``' instruction takes an :ref:`integer <t_integer>` value to
5795 cast, and a type to cast it to, which must be a :ref:`pointer <t_pointer>`
5801 The '``inttoptr``' instruction converts ``value`` to type ``ty2`` by
5802 applying either a zero extension or a truncation depending on the size
5803 of the integer ``value``. If ``value`` is larger than the size of a
5804 pointer then a truncation is done. If ``value`` is smaller than the size
5805 of a pointer then a zero extension is done. If they are the same size,
5806 nothing is done (*no-op cast*).
5811 .. code-block:: llvm
5813 %X = inttoptr i32 255 to i32* ; yields zero extension on 64-bit architecture
5814 %Y = inttoptr i32 255 to i32* ; yields no-op on 32-bit architecture
5815 %Z = inttoptr i64 0 to i32* ; yields truncation on 32-bit architecture
5816 %Z = inttoptr <4 x i32> %G to <4 x i8*>; yields truncation of vector G to four pointers
5820 '``bitcast .. to``' Instruction
5821 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5828 <result> = bitcast <ty> <value> to <ty2> ; yields ty2
5833 The '``bitcast``' instruction converts ``value`` to type ``ty2`` without
5839 The '``bitcast``' instruction takes a value to cast, which must be a
5840 non-aggregate first class value, and a type to cast it to, which must
5841 also be a non-aggregate :ref:`first class <t_firstclass>` type. The
5842 bit sizes of ``value`` and the destination type, ``ty2``, must be
5843 identical. If the source type is a pointer, the destination type must
5844 also be a pointer of the same size. This instruction supports bitwise
5845 conversion of vectors to integers and to vectors of other types (as
5846 long as they have the same size).
5851 The '``bitcast``' instruction converts ``value`` to type ``ty2``. It
5852 is always a *no-op cast* because no bits change with this
5853 conversion. The conversion is done as if the ``value`` had been stored
5854 to memory and read back as type ``ty2``. Pointer (or vector of
5855 pointers) types may only be converted to other pointer (or vector of
5856 pointers) types with the same address space through this instruction.
5857 To convert pointers to other types, use the :ref:`inttoptr <i_inttoptr>`
5858 or :ref:`ptrtoint <i_ptrtoint>` instructions first.
5863 .. code-block:: llvm
5865 %X = bitcast i8 255 to i8 ; yields i8 :-1
5866 %Y = bitcast i32* %x to sint* ; yields sint*:%x
5867 %Z = bitcast <2 x int> %V to i64; ; yields i64: %V
5868 %Z = bitcast <2 x i32*> %V to <2 x i64*> ; yields <2 x i64*>
5870 .. _i_addrspacecast:
5872 '``addrspacecast .. to``' Instruction
5873 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5880 <result> = addrspacecast <pty> <ptrval> to <pty2> ; yields pty2
5885 The '``addrspacecast``' instruction converts ``ptrval`` from ``pty`` in
5886 address space ``n`` to type ``pty2`` in address space ``m``.
5891 The '``addrspacecast``' instruction takes a pointer or vector of pointer value
5892 to cast and a pointer type to cast it to, which must have a different
5898 The '``addrspacecast``' instruction converts the pointer value
5899 ``ptrval`` to type ``pty2``. It can be a *no-op cast* or a complex
5900 value modification, depending on the target and the address space
5901 pair. Pointer conversions within the same address space must be
5902 performed with the ``bitcast`` instruction. Note that if the address space
5903 conversion is legal then both result and operand refer to the same memory
5909 .. code-block:: llvm
5911 %X = addrspacecast i32* %x to i32 addrspace(1)* ; yields i32 addrspace(1)*:%x
5912 %Y = addrspacecast i32 addrspace(1)* %y to i64 addrspace(2)* ; yields i64 addrspace(2)*:%y
5913 %Z = addrspacecast <4 x i32*> %z to <4 x float addrspace(3)*> ; yields <4 x float addrspace(3)*>:%z
5920 The instructions in this category are the "miscellaneous" instructions,
5921 which defy better classification.
5925 '``icmp``' Instruction
5926 ^^^^^^^^^^^^^^^^^^^^^^
5933 <result> = icmp <cond> <ty> <op1>, <op2> ; yields {i1} or {<N x i1>}:result
5938 The '``icmp``' instruction returns a boolean value or a vector of
5939 boolean values based on comparison of its two integer, integer vector,
5940 pointer, or pointer vector operands.
5945 The '``icmp``' instruction takes three operands. The first operand is
5946 the condition code indicating the kind of comparison to perform. It is
5947 not a value, just a keyword. The possible condition code are:
5950 #. ``ne``: not equal
5951 #. ``ugt``: unsigned greater than
5952 #. ``uge``: unsigned greater or equal
5953 #. ``ult``: unsigned less than
5954 #. ``ule``: unsigned less or equal
5955 #. ``sgt``: signed greater than
5956 #. ``sge``: signed greater or equal
5957 #. ``slt``: signed less than
5958 #. ``sle``: signed less or equal
5960 The remaining two arguments must be :ref:`integer <t_integer>` or
5961 :ref:`pointer <t_pointer>` or integer :ref:`vector <t_vector>` typed. They
5962 must also be identical types.
5967 The '``icmp``' compares ``op1`` and ``op2`` according to the condition
5968 code given as ``cond``. The comparison performed always yields either an
5969 :ref:`i1 <t_integer>` or vector of ``i1`` result, as follows:
5971 #. ``eq``: yields ``true`` if the operands are equal, ``false``
5972 otherwise. No sign interpretation is necessary or performed.
5973 #. ``ne``: yields ``true`` if the operands are unequal, ``false``
5974 otherwise. No sign interpretation is necessary or performed.
5975 #. ``ugt``: interprets the operands as unsigned values and yields
5976 ``true`` if ``op1`` is greater than ``op2``.
5977 #. ``uge``: interprets the operands as unsigned values and yields
5978 ``true`` if ``op1`` is greater than or equal to ``op2``.
5979 #. ``ult``: interprets the operands as unsigned values and yields
5980 ``true`` if ``op1`` is less than ``op2``.
5981 #. ``ule``: interprets the operands as unsigned values and yields
5982 ``true`` if ``op1`` is less than or equal to ``op2``.
5983 #. ``sgt``: interprets the operands as signed values and yields ``true``
5984 if ``op1`` is greater than ``op2``.
5985 #. ``sge``: interprets the operands as signed values and yields ``true``
5986 if ``op1`` is greater than or equal to ``op2``.
5987 #. ``slt``: interprets the operands as signed values and yields ``true``
5988 if ``op1`` is less than ``op2``.
5989 #. ``sle``: interprets the operands as signed values and yields ``true``
5990 if ``op1`` is less than or equal to ``op2``.
5992 If the operands are :ref:`pointer <t_pointer>` typed, the pointer values
5993 are compared as if they were integers.
5995 If the operands are integer vectors, then they are compared element by
5996 element. The result is an ``i1`` vector with the same number of elements
5997 as the values being compared. Otherwise, the result is an ``i1``.
6002 .. code-block:: llvm
6004 <result> = icmp eq i32 4, 5 ; yields: result=false
6005 <result> = icmp ne float* %X, %X ; yields: result=false
6006 <result> = icmp ult i16 4, 5 ; yields: result=true
6007 <result> = icmp sgt i16 4, 5 ; yields: result=false
6008 <result> = icmp ule i16 -4, 5 ; yields: result=false
6009 <result> = icmp sge i16 4, 5 ; yields: result=false
6011 Note that the code generator does not yet support vector types with the
6012 ``icmp`` instruction.
6016 '``fcmp``' Instruction
6017 ^^^^^^^^^^^^^^^^^^^^^^
6024 <result> = fcmp <cond> <ty> <op1>, <op2> ; yields {i1} or {<N x i1>}:result
6029 The '``fcmp``' instruction returns a boolean value or vector of boolean
6030 values based on comparison of its operands.
6032 If the operands are floating point scalars, then the result type is a
6033 boolean (:ref:`i1 <t_integer>`).
6035 If the operands are floating point vectors, then the result type is a
6036 vector of boolean with the same number of elements as the operands being
6042 The '``fcmp``' instruction takes three operands. The first operand is
6043 the condition code indicating the kind of comparison to perform. It is
6044 not a value, just a keyword. The possible condition code are:
6046 #. ``false``: no comparison, always returns false
6047 #. ``oeq``: ordered and equal
6048 #. ``ogt``: ordered and greater than
6049 #. ``oge``: ordered and greater than or equal
6050 #. ``olt``: ordered and less than
6051 #. ``ole``: ordered and less than or equal
6052 #. ``one``: ordered and not equal
6053 #. ``ord``: ordered (no nans)
6054 #. ``ueq``: unordered or equal
6055 #. ``ugt``: unordered or greater than
6056 #. ``uge``: unordered or greater than or equal
6057 #. ``ult``: unordered or less than
6058 #. ``ule``: unordered or less than or equal
6059 #. ``une``: unordered or not equal
6060 #. ``uno``: unordered (either nans)
6061 #. ``true``: no comparison, always returns true
6063 *Ordered* means that neither operand is a QNAN while *unordered* means
6064 that either operand may be a QNAN.
6066 Each of ``val1`` and ``val2`` arguments must be either a :ref:`floating
6067 point <t_floating>` type or a :ref:`vector <t_vector>` of floating point
6068 type. They must have identical types.
6073 The '``fcmp``' instruction compares ``op1`` and ``op2`` according to the
6074 condition code given as ``cond``. If the operands are vectors, then the
6075 vectors are compared element by element. Each comparison performed
6076 always yields an :ref:`i1 <t_integer>` result, as follows:
6078 #. ``false``: always yields ``false``, regardless of operands.
6079 #. ``oeq``: yields ``true`` if both operands are not a QNAN and ``op1``
6080 is equal to ``op2``.
6081 #. ``ogt``: yields ``true`` if both operands are not a QNAN and ``op1``
6082 is greater than ``op2``.
6083 #. ``oge``: yields ``true`` if both operands are not a QNAN and ``op1``
6084 is greater than or equal to ``op2``.
6085 #. ``olt``: yields ``true`` if both operands are not a QNAN and ``op1``
6086 is less than ``op2``.
6087 #. ``ole``: yields ``true`` if both operands are not a QNAN and ``op1``
6088 is less than or equal to ``op2``.
6089 #. ``one``: yields ``true`` if both operands are not a QNAN and ``op1``
6090 is not equal to ``op2``.
6091 #. ``ord``: yields ``true`` if both operands are not a QNAN.
6092 #. ``ueq``: yields ``true`` if either operand is a QNAN or ``op1`` is
6094 #. ``ugt``: yields ``true`` if either operand is a QNAN or ``op1`` is
6095 greater than ``op2``.
6096 #. ``uge``: yields ``true`` if either operand is a QNAN or ``op1`` is
6097 greater than or equal to ``op2``.
6098 #. ``ult``: yields ``true`` if either operand is a QNAN or ``op1`` is
6100 #. ``ule``: yields ``true`` if either operand is a QNAN or ``op1`` is
6101 less than or equal to ``op2``.
6102 #. ``une``: yields ``true`` if either operand is a QNAN or ``op1`` is
6103 not equal to ``op2``.
6104 #. ``uno``: yields ``true`` if either operand is a QNAN.
6105 #. ``true``: always yields ``true``, regardless of operands.
6110 .. code-block:: llvm
6112 <result> = fcmp oeq float 4.0, 5.0 ; yields: result=false
6113 <result> = fcmp one float 4.0, 5.0 ; yields: result=true
6114 <result> = fcmp olt float 4.0, 5.0 ; yields: result=true
6115 <result> = fcmp ueq double 1.0, 2.0 ; yields: result=false
6117 Note that the code generator does not yet support vector types with the
6118 ``fcmp`` instruction.
