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.
565 Additionally, the global can placed in a comdat if the target has the necessary
568 By default, global initializers are optimized by assuming that global
569 variables defined within the module are not modified from their
570 initial values before the start of the global initializer. This is
571 true even for variables potentially accessible from outside the
572 module, including those with external linkage or appearing in
573 ``@llvm.used`` or dllexported variables. This assumption may be suppressed
574 by marking the variable with ``externally_initialized``.
576 An explicit alignment may be specified for a global, which must be a
577 power of 2. If not present, or if the alignment is set to zero, the
578 alignment of the global is set by the target to whatever it feels
579 convenient. If an explicit alignment is specified, the global is forced
580 to have exactly that alignment. Targets and optimizers are not allowed
581 to over-align the global if the global has an assigned section. In this
582 case, the extra alignment could be observable: for example, code could
583 assume that the globals are densely packed in their section and try to
584 iterate over them as an array, alignment padding would break this
585 iteration. The maximum alignment is ``1 << 29``.
587 Globals can also have a :ref:`DLL storage class <dllstorageclass>`.
589 Variables and aliasaes can have a
590 :ref:`Thread Local Storage Model <tls_model>`.
594 [@<GlobalVarName> =] [Linkage] [Visibility] [DLLStorageClass] [ThreadLocal]
595 [unnamed_addr] [AddrSpace] [ExternallyInitialized]
596 <global | constant> <Type> [<InitializerConstant>]
597 [, section "name"] [, align <Alignment>]
599 For example, the following defines a global in a numbered address space
600 with an initializer, section, and alignment:
604 @G = addrspace(5) constant float 1.0, section "foo", align 4
606 The following example just declares a global variable
610 @G = external global i32
612 The following example defines a thread-local global with the
613 ``initialexec`` TLS model:
617 @G = thread_local(initialexec) global i32 0, align 4
619 .. _functionstructure:
624 LLVM function definitions consist of the "``define``" keyword, an
625 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
626 style <visibility>`, an optional :ref:`DLL storage class <dllstorageclass>`,
627 an optional :ref:`calling convention <callingconv>`,
628 an optional ``unnamed_addr`` attribute, a return type, an optional
629 :ref:`parameter attribute <paramattrs>` for the return type, a function
630 name, a (possibly empty) argument list (each with optional :ref:`parameter
631 attributes <paramattrs>`), optional :ref:`function attributes <fnattrs>`,
632 an optional section, an optional alignment,
633 an optional :ref:`comdat <langref_comdats>`,
634 an optional :ref:`garbage collector name <gc>`, an optional :ref:`prefix <prefixdata>`, an opening
635 curly brace, a list of basic blocks, and a closing curly brace.
637 LLVM function declarations consist of the "``declare``" keyword, an
638 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
639 style <visibility>`, an optional :ref:`DLL storage class <dllstorageclass>`,
640 an optional :ref:`calling convention <callingconv>`,
641 an optional ``unnamed_addr`` attribute, a return type, an optional
642 :ref:`parameter attribute <paramattrs>` for the return type, a function
643 name, a possibly empty list of arguments, an optional alignment, an optional
644 :ref:`garbage collector name <gc>` and an optional :ref:`prefix <prefixdata>`.
646 A function definition contains a list of basic blocks, forming the CFG (Control
647 Flow Graph) for the function. Each basic block may optionally start with a label
648 (giving the basic block a symbol table entry), contains a list of instructions,
649 and ends with a :ref:`terminator <terminators>` instruction (such as a branch or
650 function return). If an explicit label is not provided, a block is assigned an
651 implicit numbered label, using the next value from the same counter as used for
652 unnamed temporaries (:ref:`see above<identifiers>`). For example, if a function
653 entry block does not have an explicit label, it will be assigned label "%0",
654 then the first unnamed temporary in that block will be "%1", etc.
656 The first basic block in a function is special in two ways: it is
657 immediately executed on entrance to the function, and it is not allowed
658 to have predecessor basic blocks (i.e. there can not be any branches to
659 the entry block of a function). Because the block can have no
660 predecessors, it also cannot have any :ref:`PHI nodes <i_phi>`.
662 LLVM allows an explicit section to be specified for functions. If the
663 target supports it, it will emit functions to the section specified.
664 Additionally, the function can placed in a COMDAT.
666 An explicit alignment may be specified for a function. If not present,
667 or if the alignment is set to zero, the alignment of the function is set
668 by the target to whatever it feels convenient. If an explicit alignment
669 is specified, the function is forced to have at least that much
670 alignment. All alignments must be a power of 2.
672 If the ``unnamed_addr`` attribute is given, the address is know to not
673 be significant and two identical functions can be merged.
677 define [linkage] [visibility] [DLLStorageClass]
679 <ResultType> @<FunctionName> ([argument list])
680 [unnamed_addr] [fn Attrs] [section "name"] [comdat $<ComdatName>]
681 [align N] [gc] [prefix Constant] { ... }
688 Aliases, unlike function or variables, don't create any new data. They
689 are just a new symbol and metadata for an existing position.
691 Aliases have a name and an aliasee that is either a global value or a
694 Aliases may have an optional :ref:`linkage type <linkage>`, an optional
695 :ref:`visibility style <visibility>`, an optional :ref:`DLL storage class
696 <dllstorageclass>` and an optional :ref:`tls model <tls_model>`.
700 @<Name> = [Visibility] [DLLStorageClass] [ThreadLocal] [unnamed_addr] alias [Linkage] <AliaseeTy> @<Aliasee>
702 The linkage must be one of ``private``, ``internal``, ``linkonce``, ``weak``,
703 ``linkonce_odr``, ``weak_odr``, ``external``. Note that some system linkers
704 might not correctly handle dropping a weak symbol that is aliased.
706 Alias that are not ``unnamed_addr`` are guaranteed to have the same address as
707 the aliasee expression. ``unnamed_addr`` ones are only guaranteed to point
710 Since aliases are only a second name, some restrictions apply, of which
711 some can only be checked when producing an object file:
713 * The expression defining the aliasee must be computable at assembly
714 time. Since it is just a name, no relocations can be used.
716 * No alias in the expression can be weak as the possibility of the
717 intermediate alias being overridden cannot be represented in an
720 * No global value in the expression can be a declaration, since that
721 would require a relocation, which is not possible.
728 Comdat IR provides access to COFF and ELF object file COMDAT functionality.
730 Comdats have a name which represents the COMDAT key. All global objects which
731 specify this key will only end up in the final object file if the linker chooses
732 that key over some other key. Aliases are placed in the same COMDAT that their
733 aliasee computes to, if any.
735 Comdats have a selection kind to provide input on how the linker should
736 choose between keys in two different object files.
740 $<Name> = comdat SelectionKind
742 The selection kind must be one of the following:
745 The linker may choose any COMDAT key, the choice is arbitrary.
747 The linker may choose any COMDAT key but the sections must contain the
750 The linker will choose the section containing the largest COMDAT key.
752 The linker requires that only section with this COMDAT key exist.
754 The linker may choose any COMDAT key but the sections must contain the
757 Note that the Mach-O platform doesn't support COMDATs and ELF only supports
758 ``any`` as a selection kind.
760 Here is an example of a COMDAT group where a function will only be selected if
761 the COMDAT key's section is the largest:
765 $foo = comdat largest
766 @foo = global i32 2, comdat $foo
768 define void @bar() comdat $foo {
772 In a COFF object file, this will create a COMDAT section with selection kind
773 ``IMAGE_COMDAT_SELECT_LARGEST`` containing the contents of the ``@foo`` symbol
774 and another COMDAT section with selection kind
775 ``IMAGE_COMDAT_SELECT_ASSOCIATIVE`` which is associated with the first COMDAT
776 section and contains the contents of the ``@baz`` symbol.
778 There are some restrictions on the properties of the global object.
779 It, or an alias to it, must have the same name as the COMDAT group when
781 The contents and size of this object may be used during link-time to determine
782 which COMDAT groups get selected depending on the selection kind.
783 Because the name of the object must match the name of the COMDAT group, the
784 linkage of the global object must not be local; local symbols can get renamed
785 if a collision occurs in the symbol table.
787 The combined use of COMDATS and section attributes may yield surprising results.
794 @g1 = global i32 42, section "sec", comdat $foo
795 @g2 = global i32 42, section "sec", comdat $bar
797 From the object file perspective, this requires the creation of two sections
798 with the same name. This is necessary because both globals belong to different
799 COMDAT groups and COMDATs, at the object file level, are represented by
802 Note that certain IR constructs like global variables and functions may create
803 COMDATs in the object file in addition to any which are specified using COMDAT
804 IR. This arises, for example, when a global variable has linkonce_odr linkage.
806 .. _namedmetadatastructure:
811 Named metadata is a collection of metadata. :ref:`Metadata
812 nodes <metadata>` (but not metadata strings) are the only valid
813 operands for a named metadata.
817 ; Some unnamed metadata nodes, which are referenced by the named metadata.
818 !0 = metadata !{metadata !"zero"}
819 !1 = metadata !{metadata !"one"}
820 !2 = metadata !{metadata !"two"}
822 !name = !{!0, !1, !2}
829 The return type and each parameter of a function type may have a set of
830 *parameter attributes* associated with them. Parameter attributes are
831 used to communicate additional information about the result or
832 parameters of a function. Parameter attributes are considered to be part
833 of the function, not of the function type, so functions with different
834 parameter attributes can have the same function type.
836 Parameter attributes are simple keywords that follow the type specified.
837 If multiple parameter attributes are needed, they are space separated.
842 declare i32 @printf(i8* noalias nocapture, ...)
843 declare i32 @atoi(i8 zeroext)
844 declare signext i8 @returns_signed_char()
846 Note that any attributes for the function result (``nounwind``,
847 ``readonly``) come immediately after the argument list.
849 Currently, only the following parameter attributes are defined:
852 This indicates to the code generator that the parameter or return
853 value should be zero-extended to the extent required by the target's
854 ABI (which is usually 32-bits, but is 8-bits for a i1 on x86-64) by
855 the caller (for a parameter) or the callee (for a return value).
857 This indicates to the code generator that the parameter or return
858 value should be sign-extended to the extent required by the target's
859 ABI (which is usually 32-bits) by the caller (for a parameter) or
860 the callee (for a return value).
862 This indicates that this parameter or return value should be treated
863 in a special target-dependent fashion during while emitting code for
864 a function call or return (usually, by putting it in a register as
865 opposed to memory, though some targets use it to distinguish between
866 two different kinds of registers). Use of this attribute is
869 This indicates that the pointer parameter should really be passed by
870 value to the function. The attribute implies that a hidden copy of
871 the pointee is made between the caller and the callee, so the callee
872 is unable to modify the value in the caller. This attribute is only
873 valid on LLVM pointer arguments. It is generally used to pass
874 structs and arrays by value, but is also valid on pointers to
875 scalars. The copy is considered to belong to the caller not the
876 callee (for example, ``readonly`` functions should not write to
877 ``byval`` parameters). This is not a valid attribute for return
880 The byval attribute also supports specifying an alignment with the
881 align attribute. It indicates the alignment of the stack slot to
882 form and the known alignment of the pointer specified to the call
883 site. If the alignment is not specified, then the code generator
884 makes a target-specific assumption.
890 The ``inalloca`` argument attribute allows the caller to take the
891 address of outgoing stack arguments. An ``inalloca`` argument must
892 be a pointer to stack memory produced by an ``alloca`` instruction.
893 The alloca, or argument allocation, must also be tagged with the
894 inalloca keyword. Only the last argument may have the ``inalloca``
895 attribute, and that argument is guaranteed to be passed in memory.
897 An argument allocation may be used by a call at most once because
898 the call may deallocate it. The ``inalloca`` attribute cannot be
899 used in conjunction with other attributes that affect argument
900 storage, like ``inreg``, ``nest``, ``sret``, or ``byval``. The
901 ``inalloca`` attribute also disables LLVM's implicit lowering of
902 large aggregate return values, which means that frontend authors
903 must lower them with ``sret`` pointers.
905 When the call site is reached, the argument allocation must have
906 been the most recent stack allocation that is still live, or the
907 results are undefined. It is possible to allocate additional stack
908 space after an argument allocation and before its call site, but it
909 must be cleared off with :ref:`llvm.stackrestore
912 See :doc:`InAlloca` for more information on how to use this
916 This indicates that the pointer parameter specifies the address of a
917 structure that is the return value of the function in the source
918 program. This pointer must be guaranteed by the caller to be valid:
919 loads and stores to the structure may be assumed by the callee
920 not to trap and to be properly aligned. This may only be applied to
921 the first parameter. This is not a valid attribute for return
927 This indicates that pointer values :ref:`based <pointeraliasing>` on
928 the argument or return value do not alias pointer values which are
929 not *based* on it, ignoring certain "irrelevant" dependencies. For a
930 call to the parent function, dependencies between memory references
931 from before or after the call and from those during the call are
932 "irrelevant" to the ``noalias`` keyword for the arguments and return
933 value used in that call. The caller shares the responsibility with
934 the callee for ensuring that these requirements are met. For further
935 details, please see the discussion of the NoAlias response in :ref:`alias
936 analysis <Must, May, or No>`.
938 Note that this definition of ``noalias`` is intentionally similar
939 to the definition of ``restrict`` in C99 for function arguments,
940 though it is slightly weaker.
942 For function return values, C99's ``restrict`` is not meaningful,
943 while LLVM's ``noalias`` is.
945 This indicates that the callee does not make any copies of the
946 pointer that outlive the callee itself. This is not a valid
947 attribute for return values.
952 This indicates that the pointer parameter can be excised using the
953 :ref:`trampoline intrinsics <int_trampoline>`. This is not a valid
954 attribute for return values and can only be applied to one parameter.
957 This indicates that the function always returns the argument as its return
958 value. This is an optimization hint to the code generator when generating
959 the caller, allowing tail call optimization and omission of register saves
960 and restores in some cases; it is not checked or enforced when generating
961 the callee. The parameter and the function return type must be valid
962 operands for the :ref:`bitcast instruction <i_bitcast>`. This is not a
963 valid attribute for return values and can only be applied to one parameter.
966 This indicates that the parameter or return pointer is not null. This
967 attribute may only be applied to pointer typed parameters. This is not
968 checked or enforced by LLVM, the caller must ensure that the pointer
969 passed in is non-null, or the callee must ensure that the returned pointer
974 Garbage Collector Names
975 -----------------------
977 Each function may specify a garbage collector name, which is simply a
982 define void @f() gc "name" { ... }
984 The compiler declares the supported values of *name*. Specifying a
985 collector which will cause the compiler to alter its output in order to
986 support the named garbage collection algorithm.
993 Prefix data is data associated with a function which the code generator
994 will emit immediately before the function body. The purpose of this feature
995 is to allow frontends to associate language-specific runtime metadata with
996 specific functions and make it available through the function pointer while
997 still allowing the function pointer to be called. To access the data for a
998 given function, a program may bitcast the function pointer to a pointer to
999 the constant's type. This implies that the IR symbol points to the start
1002 To maintain the semantics of ordinary function calls, the prefix data must
1003 have a particular format. Specifically, it must begin with a sequence of
1004 bytes which decode to a sequence of machine instructions, valid for the
1005 module's target, which transfer control to the point immediately succeeding
1006 the prefix data, without performing any other visible action. This allows
1007 the inliner and other passes to reason about the semantics of the function
1008 definition without needing to reason about the prefix data. Obviously this
1009 makes the format of the prefix data highly target dependent.
1011 Prefix data is laid out as if it were an initializer for a global variable
1012 of the prefix data's type. No padding is automatically placed between the
1013 prefix data and the function body. If padding is required, it must be part
1016 A trivial example of valid prefix data for the x86 architecture is ``i8 144``,
1017 which encodes the ``nop`` instruction:
1019 .. code-block:: llvm
1021 define void @f() prefix i8 144 { ... }
1023 Generally prefix data can be formed by encoding a relative branch instruction
1024 which skips the metadata, as in this example of valid prefix data for the
1025 x86_64 architecture, where the first two bytes encode ``jmp .+10``:
1027 .. code-block:: llvm
1029 %0 = type <{ i8, i8, i8* }>
1031 define void @f() prefix %0 <{ i8 235, i8 8, i8* @md}> { ... }
1033 A function may have prefix data but no body. This has similar semantics
1034 to the ``available_externally`` linkage in that the data may be used by the
1035 optimizers but will not be emitted in the object file.
1042 Attribute groups are groups of attributes that are referenced by objects within
1043 the IR. They are important for keeping ``.ll`` files readable, because a lot of
1044 functions will use the same set of attributes. In the degenerative case of a
1045 ``.ll`` file that corresponds to a single ``.c`` file, the single attribute
1046 group will capture the important command line flags used to build that file.
1048 An attribute group is a module-level object. To use an attribute group, an
1049 object references the attribute group's ID (e.g. ``#37``). An object may refer
1050 to more than one attribute group. In that situation, the attributes from the
1051 different groups are merged.
1053 Here is an example of attribute groups for a function that should always be
1054 inlined, has a stack alignment of 4, and which shouldn't use SSE instructions:
1056 .. code-block:: llvm
1058 ; Target-independent attributes:
1059 attributes #0 = { alwaysinline alignstack=4 }
1061 ; Target-dependent attributes:
1062 attributes #1 = { "no-sse" }
1064 ; Function @f has attributes: alwaysinline, alignstack=4, and "no-sse".
1065 define void @f() #0 #1 { ... }
1072 Function attributes are set to communicate additional information about
1073 a function. Function attributes are considered to be part of the
1074 function, not of the function type, so functions with different function
1075 attributes can have the same function type.
1077 Function attributes are simple keywords that follow the type specified.
1078 If multiple attributes are needed, they are space separated. For
1081 .. code-block:: llvm
1083 define void @f() noinline { ... }
1084 define void @f() alwaysinline { ... }
1085 define void @f() alwaysinline optsize { ... }
1086 define void @f() optsize { ... }
1089 This attribute indicates that, when emitting the prologue and
1090 epilogue, the backend should forcibly align the stack pointer.
1091 Specify the desired alignment, which must be a power of two, in
1094 This attribute indicates that the inliner should attempt to inline
1095 this function into callers whenever possible, ignoring any active
1096 inlining size threshold for this caller.
1098 This indicates that the callee function at a call site should be
1099 recognized as a built-in function, even though the function's declaration
1100 uses the ``nobuiltin`` attribute. This is only valid at call sites for
1101 direct calls to functions which are declared with the ``nobuiltin``
1104 This attribute indicates that this function is rarely called. When
1105 computing edge weights, basic blocks post-dominated by a cold
1106 function call are also considered to be cold; and, thus, given low
1109 This attribute indicates that the source code contained a hint that
1110 inlining this function is desirable (such as the "inline" keyword in
1111 C/C++). It is just a hint; it imposes no requirements on the
1114 This attribute indicates that the function should be added to a
1115 jump-instruction table at code-generation time, and that all address-taken
1116 references to this function should be replaced with a reference to the
1117 appropriate jump-instruction-table function pointer. Note that this creates
1118 a new pointer for the original function, which means that code that depends
1119 on function-pointer identity can break. So, any function annotated with
1120 ``jumptable`` must also be ``unnamed_addr``.
1122 This attribute suggests that optimization passes and code generator
1123 passes make choices that keep the code size of this function as small
1124 as possible and perform optimizations that may sacrifice runtime
1125 performance in order to minimize the size of the generated code.
1127 This attribute disables prologue / epilogue emission for the
1128 function. This can have very system-specific consequences.
1130 This indicates that the callee function at a call site is not recognized as
1131 a built-in function. LLVM will retain the original call and not replace it
1132 with equivalent code based on the semantics of the built-in function, unless
1133 the call site uses the ``builtin`` attribute. This is valid at call sites
1134 and on function declarations and definitions.
1136 This attribute indicates that calls to the function cannot be
1137 duplicated. A call to a ``noduplicate`` function may be moved
1138 within its parent function, but may not be duplicated within
1139 its parent function.
1141 A function containing a ``noduplicate`` call may still
1142 be an inlining candidate, provided that the call is not
1143 duplicated by inlining. That implies that the function has
1144 internal linkage and only has one call site, so the original
1145 call is dead after inlining.
1147 This attributes disables implicit floating point instructions.
1149 This attribute indicates that the inliner should never inline this
1150 function in any situation. This attribute may not be used together
1151 with the ``alwaysinline`` attribute.
1153 This attribute suppresses lazy symbol binding for the function. This
1154 may make calls to the function faster, at the cost of extra program
1155 startup time if the function is not called during program startup.
1157 This attribute indicates that the code generator should not use a
1158 red zone, even if the target-specific ABI normally permits it.
1160 This function attribute indicates that the function never returns
1161 normally. This produces undefined behavior at runtime if the
1162 function ever does dynamically return.
1164 This function attribute indicates that the function never returns
1165 with an unwind or exceptional control flow. If the function does
1166 unwind, its runtime behavior is undefined.
1168 This function attribute indicates that the function is not optimized
1169 by any optimization or code generator passes with the
1170 exception of interprocedural optimization passes.
1171 This attribute cannot be used together with the ``alwaysinline``
1172 attribute; this attribute is also incompatible
1173 with the ``minsize`` attribute and the ``optsize`` attribute.
1175 This attribute requires the ``noinline`` attribute to be specified on
1176 the function as well, so the function is never inlined into any caller.
1177 Only functions with the ``alwaysinline`` attribute are valid
1178 candidates for inlining into the body of this function.
1180 This attribute suggests that optimization passes and code generator
1181 passes make choices that keep the code size of this function low,
1182 and otherwise do optimizations specifically to reduce code size as
1183 long as they do not significantly impact runtime performance.
1185 On a function, this attribute indicates that the function computes its
1186 result (or decides to unwind an exception) based strictly on its arguments,
1187 without dereferencing any pointer arguments or otherwise accessing
1188 any mutable state (e.g. memory, control registers, etc) visible to
1189 caller functions. It does not write through any pointer arguments
1190 (including ``byval`` arguments) and never changes any state visible
1191 to callers. This means that it cannot unwind exceptions by calling
1192 the ``C++`` exception throwing methods.
1194 On an argument, this attribute indicates that the function does not
1195 dereference that pointer argument, even though it may read or write the
1196 memory that the pointer points to if accessed through other pointers.
1198 On a function, this attribute indicates that the function does not write
1199 through any pointer arguments (including ``byval`` arguments) or otherwise
1200 modify any state (e.g. memory, control registers, etc) visible to
1201 caller functions. It may dereference pointer arguments and read
1202 state that may be set in the caller. A readonly function always
1203 returns the same value (or unwinds an exception identically) when
1204 called with the same set of arguments and global state. It cannot
1205 unwind an exception by calling the ``C++`` exception throwing
1208 On an argument, this attribute indicates that the function does not write
1209 through this pointer argument, even though it may write to the memory that
1210 the pointer points to.
1212 This attribute indicates that this function can return twice. The C
1213 ``setjmp`` is an example of such a function. The compiler disables
1214 some optimizations (like tail calls) in the caller of these
1216 ``sanitize_address``
1217 This attribute indicates that AddressSanitizer checks
1218 (dynamic address safety analysis) are enabled for this function.
