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.
524 Global variables in other translation units can also be declared, in which
525 case they don't have an initializer.
527 Either global variable definitions or declarations may have an explicit section
528 to be placed in and may have an optional explicit alignment specified.
530 A variable may be defined as a global ``constant``, which indicates that
531 the contents of the variable will **never** be modified (enabling better
532 optimization, allowing the global data to be placed in the read-only
533 section of an executable, etc). Note that variables that need runtime
534 initialization cannot be marked ``constant`` as there is a store to the
537 LLVM explicitly allows *declarations* of global variables to be marked
538 constant, even if the final definition of the global is not. This
539 capability can be used to enable slightly better optimization of the
540 program, but requires the language definition to guarantee that
541 optimizations based on the 'constantness' are valid for the translation
542 units that do not include the definition.
544 As SSA values, global variables define pointer values that are in scope
545 (i.e. they dominate) all basic blocks in the program. Global variables
546 always define a pointer to their "content" type because they describe a
547 region of memory, and all memory objects in LLVM are accessed through
550 Global variables can be marked with ``unnamed_addr`` which indicates
551 that the address is not significant, only the content. Constants marked
552 like this can be merged with other constants if they have the same
553 initializer. Note that a constant with significant address *can* be
554 merged with a ``unnamed_addr`` constant, the result being a constant
555 whose address is significant.
557 A global variable may be declared to reside in a target-specific
558 numbered address space. For targets that support them, address spaces
559 may affect how optimizations are performed and/or what target
560 instructions are used to access the variable. The default address space
561 is zero. The address space qualifier must precede any other attributes.
563 LLVM allows an explicit section to be specified for globals. If the
564 target supports it, it will emit globals to the section specified.
566 By default, global initializers are optimized by assuming that global
567 variables defined within the module are not modified from their
568 initial values before the start of the global initializer. This is
569 true even for variables potentially accessible from outside the
570 module, including those with external linkage or appearing in
571 ``@llvm.used`` or dllexported variables. This assumption may be suppressed
572 by marking the variable with ``externally_initialized``.
574 An explicit alignment may be specified for a global, which must be a
575 power of 2. If not present, or if the alignment is set to zero, the
576 alignment of the global is set by the target to whatever it feels
577 convenient. If an explicit alignment is specified, the global is forced
578 to have exactly that alignment. Targets and optimizers are not allowed
579 to over-align the global if the global has an assigned section. In this
580 case, the extra alignment could be observable: for example, code could
581 assume that the globals are densely packed in their section and try to
582 iterate over them as an array, alignment padding would break this
585 Globals can also have a :ref:`DLL storage class <dllstorageclass>`.
587 Variables and aliasaes can have a
588 :ref:`Thread Local Storage Model <tls_model>`.
592 [@<GlobalVarName> =] [Linkage] [Visibility] [DLLStorageClass] [ThreadLocal]
593 [unnamed_addr] [AddrSpace] [ExternallyInitialized]
594 <global | constant> <Type> [<InitializerConstant>]
595 [, section "name"] [, align <Alignment>]
597 For example, the following defines a global in a numbered address space
598 with an initializer, section, and alignment:
602 @G = addrspace(5) constant float 1.0, section "foo", align 4
604 The following example just declares a global variable
608 @G = external global i32
610 The following example defines a thread-local global with the
611 ``initialexec`` TLS model:
615 @G = thread_local(initialexec) global i32 0, align 4
617 .. _functionstructure:
622 LLVM function definitions consist of the "``define``" keyword, an
623 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
624 style <visibility>`, an optional :ref:`DLL storage class <dllstorageclass>`,
625 an optional :ref:`calling convention <callingconv>`,
626 an optional ``unnamed_addr`` attribute, a return type, an optional
627 :ref:`parameter attribute <paramattrs>` for the return type, a function
628 name, a (possibly empty) argument list (each with optional :ref:`parameter
629 attributes <paramattrs>`), optional :ref:`function attributes <fnattrs>`,
630 an optional section, an optional alignment, an optional :ref:`garbage
631 collector name <gc>`, an optional :ref:`prefix <prefixdata>`, an opening
632 curly brace, a list of basic blocks, and a closing curly brace.
634 LLVM function declarations consist of the "``declare``" keyword, an
635 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
636 style <visibility>`, an optional :ref:`DLL storage class <dllstorageclass>`,
637 an optional :ref:`calling convention <callingconv>`,
638 an optional ``unnamed_addr`` attribute, a return type, an optional
639 :ref:`parameter attribute <paramattrs>` for the return type, a function
640 name, a possibly empty list of arguments, an optional alignment, an optional
641 :ref:`garbage collector name <gc>` and an optional :ref:`prefix <prefixdata>`.
643 A function definition contains a list of basic blocks, forming the CFG (Control
644 Flow Graph) for the function. Each basic block may optionally start with a label
645 (giving the basic block a symbol table entry), contains a list of instructions,
646 and ends with a :ref:`terminator <terminators>` instruction (such as a branch or
647 function return). If an explicit label is not provided, a block is assigned an
648 implicit numbered label, using the next value from the same counter as used for
649 unnamed temporaries (:ref:`see above<identifiers>`). For example, if a function
650 entry block does not have an explicit label, it will be assigned label "%0",
651 then the first unnamed temporary in that block will be "%1", etc.
653 The first basic block in a function is special in two ways: it is
654 immediately executed on entrance to the function, and it is not allowed
655 to have predecessor basic blocks (i.e. there can not be any branches to
656 the entry block of a function). Because the block can have no
657 predecessors, it also cannot have any :ref:`PHI nodes <i_phi>`.
659 LLVM allows an explicit section to be specified for functions. If the
660 target supports it, it will emit functions to the section specified.
662 An explicit alignment may be specified for a function. If not present,
663 or if the alignment is set to zero, the alignment of the function is set
664 by the target to whatever it feels convenient. If an explicit alignment
665 is specified, the function is forced to have at least that much
666 alignment. All alignments must be a power of 2.
668 If the ``unnamed_addr`` attribute is given, the address is know to not
669 be significant and two identical functions can be merged.
673 define [linkage] [visibility] [DLLStorageClass]
675 <ResultType> @<FunctionName> ([argument list])
676 [unnamed_addr] [fn Attrs] [section "name"] [align N]
677 [gc] [prefix Constant] { ... }
684 Aliases, unlike function or variables, don't create any new data. They
685 are just a new symbol and metadata for an existing position.
687 Aliases have a name and an aliasee that is either a global value or a
690 Aliases may have an optional :ref:`linkage type <linkage>`, an optional
691 :ref:`visibility style <visibility>`, an optional :ref:`DLL storage class
692 <dllstorageclass>` and an optional :ref:`tls model <tls_model>`.
696 @<Name> = [Visibility] [DLLStorageClass] [ThreadLocal] [unnamed_addr] alias [Linkage] <AliaseeTy> @<Aliasee>
698 The linkage must be one of ``private``, ``internal``, ``linkonce``, ``weak``,
699 ``linkonce_odr``, ``weak_odr``, ``external``. Note that some system linkers
700 might not correctly handle dropping a weak symbol that is aliased.
702 Alias that are not ``unnamed_addr`` are guaranteed to have the same address as
703 the aliasee expression. ``unnamed_addr`` ones are only guaranteed to point
706 Since aliases are only a second name, some restrictions apply, of which
707 some can only be checked when producing an object file:
709 * The expression defining the aliasee must be computable at assembly
710 time. Since it is just a name, no relocations can be used.
712 * No alias in the expression can be weak as the possibility of the
713 intermediate alias being overridden cannot be represented in an
716 * No global value in the expression can be a declaration, since that
717 would require a relocation, which is not possible.
719 .. _namedmetadatastructure:
724 Named metadata is a collection of metadata. :ref:`Metadata
725 nodes <metadata>` (but not metadata strings) are the only valid
726 operands for a named metadata.
730 ; Some unnamed metadata nodes, which are referenced by the named metadata.
731 !0 = metadata !{metadata !"zero"}
732 !1 = metadata !{metadata !"one"}
733 !2 = metadata !{metadata !"two"}
735 !name = !{!0, !1, !2}
742 The return type and each parameter of a function type may have a set of
743 *parameter attributes* associated with them. Parameter attributes are
744 used to communicate additional information about the result or
745 parameters of a function. Parameter attributes are considered to be part
746 of the function, not of the function type, so functions with different
747 parameter attributes can have the same function type.
749 Parameter attributes are simple keywords that follow the type specified.
750 If multiple parameter attributes are needed, they are space separated.
755 declare i32 @printf(i8* noalias nocapture, ...)
756 declare i32 @atoi(i8 zeroext)
757 declare signext i8 @returns_signed_char()
759 Note that any attributes for the function result (``nounwind``,
760 ``readonly``) come immediately after the argument list.
762 Currently, only the following parameter attributes are defined:
765 This indicates to the code generator that the parameter or return
766 value should be zero-extended to the extent required by the target's
767 ABI (which is usually 32-bits, but is 8-bits for a i1 on x86-64) by
768 the caller (for a parameter) or the callee (for a return value).
770 This indicates to the code generator that the parameter or return
771 value should be sign-extended to the extent required by the target's
772 ABI (which is usually 32-bits) by the caller (for a parameter) or
773 the callee (for a return value).
775 This indicates that this parameter or return value should be treated
776 in a special target-dependent fashion during while emitting code for
777 a function call or return (usually, by putting it in a register as
778 opposed to memory, though some targets use it to distinguish between
779 two different kinds of registers). Use of this attribute is
782 This indicates that the pointer parameter should really be passed by
783 value to the function. The attribute implies that a hidden copy of
784 the pointee is made between the caller and the callee, so the callee
785 is unable to modify the value in the caller. This attribute is only
786 valid on LLVM pointer arguments. It is generally used to pass
787 structs and arrays by value, but is also valid on pointers to
788 scalars. The copy is considered to belong to the caller not the
789 callee (for example, ``readonly`` functions should not write to
790 ``byval`` parameters). This is not a valid attribute for return
793 The byval attribute also supports specifying an alignment with the
794 align attribute. It indicates the alignment of the stack slot to
795 form and the known alignment of the pointer specified to the call
796 site. If the alignment is not specified, then the code generator
797 makes a target-specific assumption.
803 The ``inalloca`` argument attribute allows the caller to take the
804 address of outgoing stack arguments. An ``inalloca`` argument must
805 be a pointer to stack memory produced by an ``alloca`` instruction.
806 The alloca, or argument allocation, must also be tagged with the
807 inalloca keyword. Only the past argument may have the ``inalloca``
808 attribute, and that argument is guaranteed to be passed in memory.
810 An argument allocation may be used by a call at most once because
811 the call may deallocate it. The ``inalloca`` attribute cannot be
812 used in conjunction with other attributes that affect argument
813 storage, like ``inreg``, ``nest``, ``sret``, or ``byval``. The
814 ``inalloca`` attribute also disables LLVM's implicit lowering of
815 large aggregate return values, which means that frontend authors
816 must lower them with ``sret`` pointers.
818 When the call site is reached, the argument allocation must have
819 been the most recent stack allocation that is still live, or the
820 results are undefined. It is possible to allocate additional stack
821 space after an argument allocation and before its call site, but it
822 must be cleared off with :ref:`llvm.stackrestore
825 See :doc:`InAlloca` for more information on how to use this
829 This indicates that the pointer parameter specifies the address of a
830 structure that is the return value of the function in the source
831 program. This pointer must be guaranteed by the caller to be valid:
832 loads and stores to the structure may be assumed by the callee
833 not to trap and to be properly aligned. This may only be applied to
834 the first parameter. This is not a valid attribute for return
840 This indicates that pointer values :ref:`based <pointeraliasing>` on
841 the argument or return value do not alias pointer values which are
842 not *based* on it, ignoring certain "irrelevant" dependencies. For a
843 call to the parent function, dependencies between memory references
844 from before or after the call and from those during the call are
845 "irrelevant" to the ``noalias`` keyword for the arguments and return
846 value used in that call. The caller shares the responsibility with
847 the callee for ensuring that these requirements are met. For further
848 details, please see the discussion of the NoAlias response in :ref:`alias
849 analysis <Must, May, or No>`.
851 Note that this definition of ``noalias`` is intentionally similar
852 to the definition of ``restrict`` in C99 for function arguments,
853 though it is slightly weaker.
855 For function return values, C99's ``restrict`` is not meaningful,
856 while LLVM's ``noalias`` is.
858 This indicates that the callee does not make any copies of the
859 pointer that outlive the callee itself. This is not a valid
860 attribute for return values.
865 This indicates that the pointer parameter can be excised using the
866 :ref:`trampoline intrinsics <int_trampoline>`. This is not a valid
867 attribute for return values and can only be applied to one parameter.
870 This indicates that the function always returns the argument as its return
871 value. This is an optimization hint to the code generator when generating
872 the caller, allowing tail call optimization and omission of register saves
873 and restores in some cases; it is not checked or enforced when generating
874 the callee. The parameter and the function return type must be valid
875 operands for the :ref:`bitcast instruction <i_bitcast>`. This is not a
876 valid attribute for return values and can only be applied to one parameter.
879 This indicates that the parameter or return pointer is not null. This
880 attribute may only be applied to pointer typed parameters. This is not
881 checked or enforced by LLVM, the caller must ensure that the pointer
882 passed in is non-null, or the callee must ensure that the returned pointer
887 Garbage Collector Names
888 -----------------------
890 Each function may specify a garbage collector name, which is simply a
895 define void @f() gc "name" { ... }
897 The compiler declares the supported values of *name*. Specifying a
898 collector which will cause the compiler to alter its output in order to
899 support the named garbage collection algorithm.
906 Prefix data is data associated with a function which the code generator
907 will emit immediately before the function body. The purpose of this feature
908 is to allow frontends to associate language-specific runtime metadata with
909 specific functions and make it available through the function pointer while
910 still allowing the function pointer to be called. To access the data for a
911 given function, a program may bitcast the function pointer to a pointer to
912 the constant's type. This implies that the IR symbol points to the start
915 To maintain the semantics of ordinary function calls, the prefix data must
916 have a particular format. Specifically, it must begin with a sequence of
917 bytes which decode to a sequence of machine instructions, valid for the
918 module's target, which transfer control to the point immediately succeeding
919 the prefix data, without performing any other visible action. This allows
920 the inliner and other passes to reason about the semantics of the function
921 definition without needing to reason about the prefix data. Obviously this
922 makes the format of the prefix data highly target dependent.
924 Prefix data is laid out as if it were an initializer for a global variable
925 of the prefix data's type. No padding is automatically placed between the
926 prefix data and the function body. If padding is required, it must be part
929 A trivial example of valid prefix data for the x86 architecture is ``i8 144``,
930 which encodes the ``nop`` instruction:
934 define void @f() prefix i8 144 { ... }
936 Generally prefix data can be formed by encoding a relative branch instruction
937 which skips the metadata, as in this example of valid prefix data for the
938 x86_64 architecture, where the first two bytes encode ``jmp .+10``:
942 %0 = type <{ i8, i8, i8* }>
944 define void @f() prefix %0 <{ i8 235, i8 8, i8* @md}> { ... }
946 A function may have prefix data but no body. This has similar semantics
947 to the ``available_externally`` linkage in that the data may be used by the
948 optimizers but will not be emitted in the object file.
955 Attribute groups are groups of attributes that are referenced by objects within
956 the IR. They are important for keeping ``.ll`` files readable, because a lot of
957 functions will use the same set of attributes. In the degenerative case of a
958 ``.ll`` file that corresponds to a single ``.c`` file, the single attribute
959 group will capture the important command line flags used to build that file.
961 An attribute group is a module-level object. To use an attribute group, an
962 object references the attribute group's ID (e.g. ``#37``). An object may refer
963 to more than one attribute group. In that situation, the attributes from the
964 different groups are merged.
966 Here is an example of attribute groups for a function that should always be
967 inlined, has a stack alignment of 4, and which shouldn't use SSE instructions:
971 ; Target-independent attributes:
972 attributes #0 = { alwaysinline alignstack=4 }
974 ; Target-dependent attributes:
975 attributes #1 = { "no-sse" }
977 ; Function @f has attributes: alwaysinline, alignstack=4, and "no-sse".
978 define void @f() #0 #1 { ... }
985 Function attributes are set to communicate additional information about
986 a function. Function attributes are considered to be part of the
987 function, not of the function type, so functions with different function
988 attributes can have the same function type.
990 Function attributes are simple keywords that follow the type specified.
991 If multiple attributes are needed, they are space separated. For
996 define void @f() noinline { ... }
997 define void @f() alwaysinline { ... }
998 define void @f() alwaysinline optsize { ... }
999 define void @f() optsize { ... }
1002 This attribute indicates that, when emitting the prologue and
1003 epilogue, the backend should forcibly align the stack pointer.
1004 Specify the desired alignment, which must be a power of two, in
1007 This attribute indicates that the inliner should attempt to inline
1008 this function into callers whenever possible, ignoring any active
1009 inlining size threshold for this caller.
1011 This indicates that the callee function at a call site should be
1012 recognized as a built-in function, even though the function's declaration
1013 uses the ``nobuiltin`` attribute. This is only valid at call sites for
1014 direct calls to functions which are declared with the ``nobuiltin``
1017 This attribute indicates that this function is rarely called. When
1018 computing edge weights, basic blocks post-dominated by a cold
1019 function call are also considered to be cold; and, thus, given low
1022 This attribute indicates that the source code contained a hint that
1023 inlining this function is desirable (such as the "inline" keyword in
1024 C/C++). It is just a hint; it imposes no requirements on the
1027 This attribute indicates that the function should be added to a
1028 jump-instruction table at code-generation time, and that all address-taken
1029 references to this function should be replaced with a reference to the
1030 appropriate jump-instruction-table function pointer. Note that this creates
1031 a new pointer for the original function, which means that code that depends
1032 on function-pointer identity can break. So, any function annotated with
1033 ``jumptable`` must also be ``unnamed_addr``.
1035 This attribute suggests that optimization passes and code generator
1036 passes make choices that keep the code size of this function as small
1037 as possible and perform optimizations that may sacrifice runtime
1038 performance in order to minimize the size of the generated code.
1040 This attribute disables prologue / epilogue emission for the
1041 function. This can have very system-specific consequences.
1043 This indicates that the callee function at a call site is not recognized as
1044 a built-in function. LLVM will retain the original call and not replace it
1045 with equivalent code based on the semantics of the built-in function, unless
1046 the call site uses the ``builtin`` attribute. This is valid at call sites
1047 and on function declarations and definitions.
1049 This attribute indicates that calls to the function cannot be
1050 duplicated. A call to a ``noduplicate`` function may be moved
1051 within its parent function, but may not be duplicated within
1052 its parent function.
1054 A function containing a ``noduplicate`` call may still
1055 be an inlining candidate, provided that the call is not
1056 duplicated by inlining. That implies that the function has
1057 internal linkage and only has one call site, so the original
1058 call is dead after inlining.
1060 This attributes disables implicit floating point instructions.
1062 This attribute indicates that the inliner should never inline this
1063 function in any situation. This attribute may not be used together
1064 with the ``alwaysinline`` attribute.
1066 This attribute suppresses lazy symbol binding for the function. This
1067 may make calls to the function faster, at the cost of extra program
1068 startup time if the function is not called during program startup.
1070 This attribute indicates that the code generator should not use a
1071 red zone, even if the target-specific ABI normally permits it.
1073 This function attribute indicates that the function never returns
1074 normally. This produces undefined behavior at runtime if the
1075 function ever does dynamically return.
1077 This function attribute indicates that the function never returns
1078 with an unwind or exceptional control flow. If the function does
1079 unwind, its runtime behavior is undefined.
1081 This function attribute indicates that the function is not optimized
1082 by any optimization or code generator passes with the
1083 exception of interprocedural optimization passes.
1084 This attribute cannot be used together with the ``alwaysinline``
1085 attribute; this attribute is also incompatible
1086 with the ``minsize`` attribute and the ``optsize`` attribute.
1088 This attribute requires the ``noinline`` attribute to be specified on
1089 the function as well, so the function is never inlined into any caller.
1090 Only functions with the ``alwaysinline`` attribute are valid
1091 candidates for inlining into the body of this function.
1093 This attribute suggests that optimization passes and code generator
1094 passes make choices that keep the code size of this function low,
1095 and otherwise do optimizations specifically to reduce code size as
1096 long as they do not significantly impact runtime performance.
1098 On a function, this attribute indicates that the function computes its
1099 result (or decides to unwind an exception) based strictly on its arguments,
1100 without dereferencing any pointer arguments or otherwise accessing
1101 any mutable state (e.g. memory, control registers, etc) visible to
1102 caller functions. It does not write through any pointer arguments
1103 (including ``byval`` arguments) and never changes any state visible
1104 to callers. This means that it cannot unwind exceptions by calling
1105 the ``C++`` exception throwing methods.
1107 On an argument, this attribute indicates that the function does not
1108 dereference that pointer argument, even though it may read or write the
1109 memory that the pointer points to if accessed through other pointers.
1111 On a function, this attribute indicates that the function does not write
1112 through any pointer arguments (including ``byval`` arguments) or otherwise
1113 modify any state (e.g. memory, control registers, etc) visible to
1114 caller functions. It may dereference pointer arguments and read
1115 state that may be set in the caller. A readonly function always
1116 returns the same value (or unwinds an exception identically) when
1117 called with the same set of arguments and global state. It cannot
1118 unwind an exception by calling the ``C++`` exception throwing
1121 On an argument, this attribute indicates that the function does not write
1122 through this pointer argument, even though it may write to the memory that
1123 the pointer points to.
1125 This attribute indicates that this function can return twice. The C
1126 ``setjmp`` is an example of such a function. The compiler disables
1127 some optimizations (like tail calls) in the caller of these
1129 ``sanitize_address``
1130 This attribute indicates that AddressSanitizer checks
1131 (dynamic address safety analysis) are enabled for this function.
1133 This attribute indicates that MemorySanitizer checks (dynamic detection
1134 of accesses to uninitialized memory) are enabled for this function.
1136 This attribute indicates that ThreadSanitizer checks
1137 (dynamic thread safety analysis) are enabled for this function.
1139 This attribute indicates that the function should emit a stack
1140 smashing protector. It is in the form of a "canary" --- a random value
1141 placed on the stack before the local variables that's checked upon
1142 return from the function to see if it has been overwritten. A
1143 heuristic is used to determine if a function needs stack protectors
1144 or not. The heuristic used will enable protectors for functions with:
1146 - Character arrays larger than ``ssp-buffer-size`` (default 8).
1147 - Aggregates containing character arrays larger than ``ssp-buffer-size``.
1148 - Calls to alloca() with variable sizes or constant sizes greater than
1149 ``ssp-buffer-size``.
1151 Variables that are identified as requiring a protector will be arranged
1152 on the stack such that they are adjacent to the stack protector guard.
1154 If a function that has an ``ssp`` attribute is inlined into a
1155 function that doesn't have an ``ssp`` attribute, then the resulting
1156 function will have an ``ssp`` attribute.
1158 This attribute indicates that the function should *always* emit a
1159 stack smashing protector. This overrides the ``ssp`` function
1162 Variables that are identified as requiring a protector will be arranged
1163 on the stack such that they are adjacent to the stack protector guard.
1164 The specific layout rules are:
1166 #. Large arrays and structures containing large arrays
1167 (``>= ssp-buffer-size``) are closest to the stack protector.
1168 #. Small arrays and structures containing small arrays
1169 (``< ssp-buffer-size``) are 2nd closest to the protector.
1170 #. Variables that have had their address taken are 3rd closest to the
1173 If a function that has an ``sspreq`` attribute is inlined into a
1174 function that doesn't have an ``sspreq`` attribute or which has an
1175 ``ssp`` or ``sspstrong`` attribute, then the resulting function will have
1176 an ``sspreq`` attribute.
1178 This attribute indicates that the function should emit a stack smashing
1179 protector. This attribute causes a strong heuristic to be used when
1180 determining if a function needs stack protectors. The strong heuristic
1181 will enable protectors for functions with:
1183 - Arrays of any size and type
1184 - Aggregates containing an array of any size and type.
1185 - Calls to alloca().
1186 - Local variables that have had their address taken.
1188 Variables that are identified as requiring a protector will be arranged
1189 on the stack such that they are adjacent to the stack protector guard.
1190 The specific layout rules are:
1192 #. Large arrays and structures containing large arrays
1193 (``>= ssp-buffer-size``) are closest to the stack protector.
1194 #. Small arrays and structures containing small arrays
1195 (``< ssp-buffer-size``) are 2nd closest to the protector.
1196 #. Variables that have had their address taken are 3rd closest to the
1199 This overrides the ``ssp`` function attribute.
1201 If a function that has an ``sspstrong`` attribute is inlined into a
1202 function that doesn't have an ``sspstrong`` attribute, then the
1203 resulting function will have an ``sspstrong`` attribute.
1205 This attribute indicates that the ABI being targeted requires that
1206 an unwind table entry be produce for this function even if we can
1207 show that no exceptions passes by it. This is normally the case for
1208 the ELF x86-64 abi, but it can be disabled for some compilation
1213 Module-Level Inline Assembly
1214 ----------------------------
1216 Modules may contain "module-level inline asm" blocks, which corresponds
1217 to the GCC "file scope inline asm" blocks. These blocks are internally
1218 concatenated by LLVM and treated as a single unit, but may be separated
1219 in the ``.ll`` file if desired. The syntax is very simple:
1221 .. code-block:: llvm
1223 module asm "inline asm code goes here"
1224 module asm "more can go here"
1226 The strings can contain any character by escaping non-printable
1227 characters. The escape sequence used is simply "\\xx" where "xx" is the
1228 two digit hex code for the number.
1230 The inline asm code is simply printed to the machine code .s file when
1231 assembly code is generated.
1233 .. _langref_datalayout:
1238 A module may specify a target specific data layout string that specifies
1239 how data is to be laid out in memory. The syntax for the data layout is
1242 .. code-block:: llvm
1244 target datalayout = "layout specification"
1246 The *layout specification* consists of a list of specifications
1247 separated by the minus sign character ('-'). Each specification starts
1248 with a letter and may include other information after the letter to
1249 define some aspect of the data layout. The specifications accepted are
1253 Specifies that the target lays out data in big-endian form. That is,
1254 the bits with the most significance have the lowest address
1257 Specifies that the target lays out data in little-endian form. That
1258 is, the bits with the least significance have the lowest address
1261 Specifies the natural alignment of the stack in bits. Alignment
1262 promotion of stack variables is limited to the natural stack
1263 alignment to avoid dynamic stack realignment. The stack alignment
1264 must be a multiple of 8-bits. If omitted, the natural stack
1265 alignment defaults to "unspecified", which does not prevent any
1266 alignment promotions.
1267 ``p[n]:<size>:<abi>:<pref>``
1268 This specifies the *size* of a pointer and its ``<abi>`` and
1269 ``<pref>``\erred alignments for address space ``n``. All sizes are in
1270 bits. The address space, ``n`` is optional, and if not specified,
1271 denotes the default address space 0. The value of ``n`` must be
1272 in the range [1,2^23).
1273 ``i<size>:<abi>:<pref>``
1274 This specifies the alignment for an integer type of a given bit
1275 ``<size>``. The value of ``<size>`` must be in the range [1,2^23).
1276 ``v<size>:<abi>:<pref>``
1277 This specifies the alignment for a vector type of a given bit
1279 ``f<size>:<abi>:<pref>``
1280 This specifies the alignment for a floating point type of a given bit
1281 ``<size>``. Only values of ``<size>`` that are supported by the target
1282 will work. 32 (float) and 64 (double) are supported on all targets; 80
1283 or 128 (different flavors of long double) are also supported on some
1286 This specifies the alignment for an object of aggregate type.
1288 If present, specifies that llvm names are mangled in the output. The
1291 * ``e``: ELF mangling: Private symbols get a ``.L`` prefix.
1292 * ``m``: Mips mangling: Private symbols get a ``$`` prefix.
1293 * ``o``: Mach-O mangling: Private symbols get ``L`` prefix. Other
1294 symbols get a ``_`` prefix.
1295 * ``w``: Windows COFF prefix: Similar to Mach-O, but stdcall and fastcall
1296 functions also get a suffix based on the frame size.
1297 ``n<size1>:<size2>:<size3>...``
1298 This specifies a set of native integer widths for the target CPU in
1299 bits. For example, it might contain ``n32`` for 32-bit PowerPC,
1300 ``n32:64`` for PowerPC 64, or ``n8:16:32:64`` for X86-64. Elements of
1301 this set are considered to support most general arithmetic operations
1304 On every specification that takes a ``<abi>:<pref>``, specifying the
1305 ``<pref>`` alignment is optional. If omitted, the preceding ``:``
1306 should be omitted too and ``<pref>`` will be equal to ``<abi>``.
1308 When constructing the data layout for a given target, LLVM starts with a
1309 default set of specifications which are then (possibly) overridden by
1310 the specifications in the ``datalayout`` keyword. The default
1311 specifications are given in this list:
1313 - ``E`` - big endian
1314 - ``p:64:64:64`` - 64-bit pointers with 64-bit alignment.
1315 - ``p[n]:64:64:64`` - Other address spaces are assumed to be the
1316 same as the default address space.
1317 - ``S0`` - natural stack alignment is unspecified
1318 - ``i1:8:8`` - i1 is 8-bit (byte) aligned
1319 - ``i8:8:8`` - i8 is 8-bit (byte) aligned
1320 - ``i16:16:16`` - i16 is 16-bit aligned
1321 - ``i32:32:32`` - i32 is 32-bit aligned
1322 - ``i64:32:64`` - i64 has ABI alignment of 32-bits but preferred
1323 alignment of 64-bits
1324 - ``f16:16:16`` - half is 16-bit aligned
1325 - ``f32:32:32`` - float is 32-bit aligned
1326 - ``f64:64:64`` - double is 64-bit aligned
1327 - ``f128:128:128`` - quad is 128-bit aligned
1328 - ``v64:64:64`` - 64-bit vector is 64-bit aligned
1329 - ``v128:128:128`` - 128-bit vector is 128-bit aligned
1330 - ``a:0:64`` - aggregates are 64-bit aligned
1332 When LLVM is determining the alignment for a given type, it uses the
1335 #. If the type sought is an exact match for one of the specifications,
1336 that specification is used.
1337 #. If no match is found, and the type sought is an integer type, then
1338 the smallest integer type that is larger than the bitwidth of the
1339 sought type is used. If none of the specifications are larger than
1340 the bitwidth then the largest integer type is used. For example,
1341 given the default specifications above, the i7 type will use the
1342 alignment of i8 (next largest) while both i65 and i256 will use the
1343 alignment of i64 (largest specified).
1344 #. If no match is found, and the type sought is a vector type, then the
1345 largest vector type that is smaller than the sought vector type will
1346 be used as a fall back. This happens because <128 x double> can be
1347 implemented in terms of 64 <2 x double>, for example.
1349 The function of the data layout string may not be what you expect.
1350 Notably, this is not a specification from the frontend of what alignment
1351 the code generator should use.
1353 Instead, if specified, the target data layout is required to match what
1354 the ultimate *code generator* expects. This string is used by the
1355 mid-level optimizers to improve code, and this only works if it matches
1356 what the ultimate code generator uses. If you would like to generate IR
1357 that does not embed this target-specific detail into the IR, then you
1358 don't have to specify the string. This will disable some optimizations
1359 that require precise layout information, but this also prevents those
1360 optimizations from introducing target specificity into the IR.
1367 A module may specify a target triple string that describes the target
1368 host. The syntax for the target triple is simply:
1370 .. code-block:: llvm
1372 target triple = "x86_64-apple-macosx10.7.0"
1374 The *target triple* string consists of a series of identifiers delimited
1375 by the minus sign character ('-'). The canonical forms are:
1379 ARCHITECTURE-VENDOR-OPERATING_SYSTEM
1380 ARCHITECTURE-VENDOR-OPERATING_SYSTEM-ENVIRONMENT
1382 This information is passed along to the backend so that it generates
1383 code for the proper architecture. It's possible to override this on the
1384 command line with the ``-mtriple`` command line option.