6122 '``phi``' Instruction
6123 ^^^^^^^^^^^^^^^^^^^^^
6130 <result> = phi <ty> [ <val0>, <label0>], ...
6135 The '``phi``' instruction is used to implement the φ node in the SSA
6136 graph representing the function.
6141 The type of the incoming values is specified with the first type field.
6142 After this, the '``phi``' instruction takes a list of pairs as
6143 arguments, with one pair for each predecessor basic block of the current
6144 block. Only values of :ref:`first class <t_firstclass>` type may be used as
6145 the value arguments to the PHI node. Only labels may be used as the
6148 There must be no non-phi instructions between the start of a basic block
6149 and the PHI instructions: i.e. PHI instructions must be first in a basic
6152 For the purposes of the SSA form, the use of each incoming value is
6153 deemed to occur on the edge from the corresponding predecessor block to
6154 the current block (but after any definition of an '``invoke``'
6155 instruction's return value on the same edge).
6160 At runtime, the '``phi``' instruction logically takes on the value
6161 specified by the pair corresponding to the predecessor basic block that
6162 executed just prior to the current block.
6167 .. code-block:: llvm
6169 Loop: ; Infinite loop that counts from 0 on up...
6170 %indvar = phi i32 [ 0, %LoopHeader ], [ %nextindvar, %Loop ]
6171 %nextindvar = add i32 %indvar, 1
6176 '``select``' Instruction
6177 ^^^^^^^^^^^^^^^^^^^^^^^^
6184 <result> = select selty <cond>, <ty> <val1>, <ty> <val2> ; yields ty
6186 selty is either i1 or {<N x i1>}
6191 The '``select``' instruction is used to choose one value based on a
6192 condition, without IR-level branching.
6197 The '``select``' instruction requires an 'i1' value or a vector of 'i1'
6198 values indicating the condition, and two values of the same :ref:`first
6199 class <t_firstclass>` type. If the val1/val2 are vectors and the
6200 condition is a scalar, then entire vectors are selected, not individual
6206 If the condition is an i1 and it evaluates to 1, the instruction returns
6207 the first value argument; otherwise, it returns the second value
6210 If the condition is a vector of i1, then the value arguments must be
6211 vectors of the same size, and the selection is done element by element.
6216 .. code-block:: llvm
6218 %X = select i1 true, i8 17, i8 42 ; yields i8:17
6222 '``call``' Instruction
6223 ^^^^^^^^^^^^^^^^^^^^^^
6230 <result> = [tail | musttail] call [cconv] [ret attrs] <ty> [<fnty>*] <fnptrval>(<function args>) [fn attrs]
6235 The '``call``' instruction represents a simple function call.
6240 This instruction requires several arguments:
6242 #. The optional ``tail`` and ``musttail`` markers indicate that the optimizers
6243 should perform tail call optimization. The ``tail`` marker is a hint that
6244 `can be ignored <CodeGenerator.html#sibcallopt>`_. The ``musttail`` marker
6245 means that the call must be tail call optimized in order for the program to
6246 be correct. The ``musttail`` marker provides these guarantees:
6248 #. The call will not cause unbounded stack growth if it is part of a
6249 recursive cycle in the call graph.
6250 #. Arguments with the :ref:`inalloca <attr_inalloca>` attribute are
6253 Both markers imply that the callee does not access allocas or varargs from
6254 the caller. Calls marked ``musttail`` must obey the following additional
6257 - The call must immediately precede a :ref:`ret <i_ret>` instruction,
6258 or a pointer bitcast followed by a ret instruction.
6259 - The ret instruction must return the (possibly bitcasted) value
6260 produced by the call or void.
6261 - The caller and callee prototypes must match. Pointer types of
6262 parameters or return types may differ in pointee type, but not
6264 - The calling conventions of the caller and callee must match.
6265 - All ABI-impacting function attributes, such as sret, byval, inreg,
6266 returned, and inalloca, must match.
6268 Tail call optimization for calls marked ``tail`` is guaranteed to occur if
6269 the following conditions are met:
6271 - Caller and callee both have the calling convention ``fastcc``.
6272 - The call is in tail position (ret immediately follows call and ret
6273 uses value of call or is void).
6274 - Option ``-tailcallopt`` is enabled, or
6275 ``llvm::GuaranteedTailCallOpt`` is ``true``.
6276 - `Platform specific constraints are
6277 met. <CodeGenerator.html#tailcallopt>`_
6279 #. The optional "cconv" marker indicates which :ref:`calling
6280 convention <callingconv>` the call should use. If none is
6281 specified, the call defaults to using C calling conventions. The
6282 calling convention of the call must match the calling convention of
6283 the target function, or else the behavior is undefined.
6284 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
6285 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
6287 #. '``ty``': the type of the call instruction itself which is also the
6288 type of the return value. Functions that return no value are marked
6290 #. '``fnty``': shall be the signature of the pointer to function value
6291 being invoked. The argument types must match the types implied by
6292 this signature. This type can be omitted if the function is not
6293 varargs and if the function type does not return a pointer to a
6295 #. '``fnptrval``': An LLVM value containing a pointer to a function to
6296 be invoked. In most cases, this is a direct function invocation, but
6297 indirect ``call``'s are just as possible, calling an arbitrary pointer
6299 #. '``function args``': argument list whose types match the function
6300 signature argument types and parameter attributes. All arguments must
6301 be of :ref:`first class <t_firstclass>` type. If the function signature
6302 indicates the function accepts a variable number of arguments, the
6303 extra arguments can be specified.
6304 #. The optional :ref:`function attributes <fnattrs>` list. Only
6305 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
6306 attributes are valid here.
6311 The '``call``' instruction is used to cause control flow to transfer to
6312 a specified function, with its incoming arguments bound to the specified
6313 values. Upon a '``ret``' instruction in the called function, control
6314 flow continues with the instruction after the function call, and the
6315 return value of the function is bound to the result argument.
6320 .. code-block:: llvm
6322 %retval = call i32 @test(i32 %argc)
6323 call i32 (i8*, ...)* @printf(i8* %msg, i32 12, i8 42) ; yields i32
6324 %X = tail call i32 @foo() ; yields i32
6325 %Y = tail call fastcc i32 @foo() ; yields i32
6326 call void %foo(i8 97 signext)
6328 %struct.A = type { i32, i8 }
6329 %r = call %struct.A @foo() ; yields { 32, i8 }
6330 %gr = extractvalue %struct.A %r, 0 ; yields i32
6331 %gr1 = extractvalue %struct.A %r, 1 ; yields i8
6332 %Z = call void @foo() noreturn ; indicates that %foo never returns normally
6333 %ZZ = call zeroext i32 @bar() ; Return value is %zero extended
6335 llvm treats calls to some functions with names and arguments that match
6336 the standard C99 library as being the C99 library functions, and may
6337 perform optimizations or generate code for them under that assumption.
6338 This is something we'd like to change in the future to provide better
6339 support for freestanding environments and non-C-based languages.
6343 '``va_arg``' Instruction
6344 ^^^^^^^^^^^^^^^^^^^^^^^^
6351 <resultval> = va_arg <va_list*> <arglist>, <argty>
6356 The '``va_arg``' instruction is used to access arguments passed through
6357 the "variable argument" area of a function call. It is used to implement
6358 the ``va_arg`` macro in C.
6363 This instruction takes a ``va_list*`` value and the type of the
6364 argument. It returns a value of the specified argument type and
6365 increments the ``va_list`` to point to the next argument. The actual
6366 type of ``va_list`` is target specific.
6371 The '``va_arg``' instruction loads an argument of the specified type
6372 from the specified ``va_list`` and causes the ``va_list`` to point to
6373 the next argument. For more information, see the variable argument
6374 handling :ref:`Intrinsic Functions <int_varargs>`.
6376 It is legal for this instruction to be called in a function which does
6377 not take a variable number of arguments, for example, the ``vfprintf``
6380 ``va_arg`` is an LLVM instruction instead of an :ref:`intrinsic
6381 function <intrinsics>` because it takes a type as an argument.
6386 See the :ref:`variable argument processing <int_varargs>` section.
6388 Note that the code generator does not yet fully support va\_arg on many
6389 targets. Also, it does not currently support va\_arg with aggregate
6390 types on any target.
6394 '``landingpad``' Instruction
6395 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6402 <resultval> = landingpad <resultty> personality <type> <pers_fn> <clause>+
6403 <resultval> = landingpad <resultty> personality <type> <pers_fn> cleanup <clause>*
6405 <clause> := catch <type> <value>
6406 <clause> := filter <array constant type> <array constant>
6411 The '``landingpad``' instruction is used by `LLVM's exception handling
6412 system <ExceptionHandling.html#overview>`_ to specify that a basic block
6413 is a landing pad --- one where the exception lands, and corresponds to the
6414 code found in the ``catch`` portion of a ``try``/``catch`` sequence. It
6415 defines values supplied by the personality function (``pers_fn``) upon
6416 re-entry to the function. The ``resultval`` has the type ``resultty``.
6421 This instruction takes a ``pers_fn`` value. This is the personality
6422 function associated with the unwinding mechanism. The optional
6423 ``cleanup`` flag indicates that the landing pad block is a cleanup.
6425 A ``clause`` begins with the clause type --- ``catch`` or ``filter`` --- and
6426 contains the global variable representing the "type" that may be caught
6427 or filtered respectively. Unlike the ``catch`` clause, the ``filter``
6428 clause takes an array constant as its argument. Use
6429 "``[0 x i8**] undef``" for a filter which cannot throw. The
6430 '``landingpad``' instruction must contain *at least* one ``clause`` or
6431 the ``cleanup`` flag.
6436 The '``landingpad``' instruction defines the values which are set by the
6437 personality function (``pers_fn``) upon re-entry to the function, and
6438 therefore the "result type" of the ``landingpad`` instruction. As with
6439 calling conventions, how the personality function results are
6440 represented in LLVM IR is target specific.
6442 The clauses are applied in order from top to bottom. If two
6443 ``landingpad`` instructions are merged together through inlining, the
6444 clauses from the calling function are appended to the list of clauses.
6445 When the call stack is being unwound due to an exception being thrown,
6446 the exception is compared against each ``clause`` in turn. If it doesn't
6447 match any of the clauses, and the ``cleanup`` flag is not set, then
6448 unwinding continues further up the call stack.
6450 The ``landingpad`` instruction has several restrictions:
6452 - A landing pad block is a basic block which is the unwind destination
6453 of an '``invoke``' instruction.
6454 - A landing pad block must have a '``landingpad``' instruction as its
6455 first non-PHI instruction.
6456 - There can be only one '``landingpad``' instruction within the landing
6458 - A basic block that is not a landing pad block may not include a
6459 '``landingpad``' instruction.
6460 - All '``landingpad``' instructions in a function must have the same
6461 personality function.
6466 .. code-block:: llvm
6468 ;; A landing pad which can catch an integer.
6469 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6471 ;; A landing pad that is a cleanup.
6472 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6474 ;; A landing pad which can catch an integer and can only throw a double.
6475 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6477 filter [1 x i8**] [@_ZTId]
6484 LLVM supports the notion of an "intrinsic function". These functions
6485 have well known names and semantics and are required to follow certain
6486 restrictions. Overall, these intrinsics represent an extension mechanism
6487 for the LLVM language that does not require changing all of the
6488 transformations in LLVM when adding to the language (or the bitcode
6489 reader/writer, the parser, etc...).