1220 This attribute indicates that MemorySanitizer checks (dynamic detection
1221 of accesses to uninitialized memory) are enabled for this function.
1223 This attribute indicates that ThreadSanitizer checks
1224 (dynamic thread safety analysis) are enabled for this function.
1226 This attribute indicates that the function should emit a stack
1227 smashing protector. It is in the form of a "canary" --- a random value
1228 placed on the stack before the local variables that's checked upon
1229 return from the function to see if it has been overwritten. A
1230 heuristic is used to determine if a function needs stack protectors
1231 or not. The heuristic used will enable protectors for functions with:
1233 - Character arrays larger than ``ssp-buffer-size`` (default 8).
1234 - Aggregates containing character arrays larger than ``ssp-buffer-size``.
1235 - Calls to alloca() with variable sizes or constant sizes greater than
1236 ``ssp-buffer-size``.
1238 Variables that are identified as requiring a protector will be arranged
1239 on the stack such that they are adjacent to the stack protector guard.
1241 If a function that has an ``ssp`` attribute is inlined into a
1242 function that doesn't have an ``ssp`` attribute, then the resulting
1243 function will have an ``ssp`` attribute.
1245 This attribute indicates that the function should *always* emit a
1246 stack smashing protector. This overrides the ``ssp`` function
1249 Variables that are identified as requiring a protector will be arranged
1250 on the stack such that they are adjacent to the stack protector guard.
1251 The specific layout rules are:
1253 #. Large arrays and structures containing large arrays
1254 (``>= ssp-buffer-size``) are closest to the stack protector.
1255 #. Small arrays and structures containing small arrays
1256 (``< ssp-buffer-size``) are 2nd closest to the protector.
1257 #. Variables that have had their address taken are 3rd closest to the
1260 If a function that has an ``sspreq`` attribute is inlined into a
1261 function that doesn't have an ``sspreq`` attribute or which has an
1262 ``ssp`` or ``sspstrong`` attribute, then the resulting function will have
1263 an ``sspreq`` attribute.
1265 This attribute indicates that the function should emit a stack smashing
1266 protector. This attribute causes a strong heuristic to be used when
1267 determining if a function needs stack protectors. The strong heuristic
1268 will enable protectors for functions with:
1270 - Arrays of any size and type
1271 - Aggregates containing an array of any size and type.
1272 - Calls to alloca().
1273 - Local variables that have had their address taken.
1275 Variables that are identified as requiring a protector will be arranged
1276 on the stack such that they are adjacent to the stack protector guard.
1277 The specific layout rules are:
1279 #. Large arrays and structures containing large arrays
1280 (``>= ssp-buffer-size``) are closest to the stack protector.
1281 #. Small arrays and structures containing small arrays
1282 (``< ssp-buffer-size``) are 2nd closest to the protector.
1283 #. Variables that have had their address taken are 3rd closest to the
1286 This overrides the ``ssp`` function attribute.
1288 If a function that has an ``sspstrong`` attribute is inlined into a
1289 function that doesn't have an ``sspstrong`` attribute, then the
1290 resulting function will have an ``sspstrong`` attribute.
1292 This attribute indicates that the ABI being targeted requires that
1293 an unwind table entry be produce for this function even if we can
1294 show that no exceptions passes by it. This is normally the case for
1295 the ELF x86-64 abi, but it can be disabled for some compilation
1300 Module-Level Inline Assembly
1301 ----------------------------
1303 Modules may contain "module-level inline asm" blocks, which corresponds
1304 to the GCC "file scope inline asm" blocks. These blocks are internally
1305 concatenated by LLVM and treated as a single unit, but may be separated
1306 in the ``.ll`` file if desired. The syntax is very simple:
1308 .. code-block:: llvm
1310 module asm "inline asm code goes here"
1311 module asm "more can go here"
1313 The strings can contain any character by escaping non-printable
1314 characters. The escape sequence used is simply "\\xx" where "xx" is the
1315 two digit hex code for the number.
1317 The inline asm code is simply printed to the machine code .s file when
1318 assembly code is generated.
1320 .. _langref_datalayout:
1325 A module may specify a target specific data layout string that specifies
1326 how data is to be laid out in memory. The syntax for the data layout is
1329 .. code-block:: llvm
1331 target datalayout = "layout specification"
1333 The *layout specification* consists of a list of specifications
1334 separated by the minus sign character ('-'). Each specification starts
1335 with a letter and may include other information after the letter to
1336 define some aspect of the data layout. The specifications accepted are
1340 Specifies that the target lays out data in big-endian form. That is,
1341 the bits with the most significance have the lowest address
1344 Specifies that the target lays out data in little-endian form. That
1345 is, the bits with the least significance have the lowest address
1348 Specifies the natural alignment of the stack in bits. Alignment
1349 promotion of stack variables is limited to the natural stack
1350 alignment to avoid dynamic stack realignment. The stack alignment
1351 must be a multiple of 8-bits. If omitted, the natural stack
1352 alignment defaults to "unspecified", which does not prevent any
1353 alignment promotions.
1354 ``p[n]:<size>:<abi>:<pref>``
1355 This specifies the *size* of a pointer and its ``<abi>`` and
1356 ``<pref>``\erred alignments for address space ``n``. All sizes are in
1357 bits. The address space, ``n`` is optional, and if not specified,
1358 denotes the default address space 0. The value of ``n`` must be
1359 in the range [1,2^23).
1360 ``i<size>:<abi>:<pref>``
1361 This specifies the alignment for an integer type of a given bit
1362 ``<size>``. The value of ``<size>`` must be in the range [1,2^23).
1363 ``v<size>:<abi>:<pref>``
1364 This specifies the alignment for a vector type of a given bit
1366 ``f<size>:<abi>:<pref>``
1367 This specifies the alignment for a floating point type of a given bit
1368 ``<size>``. Only values of ``<size>`` that are supported by the target
1369 will work. 32 (float) and 64 (double) are supported on all targets; 80
1370 or 128 (different flavors of long double) are also supported on some
1373 This specifies the alignment for an object of aggregate type.
1375 If present, specifies that llvm names are mangled in the output. The
1378 * ``e``: ELF mangling: Private symbols get a ``.L`` prefix.
1379 * ``m``: Mips mangling: Private symbols get a ``$`` prefix.
1380 * ``o``: Mach-O mangling: Private symbols get ``L`` prefix. Other
1381 symbols get a ``_`` prefix.
1382 * ``w``: Windows COFF prefix: Similar to Mach-O, but stdcall and fastcall
1383 functions also get a suffix based on the frame size.
1384 ``n<size1>:<size2>:<size3>...``
1385 This specifies a set of native integer widths for the target CPU in
1386 bits. For example, it might contain ``n32`` for 32-bit PowerPC,
1387 ``n32:64`` for PowerPC 64, or ``n8:16:32:64`` for X86-64. Elements of
1388 this set are considered to support most general arithmetic operations
1391 On every specification that takes a ``<abi>:<pref>``, specifying the
1392 ``<pref>`` alignment is optional. If omitted, the preceding ``:``
1393 should be omitted too and ``<pref>`` will be equal to ``<abi>``.
1395 When constructing the data layout for a given target, LLVM starts with a
1396 default set of specifications which are then (possibly) overridden by
1397 the specifications in the ``datalayout`` keyword. The default
1398 specifications are given in this list:
1400 - ``E`` - big endian
1401 - ``p:64:64:64`` - 64-bit pointers with 64-bit alignment.
1402 - ``p[n]:64:64:64`` - Other address spaces are assumed to be the
1403 same as the default address space.
1404 - ``S0`` - natural stack alignment is unspecified
1405 - ``i1:8:8`` - i1 is 8-bit (byte) aligned
1406 - ``i8:8:8`` - i8 is 8-bit (byte) aligned
1407 - ``i16:16:16`` - i16 is 16-bit aligned
1408 - ``i32:32:32`` - i32 is 32-bit aligned
1409 - ``i64:32:64`` - i64 has ABI alignment of 32-bits but preferred
1410 alignment of 64-bits
1411 - ``f16:16:16`` - half is 16-bit aligned
1412 - ``f32:32:32`` - float is 32-bit aligned
1413 - ``f64:64:64`` - double is 64-bit aligned
1414 - ``f128:128:128`` - quad is 128-bit aligned
1415 - ``v64:64:64`` - 64-bit vector is 64-bit aligned
1416 - ``v128:128:128`` - 128-bit vector is 128-bit aligned
1417 - ``a:0:64`` - aggregates are 64-bit aligned
1419 When LLVM is determining the alignment for a given type, it uses the
1422 #. If the type sought is an exact match for one of the specifications,
1423 that specification is used.
1424 #. If no match is found, and the type sought is an integer type, then
1425 the smallest integer type that is larger than the bitwidth of the
1426 sought type is used. If none of the specifications are larger than
1427 the bitwidth then the largest integer type is used. For example,
1428 given the default specifications above, the i7 type will use the
1429 alignment of i8 (next largest) while both i65 and i256 will use the
1430 alignment of i64 (largest specified).
1431 #. If no match is found, and the type sought is a vector type, then the
1432 largest vector type that is smaller than the sought vector type will
1433 be used as a fall back. This happens because <128 x double> can be
1434 implemented in terms of 64 <2 x double>, for example.
1436 The function of the data layout string may not be what you expect.
1437 Notably, this is not a specification from the frontend of what alignment
1438 the code generator should use.
1440 Instead, if specified, the target data layout is required to match what
1441 the ultimate *code generator* expects. This string is used by the
1442 mid-level optimizers to improve code, and this only works if it matches
1443 what the ultimate code generator uses. If you would like to generate IR
1444 that does not embed this target-specific detail into the IR, then you
1445 don't have to specify the string. This will disable some optimizations
1446 that require precise layout information, but this also prevents those
1447 optimizations from introducing target specificity into the IR.
1454 A module may specify a target triple string that describes the target
1455 host. The syntax for the target triple is simply:
1457 .. code-block:: llvm
1459 target triple = "x86_64-apple-macosx10.7.0"
1461 The *target triple* string consists of a series of identifiers delimited
1462 by the minus sign character ('-'). The canonical forms are:
1466 ARCHITECTURE-VENDOR-OPERATING_SYSTEM
1467 ARCHITECTURE-VENDOR-OPERATING_SYSTEM-ENVIRONMENT
1469 This information is passed along to the backend so that it generates
1470 code for the proper architecture. It's possible to override this on the
1471 command line with the ``-mtriple`` command line option.
1473 .. _pointeraliasing:
1475 Pointer Aliasing Rules
1476 ----------------------
1478 Any memory access must be done through a pointer value associated with
1479 an address range of the memory access, otherwise the behavior is
1480 undefined. Pointer values are associated with address ranges according
1481 to the following rules:
1483 - A pointer value is associated with the addresses associated with any
1484 value it is *based* on.
1485 - An address of a global variable is associated with the address range
1486 of the variable's storage.
1487 - The result value of an allocation instruction is associated with the
1488 address range of the allocated storage.
1489 - A null pointer in the default address-space is associated with no
1491 - An integer constant other than zero or a pointer value returned from
1492 a function not defined within LLVM may be associated with address
1493 ranges allocated through mechanisms other than those provided by
1494 LLVM. Such ranges shall not overlap with any ranges of addresses
1495 allocated by mechanisms provided by LLVM.
1497 A pointer value is *based* on another pointer value according to the
1500 - A pointer value formed from a ``getelementptr`` operation is *based*
1501 on the first operand of the ``getelementptr``.
1502 - The result value of a ``bitcast`` is *based* on the operand of the
1504 - A pointer value formed by an ``inttoptr`` is *based* on all pointer
1505 values that contribute (directly or indirectly) to the computation of
1506 the pointer's value.
1507 - The "*based* on" relationship is transitive.
1509 Note that this definition of *"based"* is intentionally similar to the
1510 definition of *"based"* in C99, though it is slightly weaker.
1512 LLVM IR does not associate types with memory. The result type of a
1513 ``load`` merely indicates the size and alignment of the memory from
1514 which to load, as well as the interpretation of the value. The first
1515 operand type of a ``store`` similarly only indicates the size and
1516 alignment of the store.
1518 Consequently, type-based alias analysis, aka TBAA, aka
1519 ``-fstrict-aliasing``, is not applicable to general unadorned LLVM IR.
1520 :ref:`Metadata <metadata>` may be used to encode additional information
1521 which specialized optimization passes may use to implement type-based
1526 Volatile Memory Accesses
1527 ------------------------
1529 Certain memory accesses, such as :ref:`load <i_load>`'s,
1530 :ref:`store <i_store>`'s, and :ref:`llvm.memcpy <int_memcpy>`'s may be
1531 marked ``volatile``. The optimizers must not change the number of
1532 volatile operations or change their order of execution relative to other
1533 volatile operations. The optimizers *may* change the order of volatile
1534 operations relative to non-volatile operations. This is not Java's
1535 "volatile" and has no cross-thread synchronization behavior.
1537 IR-level volatile loads and stores cannot safely be optimized into
1538 llvm.memcpy or llvm.memmove intrinsics even when those intrinsics are
1539 flagged volatile. Likewise, the backend should never split or merge
1540 target-legal volatile load/store instructions.
1542 .. admonition:: Rationale
1544 Platforms may rely on volatile loads and stores of natively supported
1545 data width to be executed as single instruction. For example, in C
1546 this holds for an l-value of volatile primitive type with native
1547 hardware support, but not necessarily for aggregate types. The
1548 frontend upholds these expectations, which are intentionally
1549 unspecified in the IR. The rules above ensure that IR transformation
1550 do not violate the frontend's contract with the language.
1554 Memory Model for Concurrent Operations
1555 --------------------------------------
1557 The LLVM IR does not define any way to start parallel threads of
1558 execution or to register signal handlers. Nonetheless, there are
1559 platform-specific ways to create them, and we define LLVM IR's behavior
1560 in their presence. This model is inspired by the C++0x memory model.
1562 For a more informal introduction to this model, see the :doc:`Atomics`.
1564 We define a *happens-before* partial order as the least partial order
1567 - Is a superset of single-thread program order, and
1568 - When a *synchronizes-with* ``b``, includes an edge from ``a`` to
1569 ``b``. *Synchronizes-with* pairs are introduced by platform-specific
1570 techniques, like pthread locks, thread creation, thread joining,
1571 etc., and by atomic instructions. (See also :ref:`Atomic Memory Ordering
1572 Constraints <ordering>`).
1574 Note that program order does not introduce *happens-before* edges
1575 between a thread and signals executing inside that thread.
1577 Every (defined) read operation (load instructions, memcpy, atomic
1578 loads/read-modify-writes, etc.) R reads a series of bytes written by
1579 (defined) write operations (store instructions, atomic
1580 stores/read-modify-writes, memcpy, etc.). For the purposes of this
1581 section, initialized globals are considered to have a write of the
1582 initializer which is atomic and happens before any other read or write
1583 of the memory in question. For each byte of a read R, R\ :sub:`byte`
1584 may see any write to the same byte, except:
1586 - If write\ :sub:`1` happens before write\ :sub:`2`, and
1587 write\ :sub:`2` happens before R\ :sub:`byte`, then
1588 R\ :sub:`byte` does not see write\ :sub:`1`.
1589 - If R\ :sub:`byte` happens before write\ :sub:`3`, then
1590 R\ :sub:`byte` does not see write\ :sub:`3`.
1592 Given that definition, R\ :sub:`byte` is defined as follows:
1594 - If R is volatile, the result is target-dependent. (Volatile is
1595 supposed to give guarantees which can support ``sig_atomic_t`` in
1596 C/C++, and may be used for accesses to addresses which do not behave
1597 like normal memory. It does not generally provide cross-thread
1599 - Otherwise, if there is no write to the same byte that happens before
1600 R\ :sub:`byte`, R\ :sub:`byte` returns ``undef`` for that byte.
1601 - Otherwise, if R\ :sub:`byte` may see exactly one write,
1602 R\ :sub:`byte` returns the value written by that write.
1603 - Otherwise, if R is atomic, and all the writes R\ :sub:`byte` may
1604 see are atomic, it chooses one of the values written. See the :ref:`Atomic
1605 Memory Ordering Constraints <ordering>` section for additional
1606 constraints on how the choice is made.
1607 - Otherwise R\ :sub:`byte` returns ``undef``.
1609 R returns the value composed of the series of bytes it read. This
1610 implies that some bytes within the value may be ``undef`` **without**
1611 the entire value being ``undef``. Note that this only defines the
1612 semantics of the operation; it doesn't mean that targets will emit more
1613 than one instruction to read the series of bytes.
1615 Note that in cases where none of the atomic intrinsics are used, this
1616 model places only one restriction on IR transformations on top of what
1617 is required for single-threaded execution: introducing a store to a byte
1618 which might not otherwise be stored is not allowed in general.
1619 (Specifically, in the case where another thread might write to and read
1620 from an address, introducing a store can change a load that may see
1621 exactly one write into a load that may see multiple writes.)
1625 Atomic Memory Ordering Constraints
1626 ----------------------------------
1628 Atomic instructions (:ref:`cmpxchg <i_cmpxchg>`,
1629 :ref:`atomicrmw <i_atomicrmw>`, :ref:`fence <i_fence>`,
1630 :ref:`atomic load <i_load>`, and :ref:`atomic store <i_store>`) take
1631 ordering parameters that determine which other atomic instructions on
1632 the same address they *synchronize with*. These semantics are borrowed
1633 from Java and C++0x, but are somewhat more colloquial. If these
1634 descriptions aren't precise enough, check those specs (see spec
1635 references in the :doc:`atomics guide <Atomics>`).
1636 :ref:`fence <i_fence>` instructions treat these orderings somewhat
1637 differently since they don't take an address. See that instruction's
1638 documentation for details.
1640 For a simpler introduction to the ordering constraints, see the
1644 The set of values that can be read is governed by the happens-before
1645 partial order. A value cannot be read unless some operation wrote
1646 it. This is intended to provide a guarantee strong enough to model
1647 Java's non-volatile shared variables. This ordering cannot be
1648 specified for read-modify-write operations; it is not strong enough
1649 to make them atomic in any interesting way.
1651 In addition to the guarantees of ``unordered``, there is a single
1652 total order for modifications by ``monotonic`` operations on each
1653 address. All modification orders must be compatible with the
1654 happens-before order. There is no guarantee that the modification
1655 orders can be combined to a global total order for the whole program
1656 (and this often will not be possible). The read in an atomic
1657 read-modify-write operation (:ref:`cmpxchg <i_cmpxchg>` and
1658 :ref:`atomicrmw <i_atomicrmw>`) reads the value in the modification
1659 order immediately before the value it writes. If one atomic read
1660 happens before another atomic read of the same address, the later
1661 read must see the same value or a later value in the address's
1662 modification order. This disallows reordering of ``monotonic`` (or
1663 stronger) operations on the same address. If an address is written
1664 ``monotonic``-ally by one thread, and other threads ``monotonic``-ally
1665 read that address repeatedly, the other threads must eventually see
1666 the write. This corresponds to the C++0x/C1x
1667 ``memory_order_relaxed``.
1669 In addition to the guarantees of ``monotonic``, a
1670 *synchronizes-with* edge may be formed with a ``release`` operation.
1671 This is intended to model C++'s ``memory_order_acquire``.
1673 In addition to the guarantees of ``monotonic``, if this operation
1674 writes a value which is subsequently read by an ``acquire``
1675 operation, it *synchronizes-with* that operation. (This isn't a
1676 complete description; see the C++0x definition of a release
1677 sequence.) This corresponds to the C++0x/C1x
1678 ``memory_order_release``.
1679 ``acq_rel`` (acquire+release)
1680 Acts as both an ``acquire`` and ``release`` operation on its
1681 address. This corresponds to the C++0x/C1x ``memory_order_acq_rel``.
1682 ``seq_cst`` (sequentially consistent)
1683 In addition to the guarantees of ``acq_rel`` (``acquire`` for an
1684 operation which only reads, ``release`` for an operation which only
1685 writes), there is a global total order on all
1686 sequentially-consistent operations on all addresses, which is
1687 consistent with the *happens-before* partial order and with the
1688 modification orders of all the affected addresses. Each
1689 sequentially-consistent read sees the last preceding write to the
1690 same address in this global order. This corresponds to the C++0x/C1x
1691 ``memory_order_seq_cst`` and Java volatile.
1695 If an atomic operation is marked ``singlethread``, it only *synchronizes
1696 with* or participates in modification and seq\_cst total orderings with
1697 other operations running in the same thread (for example, in signal
1705 LLVM IR floating-point binary ops (:ref:`fadd <i_fadd>`,
1706 :ref:`fsub <i_fsub>`, :ref:`fmul <i_fmul>`, :ref:`fdiv <i_fdiv>`,
1707 :ref:`frem <i_frem>`) have the following flags that can set to enable
1708 otherwise unsafe floating point operations
1711 No NaNs - Allow optimizations to assume the arguments and result are not
1712 NaN. Such optimizations are required to retain defined behavior over
1713 NaNs, but the value of the result is undefined.
1716 No Infs - Allow optimizations to assume the arguments and result are not
1717 +/-Inf. Such optimizations are required to retain defined behavior over
1718 +/-Inf, but the value of the result is undefined.
1721 No Signed Zeros - Allow optimizations to treat the sign of a zero
1722 argument or result as insignificant.
1725 Allow Reciprocal - Allow optimizations to use the reciprocal of an
1726 argument rather than perform division.
1729 Fast - Allow algebraically equivalent transformations that may
1730 dramatically change results in floating point (e.g. reassociate). This
1731 flag implies all the others.
1738 The LLVM type system is one of the most important features of the
1739 intermediate representation. Being typed enables a number of
1740 optimizations to be performed on the intermediate representation
1741 directly, without having to do extra analyses on the side before the
1742 transformation. A strong type system makes it easier to read the
1743 generated code and enables novel analyses and transformations that are
1744 not feasible to perform on normal three address code representations.
1754 The void type does not represent any value and has no size.
1772 The function type can be thought of as a function signature. It consists of a
1773 return type and a list of formal parameter types. The return type of a function
1774 type is a void type or first class type --- except for :ref:`label <t_label>`
1775 and :ref:`metadata <t_metadata>` types.
1781 <returntype> (<parameter list>)
1783 ...where '``<parameter list>``' is a comma-separated list of type
1784 specifiers. Optionally, the parameter list may include a type ``...``, which
1785 indicates that the function takes a variable number of arguments. Variable
1786 argument functions can access their arguments with the :ref:`variable argument
1787 handling intrinsic <int_varargs>` functions. '``<returntype>``' is any type
1788 except :ref:`label <t_label>` and :ref:`metadata <t_metadata>`.