1386 .. _pointeraliasing:
1388 Pointer Aliasing Rules
1389 ----------------------
1391 Any memory access must be done through a pointer value associated with
1392 an address range of the memory access, otherwise the behavior is
1393 undefined. Pointer values are associated with address ranges according
1394 to the following rules:
1396 - A pointer value is associated with the addresses associated with any
1397 value it is *based* on.
1398 - An address of a global variable is associated with the address range
1399 of the variable's storage.
1400 - The result value of an allocation instruction is associated with the
1401 address range of the allocated storage.
1402 - A null pointer in the default address-space is associated with no
1404 - An integer constant other than zero or a pointer value returned from
1405 a function not defined within LLVM may be associated with address
1406 ranges allocated through mechanisms other than those provided by
1407 LLVM. Such ranges shall not overlap with any ranges of addresses
1408 allocated by mechanisms provided by LLVM.
1410 A pointer value is *based* on another pointer value according to the
1413 - A pointer value formed from a ``getelementptr`` operation is *based*
1414 on the first operand of the ``getelementptr``.
1415 - The result value of a ``bitcast`` is *based* on the operand of the
1417 - A pointer value formed by an ``inttoptr`` is *based* on all pointer
1418 values that contribute (directly or indirectly) to the computation of
1419 the pointer's value.
1420 - The "*based* on" relationship is transitive.
1422 Note that this definition of *"based"* is intentionally similar to the
1423 definition of *"based"* in C99, though it is slightly weaker.
1425 LLVM IR does not associate types with memory. The result type of a
1426 ``load`` merely indicates the size and alignment of the memory from
1427 which to load, as well as the interpretation of the value. The first
1428 operand type of a ``store`` similarly only indicates the size and
1429 alignment of the store.
1431 Consequently, type-based alias analysis, aka TBAA, aka
1432 ``-fstrict-aliasing``, is not applicable to general unadorned LLVM IR.
1433 :ref:`Metadata <metadata>` may be used to encode additional information
1434 which specialized optimization passes may use to implement type-based
1439 Volatile Memory Accesses
1440 ------------------------
1442 Certain memory accesses, such as :ref:`load <i_load>`'s,
1443 :ref:`store <i_store>`'s, and :ref:`llvm.memcpy <int_memcpy>`'s may be
1444 marked ``volatile``. The optimizers must not change the number of
1445 volatile operations or change their order of execution relative to other
1446 volatile operations. The optimizers *may* change the order of volatile
1447 operations relative to non-volatile operations. This is not Java's
1448 "volatile" and has no cross-thread synchronization behavior.
1450 IR-level volatile loads and stores cannot safely be optimized into
1451 llvm.memcpy or llvm.memmove intrinsics even when those intrinsics are
1452 flagged volatile. Likewise, the backend should never split or merge
1453 target-legal volatile load/store instructions.
1455 .. admonition:: Rationale
1457 Platforms may rely on volatile loads and stores of natively supported
1458 data width to be executed as single instruction. For example, in C
1459 this holds for an l-value of volatile primitive type with native
1460 hardware support, but not necessarily for aggregate types. The
1461 frontend upholds these expectations, which are intentionally
1462 unspecified in the IR. The rules above ensure that IR transformation
1463 do not violate the frontend's contract with the language.
1467 Memory Model for Concurrent Operations
1468 --------------------------------------
1470 The LLVM IR does not define any way to start parallel threads of
1471 execution or to register signal handlers. Nonetheless, there are
1472 platform-specific ways to create them, and we define LLVM IR's behavior
1473 in their presence. This model is inspired by the C++0x memory model.
1475 For a more informal introduction to this model, see the :doc:`Atomics`.
1477 We define a *happens-before* partial order as the least partial order
1480 - Is a superset of single-thread program order, and
1481 - When a *synchronizes-with* ``b``, includes an edge from ``a`` to
1482 ``b``. *Synchronizes-with* pairs are introduced by platform-specific
1483 techniques, like pthread locks, thread creation, thread joining,
1484 etc., and by atomic instructions. (See also :ref:`Atomic Memory Ordering
1485 Constraints <ordering>`).
1487 Note that program order does not introduce *happens-before* edges
1488 between a thread and signals executing inside that thread.
1490 Every (defined) read operation (load instructions, memcpy, atomic
1491 loads/read-modify-writes, etc.) R reads a series of bytes written by
1492 (defined) write operations (store instructions, atomic
1493 stores/read-modify-writes, memcpy, etc.). For the purposes of this
1494 section, initialized globals are considered to have a write of the
1495 initializer which is atomic and happens before any other read or write
1496 of the memory in question. For each byte of a read R, R\ :sub:`byte`
1497 may see any write to the same byte, except:
1499 - If write\ :sub:`1` happens before write\ :sub:`2`, and
1500 write\ :sub:`2` happens before R\ :sub:`byte`, then
1501 R\ :sub:`byte` does not see write\ :sub:`1`.
1502 - If R\ :sub:`byte` happens before write\ :sub:`3`, then
1503 R\ :sub:`byte` does not see write\ :sub:`3`.
1505 Given that definition, R\ :sub:`byte` is defined as follows:
1507 - If R is volatile, the result is target-dependent. (Volatile is
1508 supposed to give guarantees which can support ``sig_atomic_t`` in
1509 C/C++, and may be used for accesses to addresses which do not behave
1510 like normal memory. It does not generally provide cross-thread
1512 - Otherwise, if there is no write to the same byte that happens before
1513 R\ :sub:`byte`, R\ :sub:`byte` returns ``undef`` for that byte.
1514 - Otherwise, if R\ :sub:`byte` may see exactly one write,
1515 R\ :sub:`byte` returns the value written by that write.
1516 - Otherwise, if R is atomic, and all the writes R\ :sub:`byte` may
1517 see are atomic, it chooses one of the values written. See the :ref:`Atomic
1518 Memory Ordering Constraints <ordering>` section for additional
1519 constraints on how the choice is made.
1520 - Otherwise R\ :sub:`byte` returns ``undef``.
1522 R returns the value composed of the series of bytes it read. This
1523 implies that some bytes within the value may be ``undef`` **without**
1524 the entire value being ``undef``. Note that this only defines the
1525 semantics of the operation; it doesn't mean that targets will emit more
1526 than one instruction to read the series of bytes.
1528 Note that in cases where none of the atomic intrinsics are used, this
1529 model places only one restriction on IR transformations on top of what
1530 is required for single-threaded execution: introducing a store to a byte
1531 which might not otherwise be stored is not allowed in general.
1532 (Specifically, in the case where another thread might write to and read
1533 from an address, introducing a store can change a load that may see
1534 exactly one write into a load that may see multiple writes.)
1538 Atomic Memory Ordering Constraints
1539 ----------------------------------
1541 Atomic instructions (:ref:`cmpxchg <i_cmpxchg>`,
1542 :ref:`atomicrmw <i_atomicrmw>`, :ref:`fence <i_fence>`,
1543 :ref:`atomic load <i_load>`, and :ref:`atomic store <i_store>`) take
1544 ordering parameters that determine which other atomic instructions on
1545 the same address they *synchronize with*. These semantics are borrowed
1546 from Java and C++0x, but are somewhat more colloquial. If these
1547 descriptions aren't precise enough, check those specs (see spec
1548 references in the :doc:`atomics guide <Atomics>`).
1549 :ref:`fence <i_fence>` instructions treat these orderings somewhat
1550 differently since they don't take an address. See that instruction's
1551 documentation for details.
1553 For a simpler introduction to the ordering constraints, see the
1557 The set of values that can be read is governed by the happens-before
1558 partial order. A value cannot be read unless some operation wrote
1559 it. This is intended to provide a guarantee strong enough to model
1560 Java's non-volatile shared variables. This ordering cannot be
1561 specified for read-modify-write operations; it is not strong enough
1562 to make them atomic in any interesting way.
1564 In addition to the guarantees of ``unordered``, there is a single
1565 total order for modifications by ``monotonic`` operations on each
1566 address. All modification orders must be compatible with the
1567 happens-before order. There is no guarantee that the modification
1568 orders can be combined to a global total order for the whole program
1569 (and this often will not be possible). The read in an atomic
1570 read-modify-write operation (:ref:`cmpxchg <i_cmpxchg>` and
1571 :ref:`atomicrmw <i_atomicrmw>`) reads the value in the modification
1572 order immediately before the value it writes. If one atomic read
1573 happens before another atomic read of the same address, the later
1574 read must see the same value or a later value in the address's
1575 modification order. This disallows reordering of ``monotonic`` (or
1576 stronger) operations on the same address. If an address is written
1577 ``monotonic``-ally by one thread, and other threads ``monotonic``-ally
1578 read that address repeatedly, the other threads must eventually see
1579 the write. This corresponds to the C++0x/C1x
1580 ``memory_order_relaxed``.
1582 In addition to the guarantees of ``monotonic``, a
1583 *synchronizes-with* edge may be formed with a ``release`` operation.
1584 This is intended to model C++'s ``memory_order_acquire``.
1586 In addition to the guarantees of ``monotonic``, if this operation
1587 writes a value which is subsequently read by an ``acquire``
1588 operation, it *synchronizes-with* that operation. (This isn't a
1589 complete description; see the C++0x definition of a release
1590 sequence.) This corresponds to the C++0x/C1x
1591 ``memory_order_release``.
1592 ``acq_rel`` (acquire+release)
1593 Acts as both an ``acquire`` and ``release`` operation on its
1594 address. This corresponds to the C++0x/C1x ``memory_order_acq_rel``.
1595 ``seq_cst`` (sequentially consistent)
1596 In addition to the guarantees of ``acq_rel`` (``acquire`` for an
1597 operation which only reads, ``release`` for an operation which only
1598 writes), there is a global total order on all
1599 sequentially-consistent operations on all addresses, which is
1600 consistent with the *happens-before* partial order and with the
1601 modification orders of all the affected addresses. Each
1602 sequentially-consistent read sees the last preceding write to the
1603 same address in this global order. This corresponds to the C++0x/C1x
1604 ``memory_order_seq_cst`` and Java volatile.
1608 If an atomic operation is marked ``singlethread``, it only *synchronizes
1609 with* or participates in modification and seq\_cst total orderings with
1610 other operations running in the same thread (for example, in signal
1618 LLVM IR floating-point binary ops (:ref:`fadd <i_fadd>`,
1619 :ref:`fsub <i_fsub>`, :ref:`fmul <i_fmul>`, :ref:`fdiv <i_fdiv>`,
1620 :ref:`frem <i_frem>`) have the following flags that can set to enable
1621 otherwise unsafe floating point operations
1624 No NaNs - Allow optimizations to assume the arguments and result are not
1625 NaN. Such optimizations are required to retain defined behavior over
1626 NaNs, but the value of the result is undefined.
1629 No Infs - Allow optimizations to assume the arguments and result are not
1630 +/-Inf. Such optimizations are required to retain defined behavior over
1631 +/-Inf, but the value of the result is undefined.
1634 No Signed Zeros - Allow optimizations to treat the sign of a zero
1635 argument or result as insignificant.
1638 Allow Reciprocal - Allow optimizations to use the reciprocal of an
1639 argument rather than perform division.
1642 Fast - Allow algebraically equivalent transformations that may
1643 dramatically change results in floating point (e.g. reassociate). This
1644 flag implies all the others.
1651 The LLVM type system is one of the most important features of the
1652 intermediate representation. Being typed enables a number of
1653 optimizations to be performed on the intermediate representation
1654 directly, without having to do extra analyses on the side before the
1655 transformation. A strong type system makes it easier to read the
1656 generated code and enables novel analyses and transformations that are
1657 not feasible to perform on normal three address code representations.
1667 The void type does not represent any value and has no size.
1685 The function type can be thought of as a function signature. It consists of a
1686 return type and a list of formal parameter types. The return type of a function
1687 type is a void type or first class type --- except for :ref:`label <t_label>`
1688 and :ref:`metadata <t_metadata>` types.
1694 <returntype> (<parameter list>)
1696 ...where '``<parameter list>``' is a comma-separated list of type
1697 specifiers. Optionally, the parameter list may include a type ``...``, which
1698 indicates that the function takes a variable number of arguments. Variable
1699 argument functions can access their arguments with the :ref:`variable argument
1700 handling intrinsic <int_varargs>` functions. '``<returntype>``' is any type
1701 except :ref:`label <t_label>` and :ref:`metadata <t_metadata>`.
1705 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1706 | ``i32 (i32)`` | function taking an ``i32``, returning an ``i32`` |
1707 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1708 | ``float (i16, i32 *) *`` | :ref:`Pointer <t_pointer>` to a function that takes an ``i16`` and a :ref:`pointer <t_pointer>` to ``i32``, returning ``float``. |
1709 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1710 | ``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. |
1711 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1712 | ``{i32, i32} (i32)`` | A function taking an ``i32``, returning a :ref:`structure <t_struct>` containing two ``i32`` values |
1713 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1720 The :ref:`first class <t_firstclass>` types are perhaps the most important.
1721 Values of these types are the only ones which can be produced by
1729 These are the types that are valid in registers from CodeGen's perspective.
1738 The integer type is a very simple type that simply specifies an
1739 arbitrary bit width for the integer type desired. Any bit width from 1
1740 bit to 2\ :sup:`23`\ -1 (about 8 million) can be specified.
1748 The number of bits the integer will occupy is specified by the ``N``
1754 +----------------+------------------------------------------------+
1755 | ``i1`` | a single-bit integer. |
1756 +----------------+------------------------------------------------+
1757 | ``i32`` | a 32-bit integer. |
1758 +----------------+------------------------------------------------+
1759 | ``i1942652`` | a really big integer of over 1 million bits. |
1760 +----------------+------------------------------------------------+
1764 Floating Point Types
1765 """"""""""""""""""""
1774 - 16-bit floating point value
1777 - 32-bit floating point value
1780 - 64-bit floating point value
1783 - 128-bit floating point value (112-bit mantissa)
1786 - 80-bit floating point value (X87)
1789 - 128-bit floating point value (two 64-bits)
1796 The x86_mmx type represents a value held in an MMX register on an x86
1797 machine. The operations allowed on it are quite limited: parameters and
1798 return values, load and store, and bitcast. User-specified MMX
1799 instructions are represented as intrinsic or asm calls with arguments
1800 and/or results of this type. There are no arrays, vectors or constants
1817 The pointer type is used to specify memory locations. Pointers are
1818 commonly used to reference objects in memory.
1820 Pointer types may have an optional address space attribute defining the
1821 numbered address space where the pointed-to object resides. The default
1822 address space is number zero. The semantics of non-zero address spaces
1823 are target-specific.
1825 Note that LLVM does not permit pointers to void (``void*``) nor does it
1826 permit pointers to labels (``label*``). Use ``i8*`` instead.
1836 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1837 | ``[4 x i32]*`` | A :ref:`pointer <t_pointer>` to :ref:`array <t_array>` of four ``i32`` values. |
1838 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1839 | ``i32 (i32*) *`` | A :ref:`pointer <t_pointer>` to a :ref:`function <t_function>` that takes an ``i32*``, returning an ``i32``. |
1840 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1841 | ``i32 addrspace(5)*`` | A :ref:`pointer <t_pointer>` to an ``i32`` value that resides in address space #5. |
1842 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1851 A vector type is a simple derived type that represents a vector of
1852 elements. Vector types are used when multiple primitive data are
1853 operated in parallel using a single instruction (SIMD). A vector type
1854 requires a size (number of elements) and an underlying primitive data
1855 type. Vector types are considered :ref:`first class <t_firstclass>`.
1861 < <# elements> x <elementtype> >
1863 The number of elements is a constant integer value larger than 0;
1864 elementtype may be any integer or floating point type, or a pointer to
1865 these types. Vectors of size zero are not allowed.
1869 +-------------------+--------------------------------------------------+
1870 | ``<4 x i32>`` | Vector of 4 32-bit integer values. |
1871 +-------------------+--------------------------------------------------+
1872 | ``<8 x float>`` | Vector of 8 32-bit floating-point values. |
1873 +-------------------+--------------------------------------------------+
1874 | ``<2 x i64>`` | Vector of 2 64-bit integer values. |
1875 +-------------------+--------------------------------------------------+
1876 | ``<4 x i64*>`` | Vector of 4 pointers to 64-bit integer values. |
1877 +-------------------+--------------------------------------------------+
1886 The label type represents code labels.
1901 The metadata type represents embedded metadata. No derived types may be
1902 created from metadata except for :ref:`function <t_function>` arguments.
1915 Aggregate Types are a subset of derived types that can contain multiple
1916 member types. :ref:`Arrays <t_array>` and :ref:`structs <t_struct>` are
1917 aggregate types. :ref:`Vectors <t_vector>` are not considered to be
1927 The array type is a very simple derived type that arranges elements
1928 sequentially in memory. The array type requires a size (number of
1929 elements) and an underlying data type.
1935 [<# elements> x <elementtype>]
1937 The number of elements is a constant integer value; ``elementtype`` may
1938 be any type with a size.
1942 +------------------+--------------------------------------+
1943 | ``[40 x i32]`` | Array of 40 32-bit integer values. |
1944 +------------------+--------------------------------------+
1945 | ``[41 x i32]`` | Array of 41 32-bit integer values. |
1946 +------------------+--------------------------------------+
1947 | ``[4 x i8]`` | Array of 4 8-bit integer values. |
1948 +------------------+--------------------------------------+
1950 Here are some examples of multidimensional arrays:
1952 +-----------------------------+----------------------------------------------------------+
1953 | ``[3 x [4 x i32]]`` | 3x4 array of 32-bit integer values. |
1954 +-----------------------------+----------------------------------------------------------+
1955 | ``[12 x [10 x float]]`` | 12x10 array of single precision floating point values. |
1956 +-----------------------------+----------------------------------------------------------+
1957 | ``[2 x [3 x [4 x i16]]]`` | 2x3x4 array of 16-bit integer values. |
1958 +-----------------------------+----------------------------------------------------------+
1960 There is no restriction on indexing beyond the end of the array implied
1961 by a static type (though there are restrictions on indexing beyond the
1962 bounds of an allocated object in some cases). This means that
1963 single-dimension 'variable sized array' addressing can be implemented in
1964 LLVM with a zero length array type. An implementation of 'pascal style
1965 arrays' in LLVM could use the type "``{ i32, [0 x float]}``", for
1975 The structure type is used to represent a collection of data members
1976 together in memory. The elements of a structure may be any type that has
1979 Structures in memory are accessed using '``load``' and '``store``' by
1980 getting a pointer to a field with the '``getelementptr``' instruction.
1981 Structures in registers are accessed using the '``extractvalue``' and
1982 '``insertvalue``' instructions.
1984 Structures may optionally be "packed" structures, which indicate that
1985 the alignment of the struct is one byte, and that there is no padding
1986 between the elements. In non-packed structs, padding between field types
1987 is inserted as defined by the DataLayout string in the module, which is
1988 required to match what the underlying code generator expects.
1990 Structures can either be "literal" or "identified". A literal structure
1991 is defined inline with other types (e.g. ``{i32, i32}*``) whereas
1992 identified types are always defined at the top level with a name.
1993 Literal types are uniqued by their contents and can never be recursive
1994 or opaque since there is no way to write one. Identified types can be
1995 recursive, can be opaqued, and are never uniqued.
2001 %T1 = type { <type list> } ; Identified normal struct type
2002 %T2 = type <{ <type list> }> ; Identified packed struct type
2006 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2007 | ``{ i32, i32, i32 }`` | A triple of three ``i32`` values |
2008 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2009 | ``{ 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``. |
2010 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2011 | ``<{ i8, i32 }>`` | A packed struct known to be 5 bytes in size. |
2012 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2016 Opaque Structure Types
2017 """"""""""""""""""""""
2021 Opaque structure types are used to represent named structure types that
2022 do not have a body specified. This corresponds (for example) to the C
2023 notion of a forward declared structure.
2034 +--------------+-------------------+
2035 | ``opaque`` | An opaque type. |
2036 +--------------+-------------------+
2043 LLVM has several different basic types of constants. This section
2044 describes them all and their syntax.
2049 **Boolean constants**
2050 The two strings '``true``' and '``false``' are both valid constants
2052 **Integer constants**
2053 Standard integers (such as '4') are constants of the
2054 :ref:`integer <t_integer>` type. Negative numbers may be used with
2056 **Floating point constants**
2057 Floating point constants use standard decimal notation (e.g.
2058 123.421), exponential notation (e.g. 1.23421e+2), or a more precise
2059 hexadecimal notation (see below). The assembler requires the exact
2060 decimal value of a floating-point constant. For example, the
2061 assembler accepts 1.25 but rejects 1.3 because 1.3 is a repeating
2062 decimal in binary. Floating point constants must have a :ref:`floating
2063 point <t_floating>` type.
2064 **Null pointer constants**
2065 The identifier '``null``' is recognized as a null pointer constant
2066 and must be of :ref:`pointer type <t_pointer>`.
2068 The one non-intuitive notation for constants is the hexadecimal form of
2069 floating point constants. For example, the form
2070 '``double 0x432ff973cafa8000``' is equivalent to (but harder to read
2071 than) '``double 4.5e+15``'. The only time hexadecimal floating point
2072 constants are required (and the only time that they are generated by the
2073 disassembler) is when a floating point constant must be emitted but it
2074 cannot be represented as a decimal floating point number in a reasonable
2075 number of digits. For example, NaN's, infinities, and other special
2076 values are represented in their IEEE hexadecimal format so that assembly
2077 and disassembly do not cause any bits to change in the constants.
2079 When using the hexadecimal form, constants of types half, float, and
2080 double are represented using the 16-digit form shown above (which
2081 matches the IEEE754 representation for double); half and float values
2082 must, however, be exactly representable as IEEE 754 half and single
2083 precision, respectively. Hexadecimal format is always used for long
2084 double, and there are three forms of long double. The 80-bit format used
2085 by x86 is represented as ``0xK`` followed by 20 hexadecimal digits. The
2086 128-bit format used by PowerPC (two adjacent doubles) is represented by
2087 ``0xM`` followed by 32 hexadecimal digits. The IEEE 128-bit format is
2088 represented by ``0xL`` followed by 32 hexadecimal digits. Long doubles
2089 will only work if they match the long double format on your target.
2090 The IEEE 16-bit format (half precision) is represented by ``0xH``
2091 followed by 4 hexadecimal digits. All hexadecimal formats are big-endian
2092 (sign bit at the left).
2094 There are no constants of type x86_mmx.
2096 .. _complexconstants:
2101 Complex constants are a (potentially recursive) combination of simple
2102 constants and smaller complex constants.
2104 **Structure constants**
2105 Structure constants are represented with notation similar to
2106 structure type definitions (a comma separated list of elements,
2107 surrounded by braces (``{}``)). For example:
2108 "``{ i32 4, float 17.0, i32* @G }``", where "``@G``" is declared as
2109 "``@G = external global i32``". Structure constants must have
2110 :ref:`structure type <t_struct>`, and the number and types of elements
2111 must match those specified by the type.
2113 Array constants are represented with notation similar to array type
2114 definitions (a comma separated list of elements, surrounded by
2115 square brackets (``[]``)). For example:
2116 "``[ i32 42, i32 11, i32 74 ]``". Array constants must have
2117 :ref:`array type <t_array>`, and the number and types of elements must
2118 match those specified by the type.
2119 **Vector constants**
2120 Vector constants are represented with notation similar to vector
2121 type definitions (a comma separated list of elements, surrounded by
2122 less-than/greater-than's (``<>``)). For example:
2123 "``< i32 42, i32 11, i32 74, i32 100 >``". Vector constants
2124 must have :ref:`vector type <t_vector>`, and the number and types of
2125 elements must match those specified by the type.
2126 **Zero initialization**
2127 The string '``zeroinitializer``' can be used to zero initialize a
2128 value to zero of *any* type, including scalar and
2129 :ref:`aggregate <t_aggregate>` types. This is often used to avoid
2130 having to print large zero initializers (e.g. for large arrays) and
2131 is always exactly equivalent to using explicit zero initializers.
2133 A metadata node is a structure-like constant with :ref:`metadata
2134 type <t_metadata>`. For example:
2135 "``metadata !{ i32 0, metadata !"test" }``". Unlike other
2136 constants that are meant to be interpreted as part of the
2137 instruction stream, metadata is a place to attach additional
2138 information such as debug info.
2140 Global Variable and Function Addresses
2141 --------------------------------------
2143 The addresses of :ref:`global variables <globalvars>` and
2144 :ref:`functions <functionstructure>` are always implicitly valid
2145 (link-time) constants. These constants are explicitly referenced when
2146 the :ref:`identifier for the global <identifiers>` is used and always have
2147 :ref:`pointer <t_pointer>` type. For example, the following is a legal LLVM
2150 .. code-block:: llvm
2154 @Z = global [2 x i32*] [ i32* @X, i32* @Y ]
2161 The string '``undef``' can be used anywhere a constant is expected, and
2162 indicates that the user of the value may receive an unspecified
2163 bit-pattern. Undefined values may be of any type (other than '``label``'
2164 or '``void``') and be used anywhere a constant is permitted.
2166 Undefined values are useful because they indicate to the compiler that
2167 the program is well defined no matter what value is used. This gives the
2168 compiler more freedom to optimize. Here are some examples of
2169 (potentially surprising) transformations that are valid (in pseudo IR):
2171 .. code-block:: llvm
2181 This is safe because all of the output bits are affected by the undef
2182 bits. Any output bit can have a zero or one depending on the input bits.
2184 .. code-block:: llvm
2195 These logical operations have bits that are not always affected by the
2196 input. For example, if ``%X`` has a zero bit, then the output of the
2197 '``and``' operation will always be a zero for that bit, no matter what
2198 the corresponding bit from the '``undef``' is. As such, it is unsafe to
2199 optimize or assume that the result of the '``and``' is '``undef``'.
2200 However, it is safe to assume that all bits of the '``undef``' could be
2201 0, and optimize the '``and``' to 0. Likewise, it is safe to assume that
2202 all the bits of the '``undef``' operand to the '``or``' could be set,
2203 allowing the '``or``' to be folded to -1.
2205 .. code-block:: llvm
2207 %A = select undef, %X, %Y
2208 %B = select undef, 42, %Y
2209 %C = select %X, %Y, undef
2219 This set of examples shows that undefined '``select``' (and conditional
2220 branch) conditions can go *either way*, but they have to come from one
2221 of the two operands. In the ``%A`` example, if ``%X`` and ``%Y`` were
2222 both known to have a clear low bit, then ``%A`` would have to have a
2223 cleared low bit. However, in the ``%C`` example, the optimizer is
2224 allowed to assume that the '``undef``' operand could be the same as
2225 ``%Y``, allowing the whole '``select``' to be eliminated.
2227 .. code-block:: llvm
2229 %A = xor undef, undef
2246 This example points out that two '``undef``' operands are not
2247 necessarily the same. This can be surprising to people (and also matches
2248 C semantics) where they assume that "``X^X``" is always zero, even if
2249 ``X`` is undefined. This isn't true for a number of reasons, but the
2250 short answer is that an '``undef``' "variable" can arbitrarily change
2251 its value over its "live range". This is true because the variable
2252 doesn't actually *have a live range*. Instead, the value is logically
2253 read from arbitrary registers that happen to be around when needed, so
2254 the value is not necessarily consistent over time. In fact, ``%A`` and
2255 ``%C`` need to have the same semantics or the core LLVM "replace all
2256 uses with" concept would not hold.
2258 .. code-block:: llvm
2266 These examples show the crucial difference between an *undefined value*
2267 and *undefined behavior*. An undefined value (like '``undef``') is
2268 allowed to have an arbitrary bit-pattern. This means that the ``%A``
2269 operation can be constant folded to '``undef``', because the '``undef``'
2270 could be an SNaN, and ``fdiv`` is not (currently) defined on SNaN's.
2271 However, in the second example, we can make a more aggressive
2272 assumption: because the ``undef`` is allowed to be an arbitrary value,
2273 we are allowed to assume that it could be zero. Since a divide by zero
2274 has *undefined behavior*, we are allowed to assume that the operation
2275 does not execute at all. This allows us to delete the divide and all
2276 code after it. Because the undefined operation "can't happen", the
2277 optimizer can assume that it occurs in dead code.
2279 .. code-block:: llvm
2281 a: store undef -> %X
2282 b: store %X -> undef
2287 These examples reiterate the ``fdiv`` example: a store *of* an undefined
2288 value can be assumed to not have any effect; we can assume that the
2289 value is overwritten with bits that happen to match what was already
2290 there. However, a store *to* an undefined location could clobber
2291 arbitrary memory, therefore, it has undefined behavior.
2298 Poison values are similar to :ref:`undef values <undefvalues>`, however
2299 they also represent the fact that an instruction or constant expression
2300 which cannot evoke side effects has nevertheless detected a condition
2301 which results in undefined behavior.
2303 There is currently no way of representing a poison value in the IR; they
2304 only exist when produced by operations such as :ref:`add <i_add>` with
2307 Poison value behavior is defined in terms of value *dependence*:
2309 - Values other than :ref:`phi <i_phi>` nodes depend on their operands.
2310 - :ref:`Phi <i_phi>` nodes depend on the operand corresponding to
2311 their dynamic predecessor basic block.
2312 - Function arguments depend on the corresponding actual argument values
2313 in the dynamic callers of their functions.
2314 - :ref:`Call <i_call>` instructions depend on the :ref:`ret <i_ret>`
2315 instructions that dynamically transfer control back to them.
2316 - :ref:`Invoke <i_invoke>` instructions depend on the
2317 :ref:`ret <i_ret>`, :ref:`resume <i_resume>`, or exception-throwing
2318 call instructions that dynamically transfer control back to them.
2319 - Non-volatile loads and stores depend on the most recent stores to all
2320 of the referenced memory addresses, following the order in the IR
2321 (including loads and stores implied by intrinsics such as
2322 :ref:`@llvm.memcpy <int_memcpy>`.)
2323 - An instruction with externally visible side effects depends on the
2324 most recent preceding instruction with externally visible side
2325 effects, following the order in the IR. (This includes :ref:`volatile
2326 operations <volatile>`.)
2327 - An instruction *control-depends* on a :ref:`terminator
2328 instruction <terminators>` if the terminator instruction has
2329 multiple successors and the instruction is always executed when
2330 control transfers to one of the successors, and may not be executed
2331 when control is transferred to another.
2332 - Additionally, an instruction also *control-depends* on a terminator
2333 instruction if the set of instructions it otherwise depends on would
2334 be different if the terminator had transferred control to a different
2336 - Dependence is transitive.
2338 Poison Values have the same behavior as :ref:`undef values <undefvalues>`,
2339 with the additional affect that any instruction which has a *dependence*
2340 on a poison value has undefined behavior.
2342 Here are some examples:
2344 .. code-block:: llvm
2347 %poison = sub nuw i32 0, 1 ; Results in a poison value.
2348 %still_poison = and i32 %poison, 0 ; 0, but also poison.
2349 %poison_yet_again = getelementptr i32* @h, i32 %still_poison
2350 store i32 0, i32* %poison_yet_again ; memory at @h[0] is poisoned
2352 store i32 %poison, i32* @g ; Poison value stored to memory.
2353 %poison2 = load i32* @g ; Poison value loaded back from memory.
2355 store volatile i32 %poison, i32* @g ; External observation; undefined behavior.
2357 %narrowaddr = bitcast i32* @g to i16*
2358 %wideaddr = bitcast i32* @g to i64*
2359 %poison3 = load i16* %narrowaddr ; Returns a poison value.
2360 %poison4 = load i64* %wideaddr ; Returns a poison value.
2362 %cmp = icmp slt i32 %poison, 0 ; Returns a poison value.
2363 br i1 %cmp, label %true, label %end ; Branch to either destination.
2366 store volatile i32 0, i32* @g ; This is control-dependent on %cmp, so
2367 ; it has undefined behavior.
2371 %p = phi i32 [ 0, %entry ], [ 1, %true ]
2372 ; Both edges into this PHI are
2373 ; control-dependent on %cmp, so this
2374 ; always results in a poison value.