6491 Intrinsic function names must all start with an "``llvm.``" prefix. This
6492 prefix is reserved in LLVM for intrinsic names; thus, function names may
6493 not begin with this prefix. Intrinsic functions must always be external
6494 functions: you cannot define the body of intrinsic functions. Intrinsic
6495 functions may only be used in call or invoke instructions: it is illegal
6496 to take the address of an intrinsic function. Additionally, because
6497 intrinsic functions are part of the LLVM language, it is required if any
6498 are added that they be documented here.
6500 Some intrinsic functions can be overloaded, i.e., the intrinsic
6501 represents a family of functions that perform the same operation but on
6502 different data types. Because LLVM can represent over 8 million
6503 different integer types, overloading is used commonly to allow an
6504 intrinsic function to operate on any integer type. One or more of the
6505 argument types or the result type can be overloaded to accept any
6506 integer type. Argument types may also be defined as exactly matching a
6507 previous argument's type or the result type. This allows an intrinsic
6508 function which accepts multiple arguments, but needs all of them to be
6509 of the same type, to only be overloaded with respect to a single
6510 argument or the result.
6512 Overloaded intrinsics will have the names of its overloaded argument
6513 types encoded into its function name, each preceded by a period. Only
6514 those types which are overloaded result in a name suffix. Arguments
6515 whose type is matched against another type do not. For example, the
6516 ``llvm.ctpop`` function can take an integer of any width and returns an
6517 integer of exactly the same integer width. This leads to a family of
6518 functions such as ``i8 @llvm.ctpop.i8(i8 %val)`` and
6519 ``i29 @llvm.ctpop.i29(i29 %val)``. Only one type, the return type, is
6520 overloaded, and only one type suffix is required. Because the argument's
6521 type is matched against the return type, it does not require its own
6524 To learn how to add an intrinsic function, please see the `Extending
6525 LLVM Guide <ExtendingLLVM.html>`_.
6529 Variable Argument Handling Intrinsics
6530 -------------------------------------
6532 Variable argument support is defined in LLVM with the
6533 :ref:`va_arg <i_va_arg>` instruction and these three intrinsic
6534 functions. These functions are related to the similarly named macros
6535 defined in the ``<stdarg.h>`` header file.
6537 All of these functions operate on arguments that use a target-specific
6538 value type "``va_list``". The LLVM assembly language reference manual
6539 does not define what this type is, so all transformations should be
6540 prepared to handle these functions regardless of the type used.
6542 This example shows how the :ref:`va_arg <i_va_arg>` instruction and the
6543 variable argument handling intrinsic functions are used.
6545 .. code-block:: llvm
6547 define i32 @test(i32 %X, ...) {
6548 ; Initialize variable argument processing
6550 %ap2 = bitcast i8** %ap to i8*
6551 call void @llvm.va_start(i8* %ap2)
6553 ; Read a single integer argument
6554 %tmp = va_arg i8** %ap, i32
6556 ; Demonstrate usage of llvm.va_copy and llvm.va_end
6558 %aq2 = bitcast i8** %aq to i8*
6559 call void @llvm.va_copy(i8* %aq2, i8* %ap2)
6560 call void @llvm.va_end(i8* %aq2)
6562 ; Stop processing of arguments.
6563 call void @llvm.va_end(i8* %ap2)
6567 declare void @llvm.va_start(i8*)
6568 declare void @llvm.va_copy(i8*, i8*)
6569 declare void @llvm.va_end(i8*)
6573 '``llvm.va_start``' Intrinsic
6574 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6581 declare void @llvm.va_start(i8* <arglist>)
6586 The '``llvm.va_start``' intrinsic initializes ``*<arglist>`` for
6587 subsequent use by ``va_arg``.
6592 The argument is a pointer to a ``va_list`` element to initialize.
6597 The '``llvm.va_start``' intrinsic works just like the ``va_start`` macro
6598 available in C. In a target-dependent way, it initializes the
6599 ``va_list`` element to which the argument points, so that the next call
6600 to ``va_arg`` will produce the first variable argument passed to the
6601 function. Unlike the C ``va_start`` macro, this intrinsic does not need
6602 to know the last argument of the function as the compiler can figure
6605 '``llvm.va_end``' Intrinsic
6606 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6613 declare void @llvm.va_end(i8* <arglist>)
6618 The '``llvm.va_end``' intrinsic destroys ``*<arglist>``, which has been
6619 initialized previously with ``llvm.va_start`` or ``llvm.va_copy``.
6624 The argument is a pointer to a ``va_list`` to destroy.
6629 The '``llvm.va_end``' intrinsic works just like the ``va_end`` macro
6630 available in C. In a target-dependent way, it destroys the ``va_list``
6631 element to which the argument points. Calls to
6632 :ref:`llvm.va_start <int_va_start>` and
6633 :ref:`llvm.va_copy <int_va_copy>` must be matched exactly with calls to
6638 '``llvm.va_copy``' Intrinsic
6639 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6646 declare void @llvm.va_copy(i8* <destarglist>, i8* <srcarglist>)
6651 The '``llvm.va_copy``' intrinsic copies the current argument position
6652 from the source argument list to the destination argument list.
6657 The first argument is a pointer to a ``va_list`` element to initialize.
6658 The second argument is a pointer to a ``va_list`` element to copy from.
6663 The '``llvm.va_copy``' intrinsic works just like the ``va_copy`` macro
6664 available in C. In a target-dependent way, it copies the source
6665 ``va_list`` element into the destination ``va_list`` element. This
6666 intrinsic is necessary because the `` llvm.va_start`` intrinsic may be
6667 arbitrarily complex and require, for example, memory allocation.
6669 Accurate Garbage Collection Intrinsics
6670 --------------------------------------
6672 LLVM support for `Accurate Garbage Collection <GarbageCollection.html>`_
6673 (GC) requires the implementation and generation of these intrinsics.
6674 These intrinsics allow identification of :ref:`GC roots on the
6675 stack <int_gcroot>`, as well as garbage collector implementations that
6676 require :ref:`read <int_gcread>` and :ref:`write <int_gcwrite>` barriers.
6677 Front-ends for type-safe garbage collected languages should generate
6678 these intrinsics to make use of the LLVM garbage collectors. For more
6679 details, see `Accurate Garbage Collection with
6680 LLVM <GarbageCollection.html>`_.
6682 The garbage collection intrinsics only operate on objects in the generic
6683 address space (address space zero).
6687 '``llvm.gcroot``' Intrinsic
6688 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6695 declare void @llvm.gcroot(i8** %ptrloc, i8* %metadata)
6700 The '``llvm.gcroot``' intrinsic declares the existence of a GC root to
6701 the code generator, and allows some metadata to be associated with it.
6706 The first argument specifies the address of a stack object that contains
6707 the root pointer. The second pointer (which must be either a constant or
6708 a global value address) contains the meta-data to be associated with the
6714 At runtime, a call to this intrinsic stores a null pointer into the
6715 "ptrloc" location. At compile-time, the code generator generates
6716 information to allow the runtime to find the pointer at GC safe points.
6717 The '``llvm.gcroot``' intrinsic may only be used in a function which
6718 :ref:`specifies a GC algorithm <gc>`.
6722 '``llvm.gcread``' Intrinsic
6723 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6730 declare i8* @llvm.gcread(i8* %ObjPtr, i8** %Ptr)
6735 The '``llvm.gcread``' intrinsic identifies reads of references from heap
6736 locations, allowing garbage collector implementations that require read
6742 The second argument is the address to read from, which should be an
6743 address allocated from the garbage collector. The first object is a
6744 pointer to the start of the referenced object, if needed by the language
6745 runtime (otherwise null).
6750 The '``llvm.gcread``' intrinsic has the same semantics as a load
6751 instruction, but may be replaced with substantially more complex code by
6752 the garbage collector runtime, as needed. The '``llvm.gcread``'
6753 intrinsic may only be used in a function which :ref:`specifies a GC
6758 '``llvm.gcwrite``' Intrinsic
6759 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6766 declare void @llvm.gcwrite(i8* %P1, i8* %Obj, i8** %P2)
6771 The '``llvm.gcwrite``' intrinsic identifies writes of references to heap
6772 locations, allowing garbage collector implementations that require write
6773 barriers (such as generational or reference counting collectors).
6778 The first argument is the reference to store, the second is the start of
6779 the object to store it to, and the third is the address of the field of
6780 Obj to store to. If the runtime does not require a pointer to the
6781 object, Obj may be null.
6786 The '``llvm.gcwrite``' intrinsic has the same semantics as a store
6787 instruction, but may be replaced with substantially more complex code by
6788 the garbage collector runtime, as needed. The '``llvm.gcwrite``'
6789 intrinsic may only be used in a function which :ref:`specifies a GC
6792 Code Generator Intrinsics
6793 -------------------------
6795 These intrinsics are provided by LLVM to expose special features that
6796 may only be implemented with code generator support.
6798 '``llvm.returnaddress``' Intrinsic
6799 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6806 declare i8 *@llvm.returnaddress(i32 <level>)
6811 The '``llvm.returnaddress``' intrinsic attempts to compute a
6812 target-specific value indicating the return address of the current
6813 function or one of its callers.
6818 The argument to this intrinsic indicates which function to return the
6819 address for. Zero indicates the calling function, one indicates its
6820 caller, etc. The argument is **required** to be a constant integer
6826 The '``llvm.returnaddress``' intrinsic either returns a pointer
6827 indicating the return address of the specified call frame, or zero if it
6828 cannot be identified. The value returned by this intrinsic is likely to
6829 be incorrect or 0 for arguments other than zero, so it should only be
6830 used for debugging purposes.
6832 Note that calling this intrinsic does not prevent function inlining or
6833 other aggressive transformations, so the value returned may not be that
6834 of the obvious source-language caller.
6836 '``llvm.frameaddress``' Intrinsic
6837 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6844 declare i8* @llvm.frameaddress(i32 <level>)
6849 The '``llvm.frameaddress``' intrinsic attempts to return the
6850 target-specific frame pointer value for the specified stack frame.
6855 The argument to this intrinsic indicates which function to return the
6856 frame pointer for. Zero indicates the calling function, one indicates
6857 its caller, etc. The argument is **required** to be a constant integer
6863 The '``llvm.frameaddress``' intrinsic either returns a pointer
6864 indicating the frame address of the specified call frame, or zero if it
6865 cannot be identified. The value returned by this intrinsic is likely to
6866 be incorrect or 0 for arguments other than zero, so it should only be
6867 used for debugging purposes.
6869 Note that calling this intrinsic does not prevent function inlining or
6870 other aggressive transformations, so the value returned may not be that
6871 of the obvious source-language caller.
6873 .. _int_read_register:
6874 .. _int_write_register:
6876 '``llvm.read_register``' and '``llvm.write_register``' Intrinsics
6877 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6884 declare i32 @llvm.read_register.i32(metadata)
6885 declare i64 @llvm.read_register.i64(metadata)
6886 declare void @llvm.write_register.i32(metadata, i32 @value)
6887 declare void @llvm.write_register.i64(metadata, i64 @value)
6888 !0 = metadata !{metadata !"sp\00"}
6893 The '``llvm.read_register``' and '``llvm.write_register``' intrinsics
6894 provides access to the named register. The register must be valid on
6895 the architecture being compiled to. The type needs to be compatible
6896 with the register being read.
6901 The '``llvm.read_register``' intrinsic returns the current value of the
6902 register, where possible. The '``llvm.write_register``' intrinsic sets
6903 the current value of the register, where possible.
6905 This is useful to implement named register global variables that need
6906 to always be mapped to a specific register, as is common practice on
6907 bare-metal programs including OS kernels.
6909 The compiler doesn't check for register availability or use of the used
6910 register in surrounding code, including inline assembly. Because of that,
6911 allocatable registers are not supported.