1792 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1793 | ``i32 (i32)`` | function taking an ``i32``, returning an ``i32`` |
1794 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1795 | ``float (i16, i32 *) *`` | :ref:`Pointer <t_pointer>` to a function that takes an ``i16`` and a :ref:`pointer <t_pointer>` to ``i32``, returning ``float``. |
1796 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1797 | ``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. |
1798 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1799 | ``{i32, i32} (i32)`` | A function taking an ``i32``, returning a :ref:`structure <t_struct>` containing two ``i32`` values |
1800 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1807 The :ref:`first class <t_firstclass>` types are perhaps the most important.
1808 Values of these types are the only ones which can be produced by
1816 These are the types that are valid in registers from CodeGen's perspective.
1825 The integer type is a very simple type that simply specifies an
1826 arbitrary bit width for the integer type desired. Any bit width from 1
1827 bit to 2\ :sup:`23`\ -1 (about 8 million) can be specified.
1835 The number of bits the integer will occupy is specified by the ``N``
1841 +----------------+------------------------------------------------+
1842 | ``i1`` | a single-bit integer. |
1843 +----------------+------------------------------------------------+
1844 | ``i32`` | a 32-bit integer. |
1845 +----------------+------------------------------------------------+
1846 | ``i1942652`` | a really big integer of over 1 million bits. |
1847 +----------------+------------------------------------------------+
1851 Floating Point Types
1852 """"""""""""""""""""
1861 - 16-bit floating point value
1864 - 32-bit floating point value
1867 - 64-bit floating point value
1870 - 128-bit floating point value (112-bit mantissa)
1873 - 80-bit floating point value (X87)
1876 - 128-bit floating point value (two 64-bits)
1883 The x86_mmx type represents a value held in an MMX register on an x86
1884 machine. The operations allowed on it are quite limited: parameters and
1885 return values, load and store, and bitcast. User-specified MMX
1886 instructions are represented as intrinsic or asm calls with arguments
1887 and/or results of this type. There are no arrays, vectors or constants
1904 The pointer type is used to specify memory locations. Pointers are
1905 commonly used to reference objects in memory.
1907 Pointer types may have an optional address space attribute defining the
1908 numbered address space where the pointed-to object resides. The default
1909 address space is number zero. The semantics of non-zero address spaces
1910 are target-specific.
1912 Note that LLVM does not permit pointers to void (``void*``) nor does it
1913 permit pointers to labels (``label*``). Use ``i8*`` instead.
1923 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1924 | ``[4 x i32]*`` | A :ref:`pointer <t_pointer>` to :ref:`array <t_array>` of four ``i32`` values. |
1925 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1926 | ``i32 (i32*) *`` | A :ref:`pointer <t_pointer>` to a :ref:`function <t_function>` that takes an ``i32*``, returning an ``i32``. |
1927 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1928 | ``i32 addrspace(5)*`` | A :ref:`pointer <t_pointer>` to an ``i32`` value that resides in address space #5. |
1929 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1938 A vector type is a simple derived type that represents a vector of
1939 elements. Vector types are used when multiple primitive data are
1940 operated in parallel using a single instruction (SIMD). A vector type
1941 requires a size (number of elements) and an underlying primitive data
1942 type. Vector types are considered :ref:`first class <t_firstclass>`.
1948 < <# elements> x <elementtype> >
1950 The number of elements is a constant integer value larger than 0;
1951 elementtype may be any integer or floating point type, or a pointer to
1952 these types. Vectors of size zero are not allowed.
1956 +-------------------+--------------------------------------------------+
1957 | ``<4 x i32>`` | Vector of 4 32-bit integer values. |
1958 +-------------------+--------------------------------------------------+
1959 | ``<8 x float>`` | Vector of 8 32-bit floating-point values. |
1960 +-------------------+--------------------------------------------------+
1961 | ``<2 x i64>`` | Vector of 2 64-bit integer values. |
1962 +-------------------+--------------------------------------------------+
1963 | ``<4 x i64*>`` | Vector of 4 pointers to 64-bit integer values. |
1964 +-------------------+--------------------------------------------------+
1973 The label type represents code labels.
1988 The metadata type represents embedded metadata. No derived types may be
1989 created from metadata except for :ref:`function <t_function>` arguments.
2002 Aggregate Types are a subset of derived types that can contain multiple
2003 member types. :ref:`Arrays <t_array>` and :ref:`structs <t_struct>` are
2004 aggregate types. :ref:`Vectors <t_vector>` are not considered to be
2014 The array type is a very simple derived type that arranges elements
2015 sequentially in memory. The array type requires a size (number of
2016 elements) and an underlying data type.
2022 [<# elements> x <elementtype>]
2024 The number of elements is a constant integer value; ``elementtype`` may
2025 be any type with a size.
2029 +------------------+--------------------------------------+
2030 | ``[40 x i32]`` | Array of 40 32-bit integer values. |
2031 +------------------+--------------------------------------+
2032 | ``[41 x i32]`` | Array of 41 32-bit integer values. |
2033 +------------------+--------------------------------------+
2034 | ``[4 x i8]`` | Array of 4 8-bit integer values. |
2035 +------------------+--------------------------------------+
2037 Here are some examples of multidimensional arrays:
2039 +-----------------------------+----------------------------------------------------------+
2040 | ``[3 x [4 x i32]]`` | 3x4 array of 32-bit integer values. |
2041 +-----------------------------+----------------------------------------------------------+
2042 | ``[12 x [10 x float]]`` | 12x10 array of single precision floating point values. |
2043 +-----------------------------+----------------------------------------------------------+
2044 | ``[2 x [3 x [4 x i16]]]`` | 2x3x4 array of 16-bit integer values. |
2045 +-----------------------------+----------------------------------------------------------+
2047 There is no restriction on indexing beyond the end of the array implied
2048 by a static type (though there are restrictions on indexing beyond the
2049 bounds of an allocated object in some cases). This means that
2050 single-dimension 'variable sized array' addressing can be implemented in
2051 LLVM with a zero length array type. An implementation of 'pascal style
2052 arrays' in LLVM could use the type "``{ i32, [0 x float]}``", for
2062 The structure type is used to represent a collection of data members
2063 together in memory. The elements of a structure may be any type that has
2066 Structures in memory are accessed using '``load``' and '``store``' by
2067 getting a pointer to a field with the '``getelementptr``' instruction.
2068 Structures in registers are accessed using the '``extractvalue``' and
2069 '``insertvalue``' instructions.
2071 Structures may optionally be "packed" structures, which indicate that
2072 the alignment of the struct is one byte, and that there is no padding
2073 between the elements. In non-packed structs, padding between field types
2074 is inserted as defined by the DataLayout string in the module, which is
2075 required to match what the underlying code generator expects.
2077 Structures can either be "literal" or "identified". A literal structure
2078 is defined inline with other types (e.g. ``{i32, i32}*``) whereas
2079 identified types are always defined at the top level with a name.
2080 Literal types are uniqued by their contents and can never be recursive
2081 or opaque since there is no way to write one. Identified types can be
2082 recursive, can be opaqued, and are never uniqued.
2088 %T1 = type { <type list> } ; Identified normal struct type
2089 %T2 = type <{ <type list> }> ; Identified packed struct type
2093 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2094 | ``{ i32, i32, i32 }`` | A triple of three ``i32`` values |
2095 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2096 | ``{ 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``. |
2097 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2098 | ``<{ i8, i32 }>`` | A packed struct known to be 5 bytes in size. |
2099 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2103 Opaque Structure Types
2104 """"""""""""""""""""""
2108 Opaque structure types are used to represent named structure types that
2109 do not have a body specified. This corresponds (for example) to the C
2110 notion of a forward declared structure.
2121 +--------------+-------------------+
2122 | ``opaque`` | An opaque type. |
2123 +--------------+-------------------+
2130 LLVM has several different basic types of constants. This section
2131 describes them all and their syntax.
2136 **Boolean constants**
2137 The two strings '``true``' and '``false``' are both valid constants
2139 **Integer constants**
2140 Standard integers (such as '4') are constants of the
2141 :ref:`integer <t_integer>` type. Negative numbers may be used with
2143 **Floating point constants**
2144 Floating point constants use standard decimal notation (e.g.
2145 123.421), exponential notation (e.g. 1.23421e+2), or a more precise
2146 hexadecimal notation (see below). The assembler requires the exact
2147 decimal value of a floating-point constant. For example, the
2148 assembler accepts 1.25 but rejects 1.3 because 1.3 is a repeating
2149 decimal in binary. Floating point constants must have a :ref:`floating
2150 point <t_floating>` type.
2151 **Null pointer constants**
2152 The identifier '``null``' is recognized as a null pointer constant
2153 and must be of :ref:`pointer type <t_pointer>`.
2155 The one non-intuitive notation for constants is the hexadecimal form of
2156 floating point constants. For example, the form
2157 '``double 0x432ff973cafa8000``' is equivalent to (but harder to read
2158 than) '``double 4.5e+15``'. The only time hexadecimal floating point
2159 constants are required (and the only time that they are generated by the
2160 disassembler) is when a floating point constant must be emitted but it
2161 cannot be represented as a decimal floating point number in a reasonable
2162 number of digits. For example, NaN's, infinities, and other special
2163 values are represented in their IEEE hexadecimal format so that assembly
2164 and disassembly do not cause any bits to change in the constants.
2166 When using the hexadecimal form, constants of types half, float, and
2167 double are represented using the 16-digit form shown above (which
2168 matches the IEEE754 representation for double); half and float values
2169 must, however, be exactly representable as IEEE 754 half and single
2170 precision, respectively. Hexadecimal format is always used for long
2171 double, and there are three forms of long double. The 80-bit format used
2172 by x86 is represented as ``0xK`` followed by 20 hexadecimal digits. The
2173 128-bit format used by PowerPC (two adjacent doubles) is represented by
2174 ``0xM`` followed by 32 hexadecimal digits. The IEEE 128-bit format is
2175 represented by ``0xL`` followed by 32 hexadecimal digits. Long doubles
2176 will only work if they match the long double format on your target.
2177 The IEEE 16-bit format (half precision) is represented by ``0xH``
2178 followed by 4 hexadecimal digits. All hexadecimal formats are big-endian
2179 (sign bit at the left).
2181 There are no constants of type x86_mmx.
2183 .. _complexconstants:
2188 Complex constants are a (potentially recursive) combination of simple
2189 constants and smaller complex constants.
2191 **Structure constants**
2192 Structure constants are represented with notation similar to
2193 structure type definitions (a comma separated list of elements,
2194 surrounded by braces (``{}``)). For example:
2195 "``{ i32 4, float 17.0, i32* @G }``", where "``@G``" is declared as
2196 "``@G = external global i32``". Structure constants must have
2197 :ref:`structure type <t_struct>`, and the number and types of elements
2198 must match those specified by the type.
2200 Array constants are represented with notation similar to array type
2201 definitions (a comma separated list of elements, surrounded by
2202 square brackets (``[]``)). For example:
2203 "``[ i32 42, i32 11, i32 74 ]``". Array constants must have
2204 :ref:`array type <t_array>`, and the number and types of elements must
2205 match those specified by the type.
2206 **Vector constants**
2207 Vector constants are represented with notation similar to vector
2208 type definitions (a comma separated list of elements, surrounded by
2209 less-than/greater-than's (``<>``)). For example:
2210 "``< i32 42, i32 11, i32 74, i32 100 >``". Vector constants
2211 must have :ref:`vector type <t_vector>`, and the number and types of
2212 elements must match those specified by the type.
2213 **Zero initialization**
2214 The string '``zeroinitializer``' can be used to zero initialize a
2215 value to zero of *any* type, including scalar and
2216 :ref:`aggregate <t_aggregate>` types. This is often used to avoid
2217 having to print large zero initializers (e.g. for large arrays) and
2218 is always exactly equivalent to using explicit zero initializers.
2220 A metadata node is a structure-like constant with :ref:`metadata
2221 type <t_metadata>`. For example:
2222 "``metadata !{ i32 0, metadata !"test" }``". Unlike other
2223 constants that are meant to be interpreted as part of the
2224 instruction stream, metadata is a place to attach additional
2225 information such as debug info.
2227 Global Variable and Function Addresses
2228 --------------------------------------
2230 The addresses of :ref:`global variables <globalvars>` and
2231 :ref:`functions <functionstructure>` are always implicitly valid
2232 (link-time) constants. These constants are explicitly referenced when
2233 the :ref:`identifier for the global <identifiers>` is used and always have
2234 :ref:`pointer <t_pointer>` type. For example, the following is a legal LLVM
2237 .. code-block:: llvm
2241 @Z = global [2 x i32*] [ i32* @X, i32* @Y ]
2248 The string '``undef``' can be used anywhere a constant is expected, and
2249 indicates that the user of the value may receive an unspecified
2250 bit-pattern. Undefined values may be of any type (other than '``label``'
2251 or '``void``') and be used anywhere a constant is permitted.
2253 Undefined values are useful because they indicate to the compiler that
2254 the program is well defined no matter what value is used. This gives the
2255 compiler more freedom to optimize. Here are some examples of
2256 (potentially surprising) transformations that are valid (in pseudo IR):
2258 .. code-block:: llvm
2268 This is safe because all of the output bits are affected by the undef
2269 bits. Any output bit can have a zero or one depending on the input bits.
2271 .. code-block:: llvm
2282 These logical operations have bits that are not always affected by the
2283 input. For example, if ``%X`` has a zero bit, then the output of the
2284 '``and``' operation will always be a zero for that bit, no matter what
2285 the corresponding bit from the '``undef``' is. As such, it is unsafe to
2286 optimize or assume that the result of the '``and``' is '``undef``'.
2287 However, it is safe to assume that all bits of the '``undef``' could be
2288 0, and optimize the '``and``' to 0. Likewise, it is safe to assume that
2289 all the bits of the '``undef``' operand to the '``or``' could be set,
2290 allowing the '``or``' to be folded to -1.
2292 .. code-block:: llvm
2294 %A = select undef, %X, %Y
2295 %B = select undef, 42, %Y
2296 %C = select %X, %Y, undef
2306 This set of examples shows that undefined '``select``' (and conditional
2307 branch) conditions can go *either way*, but they have to come from one
2308 of the two operands. In the ``%A`` example, if ``%X`` and ``%Y`` were
2309 both known to have a clear low bit, then ``%A`` would have to have a
2310 cleared low bit. However, in the ``%C`` example, the optimizer is
2311 allowed to assume that the '``undef``' operand could be the same as
2312 ``%Y``, allowing the whole '``select``' to be eliminated.
2314 .. code-block:: llvm
2316 %A = xor undef, undef
2333 This example points out that two '``undef``' operands are not
2334 necessarily the same. This can be surprising to people (and also matches
2335 C semantics) where they assume that "``X^X``" is always zero, even if
2336 ``X`` is undefined. This isn't true for a number of reasons, but the
2337 short answer is that an '``undef``' "variable" can arbitrarily change
2338 its value over its "live range". This is true because the variable
2339 doesn't actually *have a live range*. Instead, the value is logically
2340 read from arbitrary registers that happen to be around when needed, so
2341 the value is not necessarily consistent over time. In fact, ``%A`` and
2342 ``%C`` need to have the same semantics or the core LLVM "replace all
2343 uses with" concept would not hold.
2345 .. code-block:: llvm
2353 These examples show the crucial difference between an *undefined value*
2354 and *undefined behavior*. An undefined value (like '``undef``') is
2355 allowed to have an arbitrary bit-pattern. This means that the ``%A``
2356 operation can be constant folded to '``undef``', because the '``undef``'
2357 could be an SNaN, and ``fdiv`` is not (currently) defined on SNaN's.
2358 However, in the second example, we can make a more aggressive
2359 assumption: because the ``undef`` is allowed to be an arbitrary value,
2360 we are allowed to assume that it could be zero. Since a divide by zero
2361 has *undefined behavior*, we are allowed to assume that the operation
2362 does not execute at all. This allows us to delete the divide and all
2363 code after it. Because the undefined operation "can't happen", the
2364 optimizer can assume that it occurs in dead code.
2366 .. code-block:: llvm
2368 a: store undef -> %X
2369 b: store %X -> undef
2374 These examples reiterate the ``fdiv`` example: a store *of* an undefined
2375 value can be assumed to not have any effect; we can assume that the
2376 value is overwritten with bits that happen to match what was already
2377 there. However, a store *to* an undefined location could clobber
2378 arbitrary memory, therefore, it has undefined behavior.
2385 Poison values are similar to :ref:`undef values <undefvalues>`, however
2386 they also represent the fact that an instruction or constant expression
2387 which cannot evoke side effects has nevertheless detected a condition
2388 which results in undefined behavior.
2390 There is currently no way of representing a poison value in the IR; they
2391 only exist when produced by operations such as :ref:`add <i_add>` with
2394 Poison value behavior is defined in terms of value *dependence*:
2396 - Values other than :ref:`phi <i_phi>` nodes depend on their operands.
2397 - :ref:`Phi <i_phi>` nodes depend on the operand corresponding to
2398 their dynamic predecessor basic block.
2399 - Function arguments depend on the corresponding actual argument values
2400 in the dynamic callers of their functions.
2401 - :ref:`Call <i_call>` instructions depend on the :ref:`ret <i_ret>`
2402 instructions that dynamically transfer control back to them.
2403 - :ref:`Invoke <i_invoke>` instructions depend on the
2404 :ref:`ret <i_ret>`, :ref:`resume <i_resume>`, or exception-throwing
2405 call instructions that dynamically transfer control back to them.
2406 - Non-volatile loads and stores depend on the most recent stores to all
2407 of the referenced memory addresses, following the order in the IR
2408 (including loads and stores implied by intrinsics such as
2409 :ref:`@llvm.memcpy <int_memcpy>`.)
2410 - An instruction with externally visible side effects depends on the
2411 most recent preceding instruction with externally visible side
2412 effects, following the order in the IR. (This includes :ref:`volatile
2413 operations <volatile>`.)
2414 - An instruction *control-depends* on a :ref:`terminator
2415 instruction <terminators>` if the terminator instruction has
2416 multiple successors and the instruction is always executed when
2417 control transfers to one of the successors, and may not be executed
2418 when control is transferred to another.
2419 - Additionally, an instruction also *control-depends* on a terminator
2420 instruction if the set of instructions it otherwise depends on would
2421 be different if the terminator had transferred control to a different
2423 - Dependence is transitive.
2425 Poison Values have the same behavior as :ref:`undef values <undefvalues>`,
2426 with the additional affect that any instruction which has a *dependence*
2427 on a poison value has undefined behavior.
2429 Here are some examples:
2431 .. code-block:: llvm
2434 %poison = sub nuw i32 0, 1 ; Results in a poison value.
2435 %still_poison = and i32 %poison, 0 ; 0, but also poison.
2436 %poison_yet_again = getelementptr i32* @h, i32 %still_poison
2437 store i32 0, i32* %poison_yet_again ; memory at @h[0] is poisoned
2439 store i32 %poison, i32* @g ; Poison value stored to memory.
2440 %poison2 = load i32* @g ; Poison value loaded back from memory.
2442 store volatile i32 %poison, i32* @g ; External observation; undefined behavior.
2444 %narrowaddr = bitcast i32* @g to i16*
2445 %wideaddr = bitcast i32* @g to i64*
2446 %poison3 = load i16* %narrowaddr ; Returns a poison value.
2447 %poison4 = load i64* %wideaddr ; Returns a poison value.
2449 %cmp = icmp slt i32 %poison, 0 ; Returns a poison value.
2450 br i1 %cmp, label %true, label %end ; Branch to either destination.
2453 store volatile i32 0, i32* @g ; This is control-dependent on %cmp, so
2454 ; it has undefined behavior.
2458 %p = phi i32 [ 0, %entry ], [ 1, %true ]
2459 ; Both edges into this PHI are
2460 ; control-dependent on %cmp, so this
2461 ; always results in a poison value.
2463 store volatile i32 0, i32* @g ; This would depend on the store in %true
2464 ; if %cmp is true, or the store in %entry
2465 ; otherwise, so this is undefined behavior.
2467 br i1 %cmp, label %second_true, label %second_end
2468 ; The same branch again, but this time the
2469 ; true block doesn't have side effects.
2476 store volatile i32 0, i32* @g ; This time, the instruction always depends
2477 ; on the store in %end. Also, it is
2478 ; control-equivalent to %end, so this is
2479 ; well-defined (ignoring earlier undefined
2480 ; behavior in this example).
2484 Addresses of Basic Blocks
2485 -------------------------
2487 ``blockaddress(@function, %block)``
2489 The '``blockaddress``' constant computes the address of the specified
2490 basic block in the specified function, and always has an ``i8*`` type.
2491 Taking the address of the entry block is illegal.
2493 This value only has defined behavior when used as an operand to the
2494 ':ref:`indirectbr <i_indirectbr>`' instruction, or for comparisons
2495 against null. Pointer equality tests between labels addresses results in
2496 undefined behavior --- though, again, comparison against null is ok, and
2497 no label is equal to the null pointer. This may be passed around as an
2498 opaque pointer sized value as long as the bits are not inspected. This
2499 allows ``ptrtoint`` and arithmetic to be performed on these values so
2500 long as the original value is reconstituted before the ``indirectbr``
2503 Finally, some targets may provide defined semantics when using the value
2504 as the operand to an inline assembly, but that is target specific.
2508 Constant Expressions
2509 --------------------
2511 Constant expressions are used to allow expressions involving other
2512 constants to be used as constants. Constant expressions may be of any
2513 :ref:`first class <t_firstclass>` type and may involve any LLVM operation
2514 that does not have side effects (e.g. load and call are not supported).
2515 The following is the syntax for constant expressions:
2517 ``trunc (CST to TYPE)``
2518 Truncate a constant to another type. The bit size of CST must be
2519 larger than the bit size of TYPE. Both types must be integers.
2520 ``zext (CST to TYPE)``
2521 Zero extend a constant to another type. The bit size of CST must be
2522 smaller than the bit size of TYPE. Both types must be integers.
2523 ``sext (CST to TYPE)``
2524 Sign extend a constant to another type. The bit size of CST must be
2525 smaller than the bit size of TYPE. Both types must be integers.
2526 ``fptrunc (CST to TYPE)``
2527 Truncate a floating point constant to another floating point type.
2528 The size of CST must be larger than the size of TYPE. Both types
2529 must be floating point.
2530 ``fpext (CST to TYPE)``
2531 Floating point extend a constant to another type. The size of CST
2532 must be smaller or equal to the size of TYPE. Both types must be
2534 ``fptoui (CST to TYPE)``
2535 Convert a floating point constant to the corresponding unsigned
2536 integer constant. TYPE must be a scalar or vector integer type. CST
2537 must be of scalar or vector floating point type. Both CST and TYPE
2538 must be scalars, or vectors of the same number of elements. If the
2539 value won't fit in the integer type, the results are undefined.
2540 ``fptosi (CST to TYPE)``
2541 Convert a floating point constant to the corresponding signed
2542 integer constant. TYPE must be a scalar or vector integer type. CST
2543 must be of scalar or vector floating point type. Both CST and TYPE
2544 must be scalars, or vectors of the same number of elements. If the
2545 value won't fit in the integer type, the results are undefined.