2376 store volatile i32 0, i32* @g ; This would depend on the store in %true
2377 ; if %cmp is true, or the store in %entry
2378 ; otherwise, so this is undefined behavior.
2380 br i1 %cmp, label %second_true, label %second_end
2381 ; The same branch again, but this time the
2382 ; true block doesn't have side effects.
2389 store volatile i32 0, i32* @g ; This time, the instruction always depends
2390 ; on the store in %end. Also, it is
2391 ; control-equivalent to %end, so this is
2392 ; well-defined (ignoring earlier undefined
2393 ; behavior in this example).
2397 Addresses of Basic Blocks
2398 -------------------------
2400 ``blockaddress(@function, %block)``
2402 The '``blockaddress``' constant computes the address of the specified
2403 basic block in the specified function, and always has an ``i8*`` type.
2404 Taking the address of the entry block is illegal.
2406 This value only has defined behavior when used as an operand to the
2407 ':ref:`indirectbr <i_indirectbr>`' instruction, or for comparisons
2408 against null. Pointer equality tests between labels addresses results in
2409 undefined behavior --- though, again, comparison against null is ok, and
2410 no label is equal to the null pointer. This may be passed around as an
2411 opaque pointer sized value as long as the bits are not inspected. This
2412 allows ``ptrtoint`` and arithmetic to be performed on these values so
2413 long as the original value is reconstituted before the ``indirectbr``
2416 Finally, some targets may provide defined semantics when using the value
2417 as the operand to an inline assembly, but that is target specific.
2421 Constant Expressions
2422 --------------------
2424 Constant expressions are used to allow expressions involving other
2425 constants to be used as constants. Constant expressions may be of any
2426 :ref:`first class <t_firstclass>` type and may involve any LLVM operation
2427 that does not have side effects (e.g. load and call are not supported).
2428 The following is the syntax for constant expressions:
2430 ``trunc (CST to TYPE)``
2431 Truncate a constant to another type. The bit size of CST must be
2432 larger than the bit size of TYPE. Both types must be integers.
2433 ``zext (CST to TYPE)``
2434 Zero extend a constant to another type. The bit size of CST must be
2435 smaller than the bit size of TYPE. Both types must be integers.
2436 ``sext (CST to TYPE)``
2437 Sign extend a constant to another type. The bit size of CST must be
2438 smaller than the bit size of TYPE. Both types must be integers.
2439 ``fptrunc (CST to TYPE)``
2440 Truncate a floating point constant to another floating point type.
2441 The size of CST must be larger than the size of TYPE. Both types
2442 must be floating point.
2443 ``fpext (CST to TYPE)``
2444 Floating point extend a constant to another type. The size of CST
2445 must be smaller or equal to the size of TYPE. Both types must be
2447 ``fptoui (CST to TYPE)``
2448 Convert a floating point constant to the corresponding unsigned
2449 integer constant. TYPE must be a scalar or vector integer type. CST
2450 must be of scalar or vector floating point type. Both CST and TYPE
2451 must be scalars, or vectors of the same number of elements. If the
2452 value won't fit in the integer type, the results are undefined.
2453 ``fptosi (CST to TYPE)``
2454 Convert a floating point constant to the corresponding signed
2455 integer constant. TYPE must be a scalar or vector integer type. CST
2456 must be of scalar or vector floating point type. Both CST and TYPE
2457 must be scalars, or vectors of the same number of elements. If the
2458 value won't fit in the integer type, the results are undefined.
2459 ``uitofp (CST to TYPE)``
2460 Convert an unsigned integer constant to the corresponding floating
2461 point constant. TYPE must be a scalar or vector floating point type.
2462 CST must be of scalar or vector integer type. Both CST and TYPE must
2463 be scalars, or vectors of the same number of elements. If the value
2464 won't fit in the floating point type, the results are undefined.
2465 ``sitofp (CST to TYPE)``
2466 Convert a signed integer constant to the corresponding floating
2467 point constant. TYPE must be a scalar or vector floating point type.
2468 CST must be of scalar or vector integer type. Both CST and TYPE must
2469 be scalars, or vectors of the same number of elements. If the value
2470 won't fit in the floating point type, the results are undefined.
2471 ``ptrtoint (CST to TYPE)``
2472 Convert a pointer typed constant to the corresponding integer
2473 constant. ``TYPE`` must be an integer type. ``CST`` must be of
2474 pointer type. The ``CST`` value is zero extended, truncated, or
2475 unchanged to make it fit in ``TYPE``.
2476 ``inttoptr (CST to TYPE)``
2477 Convert an integer constant to a pointer constant. TYPE must be a
2478 pointer type. CST must be of integer type. The CST value is zero
2479 extended, truncated, or unchanged to make it fit in a pointer size.
2480 This one is *really* dangerous!
2481 ``bitcast (CST to TYPE)``
2482 Convert a constant, CST, to another TYPE. The constraints of the
2483 operands are the same as those for the :ref:`bitcast
2484 instruction <i_bitcast>`.
2485 ``addrspacecast (CST to TYPE)``
2486 Convert a constant pointer or constant vector of pointer, CST, to another
2487 TYPE in a different address space. The constraints of the operands are the
2488 same as those for the :ref:`addrspacecast instruction <i_addrspacecast>`.
2489 ``getelementptr (CSTPTR, IDX0, IDX1, ...)``, ``getelementptr inbounds (CSTPTR, IDX0, IDX1, ...)``
2490 Perform the :ref:`getelementptr operation <i_getelementptr>` on
2491 constants. As with the :ref:`getelementptr <i_getelementptr>`
2492 instruction, the index list may have zero or more indexes, which are
2493 required to make sense for the type of "CSTPTR".
2494 ``select (COND, VAL1, VAL2)``
2495 Perform the :ref:`select operation <i_select>` on constants.
2496 ``icmp COND (VAL1, VAL2)``
2497 Performs the :ref:`icmp operation <i_icmp>` on constants.
2498 ``fcmp COND (VAL1, VAL2)``
2499 Performs the :ref:`fcmp operation <i_fcmp>` on constants.
2500 ``extractelement (VAL, IDX)``
2501 Perform the :ref:`extractelement operation <i_extractelement>` on
2503 ``insertelement (VAL, ELT, IDX)``
2504 Perform the :ref:`insertelement operation <i_insertelement>` on
2506 ``shufflevector (VEC1, VEC2, IDXMASK)``
2507 Perform the :ref:`shufflevector operation <i_shufflevector>` on
2509 ``extractvalue (VAL, IDX0, IDX1, ...)``
2510 Perform the :ref:`extractvalue operation <i_extractvalue>` on
2511 constants. The index list is interpreted in a similar manner as
2512 indices in a ':ref:`getelementptr <i_getelementptr>`' operation. At
2513 least one index value must be specified.
2514 ``insertvalue (VAL, ELT, IDX0, IDX1, ...)``
2515 Perform the :ref:`insertvalue operation <i_insertvalue>` on constants.
2516 The index list is interpreted in a similar manner as indices in a
2517 ':ref:`getelementptr <i_getelementptr>`' operation. At least one index
2518 value must be specified.
2519 ``OPCODE (LHS, RHS)``
2520 Perform the specified operation of the LHS and RHS constants. OPCODE
2521 may be any of the :ref:`binary <binaryops>` or :ref:`bitwise
2522 binary <bitwiseops>` operations. The constraints on operands are
2523 the same as those for the corresponding instruction (e.g. no bitwise
2524 operations on floating point values are allowed).
2531 Inline Assembler Expressions
2532 ----------------------------
2534 LLVM supports inline assembler expressions (as opposed to :ref:`Module-Level
2535 Inline Assembly <moduleasm>`) through the use of a special value. This
2536 value represents the inline assembler as a string (containing the
2537 instructions to emit), a list of operand constraints (stored as a
2538 string), a flag that indicates whether or not the inline asm expression
2539 has side effects, and a flag indicating whether the function containing
2540 the asm needs to align its stack conservatively. An example inline
2541 assembler expression is:
2543 .. code-block:: llvm
2545 i32 (i32) asm "bswap $0", "=r,r"
2547 Inline assembler expressions may **only** be used as the callee operand
2548 of a :ref:`call <i_call>` or an :ref:`invoke <i_invoke>` instruction.
2549 Thus, typically we have:
2551 .. code-block:: llvm
2553 %X = call i32 asm "bswap $0", "=r,r"(i32 %Y)
2555 Inline asms with side effects not visible in the constraint list must be
2556 marked as having side effects. This is done through the use of the
2557 '``sideeffect``' keyword, like so:
2559 .. code-block:: llvm
2561 call void asm sideeffect "eieio", ""()
2563 In some cases inline asms will contain code that will not work unless
2564 the stack is aligned in some way, such as calls or SSE instructions on
2565 x86, yet will not contain code that does that alignment within the asm.
2566 The compiler should make conservative assumptions about what the asm
2567 might contain and should generate its usual stack alignment code in the
2568 prologue if the '``alignstack``' keyword is present:
2570 .. code-block:: llvm
2572 call void asm alignstack "eieio", ""()
2574 Inline asms also support using non-standard assembly dialects. The
2575 assumed dialect is ATT. When the '``inteldialect``' keyword is present,
2576 the inline asm is using the Intel dialect. Currently, ATT and Intel are
2577 the only supported dialects. An example is:
2579 .. code-block:: llvm
2581 call void asm inteldialect "eieio", ""()
2583 If multiple keywords appear the '``sideeffect``' keyword must come
2584 first, the '``alignstack``' keyword second and the '``inteldialect``'
2590 The call instructions that wrap inline asm nodes may have a
2591 "``!srcloc``" MDNode attached to it that contains a list of constant
2592 integers. If present, the code generator will use the integer as the
2593 location cookie value when report errors through the ``LLVMContext``
2594 error reporting mechanisms. This allows a front-end to correlate backend
2595 errors that occur with inline asm back to the source code that produced
2598 .. code-block:: llvm
2600 call void asm sideeffect "something bad", ""(), !srcloc !42
2602 !42 = !{ i32 1234567 }
2604 It is up to the front-end to make sense of the magic numbers it places
2605 in the IR. If the MDNode contains multiple constants, the code generator
2606 will use the one that corresponds to the line of the asm that the error
2611 Metadata Nodes and Metadata Strings
2612 -----------------------------------
2614 LLVM IR allows metadata to be attached to instructions in the program
2615 that can convey extra information about the code to the optimizers and
2616 code generator. One example application of metadata is source-level
2617 debug information. There are two metadata primitives: strings and nodes.
2618 All metadata has the ``metadata`` type and is identified in syntax by a
2619 preceding exclamation point ('``!``').
2621 A metadata string is a string surrounded by double quotes. It can
2622 contain any character by escaping non-printable characters with
2623 "``\xx``" where "``xx``" is the two digit hex code. For example:
2626 Metadata nodes are represented with notation similar to structure
2627 constants (a comma separated list of elements, surrounded by braces and
2628 preceded by an exclamation point). Metadata nodes can have any values as
2629 their operand. For example:
2631 .. code-block:: llvm
2633 !{ metadata !"test\00", i32 10}
2635 A :ref:`named metadata <namedmetadatastructure>` is a collection of
2636 metadata nodes, which can be looked up in the module symbol table. For
2639 .. code-block:: llvm
2641 !foo = metadata !{!4, !3}
2643 Metadata can be used as function arguments. Here ``llvm.dbg.value``
2644 function is using two metadata arguments:
2646 .. code-block:: llvm
2648 call void @llvm.dbg.value(metadata !24, i64 0, metadata !25)
2650 Metadata can be attached with an instruction. Here metadata ``!21`` is
2651 attached to the ``add`` instruction using the ``!dbg`` identifier:
2653 .. code-block:: llvm
2655 %indvar.next = add i64 %indvar, 1, !dbg !21
2657 More information about specific metadata nodes recognized by the
2658 optimizers and code generator is found below.
2663 In LLVM IR, memory does not have types, so LLVM's own type system is not
2664 suitable for doing TBAA. Instead, metadata is added to the IR to
2665 describe a type system of a higher level language. This can be used to
2666 implement typical C/C++ TBAA, but it can also be used to implement
2667 custom alias analysis behavior for other languages.
2669 The current metadata format is very simple. TBAA metadata nodes have up
2670 to three fields, e.g.:
2672 .. code-block:: llvm
2674 !0 = metadata !{ metadata !"an example type tree" }
2675 !1 = metadata !{ metadata !"int", metadata !0 }
2676 !2 = metadata !{ metadata !"float", metadata !0 }
2677 !3 = metadata !{ metadata !"const float", metadata !2, i64 1 }
2679 The first field is an identity field. It can be any value, usually a
2680 metadata string, which uniquely identifies the type. The most important
2681 name in the tree is the name of the root node. Two trees with different
2682 root node names are entirely disjoint, even if they have leaves with
2685 The second field identifies the type's parent node in the tree, or is
2686 null or omitted for a root node. A type is considered to alias all of
2687 its descendants and all of its ancestors in the tree. Also, a type is
2688 considered to alias all types in other trees, so that bitcode produced
2689 from multiple front-ends is handled conservatively.
2691 If the third field is present, it's an integer which if equal to 1
2692 indicates that the type is "constant" (meaning
2693 ``pointsToConstantMemory`` should return true; see `other useful
2694 AliasAnalysis methods <AliasAnalysis.html#OtherItfs>`_).
2696 '``tbaa.struct``' Metadata
2697 ^^^^^^^^^^^^^^^^^^^^^^^^^^
2699 The :ref:`llvm.memcpy <int_memcpy>` is often used to implement
2700 aggregate assignment operations in C and similar languages, however it
2701 is defined to copy a contiguous region of memory, which is more than
2702 strictly necessary for aggregate types which contain holes due to
2703 padding. Also, it doesn't contain any TBAA information about the fields
2706 ``!tbaa.struct`` metadata can describe which memory subregions in a
2707 memcpy are padding and what the TBAA tags of the struct are.
2709 The current metadata format is very simple. ``!tbaa.struct`` metadata
2710 nodes are a list of operands which are in conceptual groups of three.
2711 For each group of three, the first operand gives the byte offset of a
2712 field in bytes, the second gives its size in bytes, and the third gives
2715 .. code-block:: llvm
2717 !4 = metadata !{ i64 0, i64 4, metadata !1, i64 8, i64 4, metadata !2 }
2719 This describes a struct with two fields. The first is at offset 0 bytes
2720 with size 4 bytes, and has tbaa tag !1. The second is at offset 8 bytes
2721 and has size 4 bytes and has tbaa tag !2.
2723 Note that the fields need not be contiguous. In this example, there is a
2724 4 byte gap between the two fields. This gap represents padding which
2725 does not carry useful data and need not be preserved.
2727 '``fpmath``' Metadata
2728 ^^^^^^^^^^^^^^^^^^^^^
2730 ``fpmath`` metadata may be attached to any instruction of floating point
2731 type. It can be used to express the maximum acceptable error in the
2732 result of that instruction, in ULPs, thus potentially allowing the
2733 compiler to use a more efficient but less accurate method of computing
2734 it. ULP is defined as follows:
2736 If ``x`` is a real number that lies between two finite consecutive
2737 floating-point numbers ``a`` and ``b``, without being equal to one
2738 of them, then ``ulp(x) = |b - a|``, otherwise ``ulp(x)`` is the
2739 distance between the two non-equal finite floating-point numbers
2740 nearest ``x``. Moreover, ``ulp(NaN)`` is ``NaN``.
2742 The metadata node shall consist of a single positive floating point
2743 number representing the maximum relative error, for example:
2745 .. code-block:: llvm
2747 !0 = metadata !{ float 2.5 } ; maximum acceptable inaccuracy is 2.5 ULPs
2749 '``range``' Metadata
2750 ^^^^^^^^^^^^^^^^^^^^
2752 ``range`` metadata may be attached only to ``load``, ``call`` and ``invoke`` of
2753 integer types. It expresses the possible ranges the loaded value or the value
2754 returned by the called function at this call site is in. The ranges are
2755 represented with a flattened list of integers. The loaded value or the value
2756 returned is known to be in the union of the ranges defined by each consecutive
2757 pair. Each pair has the following properties:
2759 - The type must match the type loaded by the instruction.
2760 - The pair ``a,b`` represents the range ``[a,b)``.
2761 - Both ``a`` and ``b`` are constants.
2762 - The range is allowed to wrap.
2763 - The range should not represent the full or empty set. That is,
2766 In addition, the pairs must be in signed order of the lower bound and
2767 they must be non-contiguous.
2771 .. code-block:: llvm
2773 %a = load i8* %x, align 1, !range !0 ; Can only be 0 or 1
2774 %b = load i8* %y, align 1, !range !1 ; Can only be 255 (-1), 0 or 1
2775 %c = call i8 @foo(), !range !2 ; Can only be 0, 1, 3, 4 or 5
2776 %d = invoke i8 @bar() to label %cont
2777 unwind label %lpad, !range !3 ; Can only be -2, -1, 3, 4 or 5
2779 !0 = metadata !{ i8 0, i8 2 }
2780 !1 = metadata !{ i8 255, i8 2 }
2781 !2 = metadata !{ i8 0, i8 2, i8 3, i8 6 }
2782 !3 = metadata !{ i8 -2, i8 0, i8 3, i8 6 }
2787 It is sometimes useful to attach information to loop constructs. Currently,
2788 loop metadata is implemented as metadata attached to the branch instruction
2789 in the loop latch block. This type of metadata refer to a metadata node that is
2790 guaranteed to be separate for each loop. The loop identifier metadata is
2791 specified with the name ``llvm.loop``.
2793 The loop identifier metadata is implemented using a metadata that refers to
2794 itself to avoid merging it with any other identifier metadata, e.g.,
2795 during module linkage or function inlining. That is, each loop should refer
2796 to their own identification metadata even if they reside in separate functions.
2797 The following example contains loop identifier metadata for two separate loop
2800 .. code-block:: llvm
2802 !0 = metadata !{ metadata !0 }
2803 !1 = metadata !{ metadata !1 }
2805 The loop identifier metadata can be used to specify additional per-loop
2806 metadata. Any operands after the first operand can be treated as user-defined
2807 metadata. For example the ``llvm.loop.vectorize.unroll`` metadata is understood
2808 by the loop vectorizer to indicate how many times to unroll the loop:
2810 .. code-block:: llvm
2812 br i1 %exitcond, label %._crit_edge, label %.lr.ph, !llvm.loop !0
2814 !0 = metadata !{ metadata !0, metadata !1 }
2815 !1 = metadata !{ metadata !"llvm.loop.vectorize.unroll", i32 2 }
2820 Metadata types used to annotate memory accesses with information helpful
2821 for optimizations are prefixed with ``llvm.mem``.
2823 '``llvm.mem.parallel_loop_access``' Metadata
2824 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2826 The ``llvm.mem.parallel_loop_access`` metadata refers to a loop identifier,
2827 or metadata containing a list of loop identifiers for nested loops.
2828 The metadata is attached to memory accessing instructions and denotes that
2829 no loop carried memory dependence exist between it and other instructions denoted
2830 with the same loop identifier.
2832 Precisely, given two instructions ``m1`` and ``m2`` that both have the
2833 ``llvm.mem.parallel_loop_access`` metadata, with ``L1`` and ``L2`` being the
2834 set of loops associated with that metadata, respectively, then there is no loop
2835 carried dependence between ``m1`` and ``m2`` for loops in both ``L1`` and
2838 As a special case, if all memory accessing instructions in a loop have
2839 ``llvm.mem.parallel_loop_access`` metadata that refers to that loop, then the
2840 loop has no loop carried memory dependences and is considered to be a parallel
2843 Note that if not all memory access instructions have such metadata referring to
2844 the loop, then the loop is considered not being trivially parallel. Additional
2845 memory dependence analysis is required to make that determination. As a fail
2846 safe mechanism, this causes loops that were originally parallel to be considered
2847 sequential (if optimization passes that are unaware of the parallel semantics
2848 insert new memory instructions into the loop body).
2850 Example of a loop that is considered parallel due to its correct use of
2851 both ``llvm.loop`` and ``llvm.mem.parallel_loop_access``
2852 metadata types that refer to the same loop identifier metadata.
2854 .. code-block:: llvm
2858 %val0 = load i32* %arrayidx, !llvm.mem.parallel_loop_access !0
2860 store i32 %val0, i32* %arrayidx1, !llvm.mem.parallel_loop_access !0
2862 br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !0
2866 !0 = metadata !{ metadata !0 }
2868 It is also possible to have nested parallel loops. In that case the
2869 memory accesses refer to a list of loop identifier metadata nodes instead of
2870 the loop identifier metadata node directly:
2872 .. code-block:: llvm
2876 %val1 = load i32* %arrayidx3, !llvm.mem.parallel_loop_access !2
2878 br label %inner.for.body
2882 %val0 = load i32* %arrayidx1, !llvm.mem.parallel_loop_access !0
2884 store i32 %val0, i32* %arrayidx2, !llvm.mem.parallel_loop_access !0
2886 br i1 %exitcond, label %inner.for.end, label %inner.for.body, !llvm.loop !1
2890 store i32 %val1, i32* %arrayidx4, !llvm.mem.parallel_loop_access !2
2892 br i1 %exitcond, label %outer.for.end, label %outer.for.body, !llvm.loop !2
2894 outer.for.end: ; preds = %for.body
2896 !0 = metadata !{ metadata !1, metadata !2 } ; a list of loop identifiers
2897 !1 = metadata !{ metadata !1 } ; an identifier for the inner loop
2898 !2 = metadata !{ metadata !2 } ; an identifier for the outer loop
2900 '``llvm.loop.vectorize``'
2901 ^^^^^^^^^^^^^^^^^^^^^^^^^
2903 Metadata prefixed with ``llvm.loop.vectorize`` is used to control per-loop
2904 vectorization parameters such as vectorization factor and unroll factor.
2906 ``llvm.loop.vectorize`` metadata should be used in conjunction with
2907 ``llvm.loop`` loop identification metadata.
2909 '``llvm.loop.vectorize.unroll``' Metadata
2910 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2912 This metadata instructs the loop vectorizer to unroll the specified
2913 loop exactly ``N`` times.
2915 The first operand is the string ``llvm.loop.vectorize.unroll`` and the second
2916 operand is an integer specifying the unroll factor. For example:
2918 .. code-block:: llvm
2920 !0 = metadata !{ metadata !"llvm.loop.vectorize.unroll", i32 4 }
2922 Note that setting ``llvm.loop.vectorize.unroll`` to 1 disables
2923 unrolling of the loop.
2925 If ``llvm.loop.vectorize.unroll`` is set to 0 then the amount of
2926 unrolling will be determined automatically.
2928 '``llvm.loop.vectorize.width``' Metadata
2929 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2931 This metadata sets the target width of the vectorizer to ``N``. Without
2932 this metadata, the vectorizer will choose a width automatically.
2933 Regardless of this metadata, the vectorizer will only vectorize loops if
2934 it believes it is valid to do so.
2936 The first operand is the string ``llvm.loop.vectorize.width`` and the
2937 second operand is an integer specifying the width. For example:
2939 .. code-block:: llvm
2941 !0 = metadata !{ metadata !"llvm.loop.vectorize.width", i32 4 }
2943 Note that setting ``llvm.loop.vectorize.width`` to 1 disables
2944 vectorization of the loop.
2946 If ``llvm.loop.vectorize.width`` is set to 0 then the width will be
2947 determined automatically.
2949 Module Flags Metadata
2950 =====================
2952 Information about the module as a whole is difficult to convey to LLVM's
2953 subsystems. The LLVM IR isn't sufficient to transmit this information.
2954 The ``llvm.module.flags`` named metadata exists in order to facilitate
2955 this. These flags are in the form of key / value pairs --- much like a
2956 dictionary --- making it easy for any subsystem who cares about a flag to
2959 The ``llvm.module.flags`` metadata contains a list of metadata triplets.
2960 Each triplet has the following form:
2962 - The first element is a *behavior* flag, which specifies the behavior
2963 when two (or more) modules are merged together, and it encounters two
2964 (or more) metadata with the same ID. The supported behaviors are
2966 - The second element is a metadata string that is a unique ID for the
2967 metadata. Each module may only have one flag entry for each unique ID (not
2968 including entries with the **Require** behavior).
2969 - The third element is the value of the flag.
2971 When two (or more) modules are merged together, the resulting
2972 ``llvm.module.flags`` metadata is the union of the modules' flags. That is, for
2973 each unique metadata ID string, there will be exactly one entry in the merged
2974 modules ``llvm.module.flags`` metadata table, and the value for that entry will
2975 be determined by the merge behavior flag, as described below. The only exception
2976 is that entries with the *Require* behavior are always preserved.
2978 The following behaviors are supported:
2989 Emits an error if two values disagree, otherwise the resulting value
2990 is that of the operands.
2994 Emits a warning if two values disagree. The result value will be the
2995 operand for the flag from the first module being linked.
2999 Adds a requirement that another module flag be present and have a
3000 specified value after linking is performed. The value must be a
3001 metadata pair, where the first element of the pair is the ID of the
3002 module flag to be restricted, and the second element of the pair is
3003 the value the module flag should be restricted to. This behavior can
3004 be used to restrict the allowable results (via triggering of an
3005 error) of linking IDs with the **Override** behavior.
3009 Uses the specified value, regardless of the behavior or value of the
3010 other module. If both modules specify **Override**, but the values
3011 differ, an error will be emitted.
3015 Appends the two values, which are required to be metadata nodes.
3019 Appends the two values, which are required to be metadata
3020 nodes. However, duplicate entries in the second list are dropped
3021 during the append operation.
3023 It is an error for a particular unique flag ID to have multiple behaviors,
3024 except in the case of **Require** (which adds restrictions on another metadata
3025 value) or **Override**.
3027 An example of module flags:
3029 .. code-block:: llvm
3031 !0 = metadata !{ i32 1, metadata !"foo", i32 1 }
3032 !1 = metadata !{ i32 4, metadata !"bar", i32 37 }
3033 !2 = metadata !{ i32 2, metadata !"qux", i32 42 }
3034 !3 = metadata !{ i32 3, metadata !"qux",
3036 metadata !"foo", i32 1
3039 !llvm.module.flags = !{ !0, !1, !2, !3 }
3041 - Metadata ``!0`` has the ID ``!"foo"`` and the value '1'. The behavior
3042 if two or more ``!"foo"`` flags are seen is to emit an error if their
3043 values are not equal.
3045 - Metadata ``!1`` has the ID ``!"bar"`` and the value '37'. The
3046 behavior if two or more ``!"bar"`` flags are seen is to use the value
3049 - Metadata ``!2`` has the ID ``!"qux"`` and the value '42'. The
3050 behavior if two or more ``!"qux"`` flags are seen is to emit a
3051 warning if their values are not equal.
3053 - Metadata ``!3`` has the ID ``!"qux"`` and the value:
3057 metadata !{ metadata !"foo", i32 1 }
3059 The behavior is to emit an error if the ``llvm.module.flags`` does not
3060 contain a flag with the ID ``!"foo"`` that has the value '1' after linking is
3063 Objective-C Garbage Collection Module Flags Metadata
3064 ----------------------------------------------------
3066 On the Mach-O platform, Objective-C stores metadata about garbage
3067 collection in a special section called "image info". The metadata
3068 consists of a version number and a bitmask specifying what types of
3069 garbage collection are supported (if any) by the file. If two or more
3070 modules are linked together their garbage collection metadata needs to
3071 be merged rather than appended together.
3073 The Objective-C garbage collection module flags metadata consists of the
3074 following key-value pairs:
3083 * - ``Objective-C Version``
3084 - **[Required]** --- The Objective-C ABI version. Valid values are 1 and 2.
3086 * - ``Objective-C Image Info Version``
3087 - **[Required]** --- The version of the image info section. Currently
3090 * - ``Objective-C Image Info Section``
3091 - **[Required]** --- The section to place the metadata. Valid values are
3092 ``"__OBJC, __image_info, regular"`` for Objective-C ABI version 1, and
3093 ``"__DATA,__objc_imageinfo, regular, no_dead_strip"`` for
3094 Objective-C ABI version 2.
3096 * - ``Objective-C Garbage Collection``
3097 - **[Required]** --- Specifies whether garbage collection is supported or
3098 not. Valid values are 0, for no garbage collection, and 2, for garbage
3099 collection supported.
3101 * - ``Objective-C GC Only``
3102 - **[Optional]** --- Specifies that only garbage collection is supported.
3103 If present, its value must be 6. This flag requires that the
3104 ``Objective-C Garbage Collection`` flag have the value 2.
3106 Some important flag interactions:
3108 - If a module with ``Objective-C Garbage Collection`` set to 0 is
3109 merged with a module with ``Objective-C Garbage Collection`` set to
3110 2, then the resulting module has the
3111 ``Objective-C Garbage Collection`` flag set to 0.
3112 - A module with ``Objective-C Garbage Collection`` set to 0 cannot be
3113 merged with a module with ``Objective-C GC Only`` set to 6.
3115 Automatic Linker Flags Module Flags Metadata
3116 --------------------------------------------
3118 Some targets support embedding flags to the linker inside individual object
3119 files. Typically this is used in conjunction with language extensions which
3120 allow source files to explicitly declare the libraries they depend on, and have
3121 these automatically be transmitted to the linker via object files.
3123 These flags are encoded in the IR using metadata in the module flags section,
3124 using the ``Linker Options`` key. The merge behavior for this flag is required
3125 to be ``AppendUnique``, and the value for the key is expected to be a metadata
3126 node which should be a list of other metadata nodes, each of which should be a
3127 list of metadata strings defining linker options.
3129 For example, the following metadata section specifies two separate sets of
3130 linker options, presumably to link against ``libz`` and the ``Cocoa``
3133 !0 = metadata !{ i32 6, metadata !"Linker Options",
3135 metadata !{ metadata !"-lz" },
3136 metadata !{ metadata !"-framework", metadata !"Cocoa" } } }
3137 !llvm.module.flags = !{ !0 }
3139 The metadata encoding as lists of lists of options, as opposed to a collapsed
3140 list of options, is chosen so that the IR encoding can use multiple option
3141 strings to specify e.g., a single library, while still having that specifier be
3142 preserved as an atomic element that can be recognized by a target specific
3143 assembly writer or object file emitter.
3145 Each individual option is required to be either a valid option for the target's
3146 linker, or an option that is reserved by the target specific assembly writer or
3147 object file emitter. No other aspect of these options is defined by the IR.
3149 C type width Module Flags Metadata
3150 ----------------------------------
3152 The ARM backend emits a section into each generated object file describing the
3153 options that it was compiled with (in a compiler-independent way) to prevent
3154 linking incompatible objects, and to allow automatic library selection. Some
3155 of these options are not visible at the IR level, namely wchar_t width and enum
3158 To pass this information to the backend, these options are encoded in module
3159 flags metadata, using the following key-value pairs:
3169 - * 0 --- sizeof(wchar_t) == 4
3170 * 1 --- sizeof(wchar_t) == 2
3173 - * 0 --- Enums are at least as large as an ``int``.
3174 * 1 --- Enums are stored in the smallest integer type which can
3175 represent all of its values.
3177 For example, the following metadata section specifies that the module was
3178 compiled with a ``wchar_t`` width of 4 bytes, and the underlying type of an
3179 enum is the smallest type which can represent all of its values::
3181 !llvm.module.flags = !{!0, !1}
3182 !0 = metadata !{i32 1, metadata !"short_wchar", i32 1}
3183 !1 = metadata !{i32 1, metadata !"short_enum", i32 0}
3185 .. _intrinsicglobalvariables:
3187 Intrinsic Global Variables
3188 ==========================
3190 LLVM has a number of "magic" global variables that contain data that
3191 affect code generation or other IR semantics. These are documented here.
3192 All globals of this sort should have a section specified as
3193 "``llvm.metadata``". This section and all globals that start with
3194 "``llvm.``" are reserved for use by LLVM.