6913 Warning: So far it only works with the stack pointer on selected
6914 architectures (ARM, AArch64, PowerPC and x86_64). Significant amount of
6915 work is needed to support other registers and even more so, allocatable
6920 '``llvm.stacksave``' Intrinsic
6921 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6928 declare i8* @llvm.stacksave()
6933 The '``llvm.stacksave``' intrinsic is used to remember the current state
6934 of the function stack, for use with
6935 :ref:`llvm.stackrestore <int_stackrestore>`. This is useful for
6936 implementing language features like scoped automatic variable sized
6942 This intrinsic returns a opaque pointer value that can be passed to
6943 :ref:`llvm.stackrestore <int_stackrestore>`. When an
6944 ``llvm.stackrestore`` intrinsic is executed with a value saved from
6945 ``llvm.stacksave``, it effectively restores the state of the stack to
6946 the state it was in when the ``llvm.stacksave`` intrinsic executed. In
6947 practice, this pops any :ref:`alloca <i_alloca>` blocks from the stack that
6948 were allocated after the ``llvm.stacksave`` was executed.
6950 .. _int_stackrestore:
6952 '``llvm.stackrestore``' Intrinsic
6953 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6960 declare void @llvm.stackrestore(i8* %ptr)
6965 The '``llvm.stackrestore``' intrinsic is used to restore the state of
6966 the function stack to the state it was in when the corresponding
6967 :ref:`llvm.stacksave <int_stacksave>` intrinsic executed. This is
6968 useful for implementing language features like scoped automatic variable
6969 sized arrays in C99.
6974 See the description for :ref:`llvm.stacksave <int_stacksave>`.
6976 '``llvm.prefetch``' Intrinsic
6977 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6984 declare void @llvm.prefetch(i8* <address>, i32 <rw>, i32 <locality>, i32 <cache type>)
6989 The '``llvm.prefetch``' intrinsic is a hint to the code generator to
6990 insert a prefetch instruction if supported; otherwise, it is a noop.
6991 Prefetches have no effect on the behavior of the program but can change
6992 its performance characteristics.
6997 ``address`` is the address to be prefetched, ``rw`` is the specifier
6998 determining if the fetch should be for a read (0) or write (1), and
6999 ``locality`` is a temporal locality specifier ranging from (0) - no
7000 locality, to (3) - extremely local keep in cache. The ``cache type``
7001 specifies whether the prefetch is performed on the data (1) or
7002 instruction (0) cache. The ``rw``, ``locality`` and ``cache type``
7003 arguments must be constant integers.
7008 This intrinsic does not modify the behavior of the program. In
7009 particular, prefetches cannot trap and do not produce a value. On
7010 targets that support this intrinsic, the prefetch can provide hints to
7011 the processor cache for better performance.
7013 '``llvm.pcmarker``' Intrinsic
7014 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7021 declare void @llvm.pcmarker(i32 <id>)
7026 The '``llvm.pcmarker``' intrinsic is a method to export a Program
7027 Counter (PC) in a region of code to simulators and other tools. The
7028 method is target specific, but it is expected that the marker will use
7029 exported symbols to transmit the PC of the marker. The marker makes no
7030 guarantees that it will remain with any specific instruction after
7031 optimizations. It is possible that the presence of a marker will inhibit
7032 optimizations. The intended use is to be inserted after optimizations to
7033 allow correlations of simulation runs.
7038 ``id`` is a numerical id identifying the marker.
7043 This intrinsic does not modify the behavior of the program. Backends
7044 that do not support this intrinsic may ignore it.
7046 '``llvm.readcyclecounter``' Intrinsic
7047 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7054 declare i64 @llvm.readcyclecounter()
7059 The '``llvm.readcyclecounter``' intrinsic provides access to the cycle
7060 counter register (or similar low latency, high accuracy clocks) on those
7061 targets that support it. On X86, it should map to RDTSC. On Alpha, it
7062 should map to RPCC. As the backing counters overflow quickly (on the
7063 order of 9 seconds on alpha), this should only be used for small
7069 When directly supported, reading the cycle counter should not modify any
7070 memory. Implementations are allowed to either return a application
7071 specific value or a system wide value. On backends without support, this
7072 is lowered to a constant 0.
7074 Note that runtime support may be conditional on the privilege-level code is
7075 running at and the host platform.
7077 '``llvm.clear_cache``' Intrinsic
7078 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7085 declare void @llvm.clear_cache(i8*, i8*)
7090 The '``llvm.clear_cache``' intrinsic ensures visibility of modifications
7091 in the specified range to the execution unit of the processor. On
7092 targets with non-unified instruction and data cache, the implementation
7093 flushes the instruction cache.
7098 On platforms with coherent instruction and data caches (e.g. x86), this
7099 intrinsic is a nop. On platforms with non-coherent instruction and data
7100 cache (e.g. ARM, MIPS), the intrinsic is lowered either to appropriate
7101 instructions or a system call, if cache flushing requires special
7104 The default behavior is to emit a call to ``__clear_cache`` from the run
7107 This instrinsic does *not* empty the instruction pipeline. Modifications
7108 of the current function are outside the scope of the intrinsic.
7110 Standard C Library Intrinsics
7111 -----------------------------
7113 LLVM provides intrinsics for a few important standard C library
7114 functions. These intrinsics allow source-language front-ends to pass
7115 information about the alignment of the pointer arguments to the code
7116 generator, providing opportunity for more efficient code generation.
7120 '``llvm.memcpy``' Intrinsic
7121 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7126 This is an overloaded intrinsic. You can use ``llvm.memcpy`` on any
7127 integer bit width and for different address spaces. Not all targets
7128 support all bit widths however.
7132 declare void @llvm.memcpy.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
7133 i32 <len>, i32 <align>, i1 <isvolatile>)
7134 declare void @llvm.memcpy.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
7135 i64 <len>, i32 <align>, i1 <isvolatile>)
7140 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
7141 source location to the destination location.
7143 Note that, unlike the standard libc function, the ``llvm.memcpy.*``
7144 intrinsics do not return a value, takes extra alignment/isvolatile
7145 arguments and the pointers can be in specified address spaces.
7150 The first argument is a pointer to the destination, the second is a
7151 pointer to the source. The third argument is an integer argument
7152 specifying the number of bytes to copy, the fourth argument is the
7153 alignment of the source and destination locations, and the fifth is a
7154 boolean indicating a volatile access.
7156 If the call to this intrinsic has an alignment value that is not 0 or 1,
7157 then the caller guarantees that both the source and destination pointers
7158 are aligned to that boundary.
7160 If the ``isvolatile`` parameter is ``true``, the ``llvm.memcpy`` call is
7161 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
7162 very cleanly specified and it is unwise to depend on it.
7167 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
7168 source location to the destination location, which are not allowed to
7169 overlap. It copies "len" bytes of memory over. If the argument is known
7170 to be aligned to some boundary, this can be specified as the fourth
7171 argument, otherwise it should be set to 0 or 1 (both meaning no alignment).
7173 '``llvm.memmove``' Intrinsic
7174 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7179 This is an overloaded intrinsic. You can use llvm.memmove on any integer
7180 bit width and for different address space. Not all targets support all
7185 declare void @llvm.memmove.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
7186 i32 <len>, i32 <align>, i1 <isvolatile>)
7187 declare void @llvm.memmove.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
7188 i64 <len>, i32 <align>, i1 <isvolatile>)
7193 The '``llvm.memmove.*``' intrinsics move a block of memory from the
7194 source location to the destination location. It is similar to the
7195 '``llvm.memcpy``' intrinsic but allows the two memory locations to
7198 Note that, unlike the standard libc function, the ``llvm.memmove.*``
7199 intrinsics do not return a value, takes extra alignment/isvolatile
7200 arguments and the pointers can be in specified address spaces.
7205 The first argument is a pointer to the destination, the second is a
7206 pointer to the source. The third argument is an integer argument
7207 specifying the number of bytes to copy, the fourth argument is the
7208 alignment of the source and destination locations, and the fifth is a
7209 boolean indicating a volatile access.
7211 If the call to this intrinsic has an alignment value that is not 0 or 1,
7212 then the caller guarantees that the source and destination pointers are
7213 aligned to that boundary.
7215 If the ``isvolatile`` parameter is ``true``, the ``llvm.memmove`` call
7216 is a :ref:`volatile operation <volatile>`. The detailed access behavior is
7217 not very cleanly specified and it is unwise to depend on it.
7222 The '``llvm.memmove.*``' intrinsics copy a block of memory from the
7223 source location to the destination location, which may overlap. It
7224 copies "len" bytes of memory over. If the argument is known to be
7225 aligned to some boundary, this can be specified as the fourth argument,
7226 otherwise it should be set to 0 or 1 (both meaning no alignment).
7228 '``llvm.memset.*``' Intrinsics
7229 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7234 This is an overloaded intrinsic. You can use llvm.memset on any integer
7235 bit width and for different address spaces. However, not all targets
7236 support all bit widths.
7240 declare void @llvm.memset.p0i8.i32(i8* <dest>, i8 <val>,
7241 i32 <len>, i32 <align>, i1 <isvolatile>)
7242 declare void @llvm.memset.p0i8.i64(i8* <dest>, i8 <val>,
7243 i64 <len>, i32 <align>, i1 <isvolatile>)
7248 The '``llvm.memset.*``' intrinsics fill a block of memory with a
7249 particular byte value.
7251 Note that, unlike the standard libc function, the ``llvm.memset``
7252 intrinsic does not return a value and takes extra alignment/volatile
7253 arguments. Also, the destination can be in an arbitrary address space.
7258 The first argument is a pointer to the destination to fill, the second
7259 is the byte value with which to fill it, the third argument is an
7260 integer argument specifying the number of bytes to fill, and the fourth
7261 argument is the known alignment of the destination location.
7263 If the call to this intrinsic has an alignment value that is not 0 or 1,
7264 then the caller guarantees that the destination pointer is aligned to
7267 If the ``isvolatile`` parameter is ``true``, the ``llvm.memset`` call is
7268 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
7269 very cleanly specified and it is unwise to depend on it.
7274 The '``llvm.memset.*``' intrinsics fill "len" bytes of memory starting
7275 at the destination location. If the argument is known to be aligned to
7276 some boundary, this can be specified as the fourth argument, otherwise
7277 it should be set to 0 or 1 (both meaning no alignment).
7279 '``llvm.sqrt.*``' Intrinsic
7280 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7285 This is an overloaded intrinsic. You can use ``llvm.sqrt`` on any
7286 floating point or vector of floating point type. Not all targets support
7291 declare float @llvm.sqrt.f32(float %Val)
7292 declare double @llvm.sqrt.f64(double %Val)
7293 declare x86_fp80 @llvm.sqrt.f80(x86_fp80 %Val)
7294 declare fp128 @llvm.sqrt.f128(fp128 %Val)
7295 declare ppc_fp128 @llvm.sqrt.ppcf128(ppc_fp128 %Val)
7300 The '``llvm.sqrt``' intrinsics return the sqrt of the specified operand,
7301 returning the same value as the libm '``sqrt``' functions would. Unlike
7302 ``sqrt`` in libm, however, ``llvm.sqrt`` has undefined behavior for
7303 negative numbers other than -0.0 (which allows for better optimization,
7304 because there is no need to worry about errno being set).
7305 ``llvm.sqrt(-0.0)`` is defined to return -0.0 like IEEE sqrt.
7310 The argument and return value are floating point numbers of the same
7316 This function returns the sqrt of the specified operand if it is a
7317 nonnegative floating point number.