2546 ``uitofp (CST to TYPE)``
2547 Convert an unsigned integer constant to the corresponding floating
2548 point constant. TYPE must be a scalar or vector floating point type.
2549 CST must be of scalar or vector integer type. Both CST and TYPE must
2550 be scalars, or vectors of the same number of elements. If the value
2551 won't fit in the floating point type, the results are undefined.
2552 ``sitofp (CST to TYPE)``
2553 Convert a signed integer constant to the corresponding floating
2554 point constant. TYPE must be a scalar or vector floating point type.
2555 CST must be of scalar or vector integer type. Both CST and TYPE must
2556 be scalars, or vectors of the same number of elements. If the value
2557 won't fit in the floating point type, the results are undefined.
2558 ``ptrtoint (CST to TYPE)``
2559 Convert a pointer typed constant to the corresponding integer
2560 constant. ``TYPE`` must be an integer type. ``CST`` must be of
2561 pointer type. The ``CST`` value is zero extended, truncated, or
2562 unchanged to make it fit in ``TYPE``.
2563 ``inttoptr (CST to TYPE)``
2564 Convert an integer constant to a pointer constant. TYPE must be a
2565 pointer type. CST must be of integer type. The CST value is zero
2566 extended, truncated, or unchanged to make it fit in a pointer size.
2567 This one is *really* dangerous!
2568 ``bitcast (CST to TYPE)``
2569 Convert a constant, CST, to another TYPE. The constraints of the
2570 operands are the same as those for the :ref:`bitcast
2571 instruction <i_bitcast>`.
2572 ``addrspacecast (CST to TYPE)``
2573 Convert a constant pointer or constant vector of pointer, CST, to another
2574 TYPE in a different address space. The constraints of the operands are the
2575 same as those for the :ref:`addrspacecast instruction <i_addrspacecast>`.
2576 ``getelementptr (CSTPTR, IDX0, IDX1, ...)``, ``getelementptr inbounds (CSTPTR, IDX0, IDX1, ...)``
2577 Perform the :ref:`getelementptr operation <i_getelementptr>` on
2578 constants. As with the :ref:`getelementptr <i_getelementptr>`
2579 instruction, the index list may have zero or more indexes, which are
2580 required to make sense for the type of "CSTPTR".
2581 ``select (COND, VAL1, VAL2)``
2582 Perform the :ref:`select operation <i_select>` on constants.
2583 ``icmp COND (VAL1, VAL2)``
2584 Performs the :ref:`icmp operation <i_icmp>` on constants.
2585 ``fcmp COND (VAL1, VAL2)``
2586 Performs the :ref:`fcmp operation <i_fcmp>` on constants.
2587 ``extractelement (VAL, IDX)``
2588 Perform the :ref:`extractelement operation <i_extractelement>` on
2590 ``insertelement (VAL, ELT, IDX)``
2591 Perform the :ref:`insertelement operation <i_insertelement>` on
2593 ``shufflevector (VEC1, VEC2, IDXMASK)``
2594 Perform the :ref:`shufflevector operation <i_shufflevector>` on
2596 ``extractvalue (VAL, IDX0, IDX1, ...)``
2597 Perform the :ref:`extractvalue operation <i_extractvalue>` on
2598 constants. The index list is interpreted in a similar manner as
2599 indices in a ':ref:`getelementptr <i_getelementptr>`' operation. At
2600 least one index value must be specified.
2601 ``insertvalue (VAL, ELT, IDX0, IDX1, ...)``
2602 Perform the :ref:`insertvalue operation <i_insertvalue>` on constants.
2603 The index list is interpreted in a similar manner as indices in a
2604 ':ref:`getelementptr <i_getelementptr>`' operation. At least one index
2605 value must be specified.
2606 ``OPCODE (LHS, RHS)``
2607 Perform the specified operation of the LHS and RHS constants. OPCODE
2608 may be any of the :ref:`binary <binaryops>` or :ref:`bitwise
2609 binary <bitwiseops>` operations. The constraints on operands are
2610 the same as those for the corresponding instruction (e.g. no bitwise
2611 operations on floating point values are allowed).
2618 Inline Assembler Expressions
2619 ----------------------------
2621 LLVM supports inline assembler expressions (as opposed to :ref:`Module-Level
2622 Inline Assembly <moduleasm>`) through the use of a special value. This
2623 value represents the inline assembler as a string (containing the
2624 instructions to emit), a list of operand constraints (stored as a
2625 string), a flag that indicates whether or not the inline asm expression
2626 has side effects, and a flag indicating whether the function containing
2627 the asm needs to align its stack conservatively. An example inline
2628 assembler expression is:
2630 .. code-block:: llvm
2632 i32 (i32) asm "bswap $0", "=r,r"
2634 Inline assembler expressions may **only** be used as the callee operand
2635 of a :ref:`call <i_call>` or an :ref:`invoke <i_invoke>` instruction.
2636 Thus, typically we have:
2638 .. code-block:: llvm
2640 %X = call i32 asm "bswap $0", "=r,r"(i32 %Y)
2642 Inline asms with side effects not visible in the constraint list must be
2643 marked as having side effects. This is done through the use of the
2644 '``sideeffect``' keyword, like so:
2646 .. code-block:: llvm
2648 call void asm sideeffect "eieio", ""()
2650 In some cases inline asms will contain code that will not work unless
2651 the stack is aligned in some way, such as calls or SSE instructions on
2652 x86, yet will not contain code that does that alignment within the asm.
2653 The compiler should make conservative assumptions about what the asm
2654 might contain and should generate its usual stack alignment code in the
2655 prologue if the '``alignstack``' keyword is present:
2657 .. code-block:: llvm
2659 call void asm alignstack "eieio", ""()
2661 Inline asms also support using non-standard assembly dialects. The
2662 assumed dialect is ATT. When the '``inteldialect``' keyword is present,
2663 the inline asm is using the Intel dialect. Currently, ATT and Intel are
2664 the only supported dialects. An example is:
2666 .. code-block:: llvm
2668 call void asm inteldialect "eieio", ""()
2670 If multiple keywords appear the '``sideeffect``' keyword must come
2671 first, the '``alignstack``' keyword second and the '``inteldialect``'
2677 The call instructions that wrap inline asm nodes may have a
2678 "``!srcloc``" MDNode attached to it that contains a list of constant
2679 integers. If present, the code generator will use the integer as the
2680 location cookie value when report errors through the ``LLVMContext``
2681 error reporting mechanisms. This allows a front-end to correlate backend
2682 errors that occur with inline asm back to the source code that produced
2685 .. code-block:: llvm
2687 call void asm sideeffect "something bad", ""(), !srcloc !42
2689 !42 = !{ i32 1234567 }
2691 It is up to the front-end to make sense of the magic numbers it places
2692 in the IR. If the MDNode contains multiple constants, the code generator
2693 will use the one that corresponds to the line of the asm that the error
2698 Metadata Nodes and Metadata Strings
2699 -----------------------------------
2701 LLVM IR allows metadata to be attached to instructions in the program
2702 that can convey extra information about the code to the optimizers and
2703 code generator. One example application of metadata is source-level
2704 debug information. There are two metadata primitives: strings and nodes.
2705 All metadata has the ``metadata`` type and is identified in syntax by a
2706 preceding exclamation point ('``!``').
2708 A metadata string is a string surrounded by double quotes. It can
2709 contain any character by escaping non-printable characters with
2710 "``\xx``" where "``xx``" is the two digit hex code. For example:
2713 Metadata nodes are represented with notation similar to structure
2714 constants (a comma separated list of elements, surrounded by braces and
2715 preceded by an exclamation point). Metadata nodes can have any values as
2716 their operand. For example:
2718 .. code-block:: llvm
2720 !{ metadata !"test\00", i32 10}
2722 A :ref:`named metadata <namedmetadatastructure>` is a collection of
2723 metadata nodes, which can be looked up in the module symbol table. For
2726 .. code-block:: llvm
2728 !foo = metadata !{!4, !3}
2730 Metadata can be used as function arguments. Here ``llvm.dbg.value``
2731 function is using two metadata arguments:
2733 .. code-block:: llvm
2735 call void @llvm.dbg.value(metadata !24, i64 0, metadata !25)
2737 Metadata can be attached with an instruction. Here metadata ``!21`` is
2738 attached to the ``add`` instruction using the ``!dbg`` identifier:
2740 .. code-block:: llvm
2742 %indvar.next = add i64 %indvar, 1, !dbg !21
2744 More information about specific metadata nodes recognized by the
2745 optimizers and code generator is found below.
2750 In LLVM IR, memory does not have types, so LLVM's own type system is not
2751 suitable for doing TBAA. Instead, metadata is added to the IR to
2752 describe a type system of a higher level language. This can be used to
2753 implement typical C/C++ TBAA, but it can also be used to implement
2754 custom alias analysis behavior for other languages.
2756 The current metadata format is very simple. TBAA metadata nodes have up
2757 to three fields, e.g.:
2759 .. code-block:: llvm
2761 !0 = metadata !{ metadata !"an example type tree" }
2762 !1 = metadata !{ metadata !"int", metadata !0 }
2763 !2 = metadata !{ metadata !"float", metadata !0 }
2764 !3 = metadata !{ metadata !"const float", metadata !2, i64 1 }
2766 The first field is an identity field. It can be any value, usually a
2767 metadata string, which uniquely identifies the type. The most important
2768 name in the tree is the name of the root node. Two trees with different
2769 root node names are entirely disjoint, even if they have leaves with
2772 The second field identifies the type's parent node in the tree, or is
2773 null or omitted for a root node. A type is considered to alias all of
2774 its descendants and all of its ancestors in the tree. Also, a type is
2775 considered to alias all types in other trees, so that bitcode produced
2776 from multiple front-ends is handled conservatively.
2778 If the third field is present, it's an integer which if equal to 1
2779 indicates that the type is "constant" (meaning
2780 ``pointsToConstantMemory`` should return true; see `other useful
2781 AliasAnalysis methods <AliasAnalysis.html#OtherItfs>`_).
2783 '``tbaa.struct``' Metadata
2784 ^^^^^^^^^^^^^^^^^^^^^^^^^^
2786 The :ref:`llvm.memcpy <int_memcpy>` is often used to implement
2787 aggregate assignment operations in C and similar languages, however it
2788 is defined to copy a contiguous region of memory, which is more than
2789 strictly necessary for aggregate types which contain holes due to
2790 padding. Also, it doesn't contain any TBAA information about the fields
2793 ``!tbaa.struct`` metadata can describe which memory subregions in a
2794 memcpy are padding and what the TBAA tags of the struct are.
2796 The current metadata format is very simple. ``!tbaa.struct`` metadata
2797 nodes are a list of operands which are in conceptual groups of three.
2798 For each group of three, the first operand gives the byte offset of a
2799 field in bytes, the second gives its size in bytes, and the third gives
2802 .. code-block:: llvm
2804 !4 = metadata !{ i64 0, i64 4, metadata !1, i64 8, i64 4, metadata !2 }
2806 This describes a struct with two fields. The first is at offset 0 bytes
2807 with size 4 bytes, and has tbaa tag !1. The second is at offset 8 bytes
2808 and has size 4 bytes and has tbaa tag !2.
2810 Note that the fields need not be contiguous. In this example, there is a
2811 4 byte gap between the two fields. This gap represents padding which
2812 does not carry useful data and need not be preserved.
2814 '``fpmath``' Metadata
2815 ^^^^^^^^^^^^^^^^^^^^^
2817 ``fpmath`` metadata may be attached to any instruction of floating point
2818 type. It can be used to express the maximum acceptable error in the
2819 result of that instruction, in ULPs, thus potentially allowing the
2820 compiler to use a more efficient but less accurate method of computing
2821 it. ULP is defined as follows:
2823 If ``x`` is a real number that lies between two finite consecutive
2824 floating-point numbers ``a`` and ``b``, without being equal to one
2825 of them, then ``ulp(x) = |b - a|``, otherwise ``ulp(x)`` is the
2826 distance between the two non-equal finite floating-point numbers
2827 nearest ``x``. Moreover, ``ulp(NaN)`` is ``NaN``.
2829 The metadata node shall consist of a single positive floating point
2830 number representing the maximum relative error, for example:
2832 .. code-block:: llvm
2834 !0 = metadata !{ float 2.5 } ; maximum acceptable inaccuracy is 2.5 ULPs
2836 '``range``' Metadata
2837 ^^^^^^^^^^^^^^^^^^^^
2839 ``range`` metadata may be attached only to ``load``, ``call`` and ``invoke`` of
2840 integer types. It expresses the possible ranges the loaded value or the value
2841 returned by the called function at this call site is in. The ranges are
2842 represented with a flattened list of integers. The loaded value or the value
2843 returned is known to be in the union of the ranges defined by each consecutive
2844 pair. Each pair has the following properties:
2846 - The type must match the type loaded by the instruction.
2847 - The pair ``a,b`` represents the range ``[a,b)``.
2848 - Both ``a`` and ``b`` are constants.
2849 - The range is allowed to wrap.
2850 - The range should not represent the full or empty set. That is,
2853 In addition, the pairs must be in signed order of the lower bound and
2854 they must be non-contiguous.
2858 .. code-block:: llvm
2860 %a = load i8* %x, align 1, !range !0 ; Can only be 0 or 1
2861 %b = load i8* %y, align 1, !range !1 ; Can only be 255 (-1), 0 or 1
2862 %c = call i8 @foo(), !range !2 ; Can only be 0, 1, 3, 4 or 5
2863 %d = invoke i8 @bar() to label %cont
2864 unwind label %lpad, !range !3 ; Can only be -2, -1, 3, 4 or 5
2866 !0 = metadata !{ i8 0, i8 2 }
2867 !1 = metadata !{ i8 255, i8 2 }
2868 !2 = metadata !{ i8 0, i8 2, i8 3, i8 6 }
2869 !3 = metadata !{ i8 -2, i8 0, i8 3, i8 6 }
2874 It is sometimes useful to attach information to loop constructs. Currently,
2875 loop metadata is implemented as metadata attached to the branch instruction
2876 in the loop latch block. This type of metadata refer to a metadata node that is
2877 guaranteed to be separate for each loop. The loop identifier metadata is
2878 specified with the name ``llvm.loop``.
2880 The loop identifier metadata is implemented using a metadata that refers to
2881 itself to avoid merging it with any other identifier metadata, e.g.,
2882 during module linkage or function inlining. That is, each loop should refer
2883 to their own identification metadata even if they reside in separate functions.
2884 The following example contains loop identifier metadata for two separate loop
2887 .. code-block:: llvm
2889 !0 = metadata !{ metadata !0 }
2890 !1 = metadata !{ metadata !1 }
2892 The loop identifier metadata can be used to specify additional per-loop
2893 metadata. Any operands after the first operand can be treated as user-defined
2894 metadata. For example the ``llvm.loop.vectorize.unroll`` metadata is understood
2895 by the loop vectorizer to indicate how many times to unroll the loop:
2897 .. code-block:: llvm
2899 br i1 %exitcond, label %._crit_edge, label %.lr.ph, !llvm.loop !0
2901 !0 = metadata !{ metadata !0, metadata !1 }
2902 !1 = metadata !{ metadata !"llvm.loop.vectorize.unroll", i32 2 }
2907 Metadata types used to annotate memory accesses with information helpful
2908 for optimizations are prefixed with ``llvm.mem``.
2910 '``llvm.mem.parallel_loop_access``' Metadata
2911 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2913 The ``llvm.mem.parallel_loop_access`` metadata refers to a loop identifier,
2914 or metadata containing a list of loop identifiers for nested loops.
2915 The metadata is attached to memory accessing instructions and denotes that
2916 no loop carried memory dependence exist between it and other instructions denoted
2917 with the same loop identifier.
2919 Precisely, given two instructions ``m1`` and ``m2`` that both have the
2920 ``llvm.mem.parallel_loop_access`` metadata, with ``L1`` and ``L2`` being the
2921 set of loops associated with that metadata, respectively, then there is no loop
2922 carried dependence between ``m1`` and ``m2`` for loops in both ``L1`` and
2925 As a special case, if all memory accessing instructions in a loop have
2926 ``llvm.mem.parallel_loop_access`` metadata that refers to that loop, then the
2927 loop has no loop carried memory dependences and is considered to be a parallel
2930 Note that if not all memory access instructions have such metadata referring to
2931 the loop, then the loop is considered not being trivially parallel. Additional
2932 memory dependence analysis is required to make that determination. As a fail
2933 safe mechanism, this causes loops that were originally parallel to be considered
2934 sequential (if optimization passes that are unaware of the parallel semantics
2935 insert new memory instructions into the loop body).
2937 Example of a loop that is considered parallel due to its correct use of
2938 both ``llvm.loop`` and ``llvm.mem.parallel_loop_access``
2939 metadata types that refer to the same loop identifier metadata.
2941 .. code-block:: llvm
2945 %val0 = load i32* %arrayidx, !llvm.mem.parallel_loop_access !0
2947 store i32 %val0, i32* %arrayidx1, !llvm.mem.parallel_loop_access !0
2949 br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !0
2953 !0 = metadata !{ metadata !0 }
2955 It is also possible to have nested parallel loops. In that case the
2956 memory accesses refer to a list of loop identifier metadata nodes instead of
2957 the loop identifier metadata node directly:
2959 .. code-block:: llvm
2963 %val1 = load i32* %arrayidx3, !llvm.mem.parallel_loop_access !2
2965 br label %inner.for.body
2969 %val0 = load i32* %arrayidx1, !llvm.mem.parallel_loop_access !0
2971 store i32 %val0, i32* %arrayidx2, !llvm.mem.parallel_loop_access !0
2973 br i1 %exitcond, label %inner.for.end, label %inner.for.body, !llvm.loop !1
2977 store i32 %val1, i32* %arrayidx4, !llvm.mem.parallel_loop_access !2
2979 br i1 %exitcond, label %outer.for.end, label %outer.for.body, !llvm.loop !2
2981 outer.for.end: ; preds = %for.body
2983 !0 = metadata !{ metadata !1, metadata !2 } ; a list of loop identifiers
2984 !1 = metadata !{ metadata !1 } ; an identifier for the inner loop
2985 !2 = metadata !{ metadata !2 } ; an identifier for the outer loop
2987 '``llvm.loop.vectorize``'
2988 ^^^^^^^^^^^^^^^^^^^^^^^^^
2990 Metadata prefixed with ``llvm.loop.vectorize`` is used to control per-loop
2991 vectorization parameters such as vectorization factor and unroll factor.
2993 ``llvm.loop.vectorize`` metadata should be used in conjunction with
2994 ``llvm.loop`` loop identification metadata.
2996 '``llvm.loop.vectorize.unroll``' Metadata
2997 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2999 This metadata instructs the loop vectorizer to unroll the specified
3000 loop exactly ``N`` times.
3002 The first operand is the string ``llvm.loop.vectorize.unroll`` and the second
3003 operand is an integer specifying the unroll factor. For example:
3005 .. code-block:: llvm
3007 !0 = metadata !{ metadata !"llvm.loop.vectorize.unroll", i32 4 }
3009 Note that setting ``llvm.loop.vectorize.unroll`` to 1 disables
3010 unrolling of the loop.
3012 If ``llvm.loop.vectorize.unroll`` is set to 0 then the amount of
3013 unrolling will be determined automatically.
3015 '``llvm.loop.vectorize.width``' Metadata
3016 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3018 This metadata sets the target width of the vectorizer to ``N``. Without
3019 this metadata, the vectorizer will choose a width automatically.
3020 Regardless of this metadata, the vectorizer will only vectorize loops if
3021 it believes it is valid to do so.
3023 The first operand is the string ``llvm.loop.vectorize.width`` and the
3024 second operand is an integer specifying the width. For example:
3026 .. code-block:: llvm
3028 !0 = metadata !{ metadata !"llvm.loop.vectorize.width", i32 4 }
3030 Note that setting ``llvm.loop.vectorize.width`` to 1 disables
3031 vectorization of the loop.
3033 If ``llvm.loop.vectorize.width`` is set to 0 then the width will be
3034 determined automatically.
3036 Module Flags Metadata
3037 =====================
3039 Information about the module as a whole is difficult to convey to LLVM's
3040 subsystems. The LLVM IR isn't sufficient to transmit this information.
3041 The ``llvm.module.flags`` named metadata exists in order to facilitate
3042 this. These flags are in the form of key / value pairs --- much like a
3043 dictionary --- making it easy for any subsystem who cares about a flag to
3046 The ``llvm.module.flags`` metadata contains a list of metadata triplets.
3047 Each triplet has the following form:
3049 - The first element is a *behavior* flag, which specifies the behavior
3050 when two (or more) modules are merged together, and it encounters two
3051 (or more) metadata with the same ID. The supported behaviors are
3053 - The second element is a metadata string that is a unique ID for the
3054 metadata. Each module may only have one flag entry for each unique ID (not
3055 including entries with the **Require** behavior).
3056 - The third element is the value of the flag.
3058 When two (or more) modules are merged together, the resulting
3059 ``llvm.module.flags`` metadata is the union of the modules' flags. That is, for
3060 each unique metadata ID string, there will be exactly one entry in the merged
3061 modules ``llvm.module.flags`` metadata table, and the value for that entry will
3062 be determined by the merge behavior flag, as described below. The only exception
3063 is that entries with the *Require* behavior are always preserved.
3065 The following behaviors are supported:
3076 Emits an error if two values disagree, otherwise the resulting value
3077 is that of the operands.
3081 Emits a warning if two values disagree. The result value will be the
3082 operand for the flag from the first module being linked.
3086 Adds a requirement that another module flag be present and have a
3087 specified value after linking is performed. The value must be a
3088 metadata pair, where the first element of the pair is the ID of the
3089 module flag to be restricted, and the second element of the pair is
3090 the value the module flag should be restricted to. This behavior can
3091 be used to restrict the allowable results (via triggering of an
3092 error) of linking IDs with the **Override** behavior.
3096 Uses the specified value, regardless of the behavior or value of the
3097 other module. If both modules specify **Override**, but the values
3098 differ, an error will be emitted.
3102 Appends the two values, which are required to be metadata nodes.
3106 Appends the two values, which are required to be metadata
3107 nodes. However, duplicate entries in the second list are dropped
3108 during the append operation.
3110 It is an error for a particular unique flag ID to have multiple behaviors,
3111 except in the case of **Require** (which adds restrictions on another metadata
3112 value) or **Override**.
3114 An example of module flags:
3116 .. code-block:: llvm
3118 !0 = metadata !{ i32 1, metadata !"foo", i32 1 }
3119 !1 = metadata !{ i32 4, metadata !"bar", i32 37 }
3120 !2 = metadata !{ i32 2, metadata !"qux", i32 42 }
3121 !3 = metadata !{ i32 3, metadata !"qux",
3123 metadata !"foo", i32 1
3126 !llvm.module.flags = !{ !0, !1, !2, !3 }
3128 - Metadata ``!0`` has the ID ``!"foo"`` and the value '1'. The behavior
3129 if two or more ``!"foo"`` flags are seen is to emit an error if their
3130 values are not equal.