3198 The '``llvm.used``' Global Variable
3199 -----------------------------------
3201 The ``@llvm.used`` global is an array which has
3202 :ref:`appending linkage <linkage_appending>`. This array contains a list of
3203 pointers to named global variables, functions and aliases which may optionally
3204 have a pointer cast formed of bitcast or getelementptr. For example, a legal
3207 .. code-block:: llvm
3212 @llvm.used = appending global [2 x i8*] [
3214 i8* bitcast (i32* @Y to i8*)
3215 ], section "llvm.metadata"
3217 If a symbol appears in the ``@llvm.used`` list, then the compiler, assembler,
3218 and linker are required to treat the symbol as if there is a reference to the
3219 symbol that it cannot see (which is why they have to be named). For example, if
3220 a variable has internal linkage and no references other than that from the
3221 ``@llvm.used`` list, it cannot be deleted. This is commonly used to represent
3222 references from inline asms and other things the compiler cannot "see", and
3223 corresponds to "``attribute((used))``" in GNU C.
3225 On some targets, the code generator must emit a directive to the
3226 assembler or object file to prevent the assembler and linker from
3227 molesting the symbol.
3229 .. _gv_llvmcompilerused:
3231 The '``llvm.compiler.used``' Global Variable
3232 --------------------------------------------
3234 The ``@llvm.compiler.used`` directive is the same as the ``@llvm.used``
3235 directive, except that it only prevents the compiler from touching the
3236 symbol. On targets that support it, this allows an intelligent linker to
3237 optimize references to the symbol without being impeded as it would be
3240 This is a rare construct that should only be used in rare circumstances,
3241 and should not be exposed to source languages.
3243 .. _gv_llvmglobalctors:
3245 The '``llvm.global_ctors``' Global Variable
3246 -------------------------------------------
3248 .. code-block:: llvm
3250 %0 = type { i32, void ()*, i8* }
3251 @llvm.global_ctors = appending global [1 x %0] [%0 { i32 65535, void ()* @ctor, i8* @data }]
3253 The ``@llvm.global_ctors`` array contains a list of constructor
3254 functions, priorities, and an optional associated global or function.
3255 The functions referenced by this array will be called in ascending order
3256 of priority (i.e. lowest first) when the module is loaded. The order of
3257 functions with the same priority is not defined.
3259 If the third field is present, non-null, and points to a global variable
3260 or function, the initializer function will only run if the associated
3261 data from the current module is not discarded.
3263 .. _llvmglobaldtors:
3265 The '``llvm.global_dtors``' Global Variable
3266 -------------------------------------------
3268 .. code-block:: llvm
3270 %0 = type { i32, void ()*, i8* }
3271 @llvm.global_dtors = appending global [1 x %0] [%0 { i32 65535, void ()* @dtor, i8* @data }]
3273 The ``@llvm.global_dtors`` array contains a list of destructor
3274 functions, priorities, and an optional associated global or function.
3275 The functions referenced by this array will be called in descending
3276 order of priority (i.e. highest first) when the module is unloaded. The
3277 order of functions with the same priority is not defined.
3279 If the third field is present, non-null, and points to a global variable
3280 or function, the destructor function will only run if the associated
3281 data from the current module is not discarded.
3283 Instruction Reference
3284 =====================
3286 The LLVM instruction set consists of several different classifications
3287 of instructions: :ref:`terminator instructions <terminators>`, :ref:`binary
3288 instructions <binaryops>`, :ref:`bitwise binary
3289 instructions <bitwiseops>`, :ref:`memory instructions <memoryops>`, and
3290 :ref:`other instructions <otherops>`.
3294 Terminator Instructions
3295 -----------------------
3297 As mentioned :ref:`previously <functionstructure>`, every basic block in a
3298 program ends with a "Terminator" instruction, which indicates which
3299 block should be executed after the current block is finished. These
3300 terminator instructions typically yield a '``void``' value: they produce
3301 control flow, not values (the one exception being the
3302 ':ref:`invoke <i_invoke>`' instruction).
3304 The terminator instructions are: ':ref:`ret <i_ret>`',
3305 ':ref:`br <i_br>`', ':ref:`switch <i_switch>`',
3306 ':ref:`indirectbr <i_indirectbr>`', ':ref:`invoke <i_invoke>`',
3307 ':ref:`resume <i_resume>`', and ':ref:`unreachable <i_unreachable>`'.
3311 '``ret``' Instruction
3312 ^^^^^^^^^^^^^^^^^^^^^
3319 ret <type> <value> ; Return a value from a non-void function
3320 ret void ; Return from void function
3325 The '``ret``' instruction is used to return control flow (and optionally
3326 a value) from a function back to the caller.
3328 There are two forms of the '``ret``' instruction: one that returns a
3329 value and then causes control flow, and one that just causes control
3335 The '``ret``' instruction optionally accepts a single argument, the
3336 return value. The type of the return value must be a ':ref:`first
3337 class <t_firstclass>`' type.
3339 A function is not :ref:`well formed <wellformed>` if it it has a non-void
3340 return type and contains a '``ret``' instruction with no return value or
3341 a return value with a type that does not match its type, or if it has a
3342 void return type and contains a '``ret``' instruction with a return
3348 When the '``ret``' instruction is executed, control flow returns back to
3349 the calling function's context. If the caller is a
3350 ":ref:`call <i_call>`" instruction, execution continues at the
3351 instruction after the call. If the caller was an
3352 ":ref:`invoke <i_invoke>`" instruction, execution continues at the
3353 beginning of the "normal" destination block. If the instruction returns
3354 a value, that value shall set the call or invoke instruction's return
3360 .. code-block:: llvm
3362 ret i32 5 ; Return an integer value of 5
3363 ret void ; Return from a void function
3364 ret { i32, i8 } { i32 4, i8 2 } ; Return a struct of values 4 and 2
3368 '``br``' Instruction
3369 ^^^^^^^^^^^^^^^^^^^^
3376 br i1 <cond>, label <iftrue>, label <iffalse>
3377 br label <dest> ; Unconditional branch
3382 The '``br``' instruction is used to cause control flow to transfer to a
3383 different basic block in the current function. There are two forms of
3384 this instruction, corresponding to a conditional branch and an
3385 unconditional branch.
3390 The conditional branch form of the '``br``' instruction takes a single
3391 '``i1``' value and two '``label``' values. The unconditional form of the
3392 '``br``' instruction takes a single '``label``' value as a target.
3397 Upon execution of a conditional '``br``' instruction, the '``i1``'
3398 argument is evaluated. If the value is ``true``, control flows to the
3399 '``iftrue``' ``label`` argument. If "cond" is ``false``, control flows
3400 to the '``iffalse``' ``label`` argument.
3405 .. code-block:: llvm
3408 %cond = icmp eq i32 %a, %b
3409 br i1 %cond, label %IfEqual, label %IfUnequal
3417 '``switch``' Instruction
3418 ^^^^^^^^^^^^^^^^^^^^^^^^
3425 switch <intty> <value>, label <defaultdest> [ <intty> <val>, label <dest> ... ]
3430 The '``switch``' instruction is used to transfer control flow to one of
3431 several different places. It is a generalization of the '``br``'
3432 instruction, allowing a branch to occur to one of many possible
3438 The '``switch``' instruction uses three parameters: an integer
3439 comparison value '``value``', a default '``label``' destination, and an
3440 array of pairs of comparison value constants and '``label``'s. The table
3441 is not allowed to contain duplicate constant entries.
3446 The ``switch`` instruction specifies a table of values and destinations.
3447 When the '``switch``' instruction is executed, this table is searched
3448 for the given value. If the value is found, control flow is transferred
3449 to the corresponding destination; otherwise, control flow is transferred
3450 to the default destination.
3455 Depending on properties of the target machine and the particular
3456 ``switch`` instruction, this instruction may be code generated in
3457 different ways. For example, it could be generated as a series of
3458 chained conditional branches or with a lookup table.
3463 .. code-block:: llvm
3465 ; Emulate a conditional br instruction
3466 %Val = zext i1 %value to i32
3467 switch i32 %Val, label %truedest [ i32 0, label %falsedest ]
3469 ; Emulate an unconditional br instruction
3470 switch i32 0, label %dest [ ]
3472 ; Implement a jump table:
3473 switch i32 %val, label %otherwise [ i32 0, label %onzero
3475 i32 2, label %ontwo ]
3479 '``indirectbr``' Instruction
3480 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3487 indirectbr <somety>* <address>, [ label <dest1>, label <dest2>, ... ]
3492 The '``indirectbr``' instruction implements an indirect branch to a
3493 label within the current function, whose address is specified by
3494 "``address``". Address must be derived from a
3495 :ref:`blockaddress <blockaddress>` constant.
3500 The '``address``' argument is the address of the label to jump to. The
3501 rest of the arguments indicate the full set of possible destinations
3502 that the address may point to. Blocks are allowed to occur multiple
3503 times in the destination list, though this isn't particularly useful.
3505 This destination list is required so that dataflow analysis has an
3506 accurate understanding of the CFG.
3511 Control transfers to the block specified in the address argument. All
3512 possible destination blocks must be listed in the label list, otherwise
3513 this instruction has undefined behavior. This implies that jumps to
3514 labels defined in other functions have undefined behavior as well.
3519 This is typically implemented with a jump through a register.
3524 .. code-block:: llvm
3526 indirectbr i8* %Addr, [ label %bb1, label %bb2, label %bb3 ]
3530 '``invoke``' Instruction
3531 ^^^^^^^^^^^^^^^^^^^^^^^^
3538 <result> = invoke [cconv] [ret attrs] <ptr to function ty> <function ptr val>(<function args>) [fn attrs]
3539 to label <normal label> unwind label <exception label>
3544 The '``invoke``' instruction causes control to transfer to a specified
3545 function, with the possibility of control flow transfer to either the
3546 '``normal``' label or the '``exception``' label. If the callee function
3547 returns with the "``ret``" instruction, control flow will return to the
3548 "normal" label. If the callee (or any indirect callees) returns via the
3549 ":ref:`resume <i_resume>`" instruction or other exception handling
3550 mechanism, control is interrupted and continued at the dynamically
3551 nearest "exception" label.
3553 The '``exception``' label is a `landing
3554 pad <ExceptionHandling.html#overview>`_ for the exception. As such,
3555 '``exception``' label is required to have the
3556 ":ref:`landingpad <i_landingpad>`" instruction, which contains the
3557 information about the behavior of the program after unwinding happens,
3558 as its first non-PHI instruction. The restrictions on the
3559 "``landingpad``" instruction's tightly couples it to the "``invoke``"
3560 instruction, so that the important information contained within the
3561 "``landingpad``" instruction can't be lost through normal code motion.
3566 This instruction requires several arguments:
3568 #. The optional "cconv" marker indicates which :ref:`calling
3569 convention <callingconv>` the call should use. If none is
3570 specified, the call defaults to using C calling conventions.
3571 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
3572 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
3574 #. '``ptr to function ty``': shall be the signature of the pointer to
3575 function value being invoked. In most cases, this is a direct
3576 function invocation, but indirect ``invoke``'s are just as possible,
3577 branching off an arbitrary pointer to function value.
3578 #. '``function ptr val``': An LLVM value containing a pointer to a
3579 function to be invoked.
3580 #. '``function args``': argument list whose types match the function
3581 signature argument types and parameter attributes. All arguments must
3582 be of :ref:`first class <t_firstclass>` type. If the function signature
3583 indicates the function accepts a variable number of arguments, the
3584 extra arguments can be specified.
3585 #. '``normal label``': the label reached when the called function
3586 executes a '``ret``' instruction.
3587 #. '``exception label``': the label reached when a callee returns via
3588 the :ref:`resume <i_resume>` instruction or other exception handling
3590 #. The optional :ref:`function attributes <fnattrs>` list. Only
3591 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
3592 attributes are valid here.
3597 This instruction is designed to operate as a standard '``call``'
3598 instruction in most regards. The primary difference is that it
3599 establishes an association with a label, which is used by the runtime
3600 library to unwind the stack.
3602 This instruction is used in languages with destructors to ensure that
3603 proper cleanup is performed in the case of either a ``longjmp`` or a
3604 thrown exception. Additionally, this is important for implementation of
3605 '``catch``' clauses in high-level languages that support them.
3607 For the purposes of the SSA form, the definition of the value returned
3608 by the '``invoke``' instruction is deemed to occur on the edge from the
3609 current block to the "normal" label. If the callee unwinds then no
3610 return value is available.
3615 .. code-block:: llvm
3617 %retval = invoke i32 @Test(i32 15) to label %Continue
3618 unwind label %TestCleanup ; i32:retval set
3619 %retval = invoke coldcc i32 %Testfnptr(i32 15) to label %Continue
3620 unwind label %TestCleanup ; i32:retval set
3624 '``resume``' Instruction
3625 ^^^^^^^^^^^^^^^^^^^^^^^^
3632 resume <type> <value>
3637 The '``resume``' instruction is a terminator instruction that has no
3643 The '``resume``' instruction requires one argument, which must have the
3644 same type as the result of any '``landingpad``' instruction in the same
3650 The '``resume``' instruction resumes propagation of an existing
3651 (in-flight) exception whose unwinding was interrupted with a
3652 :ref:`landingpad <i_landingpad>` instruction.
3657 .. code-block:: llvm
3659 resume { i8*, i32 } %exn
3663 '``unreachable``' Instruction
3664 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3676 The '``unreachable``' instruction has no defined semantics. This
3677 instruction is used to inform the optimizer that a particular portion of
3678 the code is not reachable. This can be used to indicate that the code
3679 after a no-return function cannot be reached, and other facts.
3684 The '``unreachable``' instruction has no defined semantics.
3691 Binary operators are used to do most of the computation in a program.
3692 They require two operands of the same type, execute an operation on
3693 them, and produce a single value. The operands might represent multiple
3694 data, as is the case with the :ref:`vector <t_vector>` data type. The
3695 result value has the same type as its operands.
3697 There are several different binary operators:
3701 '``add``' Instruction
3702 ^^^^^^^^^^^^^^^^^^^^^
3709 <result> = add <ty> <op1>, <op2> ; yields ty:result
3710 <result> = add nuw <ty> <op1>, <op2> ; yields ty:result
3711 <result> = add nsw <ty> <op1>, <op2> ; yields ty:result
3712 <result> = add nuw nsw <ty> <op1>, <op2> ; yields ty:result
3717 The '``add``' instruction returns the sum of its two operands.
3722 The two arguments to the '``add``' instruction must be
3723 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3724 arguments must have identical types.
3729 The value produced is the integer sum of the two operands.
3731 If the sum has unsigned overflow, the result returned is the
3732 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
3735 Because LLVM integers use a two's complement representation, this
3736 instruction is appropriate for both signed and unsigned integers.
3738 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3739 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3740 result value of the ``add`` is a :ref:`poison value <poisonvalues>` if
3741 unsigned and/or signed overflow, respectively, occurs.
3746 .. code-block:: llvm
3748 <result> = add i32 4, %var ; yields i32:result = 4 + %var
3752 '``fadd``' Instruction
3753 ^^^^^^^^^^^^^^^^^^^^^^
3760 <result> = fadd [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
3765 The '``fadd``' instruction returns the sum of its two operands.
3770 The two arguments to the '``fadd``' instruction must be :ref:`floating
3771 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3772 Both arguments must have identical types.
3777 The value produced is the floating point sum of the two operands. This
3778 instruction can also take any number of :ref:`fast-math flags <fastmath>`,
3779 which are optimization hints to enable otherwise unsafe floating point
3785 .. code-block:: llvm
3787 <result> = fadd float 4.0, %var ; yields float:result = 4.0 + %var
3789 '``sub``' Instruction
3790 ^^^^^^^^^^^^^^^^^^^^^
3797 <result> = sub <ty> <op1>, <op2> ; yields ty:result
3798 <result> = sub nuw <ty> <op1>, <op2> ; yields ty:result
3799 <result> = sub nsw <ty> <op1>, <op2> ; yields ty:result
3800 <result> = sub nuw nsw <ty> <op1>, <op2> ; yields ty:result
3805 The '``sub``' instruction returns the difference of its two operands.
3807 Note that the '``sub``' instruction is used to represent the '``neg``'
3808 instruction present in most other intermediate representations.
3813 The two arguments to the '``sub``' instruction must be
3814 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3815 arguments must have identical types.
3820 The value produced is the integer difference of the two operands.
3822 If the difference has unsigned overflow, the result returned is the
3823 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
3826 Because LLVM integers use a two's complement representation, this
3827 instruction is appropriate for both signed and unsigned integers.
3829 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3830 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3831 result value of the ``sub`` is a :ref:`poison value <poisonvalues>` if
3832 unsigned and/or signed overflow, respectively, occurs.
3837 .. code-block:: llvm
3839 <result> = sub i32 4, %var ; yields i32:result = 4 - %var
3840 <result> = sub i32 0, %val ; yields i32:result = -%var
3844 '``fsub``' Instruction
3845 ^^^^^^^^^^^^^^^^^^^^^^
3852 <result> = fsub [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
3857 The '``fsub``' instruction returns the difference of its two operands.
3859 Note that the '``fsub``' instruction is used to represent the '``fneg``'
3860 instruction present in most other intermediate representations.
3865 The two arguments to the '``fsub``' instruction must be :ref:`floating
3866 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3867 Both arguments must have identical types.
3872 The value produced is the floating point difference of the two operands.
3873 This instruction can also take any number of :ref:`fast-math
3874 flags <fastmath>`, which are optimization hints to enable otherwise
3875 unsafe floating point optimizations:
3880 .. code-block:: llvm
3882 <result> = fsub float 4.0, %var ; yields float:result = 4.0 - %var
3883 <result> = fsub float -0.0, %val ; yields float:result = -%var
3885 '``mul``' Instruction
3886 ^^^^^^^^^^^^^^^^^^^^^
3893 <result> = mul <ty> <op1>, <op2> ; yields ty:result
3894 <result> = mul nuw <ty> <op1>, <op2> ; yields ty:result
3895 <result> = mul nsw <ty> <op1>, <op2> ; yields ty:result
3896 <result> = mul nuw nsw <ty> <op1>, <op2> ; yields ty:result
3901 The '``mul``' instruction returns the product of its two operands.
3906 The two arguments to the '``mul``' instruction must be
3907 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3908 arguments must have identical types.
3913 The value produced is the integer product of the two operands.
3915 If the result of the multiplication has unsigned overflow, the result
3916 returned is the mathematical result modulo 2\ :sup:`n`\ , where n is the
3917 bit width of the result.
3919 Because LLVM integers use a two's complement representation, and the
3920 result is the same width as the operands, this instruction returns the
3921 correct result for both signed and unsigned integers. If a full product
3922 (e.g. ``i32`` * ``i32`` -> ``i64``) is needed, the operands should be
3923 sign-extended or zero-extended as appropriate to the width of the full
3926 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3927 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3928 result value of the ``mul`` is a :ref:`poison value <poisonvalues>` if
3929 unsigned and/or signed overflow, respectively, occurs.
3934 .. code-block:: llvm
3936 <result> = mul i32 4, %var ; yields i32:result = 4 * %var
3940 '``fmul``' Instruction
3941 ^^^^^^^^^^^^^^^^^^^^^^
3948 <result> = fmul [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
3953 The '``fmul``' instruction returns the product of its two operands.
3958 The two arguments to the '``fmul``' instruction must be :ref:`floating
3959 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3960 Both arguments must have identical types.
3965 The value produced is the floating point product of the two operands.
3966 This instruction can also take any number of :ref:`fast-math
3967 flags <fastmath>`, which are optimization hints to enable otherwise
3968 unsafe floating point optimizations:
3973 .. code-block:: llvm
3975 <result> = fmul float 4.0, %var ; yields float:result = 4.0 * %var
3977 '``udiv``' Instruction
3978 ^^^^^^^^^^^^^^^^^^^^^^
3985 <result> = udiv <ty> <op1>, <op2> ; yields ty:result
3986 <result> = udiv exact <ty> <op1>, <op2> ; yields ty:result
3991 The '``udiv``' instruction returns the quotient of its two operands.
3996 The two arguments to the '``udiv``' instruction must be
3997 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3998 arguments must have identical types.
4003 The value produced is the unsigned integer quotient of the two operands.
4005 Note that unsigned integer division and signed integer division are
4006 distinct operations; for signed integer division, use '``sdiv``'.
4008 Division by zero leads to undefined behavior.
4010 If the ``exact`` keyword is present, the result value of the ``udiv`` is
4011 a :ref:`poison value <poisonvalues>` if %op1 is not a multiple of %op2 (as
4012 such, "((a udiv exact b) mul b) == a").
4017 .. code-block:: llvm
4019 <result> = udiv i32 4, %var ; yields i32:result = 4 / %var
4021 '``sdiv``' Instruction
4022 ^^^^^^^^^^^^^^^^^^^^^^
4029 <result> = sdiv <ty> <op1>, <op2> ; yields ty:result
4030 <result> = sdiv exact <ty> <op1>, <op2> ; yields ty:result
4035 The '``sdiv``' instruction returns the quotient of its two operands.
4040 The two arguments to the '``sdiv``' instruction must be
4041 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4042 arguments must have identical types.
4047 The value produced is the signed integer quotient of the two operands
4048 rounded towards zero.
4050 Note that signed integer division and unsigned integer division are
4051 distinct operations; for unsigned integer division, use '``udiv``'.
4053 Division by zero leads to undefined behavior. Overflow also leads to
4054 undefined behavior; this is a rare case, but can occur, for example, by
4055 doing a 32-bit division of -2147483648 by -1.
4057 If the ``exact`` keyword is present, the result value of the ``sdiv`` is
4058 a :ref:`poison value <poisonvalues>` if the result would be rounded.
4063 .. code-block:: llvm
4065 <result> = sdiv i32 4, %var ; yields i32:result = 4 / %var
4069 '``fdiv``' Instruction
4070 ^^^^^^^^^^^^^^^^^^^^^^
4077 <result> = fdiv [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
4082 The '``fdiv``' instruction returns the quotient of its two operands.
4087 The two arguments to the '``fdiv``' instruction must be :ref:`floating
4088 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
4089 Both arguments must have identical types.
4094 The value produced is the floating point quotient of the two operands.
4095 This instruction can also take any number of :ref:`fast-math
4096 flags <fastmath>`, which are optimization hints to enable otherwise
4097 unsafe floating point optimizations:
4102 .. code-block:: llvm
4104 <result> = fdiv float 4.0, %var ; yields float:result = 4.0 / %var
4106 '``urem``' Instruction
4107 ^^^^^^^^^^^^^^^^^^^^^^
4114 <result> = urem <ty> <op1>, <op2> ; yields ty:result
4119 The '``urem``' instruction returns the remainder from the unsigned
4120 division of its two arguments.
4125 The two arguments to the '``urem``' instruction must be
4126 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4127 arguments must have identical types.
4132 This instruction returns the unsigned integer *remainder* of a division.
4133 This instruction always performs an unsigned division to get the
4136 Note that unsigned integer remainder and signed integer remainder are
4137 distinct operations; for signed integer remainder, use '``srem``'.
4139 Taking the remainder of a division by zero leads to undefined behavior.
4144 .. code-block:: llvm
4146 <result> = urem i32 4, %var ; yields i32:result = 4 % %var
4148 '``srem``' Instruction
4149 ^^^^^^^^^^^^^^^^^^^^^^
4156 <result> = srem <ty> <op1>, <op2> ; yields ty:result
4161 The '``srem``' instruction returns the remainder from the signed
4162 division of its two operands. This instruction can also take
4163 :ref:`vector <t_vector>` versions of the values in which case the elements
4169 The two arguments to the '``srem``' instruction must be
4170 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4171 arguments must have identical types.
4176 This instruction returns the *remainder* of a division (where the result
4177 is either zero or has the same sign as the dividend, ``op1``), not the
4178 *modulo* operator (where the result is either zero or has the same sign
4179 as the divisor, ``op2``) of a value. For more information about the
4180 difference, see `The Math
4181 Forum <http://mathforum.org/dr.math/problems/anne.4.28.99.html>`_. For a
4182 table of how this is implemented in various languages, please see
4184 operation <http://en.wikipedia.org/wiki/Modulo_operation>`_.
4186 Note that signed integer remainder and unsigned integer remainder are
4187 distinct operations; for unsigned integer remainder, use '``urem``'.
4189 Taking the remainder of a division by zero leads to undefined behavior.
4190 Overflow also leads to undefined behavior; this is a rare case, but can
4191 occur, for example, by taking the remainder of a 32-bit division of
4192 -2147483648 by -1. (The remainder doesn't actually overflow, but this
4193 rule lets srem be implemented using instructions that return both the
4194 result of the division and the remainder.)
4199 .. code-block:: llvm
4201 <result> = srem i32 4, %var ; yields i32:result = 4 % %var
4205 '``frem``' Instruction
4206 ^^^^^^^^^^^^^^^^^^^^^^
4213 <result> = frem [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
4218 The '``frem``' instruction returns the remainder from the division of
4224 The two arguments to the '``frem``' instruction must be :ref:`floating
4225 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
4226 Both arguments must have identical types.
4231 This instruction returns the *remainder* of a division. The remainder
4232 has the same sign as the dividend. This instruction can also take any
4233 number of :ref:`fast-math flags <fastmath>`, which are optimization hints
4234 to enable otherwise unsafe floating point optimizations:
4239 .. code-block:: llvm
4241 <result> = frem float 4.0, %var ; yields float:result = 4.0 % %var
4245 Bitwise Binary Operations
4246 -------------------------
4248 Bitwise binary operators are used to do various forms of bit-twiddling
4249 in a program. They are generally very efficient instructions and can
4250 commonly be strength reduced from other instructions. They require two
4251 operands of the same type, execute an operation on them, and produce a
4252 single value. The resulting value is the same type as its operands.
4254 '``shl``' Instruction
4255 ^^^^^^^^^^^^^^^^^^^^^
4262 <result> = shl <ty> <op1>, <op2> ; yields ty:result
4263 <result> = shl nuw <ty> <op1>, <op2> ; yields ty:result
4264 <result> = shl nsw <ty> <op1>, <op2> ; yields ty:result
4265 <result> = shl nuw nsw <ty> <op1>, <op2> ; yields ty:result
4270 The '``shl``' instruction returns the first operand shifted to the left
4271 a specified number of bits.
4276 Both arguments to the '``shl``' instruction must be the same
4277 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4278 '``op2``' is treated as an unsigned value.
4283 The value produced is ``op1`` \* 2\ :sup:`op2` mod 2\ :sup:`n`,
4284 where ``n`` is the width of the result. If ``op2`` is (statically or
4285 dynamically) negative or equal to or larger than the number of bits in
4286 ``op1``, the result is undefined. If the arguments are vectors, each
4287 vector element of ``op1`` is shifted by the corresponding shift amount
4290 If the ``nuw`` keyword is present, then the shift produces a :ref:`poison
4291 value <poisonvalues>` if it shifts out any non-zero bits. If the
4292 ``nsw`` keyword is present, then the shift produces a :ref:`poison
4293 value <poisonvalues>` if it shifts out any bits that disagree with the
4294 resultant sign bit. As such, NUW/NSW have the same semantics as they
4295 would if the shift were expressed as a mul instruction with the same
4296 nsw/nuw bits in (mul %op1, (shl 1, %op2)).
4301 .. code-block:: llvm
4303 <result> = shl i32 4, %var ; yields i32: 4 << %var
4304 <result> = shl i32 4, 2 ; yields i32: 16
4305 <result> = shl i32 1, 10 ; yields i32: 1024
4306 <result> = shl i32 1, 32 ; undefined
4307 <result> = shl <2 x i32> < i32 1, i32 1>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 2, i32 4>
4309 '``lshr``' Instruction
4310 ^^^^^^^^^^^^^^^^^^^^^^
4317 <result> = lshr <ty> <op1>, <op2> ; yields ty:result
4318 <result> = lshr exact <ty> <op1>, <op2> ; yields ty:result
4323 The '``lshr``' instruction (logical shift right) returns the first
4324 operand shifted to the right a specified number of bits with zero fill.
4329 Both arguments to the '``lshr``' instruction must be the same
4330 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4331 '``op2``' is treated as an unsigned value.
4336 This instruction always performs a logical shift right operation. The
4337 most significant bits of the result will be filled with zero bits after
4338 the shift. If ``op2`` is (statically or dynamically) equal to or larger
4339 than the number of bits in ``op1``, the result is undefined. If the
4340 arguments are vectors, each vector element of ``op1`` is shifted by the
4341 corresponding shift amount in ``op2``.
4343 If the ``exact`` keyword is present, the result value of the ``lshr`` is
4344 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
4350 .. code-block:: llvm
4352 <result> = lshr i32 4, 1 ; yields i32:result = 2
4353 <result> = lshr i32 4, 2 ; yields i32:result = 1
4354 <result> = lshr i8 4, 3 ; yields i8:result = 0
4355 <result> = lshr i8 -2, 1 ; yields i8:result = 0x7F
4356 <result> = lshr i32 1, 32 ; undefined
4357 <result> = lshr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 0x7FFFFFFF, i32 1>
4359 '``ashr``' Instruction
4360 ^^^^^^^^^^^^^^^^^^^^^^
4367 <result> = ashr <ty> <op1>, <op2> ; yields ty:result
4368 <result> = ashr exact <ty> <op1>, <op2> ; yields ty:result
4373 The '``ashr``' instruction (arithmetic shift right) returns the first
4374 operand shifted to the right a specified number of bits with sign
4380 Both arguments to the '``ashr``' instruction must be the same
4381 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4382 '``op2``' is treated as an unsigned value.
4387 This instruction always performs an arithmetic shift right operation,
4388 The most significant bits of the result will be filled with the sign bit
4389 of ``op1``. If ``op2`` is (statically or dynamically) equal to or larger
4390 than the number of bits in ``op1``, the result is undefined. If the
4391 arguments are vectors, each vector element of ``op1`` is shifted by the
4392 corresponding shift amount in ``op2``.
4394 If the ``exact`` keyword is present, the result value of the ``ashr`` is
4395 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
4401 .. code-block:: llvm
4403 <result> = ashr i32 4, 1 ; yields i32:result = 2
4404 <result> = ashr i32 4, 2 ; yields i32:result = 1
4405 <result> = ashr i8 4, 3 ; yields i8:result = 0
4406 <result> = ashr i8 -2, 1 ; yields i8:result = -1
4407 <result> = ashr i32 1, 32 ; undefined
4408 <result> = ashr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 3> ; yields: result=<2 x i32> < i32 -1, i32 0>
4410 '``and``' Instruction
4411 ^^^^^^^^^^^^^^^^^^^^^
4418 <result> = and <ty> <op1>, <op2> ; yields ty:result
4423 The '``and``' instruction returns the bitwise logical and of its two
4429 The two arguments to the '``and``' instruction must be
4430 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4431 arguments must have identical types.
4436 The truth table used for the '``and``' instruction is:
4453 .. code-block:: llvm
4455 <result> = and i32 4, %var ; yields i32:result = 4 & %var
4456 <result> = and i32 15, 40 ; yields i32:result = 8
4457 <result> = and i32 4, 8 ; yields i32:result = 0
4459 '``or``' Instruction
4460 ^^^^^^^^^^^^^^^^^^^^
4467 <result> = or <ty> <op1>, <op2> ; yields ty:result
4472 The '``or``' instruction returns the bitwise logical inclusive or of its
4478 The two arguments to the '``or``' instruction must be
4479 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4480 arguments must have identical types.
4485 The truth table used for the '``or``' instruction is:
4504 <result> = or i32 4, %var ; yields i32:result = 4 | %var
4505 <result> = or i32 15, 40 ; yields i32:result = 47
4506 <result> = or i32 4, 8 ; yields i32:result = 12
4508 '``xor``' Instruction
4509 ^^^^^^^^^^^^^^^^^^^^^
4516 <result> = xor <ty> <op1>, <op2> ; yields ty:result
4521 The '``xor``' instruction returns the bitwise logical exclusive or of
4522 its two operands. The ``xor`` is used to implement the "one's
4523 complement" operation, which is the "~" operator in C.
4528 The two arguments to the '``xor``' instruction must be
4529 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4530 arguments must have identical types.