7319 '``llvm.powi.*``' Intrinsic
7320 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7325 This is an overloaded intrinsic. You can use ``llvm.powi`` on any
7326 floating point or vector of floating point type. Not all targets support
7331 declare float @llvm.powi.f32(float %Val, i32 %power)
7332 declare double @llvm.powi.f64(double %Val, i32 %power)
7333 declare x86_fp80 @llvm.powi.f80(x86_fp80 %Val, i32 %power)
7334 declare fp128 @llvm.powi.f128(fp128 %Val, i32 %power)
7335 declare ppc_fp128 @llvm.powi.ppcf128(ppc_fp128 %Val, i32 %power)
7340 The '``llvm.powi.*``' intrinsics return the first operand raised to the
7341 specified (positive or negative) power. The order of evaluation of
7342 multiplications is not defined. When a vector of floating point type is
7343 used, the second argument remains a scalar integer value.
7348 The second argument is an integer power, and the first is a value to
7349 raise to that power.
7354 This function returns the first value raised to the second power with an
7355 unspecified sequence of rounding operations.
7357 '``llvm.sin.*``' Intrinsic
7358 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7363 This is an overloaded intrinsic. You can use ``llvm.sin`` on any
7364 floating point or vector of floating point type. Not all targets support
7369 declare float @llvm.sin.f32(float %Val)
7370 declare double @llvm.sin.f64(double %Val)
7371 declare x86_fp80 @llvm.sin.f80(x86_fp80 %Val)
7372 declare fp128 @llvm.sin.f128(fp128 %Val)
7373 declare ppc_fp128 @llvm.sin.ppcf128(ppc_fp128 %Val)
7378 The '``llvm.sin.*``' intrinsics return the sine of the operand.
7383 The argument and return value are floating point numbers of the same
7389 This function returns the sine of the specified operand, returning the
7390 same values as the libm ``sin`` functions would, and handles error
7391 conditions in the same way.
7393 '``llvm.cos.*``' Intrinsic
7394 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7399 This is an overloaded intrinsic. You can use ``llvm.cos`` on any
7400 floating point or vector of floating point type. Not all targets support
7405 declare float @llvm.cos.f32(float %Val)
7406 declare double @llvm.cos.f64(double %Val)
7407 declare x86_fp80 @llvm.cos.f80(x86_fp80 %Val)
7408 declare fp128 @llvm.cos.f128(fp128 %Val)
7409 declare ppc_fp128 @llvm.cos.ppcf128(ppc_fp128 %Val)
7414 The '``llvm.cos.*``' intrinsics return the cosine of the operand.
7419 The argument and return value are floating point numbers of the same
7425 This function returns the cosine of the specified operand, returning the
7426 same values as the libm ``cos`` functions would, and handles error
7427 conditions in the same way.
7429 '``llvm.pow.*``' Intrinsic
7430 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7435 This is an overloaded intrinsic. You can use ``llvm.pow`` on any
7436 floating point or vector of floating point type. Not all targets support
7441 declare float @llvm.pow.f32(float %Val, float %Power)
7442 declare double @llvm.pow.f64(double %Val, double %Power)
7443 declare x86_fp80 @llvm.pow.f80(x86_fp80 %Val, x86_fp80 %Power)
7444 declare fp128 @llvm.pow.f128(fp128 %Val, fp128 %Power)
7445 declare ppc_fp128 @llvm.pow.ppcf128(ppc_fp128 %Val, ppc_fp128 Power)
7450 The '``llvm.pow.*``' intrinsics return the first operand raised to the
7451 specified (positive or negative) power.
7456 The second argument is a floating point power, and the first is a value
7457 to raise to that power.
7462 This function returns the first value raised to the second power,
7463 returning the same values as the libm ``pow`` functions would, and
7464 handles error conditions in the same way.
7466 '``llvm.exp.*``' Intrinsic
7467 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7472 This is an overloaded intrinsic. You can use ``llvm.exp`` on any
7473 floating point or vector of floating point type. Not all targets support
7478 declare float @llvm.exp.f32(float %Val)
7479 declare double @llvm.exp.f64(double %Val)
7480 declare x86_fp80 @llvm.exp.f80(x86_fp80 %Val)
7481 declare fp128 @llvm.exp.f128(fp128 %Val)
7482 declare ppc_fp128 @llvm.exp.ppcf128(ppc_fp128 %Val)
7487 The '``llvm.exp.*``' intrinsics perform the exp function.
7492 The argument and return value are floating point numbers of the same
7498 This function returns the same values as the libm ``exp`` functions
7499 would, and handles error conditions in the same way.
7501 '``llvm.exp2.*``' Intrinsic
7502 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7507 This is an overloaded intrinsic. You can use ``llvm.exp2`` on any
7508 floating point or vector of floating point type. Not all targets support
7513 declare float @llvm.exp2.f32(float %Val)
7514 declare double @llvm.exp2.f64(double %Val)
7515 declare x86_fp80 @llvm.exp2.f80(x86_fp80 %Val)
7516 declare fp128 @llvm.exp2.f128(fp128 %Val)
7517 declare ppc_fp128 @llvm.exp2.ppcf128(ppc_fp128 %Val)
7522 The '``llvm.exp2.*``' intrinsics perform the exp2 function.
7527 The argument and return value are floating point numbers of the same
7533 This function returns the same values as the libm ``exp2`` functions
7534 would, and handles error conditions in the same way.
7536 '``llvm.log.*``' Intrinsic
7537 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7542 This is an overloaded intrinsic. You can use ``llvm.log`` on any
7543 floating point or vector of floating point type. Not all targets support
7548 declare float @llvm.log.f32(float %Val)
7549 declare double @llvm.log.f64(double %Val)
7550 declare x86_fp80 @llvm.log.f80(x86_fp80 %Val)
7551 declare fp128 @llvm.log.f128(fp128 %Val)
7552 declare ppc_fp128 @llvm.log.ppcf128(ppc_fp128 %Val)
7557 The '``llvm.log.*``' intrinsics perform the log function.
7562 The argument and return value are floating point numbers of the same
7568 This function returns the same values as the libm ``log`` functions
7569 would, and handles error conditions in the same way.
7571 '``llvm.log10.*``' Intrinsic
7572 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7577 This is an overloaded intrinsic. You can use ``llvm.log10`` on any
7578 floating point or vector of floating point type. Not all targets support
7583 declare float @llvm.log10.f32(float %Val)
7584 declare double @llvm.log10.f64(double %Val)
7585 declare x86_fp80 @llvm.log10.f80(x86_fp80 %Val)
7586 declare fp128 @llvm.log10.f128(fp128 %Val)
7587 declare ppc_fp128 @llvm.log10.ppcf128(ppc_fp128 %Val)
7592 The '``llvm.log10.*``' intrinsics perform the log10 function.
7597 The argument and return value are floating point numbers of the same
7603 This function returns the same values as the libm ``log10`` functions
7604 would, and handles error conditions in the same way.
7606 '``llvm.log2.*``' Intrinsic
7607 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7612 This is an overloaded intrinsic. You can use ``llvm.log2`` on any
7613 floating point or vector of floating point type. Not all targets support
7618 declare float @llvm.log2.f32(float %Val)
7619 declare double @llvm.log2.f64(double %Val)
7620 declare x86_fp80 @llvm.log2.f80(x86_fp80 %Val)
7621 declare fp128 @llvm.log2.f128(fp128 %Val)
7622 declare ppc_fp128 @llvm.log2.ppcf128(ppc_fp128 %Val)
7627 The '``llvm.log2.*``' intrinsics perform the log2 function.
7632 The argument and return value are floating point numbers of the same
7638 This function returns the same values as the libm ``log2`` functions
7639 would, and handles error conditions in the same way.
7641 '``llvm.fma.*``' Intrinsic
7642 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7647 This is an overloaded intrinsic. You can use ``llvm.fma`` on any
7648 floating point or vector of floating point type. Not all targets support
7653 declare float @llvm.fma.f32(float %a, float %b, float %c)
7654 declare double @llvm.fma.f64(double %a, double %b, double %c)
7655 declare x86_fp80 @llvm.fma.f80(x86_fp80 %a, x86_fp80 %b, x86_fp80 %c)
7656 declare fp128 @llvm.fma.f128(fp128 %a, fp128 %b, fp128 %c)
7657 declare ppc_fp128 @llvm.fma.ppcf128(ppc_fp128 %a, ppc_fp128 %b, ppc_fp128 %c)
7662 The '``llvm.fma.*``' intrinsics perform the fused multiply-add
7668 The argument and return value are floating point numbers of the same
7674 This function returns the same values as the libm ``fma`` functions
7675 would, and does not set errno.
7677 '``llvm.fabs.*``' Intrinsic
7678 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7683 This is an overloaded intrinsic. You can use ``llvm.fabs`` on any
7684 floating point or vector of floating point type. Not all targets support
7689 declare float @llvm.fabs.f32(float %Val)
7690 declare double @llvm.fabs.f64(double %Val)
7691 declare x86_fp80 @llvm.fabs.f80(x86_fp80 %Val)
7692 declare fp128 @llvm.fabs.f128(fp128 %Val)
7693 declare ppc_fp128 @llvm.fabs.ppcf128(ppc_fp128 %Val)
7698 The '``llvm.fabs.*``' intrinsics return the absolute value of the
7704 The argument and return value are floating point numbers of the same
7710 This function returns the same values as the libm ``fabs`` functions
7711 would, and handles error conditions in the same way.
7713 '``llvm.copysign.*``' Intrinsic
7714 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7719 This is an overloaded intrinsic. You can use ``llvm.copysign`` on any
7720 floating point or vector of floating point type. Not all targets support
7725 declare float @llvm.copysign.f32(float %Mag, float %Sgn)
7726 declare double @llvm.copysign.f64(double %Mag, double %Sgn)
7727 declare x86_fp80 @llvm.copysign.f80(x86_fp80 %Mag, x86_fp80 %Sgn)
7728 declare fp128 @llvm.copysign.f128(fp128 %Mag, fp128 %Sgn)
7729 declare ppc_fp128 @llvm.copysign.ppcf128(ppc_fp128 %Mag, ppc_fp128 %Sgn)
7734 The '``llvm.copysign.*``' intrinsics return a value with the magnitude of the
7735 first operand and the sign of the second operand.
7740 The arguments and return value are floating point numbers of the same
7746 This function returns the same values as the libm ``copysign``
7747 functions would, and handles error conditions in the same way.
7749 '``llvm.floor.*``' Intrinsic
7750 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7755 This is an overloaded intrinsic. You can use ``llvm.floor`` on any
7756 floating point or vector of floating point type. Not all targets support
7761 declare float @llvm.floor.f32(float %Val)
7762 declare double @llvm.floor.f64(double %Val)
7763 declare x86_fp80 @llvm.floor.f80(x86_fp80 %Val)
7764 declare fp128 @llvm.floor.f128(fp128 %Val)
7765 declare ppc_fp128 @llvm.floor.ppcf128(ppc_fp128 %Val)
7770 The '``llvm.floor.*``' intrinsics return the floor of the operand.
7775 The argument and return value are floating point numbers of the same
7781 This function returns the same values as the libm ``floor`` functions
7782 would, and handles error conditions in the same way.
7784 '``llvm.ceil.*``' Intrinsic
7785 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7790 This is an overloaded intrinsic. You can use ``llvm.ceil`` on any
7791 floating point or vector of floating point type. Not all targets support
7796 declare float @llvm.ceil.f32(float %Val)
7797 declare double @llvm.ceil.f64(double %Val)
7798 declare x86_fp80 @llvm.ceil.f80(x86_fp80 %Val)
7799 declare fp128 @llvm.ceil.f128(fp128 %Val)
7800 declare ppc_fp128 @llvm.ceil.ppcf128(ppc_fp128 %Val)
7805 The '``llvm.ceil.*``' intrinsics return the ceiling of the operand.
7810 The argument and return value are floating point numbers of the same
7816 This function returns the same values as the libm ``ceil`` functions
7817 would, and handles error conditions in the same way.