3132 - Metadata ``!1`` has the ID ``!"bar"`` and the value '37'. The
3133 behavior if two or more ``!"bar"`` flags are seen is to use the value
3136 - Metadata ``!2`` has the ID ``!"qux"`` and the value '42'. The
3137 behavior if two or more ``!"qux"`` flags are seen is to emit a
3138 warning if their values are not equal.
3140 - Metadata ``!3`` has the ID ``!"qux"`` and the value:
3144 metadata !{ metadata !"foo", i32 1 }
3146 The behavior is to emit an error if the ``llvm.module.flags`` does not
3147 contain a flag with the ID ``!"foo"`` that has the value '1' after linking is
3150 Objective-C Garbage Collection Module Flags Metadata
3151 ----------------------------------------------------
3153 On the Mach-O platform, Objective-C stores metadata about garbage
3154 collection in a special section called "image info". The metadata
3155 consists of a version number and a bitmask specifying what types of
3156 garbage collection are supported (if any) by the file. If two or more
3157 modules are linked together their garbage collection metadata needs to
3158 be merged rather than appended together.
3160 The Objective-C garbage collection module flags metadata consists of the
3161 following key-value pairs:
3170 * - ``Objective-C Version``
3171 - **[Required]** --- The Objective-C ABI version. Valid values are 1 and 2.
3173 * - ``Objective-C Image Info Version``
3174 - **[Required]** --- The version of the image info section. Currently
3177 * - ``Objective-C Image Info Section``
3178 - **[Required]** --- The section to place the metadata. Valid values are
3179 ``"__OBJC, __image_info, regular"`` for Objective-C ABI version 1, and
3180 ``"__DATA,__objc_imageinfo, regular, no_dead_strip"`` for
3181 Objective-C ABI version 2.
3183 * - ``Objective-C Garbage Collection``
3184 - **[Required]** --- Specifies whether garbage collection is supported or
3185 not. Valid values are 0, for no garbage collection, and 2, for garbage
3186 collection supported.
3188 * - ``Objective-C GC Only``
3189 - **[Optional]** --- Specifies that only garbage collection is supported.
3190 If present, its value must be 6. This flag requires that the
3191 ``Objective-C Garbage Collection`` flag have the value 2.
3193 Some important flag interactions:
3195 - If a module with ``Objective-C Garbage Collection`` set to 0 is
3196 merged with a module with ``Objective-C Garbage Collection`` set to
3197 2, then the resulting module has the
3198 ``Objective-C Garbage Collection`` flag set to 0.
3199 - A module with ``Objective-C Garbage Collection`` set to 0 cannot be
3200 merged with a module with ``Objective-C GC Only`` set to 6.
3202 Automatic Linker Flags Module Flags Metadata
3203 --------------------------------------------
3205 Some targets support embedding flags to the linker inside individual object
3206 files. Typically this is used in conjunction with language extensions which
3207 allow source files to explicitly declare the libraries they depend on, and have
3208 these automatically be transmitted to the linker via object files.
3210 These flags are encoded in the IR using metadata in the module flags section,
3211 using the ``Linker Options`` key. The merge behavior for this flag is required
3212 to be ``AppendUnique``, and the value for the key is expected to be a metadata
3213 node which should be a list of other metadata nodes, each of which should be a
3214 list of metadata strings defining linker options.
3216 For example, the following metadata section specifies two separate sets of
3217 linker options, presumably to link against ``libz`` and the ``Cocoa``
3220 !0 = metadata !{ i32 6, metadata !"Linker Options",
3222 metadata !{ metadata !"-lz" },
3223 metadata !{ metadata !"-framework", metadata !"Cocoa" } } }
3224 !llvm.module.flags = !{ !0 }
3226 The metadata encoding as lists of lists of options, as opposed to a collapsed
3227 list of options, is chosen so that the IR encoding can use multiple option
3228 strings to specify e.g., a single library, while still having that specifier be
3229 preserved as an atomic element that can be recognized by a target specific
3230 assembly writer or object file emitter.
3232 Each individual option is required to be either a valid option for the target's
3233 linker, or an option that is reserved by the target specific assembly writer or
3234 object file emitter. No other aspect of these options is defined by the IR.
3236 C type width Module Flags Metadata
3237 ----------------------------------
3239 The ARM backend emits a section into each generated object file describing the
3240 options that it was compiled with (in a compiler-independent way) to prevent
3241 linking incompatible objects, and to allow automatic library selection. Some
3242 of these options are not visible at the IR level, namely wchar_t width and enum
3245 To pass this information to the backend, these options are encoded in module
3246 flags metadata, using the following key-value pairs:
3256 - * 0 --- sizeof(wchar_t) == 4
3257 * 1 --- sizeof(wchar_t) == 2
3260 - * 0 --- Enums are at least as large as an ``int``.
3261 * 1 --- Enums are stored in the smallest integer type which can
3262 represent all of its values.
3264 For example, the following metadata section specifies that the module was
3265 compiled with a ``wchar_t`` width of 4 bytes, and the underlying type of an
3266 enum is the smallest type which can represent all of its values::
3268 !llvm.module.flags = !{!0, !1}
3269 !0 = metadata !{i32 1, metadata !"short_wchar", i32 1}
3270 !1 = metadata !{i32 1, metadata !"short_enum", i32 0}
3272 .. _intrinsicglobalvariables:
3274 Intrinsic Global Variables
3275 ==========================
3277 LLVM has a number of "magic" global variables that contain data that
3278 affect code generation or other IR semantics. These are documented here.
3279 All globals of this sort should have a section specified as
3280 "``llvm.metadata``". This section and all globals that start with
3281 "``llvm.``" are reserved for use by LLVM.
3285 The '``llvm.used``' Global Variable
3286 -----------------------------------
3288 The ``@llvm.used`` global is an array which has
3289 :ref:`appending linkage <linkage_appending>`. This array contains a list of
3290 pointers to named global variables, functions and aliases which may optionally
3291 have a pointer cast formed of bitcast or getelementptr. For example, a legal
3294 .. code-block:: llvm
3299 @llvm.used = appending global [2 x i8*] [
3301 i8* bitcast (i32* @Y to i8*)
3302 ], section "llvm.metadata"
3304 If a symbol appears in the ``@llvm.used`` list, then the compiler, assembler,
3305 and linker are required to treat the symbol as if there is a reference to the
3306 symbol that it cannot see (which is why they have to be named). For example, if
3307 a variable has internal linkage and no references other than that from the
3308 ``@llvm.used`` list, it cannot be deleted. This is commonly used to represent
3309 references from inline asms and other things the compiler cannot "see", and
3310 corresponds to "``attribute((used))``" in GNU C.
3312 On some targets, the code generator must emit a directive to the
3313 assembler or object file to prevent the assembler and linker from
3314 molesting the symbol.
3316 .. _gv_llvmcompilerused:
3318 The '``llvm.compiler.used``' Global Variable
3319 --------------------------------------------
3321 The ``@llvm.compiler.used`` directive is the same as the ``@llvm.used``
3322 directive, except that it only prevents the compiler from touching the
3323 symbol. On targets that support it, this allows an intelligent linker to
3324 optimize references to the symbol without being impeded as it would be
3327 This is a rare construct that should only be used in rare circumstances,
3328 and should not be exposed to source languages.
3330 .. _gv_llvmglobalctors:
3332 The '``llvm.global_ctors``' Global Variable
3333 -------------------------------------------
3335 .. code-block:: llvm
3337 %0 = type { i32, void ()*, i8* }
3338 @llvm.global_ctors = appending global [1 x %0] [%0 { i32 65535, void ()* @ctor, i8* @data }]
3340 The ``@llvm.global_ctors`` array contains a list of constructor
3341 functions, priorities, and an optional associated global or function.
3342 The functions referenced by this array will be called in ascending order
3343 of priority (i.e. lowest first) when the module is loaded. The order of
3344 functions with the same priority is not defined.
3346 If the third field is present, non-null, and points to a global variable
3347 or function, the initializer function will only run if the associated
3348 data from the current module is not discarded.
3350 .. _llvmglobaldtors:
3352 The '``llvm.global_dtors``' Global Variable
3353 -------------------------------------------
3355 .. code-block:: llvm
3357 %0 = type { i32, void ()*, i8* }
3358 @llvm.global_dtors = appending global [1 x %0] [%0 { i32 65535, void ()* @dtor, i8* @data }]
3360 The ``@llvm.global_dtors`` array contains a list of destructor
3361 functions, priorities, and an optional associated global or function.
3362 The functions referenced by this array will be called in descending
3363 order of priority (i.e. highest first) when the module is unloaded. The
3364 order of functions with the same priority is not defined.
3366 If the third field is present, non-null, and points to a global variable
3367 or function, the destructor function will only run if the associated
3368 data from the current module is not discarded.
3370 Instruction Reference
3371 =====================
3373 The LLVM instruction set consists of several different classifications
3374 of instructions: :ref:`terminator instructions <terminators>`, :ref:`binary
3375 instructions <binaryops>`, :ref:`bitwise binary
3376 instructions <bitwiseops>`, :ref:`memory instructions <memoryops>`, and
3377 :ref:`other instructions <otherops>`.
3381 Terminator Instructions
3382 -----------------------
3384 As mentioned :ref:`previously <functionstructure>`, every basic block in a
3385 program ends with a "Terminator" instruction, which indicates which
3386 block should be executed after the current block is finished. These
3387 terminator instructions typically yield a '``void``' value: they produce
3388 control flow, not values (the one exception being the
3389 ':ref:`invoke <i_invoke>`' instruction).
3391 The terminator instructions are: ':ref:`ret <i_ret>`',
3392 ':ref:`br <i_br>`', ':ref:`switch <i_switch>`',
3393 ':ref:`indirectbr <i_indirectbr>`', ':ref:`invoke <i_invoke>`',
3394 ':ref:`resume <i_resume>`', and ':ref:`unreachable <i_unreachable>`'.
3398 '``ret``' Instruction
3399 ^^^^^^^^^^^^^^^^^^^^^
3406 ret <type> <value> ; Return a value from a non-void function
3407 ret void ; Return from void function
3412 The '``ret``' instruction is used to return control flow (and optionally
3413 a value) from a function back to the caller.
3415 There are two forms of the '``ret``' instruction: one that returns a
3416 value and then causes control flow, and one that just causes control
3422 The '``ret``' instruction optionally accepts a single argument, the
3423 return value. The type of the return value must be a ':ref:`first
3424 class <t_firstclass>`' type.
3426 A function is not :ref:`well formed <wellformed>` if it it has a non-void
3427 return type and contains a '``ret``' instruction with no return value or
3428 a return value with a type that does not match its type, or if it has a
3429 void return type and contains a '``ret``' instruction with a return
3435 When the '``ret``' instruction is executed, control flow returns back to
3436 the calling function's context. If the caller is a
3437 ":ref:`call <i_call>`" instruction, execution continues at the
3438 instruction after the call. If the caller was an
3439 ":ref:`invoke <i_invoke>`" instruction, execution continues at the
3440 beginning of the "normal" destination block. If the instruction returns
3441 a value, that value shall set the call or invoke instruction's return
3447 .. code-block:: llvm
3449 ret i32 5 ; Return an integer value of 5
3450 ret void ; Return from a void function
3451 ret { i32, i8 } { i32 4, i8 2 } ; Return a struct of values 4 and 2
3455 '``br``' Instruction
3456 ^^^^^^^^^^^^^^^^^^^^
3463 br i1 <cond>, label <iftrue>, label <iffalse>
3464 br label <dest> ; Unconditional branch
3469 The '``br``' instruction is used to cause control flow to transfer to a
3470 different basic block in the current function. There are two forms of
3471 this instruction, corresponding to a conditional branch and an
3472 unconditional branch.
3477 The conditional branch form of the '``br``' instruction takes a single
3478 '``i1``' value and two '``label``' values. The unconditional form of the
3479 '``br``' instruction takes a single '``label``' value as a target.
3484 Upon execution of a conditional '``br``' instruction, the '``i1``'
3485 argument is evaluated. If the value is ``true``, control flows to the
3486 '``iftrue``' ``label`` argument. If "cond" is ``false``, control flows
3487 to the '``iffalse``' ``label`` argument.
3492 .. code-block:: llvm
3495 %cond = icmp eq i32 %a, %b
3496 br i1 %cond, label %IfEqual, label %IfUnequal
3504 '``switch``' Instruction
3505 ^^^^^^^^^^^^^^^^^^^^^^^^
3512 switch <intty> <value>, label <defaultdest> [ <intty> <val>, label <dest> ... ]
3517 The '``switch``' instruction is used to transfer control flow to one of
3518 several different places. It is a generalization of the '``br``'
3519 instruction, allowing a branch to occur to one of many possible
3525 The '``switch``' instruction uses three parameters: an integer
3526 comparison value '``value``', a default '``label``' destination, and an
3527 array of pairs of comparison value constants and '``label``'s. The table
3528 is not allowed to contain duplicate constant entries.
3533 The ``switch`` instruction specifies a table of values and destinations.
3534 When the '``switch``' instruction is executed, this table is searched
3535 for the given value. If the value is found, control flow is transferred
3536 to the corresponding destination; otherwise, control flow is transferred
3537 to the default destination.
3542 Depending on properties of the target machine and the particular
3543 ``switch`` instruction, this instruction may be code generated in
3544 different ways. For example, it could be generated as a series of
3545 chained conditional branches or with a lookup table.
3550 .. code-block:: llvm
3552 ; Emulate a conditional br instruction
3553 %Val = zext i1 %value to i32
3554 switch i32 %Val, label %truedest [ i32 0, label %falsedest ]
3556 ; Emulate an unconditional br instruction
3557 switch i32 0, label %dest [ ]
3559 ; Implement a jump table:
3560 switch i32 %val, label %otherwise [ i32 0, label %onzero
3562 i32 2, label %ontwo ]
3566 '``indirectbr``' Instruction
3567 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3574 indirectbr <somety>* <address>, [ label <dest1>, label <dest2>, ... ]
3579 The '``indirectbr``' instruction implements an indirect branch to a
3580 label within the current function, whose address is specified by
3581 "``address``". Address must be derived from a
3582 :ref:`blockaddress <blockaddress>` constant.
3587 The '``address``' argument is the address of the label to jump to. The
3588 rest of the arguments indicate the full set of possible destinations
3589 that the address may point to. Blocks are allowed to occur multiple
3590 times in the destination list, though this isn't particularly useful.
3592 This destination list is required so that dataflow analysis has an
3593 accurate understanding of the CFG.
3598 Control transfers to the block specified in the address argument. All
3599 possible destination blocks must be listed in the label list, otherwise
3600 this instruction has undefined behavior. This implies that jumps to
3601 labels defined in other functions have undefined behavior as well.
3606 This is typically implemented with a jump through a register.
3611 .. code-block:: llvm
3613 indirectbr i8* %Addr, [ label %bb1, label %bb2, label %bb3 ]
3617 '``invoke``' Instruction
3618 ^^^^^^^^^^^^^^^^^^^^^^^^
3625 <result> = invoke [cconv] [ret attrs] <ptr to function ty> <function ptr val>(<function args>) [fn attrs]
3626 to label <normal label> unwind label <exception label>
3631 The '``invoke``' instruction causes control to transfer to a specified
3632 function, with the possibility of control flow transfer to either the
3633 '``normal``' label or the '``exception``' label. If the callee function
3634 returns with the "``ret``" instruction, control flow will return to the
3635 "normal" label. If the callee (or any indirect callees) returns via the
3636 ":ref:`resume <i_resume>`" instruction or other exception handling
3637 mechanism, control is interrupted and continued at the dynamically
3638 nearest "exception" label.
3640 The '``exception``' label is a `landing
3641 pad <ExceptionHandling.html#overview>`_ for the exception. As such,
3642 '``exception``' label is required to have the
3643 ":ref:`landingpad <i_landingpad>`" instruction, which contains the
3644 information about the behavior of the program after unwinding happens,
3645 as its first non-PHI instruction. The restrictions on the
3646 "``landingpad``" instruction's tightly couples it to the "``invoke``"
3647 instruction, so that the important information contained within the
3648 "``landingpad``" instruction can't be lost through normal code motion.
3653 This instruction requires several arguments:
3655 #. The optional "cconv" marker indicates which :ref:`calling
3656 convention <callingconv>` the call should use. If none is
3657 specified, the call defaults to using C calling conventions.
3658 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
3659 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
3661 #. '``ptr to function ty``': shall be the signature of the pointer to
3662 function value being invoked. In most cases, this is a direct
3663 function invocation, but indirect ``invoke``'s are just as possible,
3664 branching off an arbitrary pointer to function value.
3665 #. '``function ptr val``': An LLVM value containing a pointer to a
3666 function to be invoked.
3667 #. '``function args``': argument list whose types match the function
3668 signature argument types and parameter attributes. All arguments must
3669 be of :ref:`first class <t_firstclass>` type. If the function signature
3670 indicates the function accepts a variable number of arguments, the
3671 extra arguments can be specified.
3672 #. '``normal label``': the label reached when the called function
3673 executes a '``ret``' instruction.
3674 #. '``exception label``': the label reached when a callee returns via
3675 the :ref:`resume <i_resume>` instruction or other exception handling
3677 #. The optional :ref:`function attributes <fnattrs>` list. Only
3678 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
3679 attributes are valid here.
3684 This instruction is designed to operate as a standard '``call``'
3685 instruction in most regards. The primary difference is that it
3686 establishes an association with a label, which is used by the runtime
3687 library to unwind the stack.
3689 This instruction is used in languages with destructors to ensure that
3690 proper cleanup is performed in the case of either a ``longjmp`` or a
3691 thrown exception. Additionally, this is important for implementation of
3692 '``catch``' clauses in high-level languages that support them.
3694 For the purposes of the SSA form, the definition of the value returned
3695 by the '``invoke``' instruction is deemed to occur on the edge from the
3696 current block to the "normal" label. If the callee unwinds then no
3697 return value is available.
3702 .. code-block:: llvm
3704 %retval = invoke i32 @Test(i32 15) to label %Continue
3705 unwind label %TestCleanup ; i32:retval set
3706 %retval = invoke coldcc i32 %Testfnptr(i32 15) to label %Continue
3707 unwind label %TestCleanup ; i32:retval set
3711 '``resume``' Instruction
3712 ^^^^^^^^^^^^^^^^^^^^^^^^
3719 resume <type> <value>
3724 The '``resume``' instruction is a terminator instruction that has no
3730 The '``resume``' instruction requires one argument, which must have the
3731 same type as the result of any '``landingpad``' instruction in the same
3737 The '``resume``' instruction resumes propagation of an existing
3738 (in-flight) exception whose unwinding was interrupted with a
3739 :ref:`landingpad <i_landingpad>` instruction.
3744 .. code-block:: llvm
3746 resume { i8*, i32 } %exn
3750 '``unreachable``' Instruction
3751 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3763 The '``unreachable``' instruction has no defined semantics. This
3764 instruction is used to inform the optimizer that a particular portion of
3765 the code is not reachable. This can be used to indicate that the code
3766 after a no-return function cannot be reached, and other facts.
3771 The '``unreachable``' instruction has no defined semantics.
3778 Binary operators are used to do most of the computation in a program.
3779 They require two operands of the same type, execute an operation on
3780 them, and produce a single value. The operands might represent multiple
3781 data, as is the case with the :ref:`vector <t_vector>` data type. The
3782 result value has the same type as its operands.
3784 There are several different binary operators:
3788 '``add``' Instruction
3789 ^^^^^^^^^^^^^^^^^^^^^
3796 <result> = add <ty> <op1>, <op2> ; yields ty:result
3797 <result> = add nuw <ty> <op1>, <op2> ; yields ty:result
3798 <result> = add nsw <ty> <op1>, <op2> ; yields ty:result
3799 <result> = add nuw nsw <ty> <op1>, <op2> ; yields ty:result
3804 The '``add``' instruction returns the sum of its two operands.
3809 The two arguments to the '``add``' instruction must be
3810 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3811 arguments must have identical types.
3816 The value produced is the integer sum of the two operands.
3818 If the sum has unsigned overflow, the result returned is the
3819 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
3822 Because LLVM integers use a two's complement representation, this
3823 instruction is appropriate for both signed and unsigned integers.
3825 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3826 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3827 result value of the ``add`` is a :ref:`poison value <poisonvalues>` if
3828 unsigned and/or signed overflow, respectively, occurs.
3833 .. code-block:: llvm
3835 <result> = add i32 4, %var ; yields i32:result = 4 + %var
3839 '``fadd``' Instruction
3840 ^^^^^^^^^^^^^^^^^^^^^^
3847 <result> = fadd [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
3852 The '``fadd``' instruction returns the sum of its two operands.
3857 The two arguments to the '``fadd``' instruction must be :ref:`floating
3858 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3859 Both arguments must have identical types.
3864 The value produced is the floating point sum of the two operands. This
3865 instruction can also take any number of :ref:`fast-math flags <fastmath>`,
3866 which are optimization hints to enable otherwise unsafe floating point
3872 .. code-block:: llvm
3874 <result> = fadd float 4.0, %var ; yields float:result = 4.0 + %var
3876 '``sub``' Instruction
3877 ^^^^^^^^^^^^^^^^^^^^^
3884 <result> = sub <ty> <op1>, <op2> ; yields ty:result
3885 <result> = sub nuw <ty> <op1>, <op2> ; yields ty:result
3886 <result> = sub nsw <ty> <op1>, <op2> ; yields ty:result
3887 <result> = sub nuw nsw <ty> <op1>, <op2> ; yields ty:result
3892 The '``sub``' instruction returns the difference of its two operands.
3894 Note that the '``sub``' instruction is used to represent the '``neg``'
3895 instruction present in most other intermediate representations.
3900 The two arguments to the '``sub``' instruction must be
3901 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3902 arguments must have identical types.
3907 The value produced is the integer difference of the two operands.
3909 If the difference has unsigned overflow, the result returned is the
3910 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
3913 Because LLVM integers use a two's complement representation, this
3914 instruction is appropriate for both signed and unsigned integers.
3916 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3917 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3918 result value of the ``sub`` is a :ref:`poison value <poisonvalues>` if
3919 unsigned and/or signed overflow, respectively, occurs.
3924 .. code-block:: llvm
3926 <result> = sub i32 4, %var ; yields i32:result = 4 - %var
3927 <result> = sub i32 0, %val ; yields i32:result = -%var
3931 '``fsub``' Instruction
3932 ^^^^^^^^^^^^^^^^^^^^^^
3939 <result> = fsub [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
3944 The '``fsub``' instruction returns the difference of its two operands.
3946 Note that the '``fsub``' instruction is used to represent the '``fneg``'
3947 instruction present in most other intermediate representations.
3952 The two arguments to the '``fsub``' instruction must be :ref:`floating
3953 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3954 Both arguments must have identical types.