4535 The truth table used for the '``xor``' instruction is:
4552 .. code-block:: llvm
4554 <result> = xor i32 4, %var ; yields i32:result = 4 ^ %var
4555 <result> = xor i32 15, 40 ; yields i32:result = 39
4556 <result> = xor i32 4, 8 ; yields i32:result = 12
4557 <result> = xor i32 %V, -1 ; yields i32:result = ~%V
4562 LLVM supports several instructions to represent vector operations in a
4563 target-independent manner. These instructions cover the element-access
4564 and vector-specific operations needed to process vectors effectively.
4565 While LLVM does directly support these vector operations, many
4566 sophisticated algorithms will want to use target-specific intrinsics to
4567 take full advantage of a specific target.
4569 .. _i_extractelement:
4571 '``extractelement``' Instruction
4572 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4579 <result> = extractelement <n x <ty>> <val>, <ty2> <idx> ; yields <ty>
4584 The '``extractelement``' instruction extracts a single scalar element
4585 from a vector at a specified index.
4590 The first operand of an '``extractelement``' instruction is a value of
4591 :ref:`vector <t_vector>` type. The second operand is an index indicating
4592 the position from which to extract the element. The index may be a
4593 variable of any integer type.
4598 The result is a scalar of the same type as the element type of ``val``.
4599 Its value is the value at position ``idx`` of ``val``. If ``idx``
4600 exceeds the length of ``val``, the results are undefined.
4605 .. code-block:: llvm
4607 <result> = extractelement <4 x i32> %vec, i32 0 ; yields i32
4609 .. _i_insertelement:
4611 '``insertelement``' Instruction
4612 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4619 <result> = insertelement <n x <ty>> <val>, <ty> <elt>, <ty2> <idx> ; yields <n x <ty>>
4624 The '``insertelement``' instruction inserts a scalar element into a
4625 vector at a specified index.
4630 The first operand of an '``insertelement``' instruction is a value of
4631 :ref:`vector <t_vector>` type. The second operand is a scalar value whose
4632 type must equal the element type of the first operand. The third operand
4633 is an index indicating the position at which to insert the value. The
4634 index may be a variable of any integer type.
4639 The result is a vector of the same type as ``val``. Its element values
4640 are those of ``val`` except at position ``idx``, where it gets the value
4641 ``elt``. If ``idx`` exceeds the length of ``val``, the results are
4647 .. code-block:: llvm
4649 <result> = insertelement <4 x i32> %vec, i32 1, i32 0 ; yields <4 x i32>
4651 .. _i_shufflevector:
4653 '``shufflevector``' Instruction
4654 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4661 <result> = shufflevector <n x <ty>> <v1>, <n x <ty>> <v2>, <m x i32> <mask> ; yields <m x <ty>>
4666 The '``shufflevector``' instruction constructs a permutation of elements
4667 from two input vectors, returning a vector with the same element type as
4668 the input and length that is the same as the shuffle mask.
4673 The first two operands of a '``shufflevector``' instruction are vectors
4674 with the same type. The third argument is a shuffle mask whose element
4675 type is always 'i32'. The result of the instruction is a vector whose
4676 length is the same as the shuffle mask and whose element type is the
4677 same as the element type of the first two operands.
4679 The shuffle mask operand is required to be a constant vector with either
4680 constant integer or undef values.
4685 The elements of the two input vectors are numbered from left to right
4686 across both of the vectors. The shuffle mask operand specifies, for each
4687 element of the result vector, which element of the two input vectors the
4688 result element gets. The element selector may be undef (meaning "don't
4689 care") and the second operand may be undef if performing a shuffle from
4695 .. code-block:: llvm
4697 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
4698 <4 x i32> <i32 0, i32 4, i32 1, i32 5> ; yields <4 x i32>
4699 <result> = shufflevector <4 x i32> %v1, <4 x i32> undef,
4700 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32> - Identity shuffle.
4701 <result> = shufflevector <8 x i32> %v1, <8 x i32> undef,
4702 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32>
4703 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
4704 <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7 > ; yields <8 x i32>
4706 Aggregate Operations
4707 --------------------
4709 LLVM supports several instructions for working with
4710 :ref:`aggregate <t_aggregate>` values.
4714 '``extractvalue``' Instruction
4715 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4722 <result> = extractvalue <aggregate type> <val>, <idx>{, <idx>}*
4727 The '``extractvalue``' instruction extracts the value of a member field
4728 from an :ref:`aggregate <t_aggregate>` value.
4733 The first operand of an '``extractvalue``' instruction is a value of
4734 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The operands are
4735 constant indices to specify which value to extract in a similar manner
4736 as indices in a '``getelementptr``' instruction.
4738 The major differences to ``getelementptr`` indexing are:
4740 - Since the value being indexed is not a pointer, the first index is
4741 omitted and assumed to be zero.
4742 - At least one index must be specified.
4743 - Not only struct indices but also array indices must be in bounds.
4748 The result is the value at the position in the aggregate specified by
4754 .. code-block:: llvm
4756 <result> = extractvalue {i32, float} %agg, 0 ; yields i32
4760 '``insertvalue``' Instruction
4761 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4768 <result> = insertvalue <aggregate type> <val>, <ty> <elt>, <idx>{, <idx>}* ; yields <aggregate type>
4773 The '``insertvalue``' instruction inserts a value into a member field in
4774 an :ref:`aggregate <t_aggregate>` value.
4779 The first operand of an '``insertvalue``' instruction is a value of
4780 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The second operand is
4781 a first-class value to insert. The following operands are constant
4782 indices indicating the position at which to insert the value in a
4783 similar manner as indices in a '``extractvalue``' instruction. The value
4784 to insert must have the same type as the value identified by the
4790 The result is an aggregate of the same type as ``val``. Its value is
4791 that of ``val`` except that the value at the position specified by the
4792 indices is that of ``elt``.
4797 .. code-block:: llvm
4799 %agg1 = insertvalue {i32, float} undef, i32 1, 0 ; yields {i32 1, float undef}
4800 %agg2 = insertvalue {i32, float} %agg1, float %val, 1 ; yields {i32 1, float %val}
4801 %agg3 = insertvalue {i32, {float}} %agg1, float %val, 1, 0 ; yields {i32 1, float %val}
4805 Memory Access and Addressing Operations
4806 ---------------------------------------
4808 A key design point of an SSA-based representation is how it represents
4809 memory. In LLVM, no memory locations are in SSA form, which makes things
4810 very simple. This section describes how to read, write, and allocate
4815 '``alloca``' Instruction
4816 ^^^^^^^^^^^^^^^^^^^^^^^^
4823 <result> = alloca [inalloca] <type> [, <ty> <NumElements>] [, align <alignment>] ; yields type*:result
4828 The '``alloca``' instruction allocates memory on the stack frame of the
4829 currently executing function, to be automatically released when this
4830 function returns to its caller. The object is always allocated in the
4831 generic address space (address space zero).
4836 The '``alloca``' instruction allocates ``sizeof(<type>)*NumElements``
4837 bytes of memory on the runtime stack, returning a pointer of the
4838 appropriate type to the program. If "NumElements" is specified, it is
4839 the number of elements allocated, otherwise "NumElements" is defaulted
4840 to be one. If a constant alignment is specified, the value result of the
4841 allocation is guaranteed to be aligned to at least that boundary. If not
4842 specified, or if zero, the target can choose to align the allocation on
4843 any convenient boundary compatible with the type.
4845 '``type``' may be any sized type.
4850 Memory is allocated; a pointer is returned. The operation is undefined
4851 if there is insufficient stack space for the allocation. '``alloca``'d
4852 memory is automatically released when the function returns. The
4853 '``alloca``' instruction is commonly used to represent automatic
4854 variables that must have an address available. When the function returns
4855 (either with the ``ret`` or ``resume`` instructions), the memory is
4856 reclaimed. Allocating zero bytes is legal, but the result is undefined.
4857 The order in which memory is allocated (ie., which way the stack grows)
4863 .. code-block:: llvm
4865 %ptr = alloca i32 ; yields i32*:ptr
4866 %ptr = alloca i32, i32 4 ; yields i32*:ptr
4867 %ptr = alloca i32, i32 4, align 1024 ; yields i32*:ptr
4868 %ptr = alloca i32, align 1024 ; yields i32*:ptr
4872 '``load``' Instruction
4873 ^^^^^^^^^^^^^^^^^^^^^^
4880 <result> = load [volatile] <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>][, !invariant.load !<index>]
4881 <result> = load atomic [volatile] <ty>* <pointer> [singlethread] <ordering>, align <alignment>
4882 !<index> = !{ i32 1 }
4887 The '``load``' instruction is used to read from memory.
4892 The argument to the ``load`` instruction specifies the memory address
4893 from which to load. The pointer must point to a :ref:`first
4894 class <t_firstclass>` type. If the ``load`` is marked as ``volatile``,
4895 then the optimizer is not allowed to modify the number or order of
4896 execution of this ``load`` with other :ref:`volatile
4897 operations <volatile>`.
4899 If the ``load`` is marked as ``atomic``, it takes an extra
4900 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
4901 ``release`` and ``acq_rel`` orderings are not valid on ``load``
4902 instructions. Atomic loads produce :ref:`defined <memmodel>` results
4903 when they may see multiple atomic stores. The type of the pointee must
4904 be an integer type whose bit width is a power of two greater than or
4905 equal to eight and less than or equal to a target-specific size limit.
4906 ``align`` must be explicitly specified on atomic loads, and the load has
4907 undefined behavior if the alignment is not set to a value which is at
4908 least the size in bytes of the pointee. ``!nontemporal`` does not have
4909 any defined semantics for atomic loads.
4911 The optional constant ``align`` argument specifies the alignment of the
4912 operation (that is, the alignment of the memory address). A value of 0
4913 or an omitted ``align`` argument means that the operation has the ABI
4914 alignment for the target. It is the responsibility of the code emitter
4915 to ensure that the alignment information is correct. Overestimating the
4916 alignment results in undefined behavior. Underestimating the alignment
4917 may produce less efficient code. An alignment of 1 is always safe.
4919 The optional ``!nontemporal`` metadata must reference a single
4920 metadata name ``<index>`` corresponding to a metadata node with one
4921 ``i32`` entry of value 1. The existence of the ``!nontemporal``
4922 metadata on the instruction tells the optimizer and code generator
4923 that this load is not expected to be reused in the cache. The code
4924 generator may select special instructions to save cache bandwidth, such
4925 as the ``MOVNT`` instruction on x86.
4927 The optional ``!invariant.load`` metadata must reference a single
4928 metadata name ``<index>`` corresponding to a metadata node with no
4929 entries. The existence of the ``!invariant.load`` metadata on the
4930 instruction tells the optimizer and code generator that this load
4931 address points to memory which does not change value during program
4932 execution. The optimizer may then move this load around, for example, by
4933 hoisting it out of loops using loop invariant code motion.
4938 The location of memory pointed to is loaded. If the value being loaded
4939 is of scalar type then the number of bytes read does not exceed the
4940 minimum number of bytes needed to hold all bits of the type. For
4941 example, loading an ``i24`` reads at most three bytes. When loading a
4942 value of a type like ``i20`` with a size that is not an integral number
4943 of bytes, the result is undefined if the value was not originally
4944 written using a store of the same type.
4949 .. code-block:: llvm
4951 %ptr = alloca i32 ; yields i32*:ptr
4952 store i32 3, i32* %ptr ; yields void
4953 %val = load i32* %ptr ; yields i32:val = i32 3
4957 '``store``' Instruction
4958 ^^^^^^^^^^^^^^^^^^^^^^^
4965 store [volatile] <ty> <value>, <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>] ; yields void
4966 store atomic [volatile] <ty> <value>, <ty>* <pointer> [singlethread] <ordering>, align <alignment> ; yields void
4971 The '``store``' instruction is used to write to memory.
4976 There are two arguments to the ``store`` instruction: a value to store
4977 and an address at which to store it. The type of the ``<pointer>``
4978 operand must be a pointer to the :ref:`first class <t_firstclass>` type of
4979 the ``<value>`` operand. If the ``store`` is marked as ``volatile``,
4980 then the optimizer is not allowed to modify the number or order of
4981 execution of this ``store`` with other :ref:`volatile
4982 operations <volatile>`.
4984 If the ``store`` is marked as ``atomic``, it takes an extra
4985 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
4986 ``acquire`` and ``acq_rel`` orderings aren't valid on ``store``
4987 instructions. Atomic loads produce :ref:`defined <memmodel>` results
4988 when they may see multiple atomic stores. The type of the pointee must
4989 be an integer type whose bit width is a power of two greater than or
4990 equal to eight and less than or equal to a target-specific size limit.
4991 ``align`` must be explicitly specified on atomic stores, and the store
4992 has undefined behavior if the alignment is not set to a value which is
4993 at least the size in bytes of the pointee. ``!nontemporal`` does not
4994 have any defined semantics for atomic stores.
4996 The optional constant ``align`` argument specifies the alignment of the
4997 operation (that is, the alignment of the memory address). A value of 0
4998 or an omitted ``align`` argument means that the operation has the ABI
4999 alignment for the target. It is the responsibility of the code emitter
5000 to ensure that the alignment information is correct. Overestimating the
5001 alignment results in undefined behavior. Underestimating the
5002 alignment may produce less efficient code. An alignment of 1 is always
5005 The optional ``!nontemporal`` metadata must reference a single metadata
5006 name ``<index>`` corresponding to a metadata node with one ``i32`` entry of
5007 value 1. The existence of the ``!nontemporal`` metadata on the instruction
5008 tells the optimizer and code generator that this load is not expected to
5009 be reused in the cache. The code generator may select special
5010 instructions to save cache bandwidth, such as the MOVNT instruction on
5016 The contents of memory are updated to contain ``<value>`` at the
5017 location specified by the ``<pointer>`` operand. If ``<value>`` is
5018 of scalar type then the number of bytes written does not exceed the
5019 minimum number of bytes needed to hold all bits of the type. For
5020 example, storing an ``i24`` writes at most three bytes. When writing a
5021 value of a type like ``i20`` with a size that is not an integral number
5022 of bytes, it is unspecified what happens to the extra bits that do not
5023 belong to the type, but they will typically be overwritten.
5028 .. code-block:: llvm
5030 %ptr = alloca i32 ; yields i32*:ptr
5031 store i32 3, i32* %ptr ; yields void
5032 %val = load i32* %ptr ; yields i32:val = i32 3
5036 '``fence``' Instruction
5037 ^^^^^^^^^^^^^^^^^^^^^^^
5044 fence [singlethread] <ordering> ; yields void
5049 The '``fence``' instruction is used to introduce happens-before edges
5055 '``fence``' instructions take an :ref:`ordering <ordering>` argument which
5056 defines what *synchronizes-with* edges they add. They can only be given
5057 ``acquire``, ``release``, ``acq_rel``, and ``seq_cst`` orderings.
5062 A fence A which has (at least) ``release`` ordering semantics
5063 *synchronizes with* a fence B with (at least) ``acquire`` ordering
5064 semantics if and only if there exist atomic operations X and Y, both
5065 operating on some atomic object M, such that A is sequenced before X, X
5066 modifies M (either directly or through some side effect of a sequence
5067 headed by X), Y is sequenced before B, and Y observes M. This provides a
5068 *happens-before* dependency between A and B. Rather than an explicit
5069 ``fence``, one (but not both) of the atomic operations X or Y might
5070 provide a ``release`` or ``acquire`` (resp.) ordering constraint and
5071 still *synchronize-with* the explicit ``fence`` and establish the
5072 *happens-before* edge.
5074 A ``fence`` which has ``seq_cst`` ordering, in addition to having both
5075 ``acquire`` and ``release`` semantics specified above, participates in
5076 the global program order of other ``seq_cst`` operations and/or fences.
5078 The optional ":ref:`singlethread <singlethread>`" argument specifies
5079 that the fence only synchronizes with other fences in the same thread.
5080 (This is useful for interacting with signal handlers.)
5085 .. code-block:: llvm
5087 fence acquire ; yields void
5088 fence singlethread seq_cst ; yields void
5092 '``cmpxchg``' Instruction
5093 ^^^^^^^^^^^^^^^^^^^^^^^^^
5100 cmpxchg [weak] [volatile] <ty>* <pointer>, <ty> <cmp>, <ty> <new> [singlethread] <success ordering> <failure ordering> ; yields { ty, i1 }
5105 The '``cmpxchg``' instruction is used to atomically modify memory. It
5106 loads a value in memory and compares it to a given value. If they are
5107 equal, it tries to store a new value into the memory.
5112 There are three arguments to the '``cmpxchg``' instruction: an address
5113 to operate on, a value to compare to the value currently be at that
5114 address, and a new value to place at that address if the compared values
5115 are equal. The type of '<cmp>' must be an integer type whose bit width
5116 is a power of two greater than or equal to eight and less than or equal
5117 to a target-specific size limit. '<cmp>' and '<new>' must have the same
5118 type, and the type of '<pointer>' must be a pointer to that type. If the
5119 ``cmpxchg`` is marked as ``volatile``, then the optimizer is not allowed
5120 to modify the number or order of execution of this ``cmpxchg`` with
5121 other :ref:`volatile operations <volatile>`.
5123 The success and failure :ref:`ordering <ordering>` arguments specify how this
5124 ``cmpxchg`` synchronizes with other atomic operations. Both ordering parameters
5125 must be at least ``monotonic``, the ordering constraint on failure must be no
5126 stronger than that on success, and the failure ordering cannot be either
5127 ``release`` or ``acq_rel``.
5129 The optional "``singlethread``" argument declares that the ``cmpxchg``
5130 is only atomic with respect to code (usually signal handlers) running in
5131 the same thread as the ``cmpxchg``. Otherwise the cmpxchg is atomic with
5132 respect to all other code in the system.
5134 The pointer passed into cmpxchg must have alignment greater than or
5135 equal to the size in memory of the operand.
5140 The contents of memory at the location specified by the '``<pointer>``' operand
5141 is read and compared to '``<cmp>``'; if the read value is the equal, the
5142 '``<new>``' is written. The original value at the location is returned, together
5143 with a flag indicating success (true) or failure (false).
5145 If the cmpxchg operation is marked as ``weak`` then a spurious failure is
5146 permitted: the operation may not write ``<new>`` even if the comparison
5149 If the cmpxchg operation is strong (the default), the i1 value is 1 if and only
5150 if the value loaded equals ``cmp``.
5152 A successful ``cmpxchg`` is a read-modify-write instruction for the purpose of
5153 identifying release sequences. A failed ``cmpxchg`` is equivalent to an atomic
5154 load with an ordering parameter determined the second ordering parameter.
5159 .. code-block:: llvm
5162 %orig = atomic load i32* %ptr unordered ; yields i32
5166 %cmp = phi i32 [ %orig, %entry ], [%old, %loop]
5167 %squared = mul i32 %cmp, %cmp
5168 %val_success = cmpxchg i32* %ptr, i32 %cmp, i32 %squared acq_rel monotonic ; yields { i32, i1 }
5169 %value_loaded = extractvalue { i32, i1 } %val_success, 0
5170 %success = extractvalue { i32, i1 } %val_success, 1
5171 br i1 %success, label %done, label %loop
5178 '``atomicrmw``' Instruction
5179 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
5186 atomicrmw [volatile] <operation> <ty>* <pointer>, <ty> <value> [singlethread] <ordering> ; yields ty
5191 The '``atomicrmw``' instruction is used to atomically modify memory.
5196 There are three arguments to the '``atomicrmw``' instruction: an
5197 operation to apply, an address whose value to modify, an argument to the
5198 operation. The operation must be one of the following keywords:
5212 The type of '<value>' must be an integer type whose bit width is a power
5213 of two greater than or equal to eight and less than or equal to a
5214 target-specific size limit. The type of the '``<pointer>``' operand must
5215 be a pointer to that type. If the ``atomicrmw`` is marked as
5216 ``volatile``, then the optimizer is not allowed to modify the number or
5217 order of execution of this ``atomicrmw`` with other :ref:`volatile
5218 operations <volatile>`.
5223 The contents of memory at the location specified by the '``<pointer>``'
5224 operand are atomically read, modified, and written back. The original
5225 value at the location is returned. The modification is specified by the
5228 - xchg: ``*ptr = val``
5229 - add: ``*ptr = *ptr + val``
5230 - sub: ``*ptr = *ptr - val``
5231 - and: ``*ptr = *ptr & val``
5232 - nand: ``*ptr = ~(*ptr & val)``
5233 - or: ``*ptr = *ptr | val``
5234 - xor: ``*ptr = *ptr ^ val``
5235 - max: ``*ptr = *ptr > val ? *ptr : val`` (using a signed comparison)
5236 - min: ``*ptr = *ptr < val ? *ptr : val`` (using a signed comparison)
5237 - umax: ``*ptr = *ptr > val ? *ptr : val`` (using an unsigned
5239 - umin: ``*ptr = *ptr < val ? *ptr : val`` (using an unsigned
5245 .. code-block:: llvm
5247 %old = atomicrmw add i32* %ptr, i32 1 acquire ; yields i32
5249 .. _i_getelementptr:
5251 '``getelementptr``' Instruction
5252 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5259 <result> = getelementptr <pty>* <ptrval>{, <ty> <idx>}*
5260 <result> = getelementptr inbounds <pty>* <ptrval>{, <ty> <idx>}*
5261 <result> = getelementptr <ptr vector> ptrval, <vector index type> idx
5266 The '``getelementptr``' instruction is used to get the address of a
5267 subelement of an :ref:`aggregate <t_aggregate>` data structure. It performs
5268 address calculation only and does not access memory.
5273 The first argument is always a pointer or a vector of pointers, and
5274 forms the basis of the calculation. The remaining arguments are indices
5275 that indicate which of the elements of the aggregate object are indexed.
5276 The interpretation of each index is dependent on the type being indexed
5277 into. The first index always indexes the pointer value given as the
5278 first argument, the second index indexes a value of the type pointed to
5279 (not necessarily the value directly pointed to, since the first index
5280 can be non-zero), etc. The first type indexed into must be a pointer
5281 value, subsequent types can be arrays, vectors, and structs. Note that
5282 subsequent types being indexed into can never be pointers, since that
5283 would require loading the pointer before continuing calculation.
5285 The type of each index argument depends on the type it is indexing into.
5286 When indexing into a (optionally packed) structure, only ``i32`` integer
5287 **constants** are allowed (when using a vector of indices they must all
5288 be the **same** ``i32`` integer constant). When indexing into an array,
5289 pointer or vector, integers of any width are allowed, and they are not
5290 required to be constant. These integers are treated as signed values
5293 For example, let's consider a C code fragment and how it gets compiled
5309 int *foo(struct ST *s) {
5310 return &s[1].Z.B[5][13];
5313 The LLVM code generated by Clang is:
5315 .. code-block:: llvm
5317 %struct.RT = type { i8, [10 x [20 x i32]], i8 }
5318 %struct.ST = type { i32, double, %struct.RT }
5320 define i32* @foo(%struct.ST* %s) nounwind uwtable readnone optsize ssp {
5322 %arrayidx = getelementptr inbounds %struct.ST* %s, i64 1, i32 2, i32 1, i64 5, i64 13
5329 In the example above, the first index is indexing into the
5330 '``%struct.ST*``' type, which is a pointer, yielding a '``%struct.ST``'
5331 = '``{ i32, double, %struct.RT }``' type, a structure. The second index
5332 indexes into the third element of the structure, yielding a
5333 '``%struct.RT``' = '``{ i8 , [10 x [20 x i32]], i8 }``' type, another
5334 structure. The third index indexes into the second element of the
5335 structure, yielding a '``[10 x [20 x i32]]``' type, an array. The two
5336 dimensions of the array are subscripted into, yielding an '``i32``'
5337 type. The '``getelementptr``' instruction returns a pointer to this
5338 element, thus computing a value of '``i32*``' type.
5340 Note that it is perfectly legal to index partially through a structure,
5341 returning a pointer to an inner element. Because of this, the LLVM code
5342 for the given testcase is equivalent to:
5344 .. code-block:: llvm
5346 define i32* @foo(%struct.ST* %s) {
5347 %t1 = getelementptr %struct.ST* %s, i32 1 ; yields %struct.ST*:%t1
5348 %t2 = getelementptr %struct.ST* %t1, i32 0, i32 2 ; yields %struct.RT*:%t2
5349 %t3 = getelementptr %struct.RT* %t2, i32 0, i32 1 ; yields [10 x [20 x i32]]*:%t3
5350 %t4 = getelementptr [10 x [20 x i32]]* %t3, i32 0, i32 5 ; yields [20 x i32]*:%t4
5351 %t5 = getelementptr [20 x i32]* %t4, i32 0, i32 13 ; yields i32*:%t5
5355 If the ``inbounds`` keyword is present, the result value of the
5356 ``getelementptr`` is a :ref:`poison value <poisonvalues>` if the base
5357 pointer is not an *in bounds* address of an allocated object, or if any
5358 of the addresses that would be formed by successive addition of the
5359 offsets implied by the indices to the base address with infinitely
5360 precise signed arithmetic are not an *in bounds* address of that
5361 allocated object. The *in bounds* addresses for an allocated object are
5362 all the addresses that point into the object, plus the address one byte
5363 past the end. In cases where the base is a vector of pointers the
5364 ``inbounds`` keyword applies to each of the computations element-wise.
5366 If the ``inbounds`` keyword is not present, the offsets are added to the
5367 base address with silently-wrapping two's complement arithmetic. If the
5368 offsets have a different width from the pointer, they are sign-extended
5369 or truncated to the width of the pointer. The result value of the
5370 ``getelementptr`` may be outside the object pointed to by the base
5371 pointer. The result value may not necessarily be used to access memory
5372 though, even if it happens to point into allocated storage. See the
5373 :ref:`Pointer Aliasing Rules <pointeraliasing>` section for more
5376 The getelementptr instruction is often confusing. For some more insight
5377 into how it works, see :doc:`the getelementptr FAQ <GetElementPtr>`.
5382 .. code-block:: llvm
5384 ; yields [12 x i8]*:aptr
5385 %aptr = getelementptr {i32, [12 x i8]}* %saptr, i64 0, i32 1
5387 %vptr = getelementptr {i32, <2 x i8>}* %svptr, i64 0, i32 1, i32 1
5389 %eptr = getelementptr [12 x i8]* %aptr, i64 0, i32 1
5391 %iptr = getelementptr [10 x i32]* @arr, i16 0, i16 0
5393 In cases where the pointer argument is a vector of pointers, each index
5394 must be a vector with the same number of elements. For example:
5396 .. code-block:: llvm
5398 %A = getelementptr <4 x i8*> %ptrs, <4 x i64> %offsets,
5400 Conversion Operations
5401 ---------------------
5403 The instructions in this category are the conversion instructions
5404 (casting) which all take a single operand and a type. They perform
5405 various bit conversions on the operand.
5407 '``trunc .. to``' Instruction
5408 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5415 <result> = trunc <ty> <value> to <ty2> ; yields ty2
5420 The '``trunc``' instruction truncates its operand to the type ``ty2``.
5425 The '``trunc``' instruction takes a value to trunc, and a type to trunc
5426 it to. Both types must be of :ref:`integer <t_integer>` types, or vectors
5427 of the same number of integers. The bit size of the ``value`` must be
5428 larger than the bit size of the destination type, ``ty2``. Equal sized
5429 types are not allowed.
5434 The '``trunc``' instruction truncates the high order bits in ``value``
5435 and converts the remaining bits to ``ty2``. Since the source size must
5436 be larger than the destination size, ``trunc`` cannot be a *no-op cast*.
5437 It will always truncate bits.
5442 .. code-block:: llvm
5444 %X = trunc i32 257 to i8 ; yields i8:1
5445 %Y = trunc i32 123 to i1 ; yields i1:true
5446 %Z = trunc i32 122 to i1 ; yields i1:false
5447 %W = trunc <2 x i16> <i16 8, i16 7> to <2 x i8> ; yields <i8 8, i8 7>
5449 '``zext .. to``' Instruction
5450 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5457 <result> = zext <ty> <value> to <ty2> ; yields ty2
5462 The '``zext``' instruction zero extends its operand to type ``ty2``.
5467 The '``zext``' instruction takes a value to cast, and a type to cast it
5468 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
5469 the same number of integers. The bit size of the ``value`` must be
5470 smaller than the bit size of the destination type, ``ty2``.
5475 The ``zext`` fills the high order bits of the ``value`` with zero bits
5476 until it reaches the size of the destination type, ``ty2``.
5478 When zero extending from i1, the result will always be either 0 or 1.
5483 .. code-block:: llvm
5485 %X = zext i32 257 to i64 ; yields i64:257
5486 %Y = zext i1 true to i32 ; yields i32:1
5487 %Z = zext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
5489 '``sext .. to``' Instruction
5490 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5497 <result> = sext <ty> <value> to <ty2> ; yields ty2
5502 The '``sext``' sign extends ``value`` to the type ``ty2``.
5507 The '``sext``' instruction takes a value to cast, and a type to cast it
5508 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
5509 the same number of integers. The bit size of the ``value`` must be
5510 smaller than the bit size of the destination type, ``ty2``.
5515 The '``sext``' instruction performs a sign extension by copying the sign
5516 bit (highest order bit) of the ``value`` until it reaches the bit size
5517 of the type ``ty2``.
5519 When sign extending from i1, the extension always results in -1 or 0.
5524 .. code-block:: llvm
5526 %X = sext i8 -1 to i16 ; yields i16 :65535
5527 %Y = sext i1 true to i32 ; yields i32:-1
5528 %Z = sext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
5530 '``fptrunc .. to``' Instruction
5531 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5538 <result> = fptrunc <ty> <value> to <ty2> ; yields ty2
5543 The '``fptrunc``' instruction truncates ``value`` to type ``ty2``.
5548 The '``fptrunc``' instruction takes a :ref:`floating point <t_floating>`
5549 value to cast and a :ref:`floating point <t_floating>` type to cast it to.
5550 The size of ``value`` must be larger than the size of ``ty2``. This
5551 implies that ``fptrunc`` cannot be used to make a *no-op cast*.
5556 The '``fptrunc``' instruction truncates a ``value`` from a larger
5557 :ref:`floating point <t_floating>` type to a smaller :ref:`floating
5558 point <t_floating>` type. If the value cannot fit within the
5559 destination type, ``ty2``, then the results are undefined.
5564 .. code-block:: llvm
5566 %X = fptrunc double 123.0 to float ; yields float:123.0
5567 %Y = fptrunc double 1.0E+300 to float ; yields undefined
5569 '``fpext .. to``' Instruction
5570 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5577 <result> = fpext <ty> <value> to <ty2> ; yields ty2
5582 The '``fpext``' extends a floating point ``value`` to a larger floating
5588 The '``fpext``' instruction takes a :ref:`floating point <t_floating>`
5589 ``value`` to cast, and a :ref:`floating point <t_floating>` type to cast it
5590 to. The source type must be smaller than the destination type.
5595 The '``fpext``' instruction extends the ``value`` from a smaller
5596 :ref:`floating point <t_floating>` type to a larger :ref:`floating
5597 point <t_floating>` type. The ``fpext`` cannot be used to make a
5598 *no-op cast* because it always changes bits. Use ``bitcast`` to make a
5599 *no-op cast* for a floating point cast.
5604 .. code-block:: llvm
5606 %X = fpext float 3.125 to double ; yields double:3.125000e+00
5607 %Y = fpext double %X to fp128 ; yields fp128:0xL00000000000000004000900000000000
5609 '``fptoui .. to``' Instruction
5610 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5617 <result> = fptoui <ty> <value> to <ty2> ; yields ty2
5622 The '``fptoui``' converts a floating point ``value`` to its unsigned
5623 integer equivalent of type ``ty2``.
5628 The '``fptoui``' instruction takes a value to cast, which must be a
5629 scalar or vector :ref:`floating point <t_floating>` value, and a type to
5630 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
5631 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
5632 type with the same number of elements as ``ty``
5637 The '``fptoui``' instruction converts its :ref:`floating
5638 point <t_floating>` operand into the nearest (rounding towards zero)
5639 unsigned integer value. If the value cannot fit in ``ty2``, the results
5645 .. code-block:: llvm
5647 %X = fptoui double 123.0 to i32 ; yields i32:123
5648 %Y = fptoui float 1.0E+300 to i1 ; yields undefined:1
5649 %Z = fptoui float 1.04E+17 to i8 ; yields undefined:1
5651 '``fptosi .. to``' Instruction
5652 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5659 <result> = fptosi <ty> <value> to <ty2> ; yields ty2
5664 The '``fptosi``' instruction converts :ref:`floating point <t_floating>`
5665 ``value`` to type ``ty2``.