7819 '``llvm.trunc.*``' Intrinsic
7820 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7825 This is an overloaded intrinsic. You can use ``llvm.trunc`` on any
7826 floating point or vector of floating point type. Not all targets support
7831 declare float @llvm.trunc.f32(float %Val)
7832 declare double @llvm.trunc.f64(double %Val)
7833 declare x86_fp80 @llvm.trunc.f80(x86_fp80 %Val)
7834 declare fp128 @llvm.trunc.f128(fp128 %Val)
7835 declare ppc_fp128 @llvm.trunc.ppcf128(ppc_fp128 %Val)
7840 The '``llvm.trunc.*``' intrinsics returns the operand rounded to the
7841 nearest integer not larger in magnitude than the operand.
7846 The argument and return value are floating point numbers of the same
7852 This function returns the same values as the libm ``trunc`` functions
7853 would, and handles error conditions in the same way.
7855 '``llvm.rint.*``' Intrinsic
7856 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7861 This is an overloaded intrinsic. You can use ``llvm.rint`` on any
7862 floating point or vector of floating point type. Not all targets support
7867 declare float @llvm.rint.f32(float %Val)
7868 declare double @llvm.rint.f64(double %Val)
7869 declare x86_fp80 @llvm.rint.f80(x86_fp80 %Val)
7870 declare fp128 @llvm.rint.f128(fp128 %Val)
7871 declare ppc_fp128 @llvm.rint.ppcf128(ppc_fp128 %Val)
7876 The '``llvm.rint.*``' intrinsics returns the operand rounded to the
7877 nearest integer. It may raise an inexact floating-point exception if the
7878 operand isn't an integer.
7883 The argument and return value are floating point numbers of the same
7889 This function returns the same values as the libm ``rint`` functions
7890 would, and handles error conditions in the same way.
7892 '``llvm.nearbyint.*``' Intrinsic
7893 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7898 This is an overloaded intrinsic. You can use ``llvm.nearbyint`` on any
7899 floating point or vector of floating point type. Not all targets support
7904 declare float @llvm.nearbyint.f32(float %Val)
7905 declare double @llvm.nearbyint.f64(double %Val)
7906 declare x86_fp80 @llvm.nearbyint.f80(x86_fp80 %Val)
7907 declare fp128 @llvm.nearbyint.f128(fp128 %Val)
7908 declare ppc_fp128 @llvm.nearbyint.ppcf128(ppc_fp128 %Val)
7913 The '``llvm.nearbyint.*``' intrinsics returns the operand rounded to the
7919 The argument and return value are floating point numbers of the same
7925 This function returns the same values as the libm ``nearbyint``
7926 functions would, and handles error conditions in the same way.
7928 '``llvm.round.*``' Intrinsic
7929 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7934 This is an overloaded intrinsic. You can use ``llvm.round`` on any
7935 floating point or vector of floating point type. Not all targets support
7940 declare float @llvm.round.f32(float %Val)
7941 declare double @llvm.round.f64(double %Val)
7942 declare x86_fp80 @llvm.round.f80(x86_fp80 %Val)
7943 declare fp128 @llvm.round.f128(fp128 %Val)
7944 declare ppc_fp128 @llvm.round.ppcf128(ppc_fp128 %Val)
7949 The '``llvm.round.*``' intrinsics returns the operand rounded to the
7955 The argument and return value are floating point numbers of the same
7961 This function returns the same values as the libm ``round``
7962 functions would, and handles error conditions in the same way.
7964 Bit Manipulation Intrinsics
7965 ---------------------------
7967 LLVM provides intrinsics for a few important bit manipulation
7968 operations. These allow efficient code generation for some algorithms.
7970 '``llvm.bswap.*``' Intrinsics
7971 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7976 This is an overloaded intrinsic function. You can use bswap on any
7977 integer type that is an even number of bytes (i.e. BitWidth % 16 == 0).
7981 declare i16 @llvm.bswap.i16(i16 <id>)
7982 declare i32 @llvm.bswap.i32(i32 <id>)
7983 declare i64 @llvm.bswap.i64(i64 <id>)
7988 The '``llvm.bswap``' family of intrinsics is used to byte swap integer
7989 values with an even number of bytes (positive multiple of 16 bits).
7990 These are useful for performing operations on data that is not in the
7991 target's native byte order.
7996 The ``llvm.bswap.i16`` intrinsic returns an i16 value that has the high
7997 and low byte of the input i16 swapped. Similarly, the ``llvm.bswap.i32``
7998 intrinsic returns an i32 value that has the four bytes of the input i32
7999 swapped, so that if the input bytes are numbered 0, 1, 2, 3 then the
8000 returned i32 will have its bytes in 3, 2, 1, 0 order. The
8001 ``llvm.bswap.i48``, ``llvm.bswap.i64`` and other intrinsics extend this
8002 concept to additional even-byte lengths (6 bytes, 8 bytes and more,
8005 '``llvm.ctpop.*``' Intrinsic
8006 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8011 This is an overloaded intrinsic. You can use llvm.ctpop on any integer
8012 bit width, or on any vector with integer elements. Not all targets
8013 support all bit widths or vector types, however.
8017 declare i8 @llvm.ctpop.i8(i8 <src>)
8018 declare i16 @llvm.ctpop.i16(i16 <src>)
8019 declare i32 @llvm.ctpop.i32(i32 <src>)
8020 declare i64 @llvm.ctpop.i64(i64 <src>)
8021 declare i256 @llvm.ctpop.i256(i256 <src>)
8022 declare <2 x i32> @llvm.ctpop.v2i32(<2 x i32> <src>)
8027 The '``llvm.ctpop``' family of intrinsics counts the number of bits set
8033 The only argument is the value to be counted. The argument may be of any
8034 integer type, or a vector with integer elements. The return type must
8035 match the argument type.
8040 The '``llvm.ctpop``' intrinsic counts the 1's in a variable, or within
8041 each element of a vector.
8043 '``llvm.ctlz.*``' Intrinsic
8044 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8049 This is an overloaded intrinsic. You can use ``llvm.ctlz`` on any
8050 integer bit width, or any vector whose elements are integers. Not all
8051 targets support all bit widths or vector types, however.
8055 declare i8 @llvm.ctlz.i8 (i8 <src>, i1 <is_zero_undef>)
8056 declare i16 @llvm.ctlz.i16 (i16 <src>, i1 <is_zero_undef>)
8057 declare i32 @llvm.ctlz.i32 (i32 <src>, i1 <is_zero_undef>)
8058 declare i64 @llvm.ctlz.i64 (i64 <src>, i1 <is_zero_undef>)
8059 declare i256 @llvm.ctlz.i256(i256 <src>, i1 <is_zero_undef>)
8060 declase <2 x i32> @llvm.ctlz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
8065 The '``llvm.ctlz``' family of intrinsic functions counts the number of
8066 leading zeros in a variable.
8071 The first argument is the value to be counted. This argument may be of
8072 any integer type, or a vectory with integer element type. The return
8073 type must match the first argument type.
8075 The second argument must be a constant and is a flag to indicate whether
8076 the intrinsic should ensure that a zero as the first argument produces a
8077 defined result. Historically some architectures did not provide a
8078 defined result for zero values as efficiently, and many algorithms are
8079 now predicated on avoiding zero-value inputs.
8084 The '``llvm.ctlz``' intrinsic counts the leading (most significant)
8085 zeros in a variable, or within each element of the vector. If
8086 ``src == 0`` then the result is the size in bits of the type of ``src``
8087 if ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
8088 ``llvm.ctlz(i32 2) = 30``.
8090 '``llvm.cttz.*``' Intrinsic
8091 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8096 This is an overloaded intrinsic. You can use ``llvm.cttz`` on any
8097 integer bit width, or any vector of integer elements. Not all targets
8098 support all bit widths or vector types, however.
8102 declare i8 @llvm.cttz.i8 (i8 <src>, i1 <is_zero_undef>)
8103 declare i16 @llvm.cttz.i16 (i16 <src>, i1 <is_zero_undef>)
8104 declare i32 @llvm.cttz.i32 (i32 <src>, i1 <is_zero_undef>)
8105 declare i64 @llvm.cttz.i64 (i64 <src>, i1 <is_zero_undef>)
8106 declare i256 @llvm.cttz.i256(i256 <src>, i1 <is_zero_undef>)
8107 declase <2 x i32> @llvm.cttz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
8112 The '``llvm.cttz``' family of intrinsic functions counts the number of
8118 The first argument is the value to be counted. This argument may be of
8119 any integer type, or a vectory with integer element type. The return
8120 type must match the first argument type.
8122 The second argument must be a constant and is a flag to indicate whether
8123 the intrinsic should ensure that a zero as the first argument produces a
8124 defined result. Historically some architectures did not provide a
8125 defined result for zero values as efficiently, and many algorithms are
8126 now predicated on avoiding zero-value inputs.
8131 The '``llvm.cttz``' intrinsic counts the trailing (least significant)
8132 zeros in a variable, or within each element of a vector. If ``src == 0``
8133 then the result is the size in bits of the type of ``src`` if
8134 ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
8135 ``llvm.cttz(2) = 1``.
8137 Arithmetic with Overflow Intrinsics
8138 -----------------------------------
8140 LLVM provides intrinsics for some arithmetic with overflow operations.
8142 '``llvm.sadd.with.overflow.*``' Intrinsics
8143 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8148 This is an overloaded intrinsic. You can use ``llvm.sadd.with.overflow``
8149 on any integer bit width.
8153 declare {i16, i1} @llvm.sadd.with.overflow.i16(i16 %a, i16 %b)
8154 declare {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
8155 declare {i64, i1} @llvm.sadd.with.overflow.i64(i64 %a, i64 %b)
8160 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
8161 a signed addition of the two arguments, and indicate whether an overflow
8162 occurred during the signed summation.
8167 The arguments (%a and %b) and the first element of the result structure
8168 may be of integer types of any bit width, but they must have the same
8169 bit width. The second element of the result structure must be of type
8170 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
8176 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
8177 a signed addition of the two variables. They return a structure --- the
8178 first element of which is the signed summation, and the second element
8179 of which is a bit specifying if the signed summation resulted in an
8185 .. code-block:: llvm
8187 %res = call {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
8188 %sum = extractvalue {i32, i1} %res, 0
8189 %obit = extractvalue {i32, i1} %res, 1
8190 br i1 %obit, label %overflow, label %normal
8192 '``llvm.uadd.with.overflow.*``' Intrinsics
8193 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8198 This is an overloaded intrinsic. You can use ``llvm.uadd.with.overflow``
8199 on any integer bit width.
8203 declare {i16, i1} @llvm.uadd.with.overflow.i16(i16 %a, i16 %b)
8204 declare {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
8205 declare {i64, i1} @llvm.uadd.with.overflow.i64(i64 %a, i64 %b)
8210 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
8211 an unsigned addition of the two arguments, and indicate whether a carry
8212 occurred during the unsigned summation.
8217 The arguments (%a and %b) and the first element of the result structure
8218 may be of integer types of any bit width, but they must have the same
8219 bit width. The second element of the result structure must be of type
8220 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
8226 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
8227 an unsigned addition of the two arguments. They return a structure --- the
8228 first element of which is the sum, and the second element of which is a
8229 bit specifying if the unsigned summation resulted in a carry.