3959 The value produced is the floating point difference of the two operands.
3960 This instruction can also take any number of :ref:`fast-math
3961 flags <fastmath>`, which are optimization hints to enable otherwise
3962 unsafe floating point optimizations:
3967 .. code-block:: llvm
3969 <result> = fsub float 4.0, %var ; yields float:result = 4.0 - %var
3970 <result> = fsub float -0.0, %val ; yields float:result = -%var
3972 '``mul``' Instruction
3973 ^^^^^^^^^^^^^^^^^^^^^
3980 <result> = mul <ty> <op1>, <op2> ; yields ty:result
3981 <result> = mul nuw <ty> <op1>, <op2> ; yields ty:result
3982 <result> = mul nsw <ty> <op1>, <op2> ; yields ty:result
3983 <result> = mul nuw nsw <ty> <op1>, <op2> ; yields ty:result
3988 The '``mul``' instruction returns the product of its two operands.
3993 The two arguments to the '``mul``' instruction must be
3994 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3995 arguments must have identical types.
4000 The value produced is the integer product of the two operands.
4002 If the result of the multiplication has unsigned overflow, the result
4003 returned is the mathematical result modulo 2\ :sup:`n`\ , where n is the
4004 bit width of the result.
4006 Because LLVM integers use a two's complement representation, and the
4007 result is the same width as the operands, this instruction returns the
4008 correct result for both signed and unsigned integers. If a full product
4009 (e.g. ``i32`` * ``i32`` -> ``i64``) is needed, the operands should be
4010 sign-extended or zero-extended as appropriate to the width of the full
4013 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
4014 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
4015 result value of the ``mul`` is a :ref:`poison value <poisonvalues>` if
4016 unsigned and/or signed overflow, respectively, occurs.
4021 .. code-block:: llvm
4023 <result> = mul i32 4, %var ; yields i32:result = 4 * %var
4027 '``fmul``' Instruction
4028 ^^^^^^^^^^^^^^^^^^^^^^
4035 <result> = fmul [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
4040 The '``fmul``' instruction returns the product of its two operands.
4045 The two arguments to the '``fmul``' instruction must be :ref:`floating
4046 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
4047 Both arguments must have identical types.
4052 The value produced is the floating point product of the two operands.
4053 This instruction can also take any number of :ref:`fast-math
4054 flags <fastmath>`, which are optimization hints to enable otherwise
4055 unsafe floating point optimizations:
4060 .. code-block:: llvm
4062 <result> = fmul float 4.0, %var ; yields float:result = 4.0 * %var
4064 '``udiv``' Instruction
4065 ^^^^^^^^^^^^^^^^^^^^^^
4072 <result> = udiv <ty> <op1>, <op2> ; yields ty:result
4073 <result> = udiv exact <ty> <op1>, <op2> ; yields ty:result
4078 The '``udiv``' instruction returns the quotient of its two operands.
4083 The two arguments to the '``udiv``' instruction must be
4084 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4085 arguments must have identical types.
4090 The value produced is the unsigned integer quotient of the two operands.
4092 Note that unsigned integer division and signed integer division are
4093 distinct operations; for signed integer division, use '``sdiv``'.
4095 Division by zero leads to undefined behavior.
4097 If the ``exact`` keyword is present, the result value of the ``udiv`` is
4098 a :ref:`poison value <poisonvalues>` if %op1 is not a multiple of %op2 (as
4099 such, "((a udiv exact b) mul b) == a").
4104 .. code-block:: llvm
4106 <result> = udiv i32 4, %var ; yields i32:result = 4 / %var
4108 '``sdiv``' Instruction
4109 ^^^^^^^^^^^^^^^^^^^^^^
4116 <result> = sdiv <ty> <op1>, <op2> ; yields ty:result
4117 <result> = sdiv exact <ty> <op1>, <op2> ; yields ty:result
4122 The '``sdiv``' instruction returns the quotient of its two operands.
4127 The two arguments to the '``sdiv``' instruction must be
4128 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4129 arguments must have identical types.
4134 The value produced is the signed integer quotient of the two operands
4135 rounded towards zero.
4137 Note that signed integer division and unsigned integer division are
4138 distinct operations; for unsigned integer division, use '``udiv``'.
4140 Division by zero leads to undefined behavior. Overflow also leads to
4141 undefined behavior; this is a rare case, but can occur, for example, by
4142 doing a 32-bit division of -2147483648 by -1.
4144 If the ``exact`` keyword is present, the result value of the ``sdiv`` is
4145 a :ref:`poison value <poisonvalues>` if the result would be rounded.
4150 .. code-block:: llvm
4152 <result> = sdiv i32 4, %var ; yields i32:result = 4 / %var
4156 '``fdiv``' Instruction
4157 ^^^^^^^^^^^^^^^^^^^^^^
4164 <result> = fdiv [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
4169 The '``fdiv``' instruction returns the quotient of its two operands.
4174 The two arguments to the '``fdiv``' instruction must be :ref:`floating
4175 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
4176 Both arguments must have identical types.
4181 The value produced is the floating point quotient of the two operands.
4182 This instruction can also take any number of :ref:`fast-math
4183 flags <fastmath>`, which are optimization hints to enable otherwise
4184 unsafe floating point optimizations:
4189 .. code-block:: llvm
4191 <result> = fdiv float 4.0, %var ; yields float:result = 4.0 / %var
4193 '``urem``' Instruction
4194 ^^^^^^^^^^^^^^^^^^^^^^
4201 <result> = urem <ty> <op1>, <op2> ; yields ty:result
4206 The '``urem``' instruction returns the remainder from the unsigned
4207 division of its two arguments.
4212 The two arguments to the '``urem``' instruction must be
4213 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4214 arguments must have identical types.
4219 This instruction returns the unsigned integer *remainder* of a division.
4220 This instruction always performs an unsigned division to get the
4223 Note that unsigned integer remainder and signed integer remainder are
4224 distinct operations; for signed integer remainder, use '``srem``'.
4226 Taking the remainder of a division by zero leads to undefined behavior.
4231 .. code-block:: llvm
4233 <result> = urem i32 4, %var ; yields i32:result = 4 % %var
4235 '``srem``' Instruction
4236 ^^^^^^^^^^^^^^^^^^^^^^
4243 <result> = srem <ty> <op1>, <op2> ; yields ty:result
4248 The '``srem``' instruction returns the remainder from the signed
4249 division of its two operands. This instruction can also take
4250 :ref:`vector <t_vector>` versions of the values in which case the elements
4256 The two arguments to the '``srem``' instruction must be
4257 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4258 arguments must have identical types.
4263 This instruction returns the *remainder* of a division (where the result
4264 is either zero or has the same sign as the dividend, ``op1``), not the
4265 *modulo* operator (where the result is either zero or has the same sign
4266 as the divisor, ``op2``) of a value. For more information about the
4267 difference, see `The Math
4268 Forum <http://mathforum.org/dr.math/problems/anne.4.28.99.html>`_. For a
4269 table of how this is implemented in various languages, please see
4271 operation <http://en.wikipedia.org/wiki/Modulo_operation>`_.
4273 Note that signed integer remainder and unsigned integer remainder are
4274 distinct operations; for unsigned integer remainder, use '``urem``'.
4276 Taking the remainder of a division by zero leads to undefined behavior.
4277 Overflow also leads to undefined behavior; this is a rare case, but can
4278 occur, for example, by taking the remainder of a 32-bit division of
4279 -2147483648 by -1. (The remainder doesn't actually overflow, but this
4280 rule lets srem be implemented using instructions that return both the
4281 result of the division and the remainder.)
4286 .. code-block:: llvm
4288 <result> = srem i32 4, %var ; yields i32:result = 4 % %var
4292 '``frem``' Instruction
4293 ^^^^^^^^^^^^^^^^^^^^^^
4300 <result> = frem [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
4305 The '``frem``' instruction returns the remainder from the division of
4311 The two arguments to the '``frem``' instruction must be :ref:`floating
4312 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
4313 Both arguments must have identical types.
4318 This instruction returns the *remainder* of a division. The remainder
4319 has the same sign as the dividend. This instruction can also take any
4320 number of :ref:`fast-math flags <fastmath>`, which are optimization hints
4321 to enable otherwise unsafe floating point optimizations:
4326 .. code-block:: llvm
4328 <result> = frem float 4.0, %var ; yields float:result = 4.0 % %var
4332 Bitwise Binary Operations
4333 -------------------------
4335 Bitwise binary operators are used to do various forms of bit-twiddling
4336 in a program. They are generally very efficient instructions and can
4337 commonly be strength reduced from other instructions. They require two
4338 operands of the same type, execute an operation on them, and produce a
4339 single value. The resulting value is the same type as its operands.
4341 '``shl``' Instruction
4342 ^^^^^^^^^^^^^^^^^^^^^
4349 <result> = shl <ty> <op1>, <op2> ; yields ty:result
4350 <result> = shl nuw <ty> <op1>, <op2> ; yields ty:result
4351 <result> = shl nsw <ty> <op1>, <op2> ; yields ty:result
4352 <result> = shl nuw nsw <ty> <op1>, <op2> ; yields ty:result
4357 The '``shl``' instruction returns the first operand shifted to the left
4358 a specified number of bits.
4363 Both arguments to the '``shl``' instruction must be the same
4364 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4365 '``op2``' is treated as an unsigned value.
4370 The value produced is ``op1`` \* 2\ :sup:`op2` mod 2\ :sup:`n`,
4371 where ``n`` is the width of the result. If ``op2`` is (statically or
4372 dynamically) negative or equal to or larger than the number of bits in
4373 ``op1``, the result is undefined. If the arguments are vectors, each
4374 vector element of ``op1`` is shifted by the corresponding shift amount
4377 If the ``nuw`` keyword is present, then the shift produces a :ref:`poison
4378 value <poisonvalues>` if it shifts out any non-zero bits. If the
4379 ``nsw`` keyword is present, then the shift produces a :ref:`poison
4380 value <poisonvalues>` if it shifts out any bits that disagree with the
4381 resultant sign bit. As such, NUW/NSW have the same semantics as they
4382 would if the shift were expressed as a mul instruction with the same
4383 nsw/nuw bits in (mul %op1, (shl 1, %op2)).
4388 .. code-block:: llvm
4390 <result> = shl i32 4, %var ; yields i32: 4 << %var
4391 <result> = shl i32 4, 2 ; yields i32: 16
4392 <result> = shl i32 1, 10 ; yields i32: 1024
4393 <result> = shl i32 1, 32 ; undefined
4394 <result> = shl <2 x i32> < i32 1, i32 1>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 2, i32 4>
4396 '``lshr``' Instruction
4397 ^^^^^^^^^^^^^^^^^^^^^^
4404 <result> = lshr <ty> <op1>, <op2> ; yields ty:result
4405 <result> = lshr exact <ty> <op1>, <op2> ; yields ty:result
4410 The '``lshr``' instruction (logical shift right) returns the first
4411 operand shifted to the right a specified number of bits with zero fill.
4416 Both arguments to the '``lshr``' instruction must be the same
4417 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4418 '``op2``' is treated as an unsigned value.
4423 This instruction always performs a logical shift right operation. The
4424 most significant bits of the result will be filled with zero bits after
4425 the shift. If ``op2`` is (statically or dynamically) equal to or larger
4426 than the number of bits in ``op1``, the result is undefined. If the
4427 arguments are vectors, each vector element of ``op1`` is shifted by the
4428 corresponding shift amount in ``op2``.
4430 If the ``exact`` keyword is present, the result value of the ``lshr`` is
4431 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
4437 .. code-block:: llvm
4439 <result> = lshr i32 4, 1 ; yields i32:result = 2
4440 <result> = lshr i32 4, 2 ; yields i32:result = 1
4441 <result> = lshr i8 4, 3 ; yields i8:result = 0
4442 <result> = lshr i8 -2, 1 ; yields i8:result = 0x7F
4443 <result> = lshr i32 1, 32 ; undefined
4444 <result> = lshr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 0x7FFFFFFF, i32 1>
4446 '``ashr``' Instruction
4447 ^^^^^^^^^^^^^^^^^^^^^^
4454 <result> = ashr <ty> <op1>, <op2> ; yields ty:result
4455 <result> = ashr exact <ty> <op1>, <op2> ; yields ty:result
4460 The '``ashr``' instruction (arithmetic shift right) returns the first
4461 operand shifted to the right a specified number of bits with sign
4467 Both arguments to the '``ashr``' instruction must be the same
4468 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4469 '``op2``' is treated as an unsigned value.
4474 This instruction always performs an arithmetic shift right operation,
4475 The most significant bits of the result will be filled with the sign bit
4476 of ``op1``. If ``op2`` is (statically or dynamically) equal to or larger
4477 than the number of bits in ``op1``, the result is undefined. If the
4478 arguments are vectors, each vector element of ``op1`` is shifted by the
4479 corresponding shift amount in ``op2``.
4481 If the ``exact`` keyword is present, the result value of the ``ashr`` is
4482 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
4488 .. code-block:: llvm
4490 <result> = ashr i32 4, 1 ; yields i32:result = 2
4491 <result> = ashr i32 4, 2 ; yields i32:result = 1
4492 <result> = ashr i8 4, 3 ; yields i8:result = 0
4493 <result> = ashr i8 -2, 1 ; yields i8:result = -1
4494 <result> = ashr i32 1, 32 ; undefined
4495 <result> = ashr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 3> ; yields: result=<2 x i32> < i32 -1, i32 0>
4497 '``and``' Instruction
4498 ^^^^^^^^^^^^^^^^^^^^^
4505 <result> = and <ty> <op1>, <op2> ; yields ty:result
4510 The '``and``' instruction returns the bitwise logical and of its two
4516 The two arguments to the '``and``' instruction must be
4517 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4518 arguments must have identical types.
4523 The truth table used for the '``and``' instruction is:
4540 .. code-block:: llvm
4542 <result> = and i32 4, %var ; yields i32:result = 4 & %var
4543 <result> = and i32 15, 40 ; yields i32:result = 8
4544 <result> = and i32 4, 8 ; yields i32:result = 0
4546 '``or``' Instruction
4547 ^^^^^^^^^^^^^^^^^^^^
4554 <result> = or <ty> <op1>, <op2> ; yields ty:result
4559 The '``or``' instruction returns the bitwise logical inclusive or of its
4565 The two arguments to the '``or``' instruction must be
4566 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4567 arguments must have identical types.
4572 The truth table used for the '``or``' instruction is:
4591 <result> = or i32 4, %var ; yields i32:result = 4 | %var
4592 <result> = or i32 15, 40 ; yields i32:result = 47
4593 <result> = or i32 4, 8 ; yields i32:result = 12
4595 '``xor``' Instruction
4596 ^^^^^^^^^^^^^^^^^^^^^
4603 <result> = xor <ty> <op1>, <op2> ; yields ty:result
4608 The '``xor``' instruction returns the bitwise logical exclusive or of
4609 its two operands. The ``xor`` is used to implement the "one's
4610 complement" operation, which is the "~" operator in C.
4615 The two arguments to the '``xor``' instruction must be
4616 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4617 arguments must have identical types.
4622 The truth table used for the '``xor``' instruction is:
4639 .. code-block:: llvm
4641 <result> = xor i32 4, %var ; yields i32:result = 4 ^ %var
4642 <result> = xor i32 15, 40 ; yields i32:result = 39
4643 <result> = xor i32 4, 8 ; yields i32:result = 12
4644 <result> = xor i32 %V, -1 ; yields i32:result = ~%V
4649 LLVM supports several instructions to represent vector operations in a
4650 target-independent manner. These instructions cover the element-access
4651 and vector-specific operations needed to process vectors effectively.
4652 While LLVM does directly support these vector operations, many
4653 sophisticated algorithms will want to use target-specific intrinsics to
4654 take full advantage of a specific target.
4656 .. _i_extractelement:
4658 '``extractelement``' Instruction
4659 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4666 <result> = extractelement <n x <ty>> <val>, <ty2> <idx> ; yields <ty>
4671 The '``extractelement``' instruction extracts a single scalar element
4672 from a vector at a specified index.
4677 The first operand of an '``extractelement``' instruction is a value of
4678 :ref:`vector <t_vector>` type. The second operand is an index indicating
4679 the position from which to extract the element. The index may be a
4680 variable of any integer type.
4685 The result is a scalar of the same type as the element type of ``val``.
4686 Its value is the value at position ``idx`` of ``val``. If ``idx``
4687 exceeds the length of ``val``, the results are undefined.
4692 .. code-block:: llvm
4694 <result> = extractelement <4 x i32> %vec, i32 0 ; yields i32
4696 .. _i_insertelement:
4698 '``insertelement``' Instruction
4699 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4706 <result> = insertelement <n x <ty>> <val>, <ty> <elt>, <ty2> <idx> ; yields <n x <ty>>
4711 The '``insertelement``' instruction inserts a scalar element into a
4712 vector at a specified index.
4717 The first operand of an '``insertelement``' instruction is a value of
4718 :ref:`vector <t_vector>` type. The second operand is a scalar value whose
4719 type must equal the element type of the first operand. The third operand
4720 is an index indicating the position at which to insert the value. The
4721 index may be a variable of any integer type.
4726 The result is a vector of the same type as ``val``. Its element values
4727 are those of ``val`` except at position ``idx``, where it gets the value
4728 ``elt``. If ``idx`` exceeds the length of ``val``, the results are
4734 .. code-block:: llvm
4736 <result> = insertelement <4 x i32> %vec, i32 1, i32 0 ; yields <4 x i32>
4738 .. _i_shufflevector:
4740 '``shufflevector``' Instruction
4741 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4748 <result> = shufflevector <n x <ty>> <v1>, <n x <ty>> <v2>, <m x i32> <mask> ; yields <m x <ty>>
4753 The '``shufflevector``' instruction constructs a permutation of elements
4754 from two input vectors, returning a vector with the same element type as
4755 the input and length that is the same as the shuffle mask.
4760 The first two operands of a '``shufflevector``' instruction are vectors
4761 with the same type. The third argument is a shuffle mask whose element
4762 type is always 'i32'. The result of the instruction is a vector whose
4763 length is the same as the shuffle mask and whose element type is the
4764 same as the element type of the first two operands.
4766 The shuffle mask operand is required to be a constant vector with either
4767 constant integer or undef values.
4772 The elements of the two input vectors are numbered from left to right
4773 across both of the vectors. The shuffle mask operand specifies, for each
4774 element of the result vector, which element of the two input vectors the
4775 result element gets. The element selector may be undef (meaning "don't
4776 care") and the second operand may be undef if performing a shuffle from
4782 .. code-block:: llvm
4784 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
4785 <4 x i32> <i32 0, i32 4, i32 1, i32 5> ; yields <4 x i32>
4786 <result> = shufflevector <4 x i32> %v1, <4 x i32> undef,
4787 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32> - Identity shuffle.
4788 <result> = shufflevector <8 x i32> %v1, <8 x i32> undef,
4789 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32>
4790 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
4791 <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7 > ; yields <8 x i32>
4793 Aggregate Operations
4794 --------------------
4796 LLVM supports several instructions for working with
4797 :ref:`aggregate <t_aggregate>` values.
4801 '``extractvalue``' Instruction
4802 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4809 <result> = extractvalue <aggregate type> <val>, <idx>{, <idx>}*
4814 The '``extractvalue``' instruction extracts the value of a member field
4815 from an :ref:`aggregate <t_aggregate>` value.
4820 The first operand of an '``extractvalue``' instruction is a value of
4821 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The operands are
4822 constant indices to specify which value to extract in a similar manner
4823 as indices in a '``getelementptr``' instruction.
4825 The major differences to ``getelementptr`` indexing are:
4827 - Since the value being indexed is not a pointer, the first index is
4828 omitted and assumed to be zero.
4829 - At least one index must be specified.
4830 - Not only struct indices but also array indices must be in bounds.
4835 The result is the value at the position in the aggregate specified by
4841 .. code-block:: llvm
4843 <result> = extractvalue {i32, float} %agg, 0 ; yields i32
4847 '``insertvalue``' Instruction
4848 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4855 <result> = insertvalue <aggregate type> <val>, <ty> <elt>, <idx>{, <idx>}* ; yields <aggregate type>
4860 The '``insertvalue``' instruction inserts a value into a member field in
4861 an :ref:`aggregate <t_aggregate>` value.
4866 The first operand of an '``insertvalue``' instruction is a value of
4867 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The second operand is
4868 a first-class value to insert. The following operands are constant
4869 indices indicating the position at which to insert the value in a
4870 similar manner as indices in a '``extractvalue``' instruction. The value
4871 to insert must have the same type as the value identified by the
4877 The result is an aggregate of the same type as ``val``. Its value is
4878 that of ``val`` except that the value at the position specified by the
4879 indices is that of ``elt``.
4884 .. code-block:: llvm
4886 %agg1 = insertvalue {i32, float} undef, i32 1, 0 ; yields {i32 1, float undef}
4887 %agg2 = insertvalue {i32, float} %agg1, float %val, 1 ; yields {i32 1, float %val}
4888 %agg3 = insertvalue {i32, {float}} %agg1, float %val, 1, 0 ; yields {i32 1, float %val}
4892 Memory Access and Addressing Operations
4893 ---------------------------------------
4895 A key design point of an SSA-based representation is how it represents
4896 memory. In LLVM, no memory locations are in SSA form, which makes things
4897 very simple. This section describes how to read, write, and allocate
4902 '``alloca``' Instruction
4903 ^^^^^^^^^^^^^^^^^^^^^^^^
4910 <result> = alloca [inalloca] <type> [, <ty> <NumElements>] [, align <alignment>] ; yields type*:result
4915 The '``alloca``' instruction allocates memory on the stack frame of the
4916 currently executing function, to be automatically released when this
4917 function returns to its caller. The object is always allocated in the
4918 generic address space (address space zero).
4923 The '``alloca``' instruction allocates ``sizeof(<type>)*NumElements``
4924 bytes of memory on the runtime stack, returning a pointer of the
4925 appropriate type to the program. If "NumElements" is specified, it is
4926 the number of elements allocated, otherwise "NumElements" is defaulted
4927 to be one. If a constant alignment is specified, the value result of the
4928 allocation is guaranteed to be aligned to at least that boundary. The
4929 alignment may not be greater than ``1 << 29``. If not specified, or if
4930 zero, the target can choose to align the allocation on any convenient
4931 boundary compatible with the type.
4933 '``type``' may be any sized type.
4938 Memory is allocated; a pointer is returned. The operation is undefined
4939 if there is insufficient stack space for the allocation. '``alloca``'d
4940 memory is automatically released when the function returns. The
4941 '``alloca``' instruction is commonly used to represent automatic
4942 variables that must have an address available. When the function returns
4943 (either with the ``ret`` or ``resume`` instructions), the memory is
4944 reclaimed. Allocating zero bytes is legal, but the result is undefined.