5670 The '``fptosi``' instruction takes a value to cast, which must be a
5671 scalar or vector :ref:`floating point <t_floating>` value, and a type to
5672 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
5673 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
5674 type with the same number of elements as ``ty``
5679 The '``fptosi``' instruction converts its :ref:`floating
5680 point <t_floating>` operand into the nearest (rounding towards zero)
5681 signed integer value. If the value cannot fit in ``ty2``, the results
5687 .. code-block:: llvm
5689 %X = fptosi double -123.0 to i32 ; yields i32:-123
5690 %Y = fptosi float 1.0E-247 to i1 ; yields undefined:1
5691 %Z = fptosi float 1.04E+17 to i8 ; yields undefined:1
5693 '``uitofp .. to``' Instruction
5694 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5701 <result> = uitofp <ty> <value> to <ty2> ; yields ty2
5706 The '``uitofp``' instruction regards ``value`` as an unsigned integer
5707 and converts that value to the ``ty2`` type.
5712 The '``uitofp``' instruction takes a value to cast, which must be a
5713 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
5714 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
5715 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
5716 type with the same number of elements as ``ty``
5721 The '``uitofp``' instruction interprets its operand as an unsigned
5722 integer quantity and converts it to the corresponding floating point
5723 value. If the value cannot fit in the floating point value, the results
5729 .. code-block:: llvm
5731 %X = uitofp i32 257 to float ; yields float:257.0
5732 %Y = uitofp i8 -1 to double ; yields double:255.0
5734 '``sitofp .. to``' Instruction
5735 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5742 <result> = sitofp <ty> <value> to <ty2> ; yields ty2
5747 The '``sitofp``' instruction regards ``value`` as a signed integer and
5748 converts that value to the ``ty2`` type.
5753 The '``sitofp``' instruction takes a value to cast, which must be a
5754 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
5755 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
5756 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
5757 type with the same number of elements as ``ty``
5762 The '``sitofp``' instruction interprets its operand as a signed integer
5763 quantity and converts it to the corresponding floating point value. If
5764 the value cannot fit in the floating point value, the results are
5770 .. code-block:: llvm
5772 %X = sitofp i32 257 to float ; yields float:257.0
5773 %Y = sitofp i8 -1 to double ; yields double:-1.0
5777 '``ptrtoint .. to``' Instruction
5778 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5785 <result> = ptrtoint <ty> <value> to <ty2> ; yields ty2
5790 The '``ptrtoint``' instruction converts the pointer or a vector of
5791 pointers ``value`` to the integer (or vector of integers) type ``ty2``.
5796 The '``ptrtoint``' instruction takes a ``value`` to cast, which must be
5797 a a value of type :ref:`pointer <t_pointer>` or a vector of pointers, and a
5798 type to cast it to ``ty2``, which must be an :ref:`integer <t_integer>` or
5799 a vector of integers type.
5804 The '``ptrtoint``' instruction converts ``value`` to integer type
5805 ``ty2`` by interpreting the pointer value as an integer and either
5806 truncating or zero extending that value to the size of the integer type.
5807 If ``value`` is smaller than ``ty2`` then a zero extension is done. If
5808 ``value`` is larger than ``ty2`` then a truncation is done. If they are
5809 the same size, then nothing is done (*no-op cast*) other than a type
5815 .. code-block:: llvm
5817 %X = ptrtoint i32* %P to i8 ; yields truncation on 32-bit architecture
5818 %Y = ptrtoint i32* %P to i64 ; yields zero extension on 32-bit architecture
5819 %Z = ptrtoint <4 x i32*> %P to <4 x i64>; yields vector zero extension for a vector of addresses on 32-bit architecture
5823 '``inttoptr .. to``' Instruction
5824 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5831 <result> = inttoptr <ty> <value> to <ty2> ; yields ty2
5836 The '``inttoptr``' instruction converts an integer ``value`` to a
5837 pointer type, ``ty2``.
5842 The '``inttoptr``' instruction takes an :ref:`integer <t_integer>` value to
5843 cast, and a type to cast it to, which must be a :ref:`pointer <t_pointer>`
5849 The '``inttoptr``' instruction converts ``value`` to type ``ty2`` by
5850 applying either a zero extension or a truncation depending on the size
5851 of the integer ``value``. If ``value`` is larger than the size of a
5852 pointer then a truncation is done. If ``value`` is smaller than the size
5853 of a pointer then a zero extension is done. If they are the same size,
5854 nothing is done (*no-op cast*).
5859 .. code-block:: llvm
5861 %X = inttoptr i32 255 to i32* ; yields zero extension on 64-bit architecture
5862 %Y = inttoptr i32 255 to i32* ; yields no-op on 32-bit architecture
5863 %Z = inttoptr i64 0 to i32* ; yields truncation on 32-bit architecture
5864 %Z = inttoptr <4 x i32> %G to <4 x i8*>; yields truncation of vector G to four pointers
5868 '``bitcast .. to``' Instruction
5869 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5876 <result> = bitcast <ty> <value> to <ty2> ; yields ty2
5881 The '``bitcast``' instruction converts ``value`` to type ``ty2`` without
5887 The '``bitcast``' instruction takes a value to cast, which must be a
5888 non-aggregate first class value, and a type to cast it to, which must
5889 also be a non-aggregate :ref:`first class <t_firstclass>` type. The
5890 bit sizes of ``value`` and the destination type, ``ty2``, must be
5891 identical. If the source type is a pointer, the destination type must
5892 also be a pointer of the same size. This instruction supports bitwise
5893 conversion of vectors to integers and to vectors of other types (as
5894 long as they have the same size).
5899 The '``bitcast``' instruction converts ``value`` to type ``ty2``. It
5900 is always a *no-op cast* because no bits change with this
5901 conversion. The conversion is done as if the ``value`` had been stored
5902 to memory and read back as type ``ty2``. Pointer (or vector of
5903 pointers) types may only be converted to other pointer (or vector of
5904 pointers) types with the same address space through this instruction.
5905 To convert pointers to other types, use the :ref:`inttoptr <i_inttoptr>`
5906 or :ref:`ptrtoint <i_ptrtoint>` instructions first.
5911 .. code-block:: llvm
5913 %X = bitcast i8 255 to i8 ; yields i8 :-1
5914 %Y = bitcast i32* %x to sint* ; yields sint*:%x
5915 %Z = bitcast <2 x int> %V to i64; ; yields i64: %V
5916 %Z = bitcast <2 x i32*> %V to <2 x i64*> ; yields <2 x i64*>
5918 .. _i_addrspacecast:
5920 '``addrspacecast .. to``' Instruction
5921 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5928 <result> = addrspacecast <pty> <ptrval> to <pty2> ; yields pty2
5933 The '``addrspacecast``' instruction converts ``ptrval`` from ``pty`` in
5934 address space ``n`` to type ``pty2`` in address space ``m``.
5939 The '``addrspacecast``' instruction takes a pointer or vector of pointer value
5940 to cast and a pointer type to cast it to, which must have a different
5946 The '``addrspacecast``' instruction converts the pointer value
5947 ``ptrval`` to type ``pty2``. It can be a *no-op cast* or a complex
5948 value modification, depending on the target and the address space
5949 pair. Pointer conversions within the same address space must be
5950 performed with the ``bitcast`` instruction. Note that if the address space
5951 conversion is legal then both result and operand refer to the same memory
5957 .. code-block:: llvm
5959 %X = addrspacecast i32* %x to i32 addrspace(1)* ; yields i32 addrspace(1)*:%x
5960 %Y = addrspacecast i32 addrspace(1)* %y to i64 addrspace(2)* ; yields i64 addrspace(2)*:%y
5961 %Z = addrspacecast <4 x i32*> %z to <4 x float addrspace(3)*> ; yields <4 x float addrspace(3)*>:%z
5968 The instructions in this category are the "miscellaneous" instructions,
5969 which defy better classification.
5973 '``icmp``' Instruction
5974 ^^^^^^^^^^^^^^^^^^^^^^
5981 <result> = icmp <cond> <ty> <op1>, <op2> ; yields i1 or <N x i1>:result
5986 The '``icmp``' instruction returns a boolean value or a vector of
5987 boolean values based on comparison of its two integer, integer vector,
5988 pointer, or pointer vector operands.
5993 The '``icmp``' instruction takes three operands. The first operand is
5994 the condition code indicating the kind of comparison to perform. It is
5995 not a value, just a keyword. The possible condition code are:
5998 #. ``ne``: not equal
5999 #. ``ugt``: unsigned greater than
6000 #. ``uge``: unsigned greater or equal
6001 #. ``ult``: unsigned less than
6002 #. ``ule``: unsigned less or equal
6003 #. ``sgt``: signed greater than
6004 #. ``sge``: signed greater or equal
6005 #. ``slt``: signed less than
6006 #. ``sle``: signed less or equal
6008 The remaining two arguments must be :ref:`integer <t_integer>` or
6009 :ref:`pointer <t_pointer>` or integer :ref:`vector <t_vector>` typed. They
6010 must also be identical types.
6015 The '``icmp``' compares ``op1`` and ``op2`` according to the condition
6016 code given as ``cond``. The comparison performed always yields either an
6017 :ref:`i1 <t_integer>` or vector of ``i1`` result, as follows:
6019 #. ``eq``: yields ``true`` if the operands are equal, ``false``
6020 otherwise. No sign interpretation is necessary or performed.
6021 #. ``ne``: yields ``true`` if the operands are unequal, ``false``
6022 otherwise. No sign interpretation is necessary or performed.
6023 #. ``ugt``: interprets the operands as unsigned values and yields
6024 ``true`` if ``op1`` is greater than ``op2``.
6025 #. ``uge``: interprets the operands as unsigned values and yields
6026 ``true`` if ``op1`` is greater than or equal to ``op2``.
6027 #. ``ult``: interprets the operands as unsigned values and yields
6028 ``true`` if ``op1`` is less than ``op2``.
6029 #. ``ule``: interprets the operands as unsigned values and yields
6030 ``true`` if ``op1`` is less than or equal to ``op2``.
6031 #. ``sgt``: interprets the operands as signed values and yields ``true``
6032 if ``op1`` is greater than ``op2``.
6033 #. ``sge``: interprets the operands as signed values and yields ``true``
6034 if ``op1`` is greater than or equal to ``op2``.
6035 #. ``slt``: interprets the operands as signed values and yields ``true``
6036 if ``op1`` is less than ``op2``.
6037 #. ``sle``: interprets the operands as signed values and yields ``true``
6038 if ``op1`` is less than or equal to ``op2``.
6040 If the operands are :ref:`pointer <t_pointer>` typed, the pointer values
6041 are compared as if they were integers.
6043 If the operands are integer vectors, then they are compared element by
6044 element. The result is an ``i1`` vector with the same number of elements
6045 as the values being compared. Otherwise, the result is an ``i1``.
6050 .. code-block:: llvm
6052 <result> = icmp eq i32 4, 5 ; yields: result=false
6053 <result> = icmp ne float* %X, %X ; yields: result=false
6054 <result> = icmp ult i16 4, 5 ; yields: result=true
6055 <result> = icmp sgt i16 4, 5 ; yields: result=false
6056 <result> = icmp ule i16 -4, 5 ; yields: result=false
6057 <result> = icmp sge i16 4, 5 ; yields: result=false
6059 Note that the code generator does not yet support vector types with the
6060 ``icmp`` instruction.
6064 '``fcmp``' Instruction
6065 ^^^^^^^^^^^^^^^^^^^^^^
6072 <result> = fcmp <cond> <ty> <op1>, <op2> ; yields i1 or <N x i1>:result
6077 The '``fcmp``' instruction returns a boolean value or vector of boolean
6078 values based on comparison of its operands.
6080 If the operands are floating point scalars, then the result type is a
6081 boolean (:ref:`i1 <t_integer>`).
6083 If the operands are floating point vectors, then the result type is a
6084 vector of boolean with the same number of elements as the operands being
6090 The '``fcmp``' instruction takes three operands. The first operand is
6091 the condition code indicating the kind of comparison to perform. It is
6092 not a value, just a keyword. The possible condition code are:
6094 #. ``false``: no comparison, always returns false
6095 #. ``oeq``: ordered and equal
6096 #. ``ogt``: ordered and greater than
6097 #. ``oge``: ordered and greater than or equal
6098 #. ``olt``: ordered and less than
6099 #. ``ole``: ordered and less than or equal
6100 #. ``one``: ordered and not equal
6101 #. ``ord``: ordered (no nans)
6102 #. ``ueq``: unordered or equal
6103 #. ``ugt``: unordered or greater than
6104 #. ``uge``: unordered or greater than or equal
6105 #. ``ult``: unordered or less than
6106 #. ``ule``: unordered or less than or equal
6107 #. ``une``: unordered or not equal
6108 #. ``uno``: unordered (either nans)
6109 #. ``true``: no comparison, always returns true
6111 *Ordered* means that neither operand is a QNAN while *unordered* means
6112 that either operand may be a QNAN.
6114 Each of ``val1`` and ``val2`` arguments must be either a :ref:`floating
6115 point <t_floating>` type or a :ref:`vector <t_vector>` of floating point
6116 type. They must have identical types.
6121 The '``fcmp``' instruction compares ``op1`` and ``op2`` according to the
6122 condition code given as ``cond``. If the operands are vectors, then the
6123 vectors are compared element by element. Each comparison performed
6124 always yields an :ref:`i1 <t_integer>` result, as follows:
6126 #. ``false``: always yields ``false``, regardless of operands.
6127 #. ``oeq``: yields ``true`` if both operands are not a QNAN and ``op1``
6128 is equal to ``op2``.
6129 #. ``ogt``: yields ``true`` if both operands are not a QNAN and ``op1``
6130 is greater than ``op2``.
6131 #. ``oge``: yields ``true`` if both operands are not a QNAN and ``op1``
6132 is greater than or equal to ``op2``.
6133 #. ``olt``: yields ``true`` if both operands are not a QNAN and ``op1``
6134 is less than ``op2``.
6135 #. ``ole``: yields ``true`` if both operands are not a QNAN and ``op1``
6136 is less than or equal to ``op2``.
6137 #. ``one``: yields ``true`` if both operands are not a QNAN and ``op1``
6138 is not equal to ``op2``.
6139 #. ``ord``: yields ``true`` if both operands are not a QNAN.
6140 #. ``ueq``: yields ``true`` if either operand is a QNAN or ``op1`` is
6142 #. ``ugt``: yields ``true`` if either operand is a QNAN or ``op1`` is
6143 greater than ``op2``.
6144 #. ``uge``: yields ``true`` if either operand is a QNAN or ``op1`` is
6145 greater than or equal to ``op2``.
6146 #. ``ult``: yields ``true`` if either operand is a QNAN or ``op1`` is
6148 #. ``ule``: yields ``true`` if either operand is a QNAN or ``op1`` is
6149 less than or equal to ``op2``.
6150 #. ``une``: yields ``true`` if either operand is a QNAN or ``op1`` is
6151 not equal to ``op2``.
6152 #. ``uno``: yields ``true`` if either operand is a QNAN.
6153 #. ``true``: always yields ``true``, regardless of operands.
6158 .. code-block:: llvm
6160 <result> = fcmp oeq float 4.0, 5.0 ; yields: result=false
6161 <result> = fcmp one float 4.0, 5.0 ; yields: result=true
6162 <result> = fcmp olt float 4.0, 5.0 ; yields: result=true
6163 <result> = fcmp ueq double 1.0, 2.0 ; yields: result=false
6165 Note that the code generator does not yet support vector types with the
6166 ``fcmp`` instruction.
6170 '``phi``' Instruction
6171 ^^^^^^^^^^^^^^^^^^^^^
6178 <result> = phi <ty> [ <val0>, <label0>], ...
6183 The '``phi``' instruction is used to implement the φ node in the SSA
6184 graph representing the function.
6189 The type of the incoming values is specified with the first type field.
6190 After this, the '``phi``' instruction takes a list of pairs as
6191 arguments, with one pair for each predecessor basic block of the current
6192 block. Only values of :ref:`first class <t_firstclass>` type may be used as
6193 the value arguments to the PHI node. Only labels may be used as the
6196 There must be no non-phi instructions between the start of a basic block
6197 and the PHI instructions: i.e. PHI instructions must be first in a basic
6200 For the purposes of the SSA form, the use of each incoming value is
6201 deemed to occur on the edge from the corresponding predecessor block to
6202 the current block (but after any definition of an '``invoke``'
6203 instruction's return value on the same edge).
6208 At runtime, the '``phi``' instruction logically takes on the value
6209 specified by the pair corresponding to the predecessor basic block that
6210 executed just prior to the current block.
6215 .. code-block:: llvm
6217 Loop: ; Infinite loop that counts from 0 on up...
6218 %indvar = phi i32 [ 0, %LoopHeader ], [ %nextindvar, %Loop ]
6219 %nextindvar = add i32 %indvar, 1
6224 '``select``' Instruction
6225 ^^^^^^^^^^^^^^^^^^^^^^^^
6232 <result> = select selty <cond>, <ty> <val1>, <ty> <val2> ; yields ty
6234 selty is either i1 or {<N x i1>}
6239 The '``select``' instruction is used to choose one value based on a
6240 condition, without IR-level branching.
6245 The '``select``' instruction requires an 'i1' value or a vector of 'i1'
6246 values indicating the condition, and two values of the same :ref:`first
6247 class <t_firstclass>` type. If the val1/val2 are vectors and the
6248 condition is a scalar, then entire vectors are selected, not individual
6254 If the condition is an i1 and it evaluates to 1, the instruction returns
6255 the first value argument; otherwise, it returns the second value
6258 If the condition is a vector of i1, then the value arguments must be
6259 vectors of the same size, and the selection is done element by element.
6264 .. code-block:: llvm
6266 %X = select i1 true, i8 17, i8 42 ; yields i8:17
6270 '``call``' Instruction
6271 ^^^^^^^^^^^^^^^^^^^^^^
6278 <result> = [tail | musttail] call [cconv] [ret attrs] <ty> [<fnty>*] <fnptrval>(<function args>) [fn attrs]
6283 The '``call``' instruction represents a simple function call.
6288 This instruction requires several arguments:
6290 #. The optional ``tail`` and ``musttail`` markers indicate that the optimizers
6291 should perform tail call optimization. The ``tail`` marker is a hint that
6292 `can be ignored <CodeGenerator.html#sibcallopt>`_. The ``musttail`` marker
6293 means that the call must be tail call optimized in order for the program to
6294 be correct. The ``musttail`` marker provides these guarantees:
6296 #. The call will not cause unbounded stack growth if it is part of a
6297 recursive cycle in the call graph.
6298 #. Arguments with the :ref:`inalloca <attr_inalloca>` attribute are
6301 Both markers imply that the callee does not access allocas or varargs from
6302 the caller. Calls marked ``musttail`` must obey the following additional
6305 - The call must immediately precede a :ref:`ret <i_ret>` instruction,
6306 or a pointer bitcast followed by a ret instruction.
6307 - The ret instruction must return the (possibly bitcasted) value
6308 produced by the call or void.
6309 - The caller and callee prototypes must match. Pointer types of
6310 parameters or return types may differ in pointee type, but not
6312 - The calling conventions of the caller and callee must match.
6313 - All ABI-impacting function attributes, such as sret, byval, inreg,
6314 returned, and inalloca, must match.
6316 Tail call optimization for calls marked ``tail`` is guaranteed to occur if
6317 the following conditions are met:
6319 - Caller and callee both have the calling convention ``fastcc``.
6320 - The call is in tail position (ret immediately follows call and ret
6321 uses value of call or is void).
6322 - Option ``-tailcallopt`` is enabled, or
6323 ``llvm::GuaranteedTailCallOpt`` is ``true``.
6324 - `Platform specific constraints are
6325 met. <CodeGenerator.html#tailcallopt>`_
6327 #. The optional "cconv" marker indicates which :ref:`calling
6328 convention <callingconv>` the call should use. If none is
6329 specified, the call defaults to using C calling conventions. The
6330 calling convention of the call must match the calling convention of
6331 the target function, or else the behavior is undefined.
6332 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
6333 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
6335 #. '``ty``': the type of the call instruction itself which is also the
6336 type of the return value. Functions that return no value are marked
6338 #. '``fnty``': shall be the signature of the pointer to function value
6339 being invoked. The argument types must match the types implied by
6340 this signature. This type can be omitted if the function is not
6341 varargs and if the function type does not return a pointer to a
6343 #. '``fnptrval``': An LLVM value containing a pointer to a function to
6344 be invoked. In most cases, this is a direct function invocation, but
6345 indirect ``call``'s are just as possible, calling an arbitrary pointer
6347 #. '``function args``': argument list whose types match the function
6348 signature argument types and parameter attributes. All arguments must
6349 be of :ref:`first class <t_firstclass>` type. If the function signature
6350 indicates the function accepts a variable number of arguments, the
6351 extra arguments can be specified.
6352 #. The optional :ref:`function attributes <fnattrs>` list. Only
6353 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
6354 attributes are valid here.
6359 The '``call``' instruction is used to cause control flow to transfer to
6360 a specified function, with its incoming arguments bound to the specified
6361 values. Upon a '``ret``' instruction in the called function, control
6362 flow continues with the instruction after the function call, and the
6363 return value of the function is bound to the result argument.
6368 .. code-block:: llvm
6370 %retval = call i32 @test(i32 %argc)
6371 call i32 (i8*, ...)* @printf(i8* %msg, i32 12, i8 42) ; yields i32
6372 %X = tail call i32 @foo() ; yields i32
6373 %Y = tail call fastcc i32 @foo() ; yields i32
6374 call void %foo(i8 97 signext)
6376 %struct.A = type { i32, i8 }
6377 %r = call %struct.A @foo() ; yields { i32, i8 }
6378 %gr = extractvalue %struct.A %r, 0 ; yields i32
6379 %gr1 = extractvalue %struct.A %r, 1 ; yields i8
6380 %Z = call void @foo() noreturn ; indicates that %foo never returns normally
6381 %ZZ = call zeroext i32 @bar() ; Return value is %zero extended
6383 llvm treats calls to some functions with names and arguments that match
6384 the standard C99 library as being the C99 library functions, and may
6385 perform optimizations or generate code for them under that assumption.
6386 This is something we'd like to change in the future to provide better
6387 support for freestanding environments and non-C-based languages.
6391 '``va_arg``' Instruction
6392 ^^^^^^^^^^^^^^^^^^^^^^^^
6399 <resultval> = va_arg <va_list*> <arglist>, <argty>
6404 The '``va_arg``' instruction is used to access arguments passed through
6405 the "variable argument" area of a function call. It is used to implement
6406 the ``va_arg`` macro in C.
6411 This instruction takes a ``va_list*`` value and the type of the
6412 argument. It returns a value of the specified argument type and
6413 increments the ``va_list`` to point to the next argument. The actual
6414 type of ``va_list`` is target specific.
6419 The '``va_arg``' instruction loads an argument of the specified type
6420 from the specified ``va_list`` and causes the ``va_list`` to point to
6421 the next argument. For more information, see the variable argument
6422 handling :ref:`Intrinsic Functions <int_varargs>`.
6424 It is legal for this instruction to be called in a function which does
6425 not take a variable number of arguments, for example, the ``vfprintf``
6428 ``va_arg`` is an LLVM instruction instead of an :ref:`intrinsic
6429 function <intrinsics>` because it takes a type as an argument.
6434 See the :ref:`variable argument processing <int_varargs>` section.
6436 Note that the code generator does not yet fully support va\_arg on many
6437 targets. Also, it does not currently support va\_arg with aggregate
6438 types on any target.
6442 '``landingpad``' Instruction
6443 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6450 <resultval> = landingpad <resultty> personality <type> <pers_fn> <clause>+
6451 <resultval> = landingpad <resultty> personality <type> <pers_fn> cleanup <clause>*
6453 <clause> := catch <type> <value>
6454 <clause> := filter <array constant type> <array constant>
6459 The '``landingpad``' instruction is used by `LLVM's exception handling
6460 system <ExceptionHandling.html#overview>`_ to specify that a basic block
6461 is a landing pad --- one where the exception lands, and corresponds to the
6462 code found in the ``catch`` portion of a ``try``/``catch`` sequence. It
6463 defines values supplied by the personality function (``pers_fn``) upon
6464 re-entry to the function. The ``resultval`` has the type ``resultty``.
6469 This instruction takes a ``pers_fn`` value. This is the personality
6470 function associated with the unwinding mechanism. The optional
6471 ``cleanup`` flag indicates that the landing pad block is a cleanup.
6473 A ``clause`` begins with the clause type --- ``catch`` or ``filter`` --- and
6474 contains the global variable representing the "type" that may be caught
6475 or filtered respectively. Unlike the ``catch`` clause, the ``filter``
6476 clause takes an array constant as its argument. Use
6477 "``[0 x i8**] undef``" for a filter which cannot throw. The
6478 '``landingpad``' instruction must contain *at least* one ``clause`` or
6479 the ``cleanup`` flag.
6484 The '``landingpad``' instruction defines the values which are set by the
6485 personality function (``pers_fn``) upon re-entry to the function, and
6486 therefore the "result type" of the ``landingpad`` instruction. As with
6487 calling conventions, how the personality function results are
6488 represented in LLVM IR is target specific.
6490 The clauses are applied in order from top to bottom. If two
6491 ``landingpad`` instructions are merged together through inlining, the
6492 clauses from the calling function are appended to the list of clauses.
6493 When the call stack is being unwound due to an exception being thrown,
6494 the exception is compared against each ``clause`` in turn. If it doesn't
6495 match any of the clauses, and the ``cleanup`` flag is not set, then
6496 unwinding continues further up the call stack.
6498 The ``landingpad`` instruction has several restrictions:
6500 - A landing pad block is a basic block which is the unwind destination
6501 of an '``invoke``' instruction.
6502 - A landing pad block must have a '``landingpad``' instruction as its
6503 first non-PHI instruction.
6504 - There can be only one '``landingpad``' instruction within the landing
6506 - A basic block that is not a landing pad block may not include a
6507 '``landingpad``' instruction.
6508 - All '``landingpad``' instructions in a function must have the same
6509 personality function.
6514 .. code-block:: llvm
6516 ;; A landing pad which can catch an integer.
6517 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6519 ;; A landing pad that is a cleanup.
6520 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6522 ;; A landing pad which can catch an integer and can only throw a double.
6523 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6525 filter [1 x i8**] [@_ZTId]
6532 LLVM supports the notion of an "intrinsic function". These functions
6533 have well known names and semantics and are required to follow certain
6534 restrictions. Overall, these intrinsics represent an extension mechanism
6535 for the LLVM language that does not require changing all of the
6536 transformations in LLVM when adding to the language (or the bitcode
6537 reader/writer, the parser, etc...).
6539 Intrinsic function names must all start with an "``llvm.``" prefix. This
6540 prefix is reserved in LLVM for intrinsic names; thus, function names may
6541 not begin with this prefix. Intrinsic functions must always be external
6542 functions: you cannot define the body of intrinsic functions. Intrinsic
6543 functions may only be used in call or invoke instructions: it is illegal
6544 to take the address of an intrinsic function. Additionally, because
6545 intrinsic functions are part of the LLVM language, it is required if any
6546 are added that they be documented here.
6548 Some intrinsic functions can be overloaded, i.e., the intrinsic
6549 represents a family of functions that perform the same operation but on
6550 different data types. Because LLVM can represent over 8 million
6551 different integer types, overloading is used commonly to allow an
6552 intrinsic function to operate on any integer type. One or more of the
6553 argument types or the result type can be overloaded to accept any
6554 integer type. Argument types may also be defined as exactly matching a
6555 previous argument's type or the result type. This allows an intrinsic
6556 function which accepts multiple arguments, but needs all of them to be
6557 of the same type, to only be overloaded with respect to a single
6558 argument or the result.
6560 Overloaded intrinsics will have the names of its overloaded argument
6561 types encoded into its function name, each preceded by a period. Only
6562 those types which are overloaded result in a name suffix. Arguments
6563 whose type is matched against another type do not. For example, the
6564 ``llvm.ctpop`` function can take an integer of any width and returns an
6565 integer of exactly the same integer width. This leads to a family of
6566 functions such as ``i8 @llvm.ctpop.i8(i8 %val)`` and
6567 ``i29 @llvm.ctpop.i29(i29 %val)``. Only one type, the return type, is
6568 overloaded, and only one type suffix is required. Because the argument's
6569 type is matched against the return type, it does not require its own
6572 To learn how to add an intrinsic function, please see the `Extending
6573 LLVM Guide <ExtendingLLVM.html>`_.
6577 Variable Argument Handling Intrinsics
6578 -------------------------------------
6580 Variable argument support is defined in LLVM with the
6581 :ref:`va_arg <i_va_arg>` instruction and these three intrinsic
6582 functions. These functions are related to the similarly named macros
6583 defined in the ``<stdarg.h>`` header file.
6585 All of these functions operate on arguments that use a target-specific
6586 value type "``va_list``". The LLVM assembly language reference manual
6587 does not define what this type is, so all transformations should be
6588 prepared to handle these functions regardless of the type used.
6590 This example shows how the :ref:`va_arg <i_va_arg>` instruction and the
6591 variable argument handling intrinsic functions are used.
6593 .. code-block:: llvm
6595 define i32 @test(i32 %X, ...) {
6596 ; Initialize variable argument processing
6598 %ap2 = bitcast i8** %ap to i8*
6599 call void @llvm.va_start(i8* %ap2)
6601 ; Read a single integer argument
6602 %tmp = va_arg i8** %ap, i32
6604 ; Demonstrate usage of llvm.va_copy and llvm.va_end
6606 %aq2 = bitcast i8** %aq to i8*
6607 call void @llvm.va_copy(i8* %aq2, i8* %ap2)
6608 call void @llvm.va_end(i8* %aq2)
6610 ; Stop processing of arguments.
6611 call void @llvm.va_end(i8* %ap2)
6615 declare void @llvm.va_start(i8*)
6616 declare void @llvm.va_copy(i8*, i8*)
6617 declare void @llvm.va_end(i8*)
6621 '``llvm.va_start``' Intrinsic
6622 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6629 declare void @llvm.va_start(i8* <arglist>)
6634 The '``llvm.va_start``' intrinsic initializes ``*<arglist>`` for
6635 subsequent use by ``va_arg``.
6640 The argument is a pointer to a ``va_list`` element to initialize.
6645 The '``llvm.va_start``' intrinsic works just like the ``va_start`` macro
6646 available in C. In a target-dependent way, it initializes the
6647 ``va_list`` element to which the argument points, so that the next call
6648 to ``va_arg`` will produce the first variable argument passed to the
6649 function. Unlike the C ``va_start`` macro, this intrinsic does not need
6650 to know the last argument of the function as the compiler can figure
6653 '``llvm.va_end``' Intrinsic
6654 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6661 declare void @llvm.va_end(i8* <arglist>)
6666 The '``llvm.va_end``' intrinsic destroys ``*<arglist>``, which has been
6667 initialized previously with ``llvm.va_start`` or ``llvm.va_copy``.
6672 The argument is a pointer to a ``va_list`` to destroy.
6677 The '``llvm.va_end``' intrinsic works just like the ``va_end`` macro
6678 available in C. In a target-dependent way, it destroys the ``va_list``
6679 element to which the argument points. Calls to
6680 :ref:`llvm.va_start <int_va_start>` and
6681 :ref:`llvm.va_copy <int_va_copy>` must be matched exactly with calls to
6686 '``llvm.va_copy``' Intrinsic
6687 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6694 declare void @llvm.va_copy(i8* <destarglist>, i8* <srcarglist>)
6699 The '``llvm.va_copy``' intrinsic copies the current argument position
6700 from the source argument list to the destination argument list.