8234 .. code-block:: llvm
8236 %res = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
8237 %sum = extractvalue {i32, i1} %res, 0
8238 %obit = extractvalue {i32, i1} %res, 1
8239 br i1 %obit, label %carry, label %normal
8241 '``llvm.ssub.with.overflow.*``' Intrinsics
8242 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8247 This is an overloaded intrinsic. You can use ``llvm.ssub.with.overflow``
8248 on any integer bit width.
8252 declare {i16, i1} @llvm.ssub.with.overflow.i16(i16 %a, i16 %b)
8253 declare {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
8254 declare {i64, i1} @llvm.ssub.with.overflow.i64(i64 %a, i64 %b)
8259 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
8260 a signed subtraction of the two arguments, and indicate whether an
8261 overflow occurred during the signed subtraction.
8266 The arguments (%a and %b) and the first element of the result structure
8267 may be of integer types of any bit width, but they must have the same
8268 bit width. The second element of the result structure must be of type
8269 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
8275 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
8276 a signed subtraction of the two arguments. They return a structure --- the
8277 first element of which is the subtraction, and the second element of
8278 which is a bit specifying if the signed subtraction resulted in an
8284 .. code-block:: llvm
8286 %res = call {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
8287 %sum = extractvalue {i32, i1} %res, 0
8288 %obit = extractvalue {i32, i1} %res, 1
8289 br i1 %obit, label %overflow, label %normal
8291 '``llvm.usub.with.overflow.*``' Intrinsics
8292 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8297 This is an overloaded intrinsic. You can use ``llvm.usub.with.overflow``
8298 on any integer bit width.
8302 declare {i16, i1} @llvm.usub.with.overflow.i16(i16 %a, i16 %b)
8303 declare {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
8304 declare {i64, i1} @llvm.usub.with.overflow.i64(i64 %a, i64 %b)
8309 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
8310 an unsigned subtraction of the two arguments, and indicate whether an
8311 overflow occurred during the unsigned subtraction.
8316 The arguments (%a and %b) and the first element of the result structure
8317 may be of integer types of any bit width, but they must have the same
8318 bit width. The second element of the result structure must be of type
8319 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
8325 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
8326 an unsigned subtraction of the two arguments. They return a structure ---
8327 the first element of which is the subtraction, and the second element of
8328 which is a bit specifying if the unsigned subtraction resulted in an
8334 .. code-block:: llvm
8336 %res = call {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
8337 %sum = extractvalue {i32, i1} %res, 0
8338 %obit = extractvalue {i32, i1} %res, 1
8339 br i1 %obit, label %overflow, label %normal
8341 '``llvm.smul.with.overflow.*``' Intrinsics
8342 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8347 This is an overloaded intrinsic. You can use ``llvm.smul.with.overflow``
8348 on any integer bit width.
8352 declare {i16, i1} @llvm.smul.with.overflow.i16(i16 %a, i16 %b)
8353 declare {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
8354 declare {i64, i1} @llvm.smul.with.overflow.i64(i64 %a, i64 %b)
8359 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
8360 a signed multiplication of the two arguments, and indicate whether an
8361 overflow occurred during the signed multiplication.
8366 The arguments (%a and %b) and the first element of the result structure
8367 may be of integer types of any bit width, but they must have the same
8368 bit width. The second element of the result structure must be of type
8369 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
8375 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
8376 a signed multiplication of the two arguments. They return a structure ---
8377 the first element of which is the multiplication, and the second element
8378 of which is a bit specifying if the signed multiplication resulted in an
8384 .. code-block:: llvm
8386 %res = call {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
8387 %sum = extractvalue {i32, i1} %res, 0
8388 %obit = extractvalue {i32, i1} %res, 1
8389 br i1 %obit, label %overflow, label %normal
8391 '``llvm.umul.with.overflow.*``' Intrinsics
8392 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8397 This is an overloaded intrinsic. You can use ``llvm.umul.with.overflow``
8398 on any integer bit width.
8402 declare {i16, i1} @llvm.umul.with.overflow.i16(i16 %a, i16 %b)
8403 declare {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
8404 declare {i64, i1} @llvm.umul.with.overflow.i64(i64 %a, i64 %b)
8409 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
8410 a unsigned multiplication of the two arguments, and indicate whether an
8411 overflow occurred during the unsigned multiplication.
8416 The arguments (%a and %b) and the first element of the result structure
8417 may be of integer types of any bit width, but they must have the same
8418 bit width. The second element of the result structure must be of type
8419 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
8425 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
8426 an unsigned multiplication of the two arguments. They return a structure ---
8427 the first element of which is the multiplication, and the second
8428 element of which is a bit specifying if the unsigned multiplication
8429 resulted in an overflow.
8434 .. code-block:: llvm
8436 %res = call {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
8437 %sum = extractvalue {i32, i1} %res, 0
8438 %obit = extractvalue {i32, i1} %res, 1
8439 br i1 %obit, label %overflow, label %normal
8441 Specialised Arithmetic Intrinsics
8442 ---------------------------------
8444 '``llvm.fmuladd.*``' Intrinsic
8445 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8452 declare float @llvm.fmuladd.f32(float %a, float %b, float %c)
8453 declare double @llvm.fmuladd.f64(double %a, double %b, double %c)
8458 The '``llvm.fmuladd.*``' intrinsic functions represent multiply-add
8459 expressions that can be fused if the code generator determines that (a) the
8460 target instruction set has support for a fused operation, and (b) that the
8461 fused operation is more efficient than the equivalent, separate pair of mul
8462 and add instructions.
8467 The '``llvm.fmuladd.*``' intrinsics each take three arguments: two
8468 multiplicands, a and b, and an addend c.
8477 %0 = call float @llvm.fmuladd.f32(%a, %b, %c)
8479 is equivalent to the expression a \* b + c, except that rounding will
8480 not be performed between the multiplication and addition steps if the
8481 code generator fuses the operations. Fusion is not guaranteed, even if
8482 the target platform supports it. If a fused multiply-add is required the
8483 corresponding llvm.fma.\* intrinsic function should be used
8484 instead. This never sets errno, just as '``llvm.fma.*``'.
8489 .. code-block:: llvm
8491 %r2 = call float @llvm.fmuladd.f32(float %a, float %b, float %c) ; yields {float}:r2 = (a * b) + c
8493 Half Precision Floating Point Intrinsics
8494 ----------------------------------------
8496 For most target platforms, half precision floating point is a
8497 storage-only format. This means that it is a dense encoding (in memory)
8498 but does not support computation in the format.
8500 This means that code must first load the half-precision floating point
8501 value as an i16, then convert it to float with
8502 :ref:`llvm.convert.from.fp16 <int_convert_from_fp16>`. Computation can
8503 then be performed on the float value (including extending to double
8504 etc). To store the value back to memory, it is first converted to float
8505 if needed, then converted to i16 with
8506 :ref:`llvm.convert.to.fp16 <int_convert_to_fp16>`, then storing as an
8509 .. _int_convert_to_fp16:
8511 '``llvm.convert.to.fp16``' Intrinsic
8512 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8519 declare i16 @llvm.convert.to.fp16(f32 %a)
8524 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion
8525 from single precision floating point format to half precision floating
8531 The intrinsic function contains single argument - the value to be
8537 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion
8538 from single precision floating point format to half precision floating
8539 point format. The return value is an ``i16`` which contains the
8545 .. code-block:: llvm
8547 %res = call i16 @llvm.convert.to.fp16(f32 %a)
8548 store i16 %res, i16* @x, align 2
8550 .. _int_convert_from_fp16:
8552 '``llvm.convert.from.fp16``' Intrinsic
8553 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8560 declare f32 @llvm.convert.from.fp16(i16 %a)
8565 The '``llvm.convert.from.fp16``' intrinsic function performs a
8566 conversion from half precision floating point format to single precision
8567 floating point format.
8572 The intrinsic function contains single argument - the value to be
8578 The '``llvm.convert.from.fp16``' intrinsic function performs a
8579 conversion from half single precision floating point format to single
8580 precision floating point format. The input half-float value is
8581 represented by an ``i16`` value.
8586 .. code-block:: llvm
8588 %a = load i16* @x, align 2
8589 %res = call f32 @llvm.convert.from.fp16(i16 %a)
8594 The LLVM debugger intrinsics (which all start with ``llvm.dbg.``
8595 prefix), are described in the `LLVM Source Level
8596 Debugging <SourceLevelDebugging.html#format_common_intrinsics>`_
8599 Exception Handling Intrinsics
8600 -----------------------------
8602 The LLVM exception handling intrinsics (which all start with
8603 ``llvm.eh.`` prefix), are described in the `LLVM Exception
8604 Handling <ExceptionHandling.html#format_common_intrinsics>`_ document.
8608 Trampoline Intrinsics
8609 ---------------------
8611 These intrinsics make it possible to excise one parameter, marked with
8612 the :ref:`nest <nest>` attribute, from a function. The result is a
8613 callable function pointer lacking the nest parameter - the caller does
8614 not need to provide a value for it. Instead, the value to use is stored
8615 in advance in a "trampoline", a block of memory usually allocated on the
8616 stack, which also contains code to splice the nest value into the
8617 argument list. This is used to implement the GCC nested function address
8620 For example, if the function is ``i32 f(i8* nest %c, i32 %x, i32 %y)``
8621 then the resulting function pointer has signature ``i32 (i32, i32)*``.
8622 It can be created as follows:
8624 .. code-block:: llvm
8626 %tramp = alloca [10 x i8], align 4 ; size and alignment only correct for X86
8627 %tramp1 = getelementptr [10 x i8]* %tramp, i32 0, i32 0
8628 call i8* @llvm.init.trampoline(i8* %tramp1, i8* bitcast (i32 (i8*, i32, i32)* @f to i8*), i8* %nval)
8629 %p = call i8* @llvm.adjust.trampoline(i8* %tramp1)
8630 %fp = bitcast i8* %p to i32 (i32, i32)*
8632 The call ``%val = call i32 %fp(i32 %x, i32 %y)`` is then equivalent to
8633 ``%val = call i32 %f(i8* %nval, i32 %x, i32 %y)``.
8637 '``llvm.init.trampoline``' Intrinsic
8638 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8645 declare void @llvm.init.trampoline(i8* <tramp>, i8* <func>, i8* <nval>)
8650 This fills the memory pointed to by ``tramp`` with executable code,
8651 turning it into a trampoline.
8656 The ``llvm.init.trampoline`` intrinsic takes three arguments, all
8657 pointers. The ``tramp`` argument must point to a sufficiently large and
8658 sufficiently aligned block of memory; this memory is written to by the
8659 intrinsic. Note that the size and the alignment are target-specific -
8660 LLVM currently provides no portable way of determining them, so a
8661 front-end that generates this intrinsic needs to have some
8662 target-specific knowledge. The ``func`` argument must hold a function
8663 bitcast to an ``i8*``.
8668 The block of memory pointed to by ``tramp`` is filled with target
8669 dependent code, turning it into a function. Then ``tramp`` needs to be
8670 passed to :ref:`llvm.adjust.trampoline <int_at>` to get a pointer which can
8671 be :ref:`bitcast (to a new function) and called <int_trampoline>`. The new
8672 function's signature is the same as that of ``func`` with any arguments
8673 marked with the ``nest`` attribute removed. At most one such ``nest``
8674 argument is allowed, and it must be of pointer type. Calling the new
8675 function is equivalent to calling ``func`` with the same argument list,
8676 but with ``nval`` used for the missing ``nest`` argument. If, after
8677 calling ``llvm.init.trampoline``, the memory pointed to by ``tramp`` is
8678 modified, then the effect of any later call to the returned function
8679 pointer is undefined.
8683 '``llvm.adjust.trampoline``' Intrinsic
8684 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8691 declare i8* @llvm.adjust.trampoline(i8* <tramp>)
8696 This performs any required machine-specific adjustment to the address of
8697 a trampoline (passed as ``tramp``).