4945 The order in which memory is allocated (ie., which way the stack grows)
4951 .. code-block:: llvm
4953 %ptr = alloca i32 ; yields i32*:ptr
4954 %ptr = alloca i32, i32 4 ; yields i32*:ptr
4955 %ptr = alloca i32, i32 4, align 1024 ; yields i32*:ptr
4956 %ptr = alloca i32, align 1024 ; yields i32*:ptr
4960 '``load``' Instruction
4961 ^^^^^^^^^^^^^^^^^^^^^^
4968 <result> = load [volatile] <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>][, !invariant.load !<index>]
4969 <result> = load atomic [volatile] <ty>* <pointer> [singlethread] <ordering>, align <alignment>
4970 !<index> = !{ i32 1 }
4975 The '``load``' instruction is used to read from memory.
4980 The argument to the ``load`` instruction specifies the memory address
4981 from which to load. The pointer must point to a :ref:`first
4982 class <t_firstclass>` type. If the ``load`` is marked as ``volatile``,
4983 then the optimizer is not allowed to modify the number or order of
4984 execution of this ``load`` with other :ref:`volatile
4985 operations <volatile>`.
4987 If the ``load`` is marked as ``atomic``, it takes an extra
4988 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
4989 ``release`` and ``acq_rel`` orderings are not valid on ``load``
4990 instructions. Atomic loads produce :ref:`defined <memmodel>` results
4991 when they may see multiple atomic stores. The type of the pointee must
4992 be an integer type whose bit width is a power of two greater than or
4993 equal to eight and less than or equal to a target-specific size limit.
4994 ``align`` must be explicitly specified on atomic loads, and the load has
4995 undefined behavior if the alignment is not set to a value which is at
4996 least the size in bytes of the pointee. ``!nontemporal`` does not have
4997 any defined semantics for atomic loads.
4999 The optional constant ``align`` argument specifies the alignment of the
5000 operation (that is, the alignment of the memory address). A value of 0
5001 or an omitted ``align`` argument means that the operation has the ABI
5002 alignment for the target. It is the responsibility of the code emitter
5003 to ensure that the alignment information is correct. Overestimating the
5004 alignment results in undefined behavior. Underestimating the alignment
5005 may produce less efficient code. An alignment of 1 is always safe. The
5006 maximum possible alignment is ``1 << 29``.
5008 The optional ``!nontemporal`` metadata must reference a single
5009 metadata name ``<index>`` corresponding to a metadata node with one
5010 ``i32`` entry of value 1. The existence of the ``!nontemporal``
5011 metadata on the instruction tells the optimizer and code generator
5012 that this load is not expected to be reused in the cache. The code
5013 generator may select special instructions to save cache bandwidth, such
5014 as the ``MOVNT`` instruction on x86.
5016 The optional ``!invariant.load`` metadata must reference a single
5017 metadata name ``<index>`` corresponding to a metadata node with no
5018 entries. The existence of the ``!invariant.load`` metadata on the
5019 instruction tells the optimizer and code generator that this load
5020 address points to memory which does not change value during program
5021 execution. The optimizer may then move this load around, for example, by
5022 hoisting it out of loops using loop invariant code motion.
5027 The location of memory pointed to is loaded. If the value being loaded
5028 is of scalar type then the number of bytes read does not exceed the
5029 minimum number of bytes needed to hold all bits of the type. For
5030 example, loading an ``i24`` reads at most three bytes. When loading a
5031 value of a type like ``i20`` with a size that is not an integral number
5032 of bytes, the result is undefined if the value was not originally
5033 written using a store of the same type.
5038 .. code-block:: llvm
5040 %ptr = alloca i32 ; yields i32*:ptr
5041 store i32 3, i32* %ptr ; yields void
5042 %val = load i32* %ptr ; yields i32:val = i32 3
5046 '``store``' Instruction
5047 ^^^^^^^^^^^^^^^^^^^^^^^
5054 store [volatile] <ty> <value>, <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>] ; yields void
5055 store atomic [volatile] <ty> <value>, <ty>* <pointer> [singlethread] <ordering>, align <alignment> ; yields void
5060 The '``store``' instruction is used to write to memory.
5065 There are two arguments to the ``store`` instruction: a value to store
5066 and an address at which to store it. The type of the ``<pointer>``
5067 operand must be a pointer to the :ref:`first class <t_firstclass>` type of
5068 the ``<value>`` operand. If the ``store`` is marked as ``volatile``,
5069 then the optimizer is not allowed to modify the number or order of
5070 execution of this ``store`` with other :ref:`volatile
5071 operations <volatile>`.
5073 If the ``store`` is marked as ``atomic``, it takes an extra
5074 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
5075 ``acquire`` and ``acq_rel`` orderings aren't valid on ``store``
5076 instructions. Atomic loads produce :ref:`defined <memmodel>` results
5077 when they may see multiple atomic stores. The type of the pointee must
5078 be an integer type whose bit width is a power of two greater than or
5079 equal to eight and less than or equal to a target-specific size limit.
5080 ``align`` must be explicitly specified on atomic stores, and the store
5081 has undefined behavior if the alignment is not set to a value which is
5082 at least the size in bytes of the pointee. ``!nontemporal`` does not
5083 have any defined semantics for atomic stores.
5085 The optional constant ``align`` argument specifies the alignment of the
5086 operation (that is, the alignment of the memory address). A value of 0
5087 or an omitted ``align`` argument means that the operation has the ABI
5088 alignment for the target. It is the responsibility of the code emitter
5089 to ensure that the alignment information is correct. Overestimating the
5090 alignment results in undefined behavior. Underestimating the
5091 alignment may produce less efficient code. An alignment of 1 is always
5092 safe. The maximum possible alignment is ``1 << 29``.
5094 The optional ``!nontemporal`` metadata must reference a single metadata
5095 name ``<index>`` corresponding to a metadata node with one ``i32`` entry of
5096 value 1. The existence of the ``!nontemporal`` metadata on the instruction
5097 tells the optimizer and code generator that this load is not expected to
5098 be reused in the cache. The code generator may select special
5099 instructions to save cache bandwidth, such as the MOVNT instruction on
5105 The contents of memory are updated to contain ``<value>`` at the
5106 location specified by the ``<pointer>`` operand. If ``<value>`` is
5107 of scalar type then the number of bytes written does not exceed the
5108 minimum number of bytes needed to hold all bits of the type. For
5109 example, storing an ``i24`` writes at most three bytes. When writing a
5110 value of a type like ``i20`` with a size that is not an integral number
5111 of bytes, it is unspecified what happens to the extra bits that do not
5112 belong to the type, but they will typically be overwritten.
5117 .. code-block:: llvm
5119 %ptr = alloca i32 ; yields i32*:ptr
5120 store i32 3, i32* %ptr ; yields void
5121 %val = load i32* %ptr ; yields i32:val = i32 3
5125 '``fence``' Instruction
5126 ^^^^^^^^^^^^^^^^^^^^^^^
5133 fence [singlethread] <ordering> ; yields void
5138 The '``fence``' instruction is used to introduce happens-before edges
5144 '``fence``' instructions take an :ref:`ordering <ordering>` argument which
5145 defines what *synchronizes-with* edges they add. They can only be given
5146 ``acquire``, ``release``, ``acq_rel``, and ``seq_cst`` orderings.
5151 A fence A which has (at least) ``release`` ordering semantics
5152 *synchronizes with* a fence B with (at least) ``acquire`` ordering
5153 semantics if and only if there exist atomic operations X and Y, both
5154 operating on some atomic object M, such that A is sequenced before X, X
5155 modifies M (either directly or through some side effect of a sequence
5156 headed by X), Y is sequenced before B, and Y observes M. This provides a
5157 *happens-before* dependency between A and B. Rather than an explicit
5158 ``fence``, one (but not both) of the atomic operations X or Y might
5159 provide a ``release`` or ``acquire`` (resp.) ordering constraint and
5160 still *synchronize-with* the explicit ``fence`` and establish the
5161 *happens-before* edge.
5163 A ``fence`` which has ``seq_cst`` ordering, in addition to having both
5164 ``acquire`` and ``release`` semantics specified above, participates in
5165 the global program order of other ``seq_cst`` operations and/or fences.
5167 The optional ":ref:`singlethread <singlethread>`" argument specifies
5168 that the fence only synchronizes with other fences in the same thread.
5169 (This is useful for interacting with signal handlers.)
5174 .. code-block:: llvm
5176 fence acquire ; yields void
5177 fence singlethread seq_cst ; yields void
5181 '``cmpxchg``' Instruction
5182 ^^^^^^^^^^^^^^^^^^^^^^^^^
5189 cmpxchg [weak] [volatile] <ty>* <pointer>, <ty> <cmp>, <ty> <new> [singlethread] <success ordering> <failure ordering> ; yields { ty, i1 }
5194 The '``cmpxchg``' instruction is used to atomically modify memory. It
5195 loads a value in memory and compares it to a given value. If they are
5196 equal, it tries to store a new value into the memory.
5201 There are three arguments to the '``cmpxchg``' instruction: an address
5202 to operate on, a value to compare to the value currently be at that
5203 address, and a new value to place at that address if the compared values
5204 are equal. The type of '<cmp>' must be an integer type whose bit width
5205 is a power of two greater than or equal to eight and less than or equal
5206 to a target-specific size limit. '<cmp>' and '<new>' must have the same
5207 type, and the type of '<pointer>' must be a pointer to that type. If the
5208 ``cmpxchg`` is marked as ``volatile``, then the optimizer is not allowed
5209 to modify the number or order of execution of this ``cmpxchg`` with
5210 other :ref:`volatile operations <volatile>`.
5212 The success and failure :ref:`ordering <ordering>` arguments specify how this
5213 ``cmpxchg`` synchronizes with other atomic operations. Both ordering parameters
5214 must be at least ``monotonic``, the ordering constraint on failure must be no
5215 stronger than that on success, and the failure ordering cannot be either
5216 ``release`` or ``acq_rel``.
5218 The optional "``singlethread``" argument declares that the ``cmpxchg``
5219 is only atomic with respect to code (usually signal handlers) running in
5220 the same thread as the ``cmpxchg``. Otherwise the cmpxchg is atomic with
5221 respect to all other code in the system.
5223 The pointer passed into cmpxchg must have alignment greater than or
5224 equal to the size in memory of the operand.
5229 The contents of memory at the location specified by the '``<pointer>``' operand
5230 is read and compared to '``<cmp>``'; if the read value is the equal, the
5231 '``<new>``' is written. The original value at the location is returned, together
5232 with a flag indicating success (true) or failure (false).
5234 If the cmpxchg operation is marked as ``weak`` then a spurious failure is
5235 permitted: the operation may not write ``<new>`` even if the comparison
5238 If the cmpxchg operation is strong (the default), the i1 value is 1 if and only
5239 if the value loaded equals ``cmp``.
5241 A successful ``cmpxchg`` is a read-modify-write instruction for the purpose of
5242 identifying release sequences. A failed ``cmpxchg`` is equivalent to an atomic
5243 load with an ordering parameter determined the second ordering parameter.
5248 .. code-block:: llvm
5251 %orig = atomic load i32* %ptr unordered ; yields i32
5255 %cmp = phi i32 [ %orig, %entry ], [%old, %loop]
5256 %squared = mul i32 %cmp, %cmp
5257 %val_success = cmpxchg i32* %ptr, i32 %cmp, i32 %squared acq_rel monotonic ; yields { i32, i1 }
5258 %value_loaded = extractvalue { i32, i1 } %val_success, 0
5259 %success = extractvalue { i32, i1 } %val_success, 1
5260 br i1 %success, label %done, label %loop
5267 '``atomicrmw``' Instruction
5268 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
5275 atomicrmw [volatile] <operation> <ty>* <pointer>, <ty> <value> [singlethread] <ordering> ; yields ty
5280 The '``atomicrmw``' instruction is used to atomically modify memory.
5285 There are three arguments to the '``atomicrmw``' instruction: an
5286 operation to apply, an address whose value to modify, an argument to the
5287 operation. The operation must be one of the following keywords:
5301 The type of '<value>' must be an integer type whose bit width is a power
5302 of two greater than or equal to eight and less than or equal to a
5303 target-specific size limit. The type of the '``<pointer>``' operand must
5304 be a pointer to that type. If the ``atomicrmw`` is marked as
5305 ``volatile``, then the optimizer is not allowed to modify the number or
5306 order of execution of this ``atomicrmw`` with other :ref:`volatile
5307 operations <volatile>`.
5312 The contents of memory at the location specified by the '``<pointer>``'
5313 operand are atomically read, modified, and written back. The original
5314 value at the location is returned. The modification is specified by the
5317 - xchg: ``*ptr = val``
5318 - add: ``*ptr = *ptr + val``
5319 - sub: ``*ptr = *ptr - val``
5320 - and: ``*ptr = *ptr & val``
5321 - nand: ``*ptr = ~(*ptr & val)``
5322 - or: ``*ptr = *ptr | val``
5323 - xor: ``*ptr = *ptr ^ val``
5324 - max: ``*ptr = *ptr > val ? *ptr : val`` (using a signed comparison)
5325 - min: ``*ptr = *ptr < val ? *ptr : val`` (using a signed comparison)
5326 - umax: ``*ptr = *ptr > val ? *ptr : val`` (using an unsigned
5328 - umin: ``*ptr = *ptr < val ? *ptr : val`` (using an unsigned
5334 .. code-block:: llvm
5336 %old = atomicrmw add i32* %ptr, i32 1 acquire ; yields i32
5338 .. _i_getelementptr:
5340 '``getelementptr``' Instruction
5341 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5348 <result> = getelementptr <pty>* <ptrval>{, <ty> <idx>}*
5349 <result> = getelementptr inbounds <pty>* <ptrval>{, <ty> <idx>}*
5350 <result> = getelementptr <ptr vector> ptrval, <vector index type> idx
5355 The '``getelementptr``' instruction is used to get the address of a
5356 subelement of an :ref:`aggregate <t_aggregate>` data structure. It performs
5357 address calculation only and does not access memory.
5362 The first argument is always a pointer or a vector of pointers, and
5363 forms the basis of the calculation. The remaining arguments are indices
5364 that indicate which of the elements of the aggregate object are indexed.
5365 The interpretation of each index is dependent on the type being indexed
5366 into. The first index always indexes the pointer value given as the
5367 first argument, the second index indexes a value of the type pointed to
5368 (not necessarily the value directly pointed to, since the first index
5369 can be non-zero), etc. The first type indexed into must be a pointer
5370 value, subsequent types can be arrays, vectors, and structs. Note that
5371 subsequent types being indexed into can never be pointers, since that
5372 would require loading the pointer before continuing calculation.
5374 The type of each index argument depends on the type it is indexing into.
5375 When indexing into a (optionally packed) structure, only ``i32`` integer
5376 **constants** are allowed (when using a vector of indices they must all
5377 be the **same** ``i32`` integer constant). When indexing into an array,
5378 pointer or vector, integers of any width are allowed, and they are not
5379 required to be constant. These integers are treated as signed values
5382 For example, let's consider a C code fragment and how it gets compiled
5398 int *foo(struct ST *s) {
5399 return &s[1].Z.B[5][13];
5402 The LLVM code generated by Clang is:
5404 .. code-block:: llvm
5406 %struct.RT = type { i8, [10 x [20 x i32]], i8 }
5407 %struct.ST = type { i32, double, %struct.RT }
5409 define i32* @foo(%struct.ST* %s) nounwind uwtable readnone optsize ssp {
5411 %arrayidx = getelementptr inbounds %struct.ST* %s, i64 1, i32 2, i32 1, i64 5, i64 13
5418 In the example above, the first index is indexing into the
5419 '``%struct.ST*``' type, which is a pointer, yielding a '``%struct.ST``'
5420 = '``{ i32, double, %struct.RT }``' type, a structure. The second index
5421 indexes into the third element of the structure, yielding a
5422 '``%struct.RT``' = '``{ i8 , [10 x [20 x i32]], i8 }``' type, another
5423 structure. The third index indexes into the second element of the
5424 structure, yielding a '``[10 x [20 x i32]]``' type, an array. The two
5425 dimensions of the array are subscripted into, yielding an '``i32``'
5426 type. The '``getelementptr``' instruction returns a pointer to this
5427 element, thus computing a value of '``i32*``' type.
5429 Note that it is perfectly legal to index partially through a structure,
5430 returning a pointer to an inner element. Because of this, the LLVM code
5431 for the given testcase is equivalent to:
5433 .. code-block:: llvm
5435 define i32* @foo(%struct.ST* %s) {
5436 %t1 = getelementptr %struct.ST* %s, i32 1 ; yields %struct.ST*:%t1
5437 %t2 = getelementptr %struct.ST* %t1, i32 0, i32 2 ; yields %struct.RT*:%t2
5438 %t3 = getelementptr %struct.RT* %t2, i32 0, i32 1 ; yields [10 x [20 x i32]]*:%t3
5439 %t4 = getelementptr [10 x [20 x i32]]* %t3, i32 0, i32 5 ; yields [20 x i32]*:%t4
5440 %t5 = getelementptr [20 x i32]* %t4, i32 0, i32 13 ; yields i32*:%t5
5444 If the ``inbounds`` keyword is present, the result value of the
5445 ``getelementptr`` is a :ref:`poison value <poisonvalues>` if the base
5446 pointer is not an *in bounds* address of an allocated object, or if any
5447 of the addresses that would be formed by successive addition of the
5448 offsets implied by the indices to the base address with infinitely
5449 precise signed arithmetic are not an *in bounds* address of that
5450 allocated object. The *in bounds* addresses for an allocated object are
5451 all the addresses that point into the object, plus the address one byte
5452 past the end. In cases where the base is a vector of pointers the
5453 ``inbounds`` keyword applies to each of the computations element-wise.
5455 If the ``inbounds`` keyword is not present, the offsets are added to the
5456 base address with silently-wrapping two's complement arithmetic. If the
5457 offsets have a different width from the pointer, they are sign-extended
5458 or truncated to the width of the pointer. The result value of the
5459 ``getelementptr`` may be outside the object pointed to by the base
5460 pointer. The result value may not necessarily be used to access memory
5461 though, even if it happens to point into allocated storage. See the
5462 :ref:`Pointer Aliasing Rules <pointeraliasing>` section for more
5465 The getelementptr instruction is often confusing. For some more insight
5466 into how it works, see :doc:`the getelementptr FAQ <GetElementPtr>`.
5471 .. code-block:: llvm
5473 ; yields [12 x i8]*:aptr
5474 %aptr = getelementptr {i32, [12 x i8]}* %saptr, i64 0, i32 1
5476 %vptr = getelementptr {i32, <2 x i8>}* %svptr, i64 0, i32 1, i32 1
5478 %eptr = getelementptr [12 x i8]* %aptr, i64 0, i32 1
5480 %iptr = getelementptr [10 x i32]* @arr, i16 0, i16 0
5482 In cases where the pointer argument is a vector of pointers, each index
5483 must be a vector with the same number of elements. For example:
5485 .. code-block:: llvm
5487 %A = getelementptr <4 x i8*> %ptrs, <4 x i64> %offsets,
5489 Conversion Operations
5490 ---------------------
5492 The instructions in this category are the conversion instructions
5493 (casting) which all take a single operand and a type. They perform
5494 various bit conversions on the operand.
5496 '``trunc .. to``' Instruction
5497 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5504 <result> = trunc <ty> <value> to <ty2> ; yields ty2
5509 The '``trunc``' instruction truncates its operand to the type ``ty2``.
5514 The '``trunc``' instruction takes a value to trunc, and a type to trunc
5515 it to. Both types must be of :ref:`integer <t_integer>` types, or vectors
5516 of the same number of integers. The bit size of the ``value`` must be
5517 larger than the bit size of the destination type, ``ty2``. Equal sized
5518 types are not allowed.
5523 The '``trunc``' instruction truncates the high order bits in ``value``
5524 and converts the remaining bits to ``ty2``. Since the source size must
5525 be larger than the destination size, ``trunc`` cannot be a *no-op cast*.
5526 It will always truncate bits.
5531 .. code-block:: llvm
5533 %X = trunc i32 257 to i8 ; yields i8:1
5534 %Y = trunc i32 123 to i1 ; yields i1:true
5535 %Z = trunc i32 122 to i1 ; yields i1:false
5536 %W = trunc <2 x i16> <i16 8, i16 7> to <2 x i8> ; yields <i8 8, i8 7>
5538 '``zext .. to``' Instruction
5539 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5546 <result> = zext <ty> <value> to <ty2> ; yields ty2
5551 The '``zext``' instruction zero extends its operand to type ``ty2``.
5556 The '``zext``' instruction takes a value to cast, and a type to cast it
5557 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
5558 the same number of integers. The bit size of the ``value`` must be
5559 smaller than the bit size of the destination type, ``ty2``.
5564 The ``zext`` fills the high order bits of the ``value`` with zero bits
5565 until it reaches the size of the destination type, ``ty2``.
5567 When zero extending from i1, the result will always be either 0 or 1.
5572 .. code-block:: llvm
5574 %X = zext i32 257 to i64 ; yields i64:257
5575 %Y = zext i1 true to i32 ; yields i32:1
5576 %Z = zext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
5578 '``sext .. to``' Instruction
5579 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5586 <result> = sext <ty> <value> to <ty2> ; yields ty2
5591 The '``sext``' sign extends ``value`` to the type ``ty2``.
5596 The '``sext``' instruction takes a value to cast, and a type to cast it
5597 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
5598 the same number of integers. The bit size of the ``value`` must be
5599 smaller than the bit size of the destination type, ``ty2``.
5604 The '``sext``' instruction performs a sign extension by copying the sign
5605 bit (highest order bit) of the ``value`` until it reaches the bit size
5606 of the type ``ty2``.
5608 When sign extending from i1, the extension always results in -1 or 0.
5613 .. code-block:: llvm
5615 %X = sext i8 -1 to i16 ; yields i16 :65535
5616 %Y = sext i1 true to i32 ; yields i32:-1
5617 %Z = sext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
5619 '``fptrunc .. to``' Instruction
5620 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5627 <result> = fptrunc <ty> <value> to <ty2> ; yields ty2
5632 The '``fptrunc``' instruction truncates ``value`` to type ``ty2``.
5637 The '``fptrunc``' instruction takes a :ref:`floating point <t_floating>`
5638 value to cast and a :ref:`floating point <t_floating>` type to cast it to.
5639 The size of ``value`` must be larger than the size of ``ty2``. This
5640 implies that ``fptrunc`` cannot be used to make a *no-op cast*.
5645 The '``fptrunc``' instruction truncates a ``value`` from a larger
5646 :ref:`floating point <t_floating>` type to a smaller :ref:`floating
5647 point <t_floating>` type. If the value cannot fit within the
5648 destination type, ``ty2``, then the results are undefined.
5653 .. code-block:: llvm
5655 %X = fptrunc double 123.0 to float ; yields float:123.0
5656 %Y = fptrunc double 1.0E+300 to float ; yields undefined
5658 '``fpext .. to``' Instruction
5659 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5666 <result> = fpext <ty> <value> to <ty2> ; yields ty2
5671 The '``fpext``' extends a floating point ``value`` to a larger floating
5677 The '``fpext``' instruction takes a :ref:`floating point <t_floating>`
5678 ``value`` to cast, and a :ref:`floating point <t_floating>` type to cast it
5679 to. The source type must be smaller than the destination type.