6705 The first argument is a pointer to a ``va_list`` element to initialize.
6706 The second argument is a pointer to a ``va_list`` element to copy from.
6711 The '``llvm.va_copy``' intrinsic works just like the ``va_copy`` macro
6712 available in C. In a target-dependent way, it copies the source
6713 ``va_list`` element into the destination ``va_list`` element. This
6714 intrinsic is necessary because the `` llvm.va_start`` intrinsic may be
6715 arbitrarily complex and require, for example, memory allocation.
6717 Accurate Garbage Collection Intrinsics
6718 --------------------------------------
6720 LLVM support for `Accurate Garbage Collection <GarbageCollection.html>`_
6721 (GC) requires the implementation and generation of these intrinsics.
6722 These intrinsics allow identification of :ref:`GC roots on the
6723 stack <int_gcroot>`, as well as garbage collector implementations that
6724 require :ref:`read <int_gcread>` and :ref:`write <int_gcwrite>` barriers.
6725 Front-ends for type-safe garbage collected languages should generate
6726 these intrinsics to make use of the LLVM garbage collectors. For more
6727 details, see `Accurate Garbage Collection with
6728 LLVM <GarbageCollection.html>`_.
6730 The garbage collection intrinsics only operate on objects in the generic
6731 address space (address space zero).
6735 '``llvm.gcroot``' Intrinsic
6736 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6743 declare void @llvm.gcroot(i8** %ptrloc, i8* %metadata)
6748 The '``llvm.gcroot``' intrinsic declares the existence of a GC root to
6749 the code generator, and allows some metadata to be associated with it.
6754 The first argument specifies the address of a stack object that contains
6755 the root pointer. The second pointer (which must be either a constant or
6756 a global value address) contains the meta-data to be associated with the
6762 At runtime, a call to this intrinsic stores a null pointer into the
6763 "ptrloc" location. At compile-time, the code generator generates
6764 information to allow the runtime to find the pointer at GC safe points.
6765 The '``llvm.gcroot``' intrinsic may only be used in a function which
6766 :ref:`specifies a GC algorithm <gc>`.
6770 '``llvm.gcread``' Intrinsic
6771 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6778 declare i8* @llvm.gcread(i8* %ObjPtr, i8** %Ptr)
6783 The '``llvm.gcread``' intrinsic identifies reads of references from heap
6784 locations, allowing garbage collector implementations that require read
6790 The second argument is the address to read from, which should be an
6791 address allocated from the garbage collector. The first object is a
6792 pointer to the start of the referenced object, if needed by the language
6793 runtime (otherwise null).
6798 The '``llvm.gcread``' intrinsic has the same semantics as a load
6799 instruction, but may be replaced with substantially more complex code by
6800 the garbage collector runtime, as needed. The '``llvm.gcread``'
6801 intrinsic may only be used in a function which :ref:`specifies a GC
6806 '``llvm.gcwrite``' Intrinsic
6807 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6814 declare void @llvm.gcwrite(i8* %P1, i8* %Obj, i8** %P2)
6819 The '``llvm.gcwrite``' intrinsic identifies writes of references to heap
6820 locations, allowing garbage collector implementations that require write
6821 barriers (such as generational or reference counting collectors).
6826 The first argument is the reference to store, the second is the start of
6827 the object to store it to, and the third is the address of the field of
6828 Obj to store to. If the runtime does not require a pointer to the
6829 object, Obj may be null.
6834 The '``llvm.gcwrite``' intrinsic has the same semantics as a store
6835 instruction, but may be replaced with substantially more complex code by
6836 the garbage collector runtime, as needed. The '``llvm.gcwrite``'
6837 intrinsic may only be used in a function which :ref:`specifies a GC
6840 Code Generator Intrinsics
6841 -------------------------
6843 These intrinsics are provided by LLVM to expose special features that
6844 may only be implemented with code generator support.
6846 '``llvm.returnaddress``' Intrinsic
6847 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6854 declare i8 *@llvm.returnaddress(i32 <level>)
6859 The '``llvm.returnaddress``' intrinsic attempts to compute a
6860 target-specific value indicating the return address of the current
6861 function or one of its callers.
6866 The argument to this intrinsic indicates which function to return the
6867 address for. Zero indicates the calling function, one indicates its
6868 caller, etc. The argument is **required** to be a constant integer
6874 The '``llvm.returnaddress``' intrinsic either returns a pointer
6875 indicating the return address of the specified call frame, or zero if it
6876 cannot be identified. The value returned by this intrinsic is likely to
6877 be incorrect or 0 for arguments other than zero, so it should only be
6878 used for debugging purposes.
6880 Note that calling this intrinsic does not prevent function inlining or
6881 other aggressive transformations, so the value returned may not be that
6882 of the obvious source-language caller.
6884 '``llvm.frameaddress``' Intrinsic
6885 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6892 declare i8* @llvm.frameaddress(i32 <level>)
6897 The '``llvm.frameaddress``' intrinsic attempts to return the
6898 target-specific frame pointer value for the specified stack frame.
6903 The argument to this intrinsic indicates which function to return the
6904 frame pointer for. Zero indicates the calling function, one indicates
6905 its caller, etc. The argument is **required** to be a constant integer
6911 The '``llvm.frameaddress``' intrinsic either returns a pointer
6912 indicating the frame address of the specified call frame, or zero if it
6913 cannot be identified. The value returned by this intrinsic is likely to
6914 be incorrect or 0 for arguments other than zero, so it should only be
6915 used for debugging purposes.
6917 Note that calling this intrinsic does not prevent function inlining or
6918 other aggressive transformations, so the value returned may not be that
6919 of the obvious source-language caller.
6921 .. _int_read_register:
6922 .. _int_write_register:
6924 '``llvm.read_register``' and '``llvm.write_register``' Intrinsics
6925 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6932 declare i32 @llvm.read_register.i32(metadata)
6933 declare i64 @llvm.read_register.i64(metadata)
6934 declare void @llvm.write_register.i32(metadata, i32 @value)
6935 declare void @llvm.write_register.i64(metadata, i64 @value)
6936 !0 = metadata !{metadata !"sp\00"}
6941 The '``llvm.read_register``' and '``llvm.write_register``' intrinsics
6942 provides access to the named register. The register must be valid on
6943 the architecture being compiled to. The type needs to be compatible
6944 with the register being read.
6949 The '``llvm.read_register``' intrinsic returns the current value of the
6950 register, where possible. The '``llvm.write_register``' intrinsic sets
6951 the current value of the register, where possible.
6953 This is useful to implement named register global variables that need
6954 to always be mapped to a specific register, as is common practice on
6955 bare-metal programs including OS kernels.
6957 The compiler doesn't check for register availability or use of the used
6958 register in surrounding code, including inline assembly. Because of that,
6959 allocatable registers are not supported.
6961 Warning: So far it only works with the stack pointer on selected
6962 architectures (ARM, AArch64, PowerPC and x86_64). Significant amount of
6963 work is needed to support other registers and even more so, allocatable
6968 '``llvm.stacksave``' Intrinsic
6969 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6976 declare i8* @llvm.stacksave()
6981 The '``llvm.stacksave``' intrinsic is used to remember the current state
6982 of the function stack, for use with
6983 :ref:`llvm.stackrestore <int_stackrestore>`. This is useful for
6984 implementing language features like scoped automatic variable sized
6990 This intrinsic returns a opaque pointer value that can be passed to
6991 :ref:`llvm.stackrestore <int_stackrestore>`. When an
6992 ``llvm.stackrestore`` intrinsic is executed with a value saved from
6993 ``llvm.stacksave``, it effectively restores the state of the stack to
6994 the state it was in when the ``llvm.stacksave`` intrinsic executed. In
6995 practice, this pops any :ref:`alloca <i_alloca>` blocks from the stack that
6996 were allocated after the ``llvm.stacksave`` was executed.
6998 .. _int_stackrestore:
7000 '``llvm.stackrestore``' Intrinsic
7001 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7008 declare void @llvm.stackrestore(i8* %ptr)
7013 The '``llvm.stackrestore``' intrinsic is used to restore the state of
7014 the function stack to the state it was in when the corresponding
7015 :ref:`llvm.stacksave <int_stacksave>` intrinsic executed. This is
7016 useful for implementing language features like scoped automatic variable
7017 sized arrays in C99.
7022 See the description for :ref:`llvm.stacksave <int_stacksave>`.
7024 '``llvm.prefetch``' Intrinsic
7025 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7032 declare void @llvm.prefetch(i8* <address>, i32 <rw>, i32 <locality>, i32 <cache type>)
7037 The '``llvm.prefetch``' intrinsic is a hint to the code generator to
7038 insert a prefetch instruction if supported; otherwise, it is a noop.
7039 Prefetches have no effect on the behavior of the program but can change
7040 its performance characteristics.
7045 ``address`` is the address to be prefetched, ``rw`` is the specifier
7046 determining if the fetch should be for a read (0) or write (1), and
7047 ``locality`` is a temporal locality specifier ranging from (0) - no
7048 locality, to (3) - extremely local keep in cache. The ``cache type``
7049 specifies whether the prefetch is performed on the data (1) or
7050 instruction (0) cache. The ``rw``, ``locality`` and ``cache type``
7051 arguments must be constant integers.
7056 This intrinsic does not modify the behavior of the program. In
7057 particular, prefetches cannot trap and do not produce a value. On
7058 targets that support this intrinsic, the prefetch can provide hints to
7059 the processor cache for better performance.
7061 '``llvm.pcmarker``' Intrinsic
7062 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7069 declare void @llvm.pcmarker(i32 <id>)
7074 The '``llvm.pcmarker``' intrinsic is a method to export a Program
7075 Counter (PC) in a region of code to simulators and other tools. The
7076 method is target specific, but it is expected that the marker will use
7077 exported symbols to transmit the PC of the marker. The marker makes no
7078 guarantees that it will remain with any specific instruction after
7079 optimizations. It is possible that the presence of a marker will inhibit
7080 optimizations. The intended use is to be inserted after optimizations to
7081 allow correlations of simulation runs.
7086 ``id`` is a numerical id identifying the marker.
7091 This intrinsic does not modify the behavior of the program. Backends
7092 that do not support this intrinsic may ignore it.
7094 '``llvm.readcyclecounter``' Intrinsic
7095 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7102 declare i64 @llvm.readcyclecounter()
7107 The '``llvm.readcyclecounter``' intrinsic provides access to the cycle
7108 counter register (or similar low latency, high accuracy clocks) on those
7109 targets that support it. On X86, it should map to RDTSC. On Alpha, it
7110 should map to RPCC. As the backing counters overflow quickly (on the
7111 order of 9 seconds on alpha), this should only be used for small
7117 When directly supported, reading the cycle counter should not modify any
7118 memory. Implementations are allowed to either return a application
7119 specific value or a system wide value. On backends without support, this
7120 is lowered to a constant 0.
7122 Note that runtime support may be conditional on the privilege-level code is
7123 running at and the host platform.
7125 '``llvm.clear_cache``' Intrinsic
7126 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7133 declare void @llvm.clear_cache(i8*, i8*)
7138 The '``llvm.clear_cache``' intrinsic ensures visibility of modifications
7139 in the specified range to the execution unit of the processor. On
7140 targets with non-unified instruction and data cache, the implementation
7141 flushes the instruction cache.
7146 On platforms with coherent instruction and data caches (e.g. x86), this
7147 intrinsic is a nop. On platforms with non-coherent instruction and data
7148 cache (e.g. ARM, MIPS), the intrinsic is lowered either to appropriate
7149 instructions or a system call, if cache flushing requires special
7152 The default behavior is to emit a call to ``__clear_cache`` from the run
7155 This instrinsic does *not* empty the instruction pipeline. Modifications
7156 of the current function are outside the scope of the intrinsic.
7158 Standard C Library Intrinsics
7159 -----------------------------
7161 LLVM provides intrinsics for a few important standard C library
7162 functions. These intrinsics allow source-language front-ends to pass
7163 information about the alignment of the pointer arguments to the code
7164 generator, providing opportunity for more efficient code generation.
7168 '``llvm.memcpy``' Intrinsic
7169 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7174 This is an overloaded intrinsic. You can use ``llvm.memcpy`` on any
7175 integer bit width and for different address spaces. Not all targets
7176 support all bit widths however.
7180 declare void @llvm.memcpy.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
7181 i32 <len>, i32 <align>, i1 <isvolatile>)
7182 declare void @llvm.memcpy.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
7183 i64 <len>, i32 <align>, i1 <isvolatile>)
7188 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
7189 source location to the destination location.
7191 Note that, unlike the standard libc function, the ``llvm.memcpy.*``
7192 intrinsics do not return a value, takes extra alignment/isvolatile
7193 arguments and the pointers can be in specified address spaces.
7198 The first argument is a pointer to the destination, the second is a
7199 pointer to the source. The third argument is an integer argument
7200 specifying the number of bytes to copy, the fourth argument is the
7201 alignment of the source and destination locations, and the fifth is a
7202 boolean indicating a volatile access.
7204 If the call to this intrinsic has an alignment value that is not 0 or 1,
7205 then the caller guarantees that both the source and destination pointers
7206 are aligned to that boundary.
7208 If the ``isvolatile`` parameter is ``true``, the ``llvm.memcpy`` call is
7209 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
7210 very cleanly specified and it is unwise to depend on it.
7215 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
7216 source location to the destination location, which are not allowed to
7217 overlap. It copies "len" bytes of memory over. If the argument is known
7218 to be aligned to some boundary, this can be specified as the fourth
7219 argument, otherwise it should be set to 0 or 1 (both meaning no alignment).
7221 '``llvm.memmove``' Intrinsic
7222 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7227 This is an overloaded intrinsic. You can use llvm.memmove on any integer
7228 bit width and for different address space. Not all targets support all
7233 declare void @llvm.memmove.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
7234 i32 <len>, i32 <align>, i1 <isvolatile>)
7235 declare void @llvm.memmove.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
7236 i64 <len>, i32 <align>, i1 <isvolatile>)
7241 The '``llvm.memmove.*``' intrinsics move a block of memory from the
7242 source location to the destination location. It is similar to the
7243 '``llvm.memcpy``' intrinsic but allows the two memory locations to
7246 Note that, unlike the standard libc function, the ``llvm.memmove.*``
7247 intrinsics do not return a value, takes extra alignment/isvolatile
7248 arguments and the pointers can be in specified address spaces.
7253 The first argument is a pointer to the destination, the second is a
7254 pointer to the source. The third argument is an integer argument
7255 specifying the number of bytes to copy, the fourth argument is the
7256 alignment of the source and destination locations, and the fifth is a
7257 boolean indicating a volatile access.
7259 If the call to this intrinsic has an alignment value that is not 0 or 1,
7260 then the caller guarantees that the source and destination pointers are
7261 aligned to that boundary.
7263 If the ``isvolatile`` parameter is ``true``, the ``llvm.memmove`` call
7264 is a :ref:`volatile operation <volatile>`. The detailed access behavior is
7265 not very cleanly specified and it is unwise to depend on it.
7270 The '``llvm.memmove.*``' intrinsics copy a block of memory from the
7271 source location to the destination location, which may overlap. It
7272 copies "len" bytes of memory over. If the argument is known to be
7273 aligned to some boundary, this can be specified as the fourth argument,
7274 otherwise it should be set to 0 or 1 (both meaning no alignment).
7276 '``llvm.memset.*``' Intrinsics
7277 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7282 This is an overloaded intrinsic. You can use llvm.memset on any integer
7283 bit width and for different address spaces. However, not all targets
7284 support all bit widths.
7288 declare void @llvm.memset.p0i8.i32(i8* <dest>, i8 <val>,
7289 i32 <len>, i32 <align>, i1 <isvolatile>)
7290 declare void @llvm.memset.p0i8.i64(i8* <dest>, i8 <val>,
7291 i64 <len>, i32 <align>, i1 <isvolatile>)
7296 The '``llvm.memset.*``' intrinsics fill a block of memory with a
7297 particular byte value.
7299 Note that, unlike the standard libc function, the ``llvm.memset``
7300 intrinsic does not return a value and takes extra alignment/volatile
7301 arguments. Also, the destination can be in an arbitrary address space.
7306 The first argument is a pointer to the destination to fill, the second
7307 is the byte value with which to fill it, the third argument is an
7308 integer argument specifying the number of bytes to fill, and the fourth
7309 argument is the known alignment of the destination location.
7311 If the call to this intrinsic has an alignment value that is not 0 or 1,
7312 then the caller guarantees that the destination pointer is aligned to
7315 If the ``isvolatile`` parameter is ``true``, the ``llvm.memset`` call is
7316 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
7317 very cleanly specified and it is unwise to depend on it.
7322 The '``llvm.memset.*``' intrinsics fill "len" bytes of memory starting
7323 at the destination location. If the argument is known to be aligned to
7324 some boundary, this can be specified as the fourth argument, otherwise
7325 it should be set to 0 or 1 (both meaning no alignment).
7327 '``llvm.sqrt.*``' Intrinsic
7328 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7333 This is an overloaded intrinsic. You can use ``llvm.sqrt`` on any
7334 floating point or vector of floating point type. Not all targets support
7339 declare float @llvm.sqrt.f32(float %Val)
7340 declare double @llvm.sqrt.f64(double %Val)
7341 declare x86_fp80 @llvm.sqrt.f80(x86_fp80 %Val)
7342 declare fp128 @llvm.sqrt.f128(fp128 %Val)
7343 declare ppc_fp128 @llvm.sqrt.ppcf128(ppc_fp128 %Val)
7348 The '``llvm.sqrt``' intrinsics return the sqrt of the specified operand,
7349 returning the same value as the libm '``sqrt``' functions would. Unlike
7350 ``sqrt`` in libm, however, ``llvm.sqrt`` has undefined behavior for
7351 negative numbers other than -0.0 (which allows for better optimization,
7352 because there is no need to worry about errno being set).
7353 ``llvm.sqrt(-0.0)`` is defined to return -0.0 like IEEE sqrt.
7358 The argument and return value are floating point numbers of the same
7364 This function returns the sqrt of the specified operand if it is a
7365 nonnegative floating point number.
7367 '``llvm.powi.*``' Intrinsic
7368 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7373 This is an overloaded intrinsic. You can use ``llvm.powi`` on any
7374 floating point or vector of floating point type. Not all targets support
7379 declare float @llvm.powi.f32(float %Val, i32 %power)
7380 declare double @llvm.powi.f64(double %Val, i32 %power)
7381 declare x86_fp80 @llvm.powi.f80(x86_fp80 %Val, i32 %power)
7382 declare fp128 @llvm.powi.f128(fp128 %Val, i32 %power)
7383 declare ppc_fp128 @llvm.powi.ppcf128(ppc_fp128 %Val, i32 %power)
7388 The '``llvm.powi.*``' intrinsics return the first operand raised to the
7389 specified (positive or negative) power. The order of evaluation of
7390 multiplications is not defined. When a vector of floating point type is
7391 used, the second argument remains a scalar integer value.
7396 The second argument is an integer power, and the first is a value to
7397 raise to that power.
7402 This function returns the first value raised to the second power with an
7403 unspecified sequence of rounding operations.
7405 '``llvm.sin.*``' Intrinsic
7406 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7411 This is an overloaded intrinsic. You can use ``llvm.sin`` on any
7412 floating point or vector of floating point type. Not all targets support
7417 declare float @llvm.sin.f32(float %Val)
7418 declare double @llvm.sin.f64(double %Val)
7419 declare x86_fp80 @llvm.sin.f80(x86_fp80 %Val)
7420 declare fp128 @llvm.sin.f128(fp128 %Val)
7421 declare ppc_fp128 @llvm.sin.ppcf128(ppc_fp128 %Val)
7426 The '``llvm.sin.*``' intrinsics return the sine of the operand.
7431 The argument and return value are floating point numbers of the same
7437 This function returns the sine of the specified operand, returning the
7438 same values as the libm ``sin`` functions would, and handles error
7439 conditions in the same way.
7441 '``llvm.cos.*``' Intrinsic
7442 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7447 This is an overloaded intrinsic. You can use ``llvm.cos`` on any
7448 floating point or vector of floating point type. Not all targets support
7453 declare float @llvm.cos.f32(float %Val)
7454 declare double @llvm.cos.f64(double %Val)
7455 declare x86_fp80 @llvm.cos.f80(x86_fp80 %Val)
7456 declare fp128 @llvm.cos.f128(fp128 %Val)
7457 declare ppc_fp128 @llvm.cos.ppcf128(ppc_fp128 %Val)
7462 The '``llvm.cos.*``' intrinsics return the cosine of the operand.
7467 The argument and return value are floating point numbers of the same
7473 This function returns the cosine of the specified operand, returning the
7474 same values as the libm ``cos`` functions would, and handles error
7475 conditions in the same way.
7477 '``llvm.pow.*``' Intrinsic
7478 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7483 This is an overloaded intrinsic. You can use ``llvm.pow`` on any
7484 floating point or vector of floating point type. Not all targets support
7489 declare float @llvm.pow.f32(float %Val, float %Power)
7490 declare double @llvm.pow.f64(double %Val, double %Power)
7491 declare x86_fp80 @llvm.pow.f80(x86_fp80 %Val, x86_fp80 %Power)
7492 declare fp128 @llvm.pow.f128(fp128 %Val, fp128 %Power)
7493 declare ppc_fp128 @llvm.pow.ppcf128(ppc_fp128 %Val, ppc_fp128 Power)
7498 The '``llvm.pow.*``' intrinsics return the first operand raised to the
7499 specified (positive or negative) power.
7504 The second argument is a floating point power, and the first is a value
7505 to raise to that power.
7510 This function returns the first value raised to the second power,
7511 returning the same values as the libm ``pow`` functions would, and
7512 handles error conditions in the same way.
7514 '``llvm.exp.*``' Intrinsic
7515 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7520 This is an overloaded intrinsic. You can use ``llvm.exp`` on any
7521 floating point or vector of floating point type. Not all targets support
7526 declare float @llvm.exp.f32(float %Val)
7527 declare double @llvm.exp.f64(double %Val)
7528 declare x86_fp80 @llvm.exp.f80(x86_fp80 %Val)
7529 declare fp128 @llvm.exp.f128(fp128 %Val)
7530 declare ppc_fp128 @llvm.exp.ppcf128(ppc_fp128 %Val)
7535 The '``llvm.exp.*``' intrinsics perform the exp function.
7540 The argument and return value are floating point numbers of the same
7546 This function returns the same values as the libm ``exp`` functions
7547 would, and handles error conditions in the same way.
7549 '``llvm.exp2.*``' Intrinsic
7550 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7555 This is an overloaded intrinsic. You can use ``llvm.exp2`` on any
7556 floating point or vector of floating point type. Not all targets support
7561 declare float @llvm.exp2.f32(float %Val)
7562 declare double @llvm.exp2.f64(double %Val)
7563 declare x86_fp80 @llvm.exp2.f80(x86_fp80 %Val)
7564 declare fp128 @llvm.exp2.f128(fp128 %Val)
7565 declare ppc_fp128 @llvm.exp2.ppcf128(ppc_fp128 %Val)
7570 The '``llvm.exp2.*``' intrinsics perform the exp2 function.
7575 The argument and return value are floating point numbers of the same
7581 This function returns the same values as the libm ``exp2`` functions
7582 would, and handles error conditions in the same way.
7584 '``llvm.log.*``' Intrinsic
7585 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7590 This is an overloaded intrinsic. You can use ``llvm.log`` on any
7591 floating point or vector of floating point type. Not all targets support
7596 declare float @llvm.log.f32(float %Val)
7597 declare double @llvm.log.f64(double %Val)
7598 declare x86_fp80 @llvm.log.f80(x86_fp80 %Val)
7599 declare fp128 @llvm.log.f128(fp128 %Val)
7600 declare ppc_fp128 @llvm.log.ppcf128(ppc_fp128 %Val)
7605 The '``llvm.log.*``' intrinsics perform the log function.
7610 The argument and return value are floating point numbers of the same
7616 This function returns the same values as the libm ``log`` functions
7617 would, and handles error conditions in the same way.
7619 '``llvm.log10.*``' Intrinsic
7620 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7625 This is an overloaded intrinsic. You can use ``llvm.log10`` on any
7626 floating point or vector of floating point type. Not all targets support
7631 declare float @llvm.log10.f32(float %Val)
7632 declare double @llvm.log10.f64(double %Val)
7633 declare x86_fp80 @llvm.log10.f80(x86_fp80 %Val)
7634 declare fp128 @llvm.log10.f128(fp128 %Val)
7635 declare ppc_fp128 @llvm.log10.ppcf128(ppc_fp128 %Val)
7640 The '``llvm.log10.*``' intrinsics perform the log10 function.
7645 The argument and return value are floating point numbers of the same
7651 This function returns the same values as the libm ``log10`` functions
7652 would, and handles error conditions in the same way.
7654 '``llvm.log2.*``' Intrinsic
7655 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7660 This is an overloaded intrinsic. You can use ``llvm.log2`` on any
7661 floating point or vector of floating point type. Not all targets support
7666 declare float @llvm.log2.f32(float %Val)
7667 declare double @llvm.log2.f64(double %Val)
7668 declare x86_fp80 @llvm.log2.f80(x86_fp80 %Val)
7669 declare fp128 @llvm.log2.f128(fp128 %Val)
7670 declare ppc_fp128 @llvm.log2.ppcf128(ppc_fp128 %Val)
7675 The '``llvm.log2.*``' intrinsics perform the log2 function.
7680 The argument and return value are floating point numbers of the same
7686 This function returns the same values as the libm ``log2`` functions
7687 would, and handles error conditions in the same way.
7689 '``llvm.fma.*``' Intrinsic
7690 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7695 This is an overloaded intrinsic. You can use ``llvm.fma`` on any
7696 floating point or vector of floating point type. Not all targets support
7701 declare float @llvm.fma.f32(float %a, float %b, float %c)
7702 declare double @llvm.fma.f64(double %a, double %b, double %c)
7703 declare x86_fp80 @llvm.fma.f80(x86_fp80 %a, x86_fp80 %b, x86_fp80 %c)
7704 declare fp128 @llvm.fma.f128(fp128 %a, fp128 %b, fp128 %c)
7705 declare ppc_fp128 @llvm.fma.ppcf128(ppc_fp128 %a, ppc_fp128 %b, ppc_fp128 %c)
7710 The '``llvm.fma.*``' intrinsics perform the fused multiply-add
7716 The argument and return value are floating point numbers of the same
7722 This function returns the same values as the libm ``fma`` functions
7723 would, and does not set errno.
7725 '``llvm.fabs.*``' Intrinsic
7726 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7731 This is an overloaded intrinsic. You can use ``llvm.fabs`` on any
7732 floating point or vector of floating point type. Not all targets support
7737 declare float @llvm.fabs.f32(float %Val)
7738 declare double @llvm.fabs.f64(double %Val)
7739 declare x86_fp80 @llvm.fabs.f80(x86_fp80 %Val)
7740 declare fp128 @llvm.fabs.f128(fp128 %Val)
7741 declare ppc_fp128 @llvm.fabs.ppcf128(ppc_fp128 %Val)
7746 The '``llvm.fabs.*``' intrinsics return the absolute value of the
7752 The argument and return value are floating point numbers of the same
7758 This function returns the same values as the libm ``fabs`` functions
7759 would, and handles error conditions in the same way.
7761 '``llvm.copysign.*``' Intrinsic
7762 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7767 This is an overloaded intrinsic. You can use ``llvm.copysign`` on any
7768 floating point or vector of floating point type. Not all targets support
7773 declare float @llvm.copysign.f32(float %Mag, float %Sgn)
7774 declare double @llvm.copysign.f64(double %Mag, double %Sgn)
7775 declare x86_fp80 @llvm.copysign.f80(x86_fp80 %Mag, x86_fp80 %Sgn)
7776 declare fp128 @llvm.copysign.f128(fp128 %Mag, fp128 %Sgn)
7777 declare ppc_fp128 @llvm.copysign.ppcf128(ppc_fp128 %Mag, ppc_fp128 %Sgn)
7782 The '``llvm.copysign.*``' intrinsics return a value with the magnitude of the
7783 first operand and the sign of the second operand.
7788 The arguments and return value are floating point numbers of the same
7794 This function returns the same values as the libm ``copysign``
7795 functions would, and handles error conditions in the same way.
7797 '``llvm.floor.*``' Intrinsic
7798 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7803 This is an overloaded intrinsic. You can use ``llvm.floor`` on any
7804 floating point or vector of floating point type. Not all targets support
7809 declare float @llvm.floor.f32(float %Val)
7810 declare double @llvm.floor.f64(double %Val)
7811 declare x86_fp80 @llvm.floor.f80(x86_fp80 %Val)
7812 declare fp128 @llvm.floor.f128(fp128 %Val)
7813 declare ppc_fp128 @llvm.floor.ppcf128(ppc_fp128 %Val)
7818 The '``llvm.floor.*``' intrinsics return the floor of the operand.
7823 The argument and return value are floating point numbers of the same
7829 This function returns the same values as the libm ``floor`` functions
7830 would, and handles error conditions in the same way.
7832 '``llvm.ceil.*``' Intrinsic
7833 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7838 This is an overloaded intrinsic. You can use ``llvm.ceil`` on any
7839 floating point or vector of floating point type. Not all targets support
7844 declare float @llvm.ceil.f32(float %Val)
7845 declare double @llvm.ceil.f64(double %Val)
7846 declare x86_fp80 @llvm.ceil.f80(x86_fp80 %Val)
7847 declare fp128 @llvm.ceil.f128(fp128 %Val)
7848 declare ppc_fp128 @llvm.ceil.ppcf128(ppc_fp128 %Val)
7853 The '``llvm.ceil.*``' intrinsics return the ceiling of the operand.
7858 The argument and return value are floating point numbers of the same
7864 This function returns the same values as the libm ``ceil`` functions
7865 would, and handles error conditions in the same way.
7867 '``llvm.trunc.*``' Intrinsic
7868 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7873 This is an overloaded intrinsic. You can use ``llvm.trunc`` on any
7874 floating point or vector of floating point type. Not all targets support
7879 declare float @llvm.trunc.f32(float %Val)
7880 declare double @llvm.trunc.f64(double %Val)
7881 declare x86_fp80 @llvm.trunc.f80(x86_fp80 %Val)
7882 declare fp128 @llvm.trunc.f128(fp128 %Val)
7883 declare ppc_fp128 @llvm.trunc.ppcf128(ppc_fp128 %Val)
7888 The '``llvm.trunc.*``' intrinsics returns the operand rounded to the
7889 nearest integer not larger in magnitude than the operand.
7894 The argument and return value are floating point numbers of the same
7900 This function returns the same values as the libm ``trunc`` functions
7901 would, and handles error conditions in the same way.
7903 '``llvm.rint.*``' Intrinsic
7904 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7909 This is an overloaded intrinsic. You can use ``llvm.rint`` on any
7910 floating point or vector of floating point type. Not all targets support
7915 declare float @llvm.rint.f32(float %Val)
7916 declare double @llvm.rint.f64(double %Val)
7917 declare x86_fp80 @llvm.rint.f80(x86_fp80 %Val)
7918 declare fp128 @llvm.rint.f128(fp128 %Val)
7919 declare ppc_fp128 @llvm.rint.ppcf128(ppc_fp128 %Val)
7924 The '``llvm.rint.*``' intrinsics returns the operand rounded to the
7925 nearest integer. It may raise an inexact floating-point exception if the
7926 operand isn't an integer.
7931 The argument and return value are floating point numbers of the same
7937 This function returns the same values as the libm ``rint`` functions
7938 would, and handles error conditions in the same way.
7940 '``llvm.nearbyint.*``' Intrinsic
7941 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7946 This is an overloaded intrinsic. You can use ``llvm.nearbyint`` on any
7947 floating point or vector of floating point type. Not all targets support
7952 declare float @llvm.nearbyint.f32(float %Val)
7953 declare double @llvm.nearbyint.f64(double %Val)
7954 declare x86_fp80 @llvm.nearbyint.f80(x86_fp80 %Val)
7955 declare fp128 @llvm.nearbyint.f128(fp128 %Val)
7956 declare ppc_fp128 @llvm.nearbyint.ppcf128(ppc_fp128 %Val)
7961 The '``llvm.nearbyint.*``' intrinsics returns the operand rounded to the
7967 The argument and return value are floating point numbers of the same
7973 This function returns the same values as the libm ``nearbyint``
7974 functions would, and handles error conditions in the same way.