8702 ``tramp`` must point to a block of memory which already has trampoline
8703 code filled in by a previous call to
8704 :ref:`llvm.init.trampoline <int_it>`.
8709 On some architectures the address of the code to be executed needs to be
8710 different to the address where the trampoline is actually stored. This
8711 intrinsic returns the executable address corresponding to ``tramp``
8712 after performing the required machine specific adjustments. The pointer
8713 returned can then be :ref:`bitcast and executed <int_trampoline>`.
8718 This class of intrinsics exists to information about the lifetime of
8719 memory objects and ranges where variables are immutable.
8723 '``llvm.lifetime.start``' Intrinsic
8724 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8731 declare void @llvm.lifetime.start(i64 <size>, i8* nocapture <ptr>)
8736 The '``llvm.lifetime.start``' intrinsic specifies the start of a memory
8742 The first argument is a constant integer representing the size of the
8743 object, or -1 if it is variable sized. The second argument is a pointer
8749 This intrinsic indicates that before this point in the code, the value
8750 of the memory pointed to by ``ptr`` is dead. This means that it is known
8751 to never be used and has an undefined value. A load from the pointer
8752 that precedes this intrinsic can be replaced with ``'undef'``.
8756 '``llvm.lifetime.end``' Intrinsic
8757 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8764 declare void @llvm.lifetime.end(i64 <size>, i8* nocapture <ptr>)
8769 The '``llvm.lifetime.end``' intrinsic specifies the end of a memory
8775 The first argument is a constant integer representing the size of the
8776 object, or -1 if it is variable sized. The second argument is a pointer
8782 This intrinsic indicates that after this point in the code, the value of
8783 the memory pointed to by ``ptr`` is dead. This means that it is known to
8784 never be used and has an undefined value. Any stores into the memory
8785 object following this intrinsic may be removed as dead.
8787 '``llvm.invariant.start``' Intrinsic
8788 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8795 declare {}* @llvm.invariant.start(i64 <size>, i8* nocapture <ptr>)
8800 The '``llvm.invariant.start``' intrinsic specifies that the contents of
8801 a memory object will not change.
8806 The first argument is a constant integer representing the size of the
8807 object, or -1 if it is variable sized. The second argument is a pointer
8813 This intrinsic indicates that until an ``llvm.invariant.end`` that uses
8814 the return value, the referenced memory location is constant and
8817 '``llvm.invariant.end``' Intrinsic
8818 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8825 declare void @llvm.invariant.end({}* <start>, i64 <size>, i8* nocapture <ptr>)
8830 The '``llvm.invariant.end``' intrinsic specifies that the contents of a
8831 memory object are mutable.
8836 The first argument is the matching ``llvm.invariant.start`` intrinsic.
8837 The second argument is a constant integer representing the size of the
8838 object, or -1 if it is variable sized and the third argument is a
8839 pointer to the object.
8844 This intrinsic indicates that the memory is mutable again.
8849 This class of intrinsics is designed to be generic and has no specific
8852 '``llvm.var.annotation``' Intrinsic
8853 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8860 declare void @llvm.var.annotation(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
8865 The '``llvm.var.annotation``' intrinsic.
8870 The first argument is a pointer to a value, the second is a pointer to a
8871 global string, the third is a pointer to a global string which is the
8872 source file name, and the last argument is the line number.
8877 This intrinsic allows annotation of local variables with arbitrary
8878 strings. This can be useful for special purpose optimizations that want
8879 to look for these annotations. These have no other defined use; they are
8880 ignored by code generation and optimization.
8882 '``llvm.ptr.annotation.*``' Intrinsic
8883 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8888 This is an overloaded intrinsic. You can use '``llvm.ptr.annotation``' on a
8889 pointer to an integer of any width. *NOTE* you must specify an address space for
8890 the pointer. The identifier for the default address space is the integer
8895 declare i8* @llvm.ptr.annotation.p<address space>i8(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
8896 declare i16* @llvm.ptr.annotation.p<address space>i16(i16* <val>, i8* <str>, i8* <str>, i32 <int>)
8897 declare i32* @llvm.ptr.annotation.p<address space>i32(i32* <val>, i8* <str>, i8* <str>, i32 <int>)
8898 declare i64* @llvm.ptr.annotation.p<address space>i64(i64* <val>, i8* <str>, i8* <str>, i32 <int>)
8899 declare i256* @llvm.ptr.annotation.p<address space>i256(i256* <val>, i8* <str>, i8* <str>, i32 <int>)
8904 The '``llvm.ptr.annotation``' intrinsic.
8909 The first argument is a pointer to an integer value of arbitrary bitwidth
8910 (result of some expression), the second is a pointer to a global string, the
8911 third is a pointer to a global string which is the source file name, and the
8912 last argument is the line number. It returns the value of the first argument.
8917 This intrinsic allows annotation of a pointer to an integer with arbitrary
8918 strings. This can be useful for special purpose optimizations that want to look
8919 for these annotations. These have no other defined use; they are ignored by code
8920 generation and optimization.
8922 '``llvm.annotation.*``' Intrinsic
8923 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8928 This is an overloaded intrinsic. You can use '``llvm.annotation``' on
8929 any integer bit width.
8933 declare i8 @llvm.annotation.i8(i8 <val>, i8* <str>, i8* <str>, i32 <int>)
8934 declare i16 @llvm.annotation.i16(i16 <val>, i8* <str>, i8* <str>, i32 <int>)
8935 declare i32 @llvm.annotation.i32(i32 <val>, i8* <str>, i8* <str>, i32 <int>)
8936 declare i64 @llvm.annotation.i64(i64 <val>, i8* <str>, i8* <str>, i32 <int>)
8937 declare i256 @llvm.annotation.i256(i256 <val>, i8* <str>, i8* <str>, i32 <int>)
8942 The '``llvm.annotation``' intrinsic.
8947 The first argument is an integer value (result of some expression), the
8948 second is a pointer to a global string, the third is a pointer to a
8949 global string which is the source file name, and the last argument is
8950 the line number. It returns the value of the first argument.
8955 This intrinsic allows annotations to be put on arbitrary expressions
8956 with arbitrary strings. This can be useful for special purpose
8957 optimizations that want to look for these annotations. These have no
8958 other defined use; they are ignored by code generation and optimization.
8960 '``llvm.trap``' Intrinsic
8961 ^^^^^^^^^^^^^^^^^^^^^^^^^
8968 declare void @llvm.trap() noreturn nounwind
8973 The '``llvm.trap``' intrinsic.
8983 This intrinsic is lowered to the target dependent trap instruction. If
8984 the target does not have a trap instruction, this intrinsic will be
8985 lowered to a call of the ``abort()`` function.
8987 '``llvm.debugtrap``' Intrinsic
8988 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8995 declare void @llvm.debugtrap() nounwind
9000 The '``llvm.debugtrap``' intrinsic.
9010 This intrinsic is lowered to code which is intended to cause an
9011 execution trap with the intention of requesting the attention of a
9014 '``llvm.stackprotector``' Intrinsic
9015 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9022 declare void @llvm.stackprotector(i8* <guard>, i8** <slot>)
9027 The ``llvm.stackprotector`` intrinsic takes the ``guard`` and stores it
9028 onto the stack at ``slot``. The stack slot is adjusted to ensure that it
9029 is placed on the stack before local variables.
9034 The ``llvm.stackprotector`` intrinsic requires two pointer arguments.
9035 The first argument is the value loaded from the stack guard
9036 ``@__stack_chk_guard``. The second variable is an ``alloca`` that has
9037 enough space to hold the value of the guard.
9042 This intrinsic causes the prologue/epilogue inserter to force the position of
9043 the ``AllocaInst`` stack slot to be before local variables on the stack. This is
9044 to ensure that if a local variable on the stack is overwritten, it will destroy
9045 the value of the guard. When the function exits, the guard on the stack is
9046 checked against the original guard by ``llvm.stackprotectorcheck``. If they are
9047 different, then ``llvm.stackprotectorcheck`` causes the program to abort by
9048 calling the ``__stack_chk_fail()`` function.
9050 '``llvm.stackprotectorcheck``' Intrinsic
9051 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9058 declare void @llvm.stackprotectorcheck(i8** <guard>)
9063 The ``llvm.stackprotectorcheck`` intrinsic compares ``guard`` against an already
9064 created stack protector and if they are not equal calls the
9065 ``__stack_chk_fail()`` function.
9070 The ``llvm.stackprotectorcheck`` intrinsic requires one pointer argument, the
9071 the variable ``@__stack_chk_guard``.
9076 This intrinsic is provided to perform the stack protector check by comparing
9077 ``guard`` with the stack slot created by ``llvm.stackprotector`` and if the
9078 values do not match call the ``__stack_chk_fail()`` function.
9080 The reason to provide this as an IR level intrinsic instead of implementing it
9081 via other IR operations is that in order to perform this operation at the IR
9082 level without an intrinsic, one would need to create additional basic blocks to
9083 handle the success/failure cases. This makes it difficult to stop the stack
9084 protector check from disrupting sibling tail calls in Codegen. With this
9085 intrinsic, we are able to generate the stack protector basic blocks late in
9086 codegen after the tail call decision has occurred.
9088 '``llvm.objectsize``' Intrinsic
9089 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9096 declare i32 @llvm.objectsize.i32(i8* <object>, i1 <min>)
9097 declare i64 @llvm.objectsize.i64(i8* <object>, i1 <min>)
9102 The ``llvm.objectsize`` intrinsic is designed to provide information to
9103 the optimizers to determine at compile time whether a) an operation
9104 (like memcpy) will overflow a buffer that corresponds to an object, or
9105 b) that a runtime check for overflow isn't necessary. An object in this
9106 context means an allocation of a specific class, structure, array, or
9112 The ``llvm.objectsize`` intrinsic takes two arguments. The first
9113 argument is a pointer to or into the ``object``. The second argument is
9114 a boolean and determines whether ``llvm.objectsize`` returns 0 (if true)
9115 or -1 (if false) when the object size is unknown. The second argument
9116 only accepts constants.
9121 The ``llvm.objectsize`` intrinsic is lowered to a constant representing
9122 the size of the object concerned. If the size cannot be determined at
9123 compile time, ``llvm.objectsize`` returns ``i32/i64 -1 or 0`` (depending
9124 on the ``min`` argument).
9126 '``llvm.expect``' Intrinsic
9127 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9132 This is an overloaded intrinsic. You can use ``llvm.expect`` on any
9137 declare i1 @llvm.expect.i1(i1 <val>, i1 <expected_val>)
9138 declare i32 @llvm.expect.i32(i32 <val>, i32 <expected_val>)
9139 declare i64 @llvm.expect.i64(i64 <val>, i64 <expected_val>)
9144 The ``llvm.expect`` intrinsic provides information about expected (the
9145 most probable) value of ``val``, which can be used by optimizers.
9150 The ``llvm.expect`` intrinsic takes two arguments. The first argument is
9151 a value. The second argument is an expected value, this needs to be a
9152 constant value, variables are not allowed.
9157 This intrinsic is lowered to the ``val``.
9159 '``llvm.donothing``' Intrinsic
9160 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9167 declare void @llvm.donothing() nounwind readnone
9172 The ``llvm.donothing`` intrinsic doesn't perform any operation. It's the
9173 only intrinsic that can be called with an invoke instruction.
9183 This intrinsic does nothing, and it's removed by optimizers and ignored
9186 Stack Map Intrinsics
9187 --------------------
9189 LLVM provides experimental intrinsics to support runtime patching
9190 mechanisms commonly desired in dynamic language JITs. These intrinsics
9191 are described in :doc:`StackMaps`.