7976 '``llvm.round.*``' Intrinsic
7977 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7982 This is an overloaded intrinsic. You can use ``llvm.round`` on any
7983 floating point or vector of floating point type. Not all targets support
7988 declare float @llvm.round.f32(float %Val)
7989 declare double @llvm.round.f64(double %Val)
7990 declare x86_fp80 @llvm.round.f80(x86_fp80 %Val)
7991 declare fp128 @llvm.round.f128(fp128 %Val)
7992 declare ppc_fp128 @llvm.round.ppcf128(ppc_fp128 %Val)
7997 The '``llvm.round.*``' intrinsics returns the operand rounded to the
8003 The argument and return value are floating point numbers of the same
8009 This function returns the same values as the libm ``round``
8010 functions would, and handles error conditions in the same way.
8012 Bit Manipulation Intrinsics
8013 ---------------------------
8015 LLVM provides intrinsics for a few important bit manipulation
8016 operations. These allow efficient code generation for some algorithms.
8018 '``llvm.bswap.*``' Intrinsics
8019 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8024 This is an overloaded intrinsic function. You can use bswap on any
8025 integer type that is an even number of bytes (i.e. BitWidth % 16 == 0).
8029 declare i16 @llvm.bswap.i16(i16 <id>)
8030 declare i32 @llvm.bswap.i32(i32 <id>)
8031 declare i64 @llvm.bswap.i64(i64 <id>)
8036 The '``llvm.bswap``' family of intrinsics is used to byte swap integer
8037 values with an even number of bytes (positive multiple of 16 bits).
8038 These are useful for performing operations on data that is not in the
8039 target's native byte order.
8044 The ``llvm.bswap.i16`` intrinsic returns an i16 value that has the high
8045 and low byte of the input i16 swapped. Similarly, the ``llvm.bswap.i32``
8046 intrinsic returns an i32 value that has the four bytes of the input i32
8047 swapped, so that if the input bytes are numbered 0, 1, 2, 3 then the
8048 returned i32 will have its bytes in 3, 2, 1, 0 order. The
8049 ``llvm.bswap.i48``, ``llvm.bswap.i64`` and other intrinsics extend this
8050 concept to additional even-byte lengths (6 bytes, 8 bytes and more,
8053 '``llvm.ctpop.*``' Intrinsic
8054 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8059 This is an overloaded intrinsic. You can use llvm.ctpop on any integer
8060 bit width, or on any vector with integer elements. Not all targets
8061 support all bit widths or vector types, however.
8065 declare i8 @llvm.ctpop.i8(i8 <src>)
8066 declare i16 @llvm.ctpop.i16(i16 <src>)
8067 declare i32 @llvm.ctpop.i32(i32 <src>)
8068 declare i64 @llvm.ctpop.i64(i64 <src>)
8069 declare i256 @llvm.ctpop.i256(i256 <src>)
8070 declare <2 x i32> @llvm.ctpop.v2i32(<2 x i32> <src>)
8075 The '``llvm.ctpop``' family of intrinsics counts the number of bits set
8081 The only argument is the value to be counted. The argument may be of any
8082 integer type, or a vector with integer elements. The return type must
8083 match the argument type.
8088 The '``llvm.ctpop``' intrinsic counts the 1's in a variable, or within
8089 each element of a vector.
8091 '``llvm.ctlz.*``' Intrinsic
8092 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8097 This is an overloaded intrinsic. You can use ``llvm.ctlz`` on any
8098 integer bit width, or any vector whose elements are integers. Not all
8099 targets support all bit widths or vector types, however.
8103 declare i8 @llvm.ctlz.i8 (i8 <src>, i1 <is_zero_undef>)
8104 declare i16 @llvm.ctlz.i16 (i16 <src>, i1 <is_zero_undef>)
8105 declare i32 @llvm.ctlz.i32 (i32 <src>, i1 <is_zero_undef>)
8106 declare i64 @llvm.ctlz.i64 (i64 <src>, i1 <is_zero_undef>)
8107 declare i256 @llvm.ctlz.i256(i256 <src>, i1 <is_zero_undef>)
8108 declase <2 x i32> @llvm.ctlz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
8113 The '``llvm.ctlz``' family of intrinsic functions counts the number of
8114 leading zeros in a variable.
8119 The first argument is the value to be counted. This argument may be of
8120 any integer type, or a vectory with integer element type. The return
8121 type must match the first argument type.
8123 The second argument must be a constant and is a flag to indicate whether
8124 the intrinsic should ensure that a zero as the first argument produces a
8125 defined result. Historically some architectures did not provide a
8126 defined result for zero values as efficiently, and many algorithms are
8127 now predicated on avoiding zero-value inputs.
8132 The '``llvm.ctlz``' intrinsic counts the leading (most significant)
8133 zeros in a variable, or within each element of the vector. If
8134 ``src == 0`` then the result is the size in bits of the type of ``src``
8135 if ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
8136 ``llvm.ctlz(i32 2) = 30``.
8138 '``llvm.cttz.*``' Intrinsic
8139 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8144 This is an overloaded intrinsic. You can use ``llvm.cttz`` on any
8145 integer bit width, or any vector of integer elements. Not all targets
8146 support all bit widths or vector types, however.
8150 declare i8 @llvm.cttz.i8 (i8 <src>, i1 <is_zero_undef>)
8151 declare i16 @llvm.cttz.i16 (i16 <src>, i1 <is_zero_undef>)
8152 declare i32 @llvm.cttz.i32 (i32 <src>, i1 <is_zero_undef>)
8153 declare i64 @llvm.cttz.i64 (i64 <src>, i1 <is_zero_undef>)
8154 declare i256 @llvm.cttz.i256(i256 <src>, i1 <is_zero_undef>)
8155 declase <2 x i32> @llvm.cttz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
8160 The '``llvm.cttz``' family of intrinsic functions counts the number of
8166 The first argument is the value to be counted. This argument may be of
8167 any integer type, or a vectory with integer element type. The return
8168 type must match the first argument type.
8170 The second argument must be a constant and is a flag to indicate whether
8171 the intrinsic should ensure that a zero as the first argument produces a
8172 defined result. Historically some architectures did not provide a
8173 defined result for zero values as efficiently, and many algorithms are
8174 now predicated on avoiding zero-value inputs.
8179 The '``llvm.cttz``' intrinsic counts the trailing (least significant)
8180 zeros in a variable, or within each element of a vector. If ``src == 0``
8181 then the result is the size in bits of the type of ``src`` if
8182 ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
8183 ``llvm.cttz(2) = 1``.
8185 Arithmetic with Overflow Intrinsics
8186 -----------------------------------
8188 LLVM provides intrinsics for some arithmetic with overflow operations.
8190 '``llvm.sadd.with.overflow.*``' Intrinsics
8191 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8196 This is an overloaded intrinsic. You can use ``llvm.sadd.with.overflow``
8197 on any integer bit width.
8201 declare {i16, i1} @llvm.sadd.with.overflow.i16(i16 %a, i16 %b)
8202 declare {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
8203 declare {i64, i1} @llvm.sadd.with.overflow.i64(i64 %a, i64 %b)
8208 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
8209 a signed addition of the two arguments, and indicate whether an overflow
8210 occurred during the signed summation.
8215 The arguments (%a and %b) and the first element of the result structure
8216 may be of integer types of any bit width, but they must have the same
8217 bit width. The second element of the result structure must be of type
8218 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
8224 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
8225 a signed addition of the two variables. They return a structure --- the
8226 first element of which is the signed summation, and the second element
8227 of which is a bit specifying if the signed summation resulted in an
8233 .. code-block:: llvm
8235 %res = call {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
8236 %sum = extractvalue {i32, i1} %res, 0
8237 %obit = extractvalue {i32, i1} %res, 1
8238 br i1 %obit, label %overflow, label %normal
8240 '``llvm.uadd.with.overflow.*``' Intrinsics
8241 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8246 This is an overloaded intrinsic. You can use ``llvm.uadd.with.overflow``
8247 on any integer bit width.
8251 declare {i16, i1} @llvm.uadd.with.overflow.i16(i16 %a, i16 %b)
8252 declare {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
8253 declare {i64, i1} @llvm.uadd.with.overflow.i64(i64 %a, i64 %b)
8258 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
8259 an unsigned addition of the two arguments, and indicate whether a carry
8260 occurred during the unsigned summation.
8265 The arguments (%a and %b) and the first element of the result structure
8266 may be of integer types of any bit width, but they must have the same
8267 bit width. The second element of the result structure must be of type
8268 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
8274 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
8275 an unsigned addition of the two arguments. They return a structure --- the
8276 first element of which is the sum, and the second element of which is a
8277 bit specifying if the unsigned summation resulted in a carry.
8282 .. code-block:: llvm
8284 %res = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
8285 %sum = extractvalue {i32, i1} %res, 0
8286 %obit = extractvalue {i32, i1} %res, 1
8287 br i1 %obit, label %carry, label %normal
8289 '``llvm.ssub.with.overflow.*``' Intrinsics
8290 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8295 This is an overloaded intrinsic. You can use ``llvm.ssub.with.overflow``
8296 on any integer bit width.
8300 declare {i16, i1} @llvm.ssub.with.overflow.i16(i16 %a, i16 %b)
8301 declare {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
8302 declare {i64, i1} @llvm.ssub.with.overflow.i64(i64 %a, i64 %b)
8307 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
8308 a signed subtraction of the two arguments, and indicate whether an
8309 overflow occurred during the signed subtraction.
8314 The arguments (%a and %b) and the first element of the result structure
8315 may be of integer types of any bit width, but they must have the same
8316 bit width. The second element of the result structure must be of type
8317 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
8323 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
8324 a signed subtraction of the two arguments. They return a structure --- the
8325 first element of which is the subtraction, and the second element of
8326 which is a bit specifying if the signed subtraction resulted in an
8332 .. code-block:: llvm
8334 %res = call {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
8335 %sum = extractvalue {i32, i1} %res, 0
8336 %obit = extractvalue {i32, i1} %res, 1
8337 br i1 %obit, label %overflow, label %normal
8339 '``llvm.usub.with.overflow.*``' Intrinsics
8340 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8345 This is an overloaded intrinsic. You can use ``llvm.usub.with.overflow``
8346 on any integer bit width.
8350 declare {i16, i1} @llvm.usub.with.overflow.i16(i16 %a, i16 %b)
8351 declare {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
8352 declare {i64, i1} @llvm.usub.with.overflow.i64(i64 %a, i64 %b)
8357 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
8358 an unsigned subtraction of the two arguments, and indicate whether an
8359 overflow occurred during the unsigned subtraction.
8364 The arguments (%a and %b) and the first element of the result structure
8365 may be of integer types of any bit width, but they must have the same
8366 bit width. The second element of the result structure must be of type
8367 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
8373 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
8374 an unsigned subtraction of the two arguments. They return a structure ---
8375 the first element of which is the subtraction, and the second element of
8376 which is a bit specifying if the unsigned subtraction resulted in an
8382 .. code-block:: llvm
8384 %res = call {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
8385 %sum = extractvalue {i32, i1} %res, 0
8386 %obit = extractvalue {i32, i1} %res, 1
8387 br i1 %obit, label %overflow, label %normal
8389 '``llvm.smul.with.overflow.*``' Intrinsics
8390 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8395 This is an overloaded intrinsic. You can use ``llvm.smul.with.overflow``
8396 on any integer bit width.
8400 declare {i16, i1} @llvm.smul.with.overflow.i16(i16 %a, i16 %b)
8401 declare {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
8402 declare {i64, i1} @llvm.smul.with.overflow.i64(i64 %a, i64 %b)
8407 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
8408 a signed multiplication of the two arguments, and indicate whether an
8409 overflow occurred during the signed multiplication.
8414 The arguments (%a and %b) and the first element of the result structure
8415 may be of integer types of any bit width, but they must have the same
8416 bit width. The second element of the result structure must be of type
8417 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
8423 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
8424 a signed multiplication of the two arguments. They return a structure ---
8425 the first element of which is the multiplication, and the second element
8426 of which is a bit specifying if the signed multiplication resulted in an
8432 .. code-block:: llvm
8434 %res = call {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
8435 %sum = extractvalue {i32, i1} %res, 0
8436 %obit = extractvalue {i32, i1} %res, 1
8437 br i1 %obit, label %overflow, label %normal
8439 '``llvm.umul.with.overflow.*``' Intrinsics
8440 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8445 This is an overloaded intrinsic. You can use ``llvm.umul.with.overflow``
8446 on any integer bit width.
8450 declare {i16, i1} @llvm.umul.with.overflow.i16(i16 %a, i16 %b)
8451 declare {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
8452 declare {i64, i1} @llvm.umul.with.overflow.i64(i64 %a, i64 %b)
8457 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
8458 a unsigned multiplication of the two arguments, and indicate whether an
8459 overflow occurred during the unsigned multiplication.
8464 The arguments (%a and %b) and the first element of the result structure
8465 may be of integer types of any bit width, but they must have the same
8466 bit width. The second element of the result structure must be of type
8467 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
8473 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
8474 an unsigned multiplication of the two arguments. They return a structure ---
8475 the first element of which is the multiplication, and the second
8476 element of which is a bit specifying if the unsigned multiplication
8477 resulted in an overflow.
8482 .. code-block:: llvm
8484 %res = call {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
8485 %sum = extractvalue {i32, i1} %res, 0
8486 %obit = extractvalue {i32, i1} %res, 1
8487 br i1 %obit, label %overflow, label %normal
8489 Specialised Arithmetic Intrinsics
8490 ---------------------------------
8492 '``llvm.fmuladd.*``' Intrinsic
8493 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8500 declare float @llvm.fmuladd.f32(float %a, float %b, float %c)
8501 declare double @llvm.fmuladd.f64(double %a, double %b, double %c)
8506 The '``llvm.fmuladd.*``' intrinsic functions represent multiply-add
8507 expressions that can be fused if the code generator determines that (a) the
8508 target instruction set has support for a fused operation, and (b) that the
8509 fused operation is more efficient than the equivalent, separate pair of mul
8510 and add instructions.
8515 The '``llvm.fmuladd.*``' intrinsics each take three arguments: two
8516 multiplicands, a and b, and an addend c.
8525 %0 = call float @llvm.fmuladd.f32(%a, %b, %c)
8527 is equivalent to the expression a \* b + c, except that rounding will
8528 not be performed between the multiplication and addition steps if the
8529 code generator fuses the operations. Fusion is not guaranteed, even if
8530 the target platform supports it. If a fused multiply-add is required the
8531 corresponding llvm.fma.\* intrinsic function should be used
8532 instead. This never sets errno, just as '``llvm.fma.*``'.
8537 .. code-block:: llvm
8539 %r2 = call float @llvm.fmuladd.f32(float %a, float %b, float %c) ; yields float:r2 = (a * b) + c
8541 Half Precision Floating Point Intrinsics
8542 ----------------------------------------
8544 For most target platforms, half precision floating point is a
8545 storage-only format. This means that it is a dense encoding (in memory)
8546 but does not support computation in the format.
8548 This means that code must first load the half-precision floating point
8549 value as an i16, then convert it to float with
8550 :ref:`llvm.convert.from.fp16 <int_convert_from_fp16>`. Computation can
8551 then be performed on the float value (including extending to double
8552 etc). To store the value back to memory, it is first converted to float
8553 if needed, then converted to i16 with
8554 :ref:`llvm.convert.to.fp16 <int_convert_to_fp16>`, then storing as an
8557 .. _int_convert_to_fp16:
8559 '``llvm.convert.to.fp16``' Intrinsic
8560 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8567 declare i16 @llvm.convert.to.fp16(f32 %a)
8572 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion
8573 from single precision floating point format to half precision floating
8579 The intrinsic function contains single argument - the value to be
8585 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion
8586 from single precision floating point format to half precision floating
8587 point format. The return value is an ``i16`` which contains the
8593 .. code-block:: llvm
8595 %res = call i16 @llvm.convert.to.fp16(f32 %a)
8596 store i16 %res, i16* @x, align 2
8598 .. _int_convert_from_fp16:
8600 '``llvm.convert.from.fp16``' Intrinsic
8601 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8608 declare f32 @llvm.convert.from.fp16(i16 %a)
8613 The '``llvm.convert.from.fp16``' intrinsic function performs a
8614 conversion from half precision floating point format to single precision
8615 floating point format.
8620 The intrinsic function contains single argument - the value to be
8626 The '``llvm.convert.from.fp16``' intrinsic function performs a
8627 conversion from half single precision floating point format to single
8628 precision floating point format. The input half-float value is
8629 represented by an ``i16`` value.
8634 .. code-block:: llvm
8636 %a = load i16* @x, align 2
8637 %res = call f32 @llvm.convert.from.fp16(i16 %a)
8642 The LLVM debugger intrinsics (which all start with ``llvm.dbg.``
8643 prefix), are described in the `LLVM Source Level
8644 Debugging <SourceLevelDebugging.html#format_common_intrinsics>`_
8647 Exception Handling Intrinsics
8648 -----------------------------
8650 The LLVM exception handling intrinsics (which all start with
8651 ``llvm.eh.`` prefix), are described in the `LLVM Exception
8652 Handling <ExceptionHandling.html#format_common_intrinsics>`_ document.
8656 Trampoline Intrinsics
8657 ---------------------
8659 These intrinsics make it possible to excise one parameter, marked with
8660 the :ref:`nest <nest>` attribute, from a function. The result is a
8661 callable function pointer lacking the nest parameter - the caller does
8662 not need to provide a value for it. Instead, the value to use is stored
8663 in advance in a "trampoline", a block of memory usually allocated on the
8664 stack, which also contains code to splice the nest value into the
8665 argument list. This is used to implement the GCC nested function address
8668 For example, if the function is ``i32 f(i8* nest %c, i32 %x, i32 %y)``
8669 then the resulting function pointer has signature ``i32 (i32, i32)*``.
8670 It can be created as follows:
8672 .. code-block:: llvm
8674 %tramp = alloca [10 x i8], align 4 ; size and alignment only correct for X86
8675 %tramp1 = getelementptr [10 x i8]* %tramp, i32 0, i32 0
8676 call i8* @llvm.init.trampoline(i8* %tramp1, i8* bitcast (i32 (i8*, i32, i32)* @f to i8*), i8* %nval)
8677 %p = call i8* @llvm.adjust.trampoline(i8* %tramp1)
8678 %fp = bitcast i8* %p to i32 (i32, i32)*
8680 The call ``%val = call i32 %fp(i32 %x, i32 %y)`` is then equivalent to
8681 ``%val = call i32 %f(i8* %nval, i32 %x, i32 %y)``.
8685 '``llvm.init.trampoline``' Intrinsic
8686 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8693 declare void @llvm.init.trampoline(i8* <tramp>, i8* <func>, i8* <nval>)
8698 This fills the memory pointed to by ``tramp`` with executable code,
8699 turning it into a trampoline.
8704 The ``llvm.init.trampoline`` intrinsic takes three arguments, all
8705 pointers. The ``tramp`` argument must point to a sufficiently large and
8706 sufficiently aligned block of memory; this memory is written to by the
8707 intrinsic. Note that the size and the alignment are target-specific -
8708 LLVM currently provides no portable way of determining them, so a
8709 front-end that generates this intrinsic needs to have some
8710 target-specific knowledge. The ``func`` argument must hold a function
8711 bitcast to an ``i8*``.
8716 The block of memory pointed to by ``tramp`` is filled with target
8717 dependent code, turning it into a function. Then ``tramp`` needs to be
8718 passed to :ref:`llvm.adjust.trampoline <int_at>` to get a pointer which can
8719 be :ref:`bitcast (to a new function) and called <int_trampoline>`. The new
8720 function's signature is the same as that of ``func`` with any arguments
8721 marked with the ``nest`` attribute removed. At most one such ``nest``
8722 argument is allowed, and it must be of pointer type. Calling the new
8723 function is equivalent to calling ``func`` with the same argument list,
8724 but with ``nval`` used for the missing ``nest`` argument. If, after
8725 calling ``llvm.init.trampoline``, the memory pointed to by ``tramp`` is
8726 modified, then the effect of any later call to the returned function
8727 pointer is undefined.
8731 '``llvm.adjust.trampoline``' Intrinsic
8732 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8739 declare i8* @llvm.adjust.trampoline(i8* <tramp>)
8744 This performs any required machine-specific adjustment to the address of
8745 a trampoline (passed as ``tramp``).
8750 ``tramp`` must point to a block of memory which already has trampoline
8751 code filled in by a previous call to
8752 :ref:`llvm.init.trampoline <int_it>`.
8757 On some architectures the address of the code to be executed needs to be
8758 different to the address where the trampoline is actually stored. This
8759 intrinsic returns the executable address corresponding to ``tramp``
8760 after performing the required machine specific adjustments. The pointer
8761 returned can then be :ref:`bitcast and executed <int_trampoline>`.
8766 This class of intrinsics exists to information about the lifetime of
8767 memory objects and ranges where variables are immutable.
8771 '``llvm.lifetime.start``' Intrinsic
8772 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8779 declare void @llvm.lifetime.start(i64 <size>, i8* nocapture <ptr>)
8784 The '``llvm.lifetime.start``' intrinsic specifies the start of a memory
8790 The first argument is a constant integer representing the size of the
8791 object, or -1 if it is variable sized. The second argument is a pointer
8797 This intrinsic indicates that before this point in the code, the value
8798 of the memory pointed to by ``ptr`` is dead. This means that it is known
8799 to never be used and has an undefined value. A load from the pointer
8800 that precedes this intrinsic can be replaced with ``'undef'``.
8804 '``llvm.lifetime.end``' Intrinsic
8805 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8812 declare void @llvm.lifetime.end(i64 <size>, i8* nocapture <ptr>)
8817 The '``llvm.lifetime.end``' intrinsic specifies the end of a memory
8823 The first argument is a constant integer representing the size of the
8824 object, or -1 if it is variable sized. The second argument is a pointer
8830 This intrinsic indicates that after this point in the code, the value of
8831 the memory pointed to by ``ptr`` is dead. This means that it is known to
8832 never be used and has an undefined value. Any stores into the memory
8833 object following this intrinsic may be removed as dead.
8835 '``llvm.invariant.start``' Intrinsic
8836 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8843 declare {}* @llvm.invariant.start(i64 <size>, i8* nocapture <ptr>)
8848 The '``llvm.invariant.start``' intrinsic specifies that the contents of
8849 a memory object will not change.
8854 The first argument is a constant integer representing the size of the
8855 object, or -1 if it is variable sized. The second argument is a pointer
8861 This intrinsic indicates that until an ``llvm.invariant.end`` that uses
8862 the return value, the referenced memory location is constant and
8865 '``llvm.invariant.end``' Intrinsic
8866 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8873 declare void @llvm.invariant.end({}* <start>, i64 <size>, i8* nocapture <ptr>)
8878 The '``llvm.invariant.end``' intrinsic specifies that the contents of a
8879 memory object are mutable.
8884 The first argument is the matching ``llvm.invariant.start`` intrinsic.
8885 The second argument is a constant integer representing the size of the
8886 object, or -1 if it is variable sized and the third argument is a
8887 pointer to the object.
8892 This intrinsic indicates that the memory is mutable again.
8897 This class of intrinsics is designed to be generic and has no specific
8900 '``llvm.var.annotation``' Intrinsic
8901 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8908 declare void @llvm.var.annotation(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
8913 The '``llvm.var.annotation``' intrinsic.
8918 The first argument is a pointer to a value, the second is a pointer to a
8919 global string, the third is a pointer to a global string which is the
8920 source file name, and the last argument is the line number.
8925 This intrinsic allows annotation of local variables with arbitrary
8926 strings. This can be useful for special purpose optimizations that want
8927 to look for these annotations. These have no other defined use; they are
8928 ignored by code generation and optimization.
8930 '``llvm.ptr.annotation.*``' Intrinsic
8931 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8936 This is an overloaded intrinsic. You can use '``llvm.ptr.annotation``' on a
8937 pointer to an integer of any width. *NOTE* you must specify an address space for
8938 the pointer. The identifier for the default address space is the integer
8943 declare i8* @llvm.ptr.annotation.p<address space>i8(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
8944 declare i16* @llvm.ptr.annotation.p<address space>i16(i16* <val>, i8* <str>, i8* <str>, i32 <int>)
8945 declare i32* @llvm.ptr.annotation.p<address space>i32(i32* <val>, i8* <str>, i8* <str>, i32 <int>)
8946 declare i64* @llvm.ptr.annotation.p<address space>i64(i64* <val>, i8* <str>, i8* <str>, i32 <int>)
8947 declare i256* @llvm.ptr.annotation.p<address space>i256(i256* <val>, i8* <str>, i8* <str>, i32 <int>)
8952 The '``llvm.ptr.annotation``' intrinsic.
8957 The first argument is a pointer to an integer value of arbitrary bitwidth
8958 (result of some expression), the second is a pointer to a global string, the
8959 third is a pointer to a global string which is the source file name, and the
8960 last argument is the line number. It returns the value of the first argument.
8965 This intrinsic allows annotation of a pointer to an integer with arbitrary
8966 strings. This can be useful for special purpose optimizations that want to look
8967 for these annotations. These have no other defined use; they are ignored by code
8968 generation and optimization.
8970 '``llvm.annotation.*``' Intrinsic
8971 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8976 This is an overloaded intrinsic. You can use '``llvm.annotation``' on
8977 any integer bit width.
8981 declare i8 @llvm.annotation.i8(i8 <val>, i8* <str>, i8* <str>, i32 <int>)
8982 declare i16 @llvm.annotation.i16(i16 <val>, i8* <str>, i8* <str>, i32 <int>)
8983 declare i32 @llvm.annotation.i32(i32 <val>, i8* <str>, i8* <str>, i32 <int>)
8984 declare i64 @llvm.annotation.i64(i64 <val>, i8* <str>, i8* <str>, i32 <int>)
8985 declare i256 @llvm.annotation.i256(i256 <val>, i8* <str>, i8* <str>, i32 <int>)
8990 The '``llvm.annotation``' intrinsic.
8995 The first argument is an integer value (result of some expression), the
8996 second is a pointer to a global string, the third is a pointer to a
8997 global string which is the source file name, and the last argument is
8998 the line number. It returns the value of the first argument.
9003 This intrinsic allows annotations to be put on arbitrary expressions
9004 with arbitrary strings. This can be useful for special purpose
9005 optimizations that want to look for these annotations. These have no
9006 other defined use; they are ignored by code generation and optimization.
9008 '``llvm.trap``' Intrinsic
9009 ^^^^^^^^^^^^^^^^^^^^^^^^^
9016 declare void @llvm.trap() noreturn nounwind
9021 The '``llvm.trap``' intrinsic.
9031 This intrinsic is lowered to the target dependent trap instruction. If
9032 the target does not have a trap instruction, this intrinsic will be
9033 lowered to a call of the ``abort()`` function.
9035 '``llvm.debugtrap``' Intrinsic
9036 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9043 declare void @llvm.debugtrap() nounwind
9048 The '``llvm.debugtrap``' intrinsic.
9058 This intrinsic is lowered to code which is intended to cause an
9059 execution trap with the intention of requesting the attention of a
9062 '``llvm.stackprotector``' Intrinsic
9063 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9070 declare void @llvm.stackprotector(i8* <guard>, i8** <slot>)
9075 The ``llvm.stackprotector`` intrinsic takes the ``guard`` and stores it
9076 onto the stack at ``slot``. The stack slot is adjusted to ensure that it
9077 is placed on the stack before local variables.
9082 The ``llvm.stackprotector`` intrinsic requires two pointer arguments.
9083 The first argument is the value loaded from the stack guard
9084 ``@__stack_chk_guard``. The second variable is an ``alloca`` that has
9085 enough space to hold the value of the guard.
9090 This intrinsic causes the prologue/epilogue inserter to force the position of
9091 the ``AllocaInst`` stack slot to be before local variables on the stack. This is
9092 to ensure that if a local variable on the stack is overwritten, it will destroy
9093 the value of the guard. When the function exits, the guard on the stack is
9094 checked against the original guard by ``llvm.stackprotectorcheck``. If they are
9095 different, then ``llvm.stackprotectorcheck`` causes the program to abort by
9096 calling the ``__stack_chk_fail()`` function.
9098 '``llvm.stackprotectorcheck``' Intrinsic
9099 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9106 declare void @llvm.stackprotectorcheck(i8** <guard>)
9111 The ``llvm.stackprotectorcheck`` intrinsic compares ``guard`` against an already
9112 created stack protector and if they are not equal calls the
9113 ``__stack_chk_fail()`` function.
9118 The ``llvm.stackprotectorcheck`` intrinsic requires one pointer argument, the
9119 the variable ``@__stack_chk_guard``.
9124 This intrinsic is provided to perform the stack protector check by comparing
9125 ``guard`` with the stack slot created by ``llvm.stackprotector`` and if the
9126 values do not match call the ``__stack_chk_fail()`` function.
9128 The reason to provide this as an IR level intrinsic instead of implementing it
9129 via other IR operations is that in order to perform this operation at the IR
9130 level without an intrinsic, one would need to create additional basic blocks to
9131 handle the success/failure cases. This makes it difficult to stop the stack
9132 protector check from disrupting sibling tail calls in Codegen. With this
9133 intrinsic, we are able to generate the stack protector basic blocks late in
9134 codegen after the tail call decision has occurred.
9136 '``llvm.objectsize``' Intrinsic
9137 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9144 declare i32 @llvm.objectsize.i32(i8* <object>, i1 <min>)
9145 declare i64 @llvm.objectsize.i64(i8* <object>, i1 <min>)
9150 The ``llvm.objectsize`` intrinsic is designed to provide information to
9151 the optimizers to determine at compile time whether a) an operation
9152 (like memcpy) will overflow a buffer that corresponds to an object, or
9153 b) that a runtime check for overflow isn't necessary. An object in this
9154 context means an allocation of a specific class, structure, array, or
9160 The ``llvm.objectsize`` intrinsic takes two arguments. The first
9161 argument is a pointer to or into the ``object``. The second argument is
9162 a boolean and determines whether ``llvm.objectsize`` returns 0 (if true)
9163 or -1 (if false) when the object size is unknown. The second argument
9164 only accepts constants.
9169 The ``llvm.objectsize`` intrinsic is lowered to a constant representing
9170 the size of the object concerned. If the size cannot be determined at
9171 compile time, ``llvm.objectsize`` returns ``i32/i64 -1 or 0`` (depending
9172 on the ``min`` argument).
9174 '``llvm.expect``' Intrinsic
9175 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9180 This is an overloaded intrinsic. You can use ``llvm.expect`` on any
9185 declare i1 @llvm.expect.i1(i1 <val>, i1 <expected_val>)
9186 declare i32 @llvm.expect.i32(i32 <val>, i32 <expected_val>)
9187 declare i64 @llvm.expect.i64(i64 <val>, i64 <expected_val>)
9192 The ``llvm.expect`` intrinsic provides information about expected (the
9193 most probable) value of ``val``, which can be used by optimizers.
9198 The ``llvm.expect`` intrinsic takes two arguments. The first argument is
9199 a value. The second argument is an expected value, this needs to be a
9200 constant value, variables are not allowed.
9205 This intrinsic is lowered to the ``val``.
9207 '``llvm.donothing``' Intrinsic
9208 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9215 declare void @llvm.donothing() nounwind readnone
9220 The ``llvm.donothing`` intrinsic doesn't perform any operation. It's the
9221 only intrinsic that can be called with an invoke instruction.
9231 This intrinsic does nothing, and it's removed by optimizers and ignored
9234 Stack Map Intrinsics
9235 --------------------
9237 LLVM provides experimental intrinsics to support runtime patching
9238 mechanisms commonly desired in dynamic language JITs. These intrinsics
9239 are described in :doc:`StackMaps`.