5684 The '``fpext``' instruction extends the ``value`` from a smaller
5685 :ref:`floating point <t_floating>` type to a larger :ref:`floating
5686 point <t_floating>` type. The ``fpext`` cannot be used to make a
5687 *no-op cast* because it always changes bits. Use ``bitcast`` to make a
5688 *no-op cast* for a floating point cast.
5693 .. code-block:: llvm
5695 %X = fpext float 3.125 to double ; yields double:3.125000e+00
5696 %Y = fpext double %X to fp128 ; yields fp128:0xL00000000000000004000900000000000
5698 '``fptoui .. to``' Instruction
5699 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5706 <result> = fptoui <ty> <value> to <ty2> ; yields ty2
5711 The '``fptoui``' converts a floating point ``value`` to its unsigned
5712 integer equivalent of type ``ty2``.
5717 The '``fptoui``' instruction takes a value to cast, which must be a
5718 scalar or vector :ref:`floating point <t_floating>` value, and a type to
5719 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
5720 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
5721 type with the same number of elements as ``ty``
5726 The '``fptoui``' instruction converts its :ref:`floating
5727 point <t_floating>` operand into the nearest (rounding towards zero)
5728 unsigned integer value. If the value cannot fit in ``ty2``, the results
5734 .. code-block:: llvm
5736 %X = fptoui double 123.0 to i32 ; yields i32:123
5737 %Y = fptoui float 1.0E+300 to i1 ; yields undefined:1
5738 %Z = fptoui float 1.04E+17 to i8 ; yields undefined:1
5740 '``fptosi .. to``' Instruction
5741 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5748 <result> = fptosi <ty> <value> to <ty2> ; yields ty2
5753 The '``fptosi``' instruction converts :ref:`floating point <t_floating>`
5754 ``value`` to type ``ty2``.
5759 The '``fptosi``' instruction takes a value to cast, which must be a
5760 scalar or vector :ref:`floating point <t_floating>` value, and a type to
5761 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
5762 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
5763 type with the same number of elements as ``ty``
5768 The '``fptosi``' instruction converts its :ref:`floating
5769 point <t_floating>` operand into the nearest (rounding towards zero)
5770 signed integer value. If the value cannot fit in ``ty2``, the results
5776 .. code-block:: llvm
5778 %X = fptosi double -123.0 to i32 ; yields i32:-123
5779 %Y = fptosi float 1.0E-247 to i1 ; yields undefined:1
5780 %Z = fptosi float 1.04E+17 to i8 ; yields undefined:1
5782 '``uitofp .. to``' Instruction
5783 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5790 <result> = uitofp <ty> <value> to <ty2> ; yields ty2
5795 The '``uitofp``' instruction regards ``value`` as an unsigned integer
5796 and converts that value to the ``ty2`` type.
5801 The '``uitofp``' instruction takes a value to cast, which must be a
5802 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
5803 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
5804 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
5805 type with the same number of elements as ``ty``
5810 The '``uitofp``' instruction interprets its operand as an unsigned
5811 integer quantity and converts it to the corresponding floating point
5812 value. If the value cannot fit in the floating point value, the results
5818 .. code-block:: llvm
5820 %X = uitofp i32 257 to float ; yields float:257.0
5821 %Y = uitofp i8 -1 to double ; yields double:255.0
5823 '``sitofp .. to``' Instruction
5824 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5831 <result> = sitofp <ty> <value> to <ty2> ; yields ty2
5836 The '``sitofp``' instruction regards ``value`` as a signed integer and
5837 converts that value to the ``ty2`` type.
5842 The '``sitofp``' instruction takes a value to cast, which must be a
5843 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
5844 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
5845 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
5846 type with the same number of elements as ``ty``
5851 The '``sitofp``' instruction interprets its operand as a signed integer
5852 quantity and converts it to the corresponding floating point value. If
5853 the value cannot fit in the floating point value, the results are
5859 .. code-block:: llvm
5861 %X = sitofp i32 257 to float ; yields float:257.0
5862 %Y = sitofp i8 -1 to double ; yields double:-1.0
5866 '``ptrtoint .. to``' Instruction
5867 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5874 <result> = ptrtoint <ty> <value> to <ty2> ; yields ty2
5879 The '``ptrtoint``' instruction converts the pointer or a vector of
5880 pointers ``value`` to the integer (or vector of integers) type ``ty2``.
5885 The '``ptrtoint``' instruction takes a ``value`` to cast, which must be
5886 a a value of type :ref:`pointer <t_pointer>` or a vector of pointers, and a
5887 type to cast it to ``ty2``, which must be an :ref:`integer <t_integer>` or
5888 a vector of integers type.
5893 The '``ptrtoint``' instruction converts ``value`` to integer type
5894 ``ty2`` by interpreting the pointer value as an integer and either
5895 truncating or zero extending that value to the size of the integer type.
5896 If ``value`` is smaller than ``ty2`` then a zero extension is done. If
5897 ``value`` is larger than ``ty2`` then a truncation is done. If they are
5898 the same size, then nothing is done (*no-op cast*) other than a type
5904 .. code-block:: llvm
5906 %X = ptrtoint i32* %P to i8 ; yields truncation on 32-bit architecture
5907 %Y = ptrtoint i32* %P to i64 ; yields zero extension on 32-bit architecture
5908 %Z = ptrtoint <4 x i32*> %P to <4 x i64>; yields vector zero extension for a vector of addresses on 32-bit architecture
5912 '``inttoptr .. to``' Instruction
5913 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5920 <result> = inttoptr <ty> <value> to <ty2> ; yields ty2
5925 The '``inttoptr``' instruction converts an integer ``value`` to a
5926 pointer type, ``ty2``.
5931 The '``inttoptr``' instruction takes an :ref:`integer <t_integer>` value to
5932 cast, and a type to cast it to, which must be a :ref:`pointer <t_pointer>`
5938 The '``inttoptr``' instruction converts ``value`` to type ``ty2`` by
5939 applying either a zero extension or a truncation depending on the size
5940 of the integer ``value``. If ``value`` is larger than the size of a
5941 pointer then a truncation is done. If ``value`` is smaller than the size
5942 of a pointer then a zero extension is done. If they are the same size,
5943 nothing is done (*no-op cast*).
5948 .. code-block:: llvm
5950 %X = inttoptr i32 255 to i32* ; yields zero extension on 64-bit architecture
5951 %Y = inttoptr i32 255 to i32* ; yields no-op on 32-bit architecture
5952 %Z = inttoptr i64 0 to i32* ; yields truncation on 32-bit architecture
5953 %Z = inttoptr <4 x i32> %G to <4 x i8*>; yields truncation of vector G to four pointers
5957 '``bitcast .. to``' Instruction
5958 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5965 <result> = bitcast <ty> <value> to <ty2> ; yields ty2
5970 The '``bitcast``' instruction converts ``value`` to type ``ty2`` without
5976 The '``bitcast``' instruction takes a value to cast, which must be a
5977 non-aggregate first class value, and a type to cast it to, which must
5978 also be a non-aggregate :ref:`first class <t_firstclass>` type. The
5979 bit sizes of ``value`` and the destination type, ``ty2``, must be
5980 identical. If the source type is a pointer, the destination type must
5981 also be a pointer of the same size. This instruction supports bitwise
5982 conversion of vectors to integers and to vectors of other types (as
5983 long as they have the same size).
5988 The '``bitcast``' instruction converts ``value`` to type ``ty2``. It
5989 is always a *no-op cast* because no bits change with this
5990 conversion. The conversion is done as if the ``value`` had been stored
5991 to memory and read back as type ``ty2``. Pointer (or vector of
5992 pointers) types may only be converted to other pointer (or vector of
5993 pointers) types with the same address space through this instruction.
5994 To convert pointers to other types, use the :ref:`inttoptr <i_inttoptr>`
5995 or :ref:`ptrtoint <i_ptrtoint>` instructions first.
6000 .. code-block:: llvm
6002 %X = bitcast i8 255 to i8 ; yields i8 :-1
6003 %Y = bitcast i32* %x to sint* ; yields sint*:%x
6004 %Z = bitcast <2 x int> %V to i64; ; yields i64: %V
6005 %Z = bitcast <2 x i32*> %V to <2 x i64*> ; yields <2 x i64*>
6007 .. _i_addrspacecast:
6009 '``addrspacecast .. to``' Instruction
6010 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6017 <result> = addrspacecast <pty> <ptrval> to <pty2> ; yields pty2
6022 The '``addrspacecast``' instruction converts ``ptrval`` from ``pty`` in
6023 address space ``n`` to type ``pty2`` in address space ``m``.
6028 The '``addrspacecast``' instruction takes a pointer or vector of pointer value
6029 to cast and a pointer type to cast it to, which must have a different
6035 The '``addrspacecast``' instruction converts the pointer value
6036 ``ptrval`` to type ``pty2``. It can be a *no-op cast* or a complex
6037 value modification, depending on the target and the address space
6038 pair. Pointer conversions within the same address space must be
6039 performed with the ``bitcast`` instruction. Note that if the address space
6040 conversion is legal then both result and operand refer to the same memory
6046 .. code-block:: llvm
6048 %X = addrspacecast i32* %x to i32 addrspace(1)* ; yields i32 addrspace(1)*:%x
6049 %Y = addrspacecast i32 addrspace(1)* %y to i64 addrspace(2)* ; yields i64 addrspace(2)*:%y
6050 %Z = addrspacecast <4 x i32*> %z to <4 x float addrspace(3)*> ; yields <4 x float addrspace(3)*>:%z
6057 The instructions in this category are the "miscellaneous" instructions,
6058 which defy better classification.
6062 '``icmp``' Instruction
6063 ^^^^^^^^^^^^^^^^^^^^^^
6070 <result> = icmp <cond> <ty> <op1>, <op2> ; yields i1 or <N x i1>:result
6075 The '``icmp``' instruction returns a boolean value or a vector of
6076 boolean values based on comparison of its two integer, integer vector,
6077 pointer, or pointer vector operands.
6082 The '``icmp``' instruction takes three operands. The first operand is
6083 the condition code indicating the kind of comparison to perform. It is
6084 not a value, just a keyword. The possible condition code are:
6087 #. ``ne``: not equal
6088 #. ``ugt``: unsigned greater than
6089 #. ``uge``: unsigned greater or equal
6090 #. ``ult``: unsigned less than
6091 #. ``ule``: unsigned less or equal
6092 #. ``sgt``: signed greater than
6093 #. ``sge``: signed greater or equal
6094 #. ``slt``: signed less than
6095 #. ``sle``: signed less or equal
6097 The remaining two arguments must be :ref:`integer <t_integer>` or
6098 :ref:`pointer <t_pointer>` or integer :ref:`vector <t_vector>` typed. They
6099 must also be identical types.
6104 The '``icmp``' compares ``op1`` and ``op2`` according to the condition
6105 code given as ``cond``. The comparison performed always yields either an
6106 :ref:`i1 <t_integer>` or vector of ``i1`` result, as follows:
6108 #. ``eq``: yields ``true`` if the operands are equal, ``false``
6109 otherwise. No sign interpretation is necessary or performed.
6110 #. ``ne``: yields ``true`` if the operands are unequal, ``false``
6111 otherwise. No sign interpretation is necessary or performed.
6112 #. ``ugt``: interprets the operands as unsigned values and yields
6113 ``true`` if ``op1`` is greater than ``op2``.
6114 #. ``uge``: interprets the operands as unsigned values and yields
6115 ``true`` if ``op1`` is greater than or equal to ``op2``.
6116 #. ``ult``: interprets the operands as unsigned values and yields
6117 ``true`` if ``op1`` is less than ``op2``.
6118 #. ``ule``: interprets the operands as unsigned values and yields
6119 ``true`` if ``op1`` is less than or equal to ``op2``.
6120 #. ``sgt``: interprets the operands as signed values and yields ``true``
6121 if ``op1`` is greater than ``op2``.
6122 #. ``sge``: interprets the operands as signed values and yields ``true``
6123 if ``op1`` is greater than or equal to ``op2``.
6124 #. ``slt``: interprets the operands as signed values and yields ``true``
6125 if ``op1`` is less than ``op2``.
6126 #. ``sle``: interprets the operands as signed values and yields ``true``
6127 if ``op1`` is less than or equal to ``op2``.
6129 If the operands are :ref:`pointer <t_pointer>` typed, the pointer values
6130 are compared as if they were integers.
6132 If the operands are integer vectors, then they are compared element by
6133 element. The result is an ``i1`` vector with the same number of elements
6134 as the values being compared. Otherwise, the result is an ``i1``.
6139 .. code-block:: llvm
6141 <result> = icmp eq i32 4, 5 ; yields: result=false
6142 <result> = icmp ne float* %X, %X ; yields: result=false
6143 <result> = icmp ult i16 4, 5 ; yields: result=true
6144 <result> = icmp sgt i16 4, 5 ; yields: result=false
6145 <result> = icmp ule i16 -4, 5 ; yields: result=false
6146 <result> = icmp sge i16 4, 5 ; yields: result=false
6148 Note that the code generator does not yet support vector types with the
6149 ``icmp`` instruction.
6153 '``fcmp``' Instruction
6154 ^^^^^^^^^^^^^^^^^^^^^^
6161 <result> = fcmp <cond> <ty> <op1>, <op2> ; yields i1 or <N x i1>:result
6166 The '``fcmp``' instruction returns a boolean value or vector of boolean
6167 values based on comparison of its operands.
6169 If the operands are floating point scalars, then the result type is a
6170 boolean (:ref:`i1 <t_integer>`).
6172 If the operands are floating point vectors, then the result type is a
6173 vector of boolean with the same number of elements as the operands being
6179 The '``fcmp``' instruction takes three operands. The first operand is
6180 the condition code indicating the kind of comparison to perform. It is
6181 not a value, just a keyword. The possible condition code are:
6183 #. ``false``: no comparison, always returns false
6184 #. ``oeq``: ordered and equal
6185 #. ``ogt``: ordered and greater than
6186 #. ``oge``: ordered and greater than or equal
6187 #. ``olt``: ordered and less than
6188 #. ``ole``: ordered and less than or equal
6189 #. ``one``: ordered and not equal
6190 #. ``ord``: ordered (no nans)
6191 #. ``ueq``: unordered or equal
6192 #. ``ugt``: unordered or greater than
6193 #. ``uge``: unordered or greater than or equal
6194 #. ``ult``: unordered or less than
6195 #. ``ule``: unordered or less than or equal
6196 #. ``une``: unordered or not equal
6197 #. ``uno``: unordered (either nans)
6198 #. ``true``: no comparison, always returns true
6200 *Ordered* means that neither operand is a QNAN while *unordered* means
6201 that either operand may be a QNAN.
6203 Each of ``val1`` and ``val2`` arguments must be either a :ref:`floating
6204 point <t_floating>` type or a :ref:`vector <t_vector>` of floating point
6205 type. They must have identical types.
6210 The '``fcmp``' instruction compares ``op1`` and ``op2`` according to the
6211 condition code given as ``cond``. If the operands are vectors, then the
6212 vectors are compared element by element. Each comparison performed
6213 always yields an :ref:`i1 <t_integer>` result, as follows:
6215 #. ``false``: always yields ``false``, regardless of operands.
6216 #. ``oeq``: yields ``true`` if both operands are not a QNAN and ``op1``
6217 is equal to ``op2``.
6218 #. ``ogt``: yields ``true`` if both operands are not a QNAN and ``op1``
6219 is greater than ``op2``.
6220 #. ``oge``: yields ``true`` if both operands are not a QNAN and ``op1``
6221 is greater than or equal to ``op2``.
6222 #. ``olt``: yields ``true`` if both operands are not a QNAN and ``op1``
6223 is less than ``op2``.
6224 #. ``ole``: yields ``true`` if both operands are not a QNAN and ``op1``
6225 is less than or equal to ``op2``.
6226 #. ``one``: yields ``true`` if both operands are not a QNAN and ``op1``
6227 is not equal to ``op2``.
6228 #. ``ord``: yields ``true`` if both operands are not a QNAN.
6229 #. ``ueq``: yields ``true`` if either operand is a QNAN or ``op1`` is
6231 #. ``ugt``: yields ``true`` if either operand is a QNAN or ``op1`` is
6232 greater than ``op2``.
6233 #. ``uge``: yields ``true`` if either operand is a QNAN or ``op1`` is
6234 greater than or equal to ``op2``.
6235 #. ``ult``: yields ``true`` if either operand is a QNAN or ``op1`` is
6237 #. ``ule``: yields ``true`` if either operand is a QNAN or ``op1`` is
6238 less than or equal to ``op2``.
6239 #. ``une``: yields ``true`` if either operand is a QNAN or ``op1`` is
6240 not equal to ``op2``.
6241 #. ``uno``: yields ``true`` if either operand is a QNAN.
6242 #. ``true``: always yields ``true``, regardless of operands.
6247 .. code-block:: llvm
6249 <result> = fcmp oeq float 4.0, 5.0 ; yields: result=false
6250 <result> = fcmp one float 4.0, 5.0 ; yields: result=true
6251 <result> = fcmp olt float 4.0, 5.0 ; yields: result=true
6252 <result> = fcmp ueq double 1.0, 2.0 ; yields: result=false
6254 Note that the code generator does not yet support vector types with the
6255 ``fcmp`` instruction.
6259 '``phi``' Instruction
6260 ^^^^^^^^^^^^^^^^^^^^^
6267 <result> = phi <ty> [ <val0>, <label0>], ...
6272 The '``phi``' instruction is used to implement the φ node in the SSA
6273 graph representing the function.
6278 The type of the incoming values is specified with the first type field.
6279 After this, the '``phi``' instruction takes a list of pairs as
6280 arguments, with one pair for each predecessor basic block of the current
6281 block. Only values of :ref:`first class <t_firstclass>` type may be used as
6282 the value arguments to the PHI node. Only labels may be used as the
6285 There must be no non-phi instructions between the start of a basic block
6286 and the PHI instructions: i.e. PHI instructions must be first in a basic
6289 For the purposes of the SSA form, the use of each incoming value is
6290 deemed to occur on the edge from the corresponding predecessor block to
6291 the current block (but after any definition of an '``invoke``'
6292 instruction's return value on the same edge).
6297 At runtime, the '``phi``' instruction logically takes on the value
6298 specified by the pair corresponding to the predecessor basic block that
6299 executed just prior to the current block.
6304 .. code-block:: llvm
6306 Loop: ; Infinite loop that counts from 0 on up...
6307 %indvar = phi i32 [ 0, %LoopHeader ], [ %nextindvar, %Loop ]
6308 %nextindvar = add i32 %indvar, 1
6313 '``select``' Instruction
6314 ^^^^^^^^^^^^^^^^^^^^^^^^
6321 <result> = select selty <cond>, <ty> <val1>, <ty> <val2> ; yields ty
6323 selty is either i1 or {<N x i1>}
6328 The '``select``' instruction is used to choose one value based on a
6329 condition, without IR-level branching.
6334 The '``select``' instruction requires an 'i1' value or a vector of 'i1'
6335 values indicating the condition, and two values of the same :ref:`first
6336 class <t_firstclass>` type. If the val1/val2 are vectors and the
6337 condition is a scalar, then entire vectors are selected, not individual
6343 If the condition is an i1 and it evaluates to 1, the instruction returns
6344 the first value argument; otherwise, it returns the second value
6347 If the condition is a vector of i1, then the value arguments must be
6348 vectors of the same size, and the selection is done element by element.
6353 .. code-block:: llvm
6355 %X = select i1 true, i8 17, i8 42 ; yields i8:17
6359 '``call``' Instruction
6360 ^^^^^^^^^^^^^^^^^^^^^^
6367 <result> = [tail | musttail] call [cconv] [ret attrs] <ty> [<fnty>*] <fnptrval>(<function args>) [fn attrs]
6372 The '``call``' instruction represents a simple function call.
6377 This instruction requires several arguments:
6379 #. The optional ``tail`` and ``musttail`` markers indicate that the optimizers
6380 should perform tail call optimization. The ``tail`` marker is a hint that
6381 `can be ignored <CodeGenerator.html#sibcallopt>`_. The ``musttail`` marker
6382 means that the call must be tail call optimized in order for the program to
6383 be correct. The ``musttail`` marker provides these guarantees:
6385 #. The call will not cause unbounded stack growth if it is part of a
6386 recursive cycle in the call graph.
6387 #. Arguments with the :ref:`inalloca <attr_inalloca>` attribute are
6390 Both markers imply that the callee does not access allocas or varargs from
6391 the caller. Calls marked ``musttail`` must obey the following additional
6394 - The call must immediately precede a :ref:`ret <i_ret>` instruction,
6395 or a pointer bitcast followed by a ret instruction.
6396 - The ret instruction must return the (possibly bitcasted) value
6397 produced by the call or void.
6398 - The caller and callee prototypes must match. Pointer types of
6399 parameters or return types may differ in pointee type, but not
6401 - The calling conventions of the caller and callee must match.
6402 - All ABI-impacting function attributes, such as sret, byval, inreg,
6403 returned, and inalloca, must match.
6405 Tail call optimization for calls marked ``tail`` is guaranteed to occur if
6406 the following conditions are met:
6408 - Caller and callee both have the calling convention ``fastcc``.
6409 - The call is in tail position (ret immediately follows call and ret
6410 uses value of call or is void).
6411 - Option ``-tailcallopt`` is enabled, or
6412 ``llvm::GuaranteedTailCallOpt`` is ``true``.
6413 - `Platform-specific constraints are
6414 met. <CodeGenerator.html#tailcallopt>`_
6416 #. The optional "cconv" marker indicates which :ref:`calling
6417 convention <callingconv>` the call should use. If none is
6418 specified, the call defaults to using C calling conventions. The
6419 calling convention of the call must match the calling convention of
6420 the target function, or else the behavior is undefined.
6421 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
6422 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
6424 #. '``ty``': the type of the call instruction itself which is also the
6425 type of the return value. Functions that return no value are marked
6427 #. '``fnty``': shall be the signature of the pointer to function value
6428 being invoked. The argument types must match the types implied by
6429 this signature. This type can be omitted if the function is not
6430 varargs and if the function type does not return a pointer to a
6432 #. '``fnptrval``': An LLVM value containing a pointer to a function to
6433 be invoked. In most cases, this is a direct function invocation, but
6434 indirect ``call``'s are just as possible, calling an arbitrary pointer
6436 #. '``function args``': argument list whose types match the function
6437 signature argument types and parameter attributes. All arguments must
6438 be of :ref:`first class <t_firstclass>` type. If the function signature
6439 indicates the function accepts a variable number of arguments, the
6440 extra arguments can be specified.
6441 #. The optional :ref:`function attributes <fnattrs>` list. Only
6442 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
6443 attributes are valid here.
6448 The '``call``' instruction is used to cause control flow to transfer to
6449 a specified function, with its incoming arguments bound to the specified
6450 values. Upon a '``ret``' instruction in the called function, control
6451 flow continues with the instruction after the function call, and the
6452 return value of the function is bound to the result argument.
6457 .. code-block:: llvm
6459 %retval = call i32 @test(i32 %argc)
6460 call i32 (i8*, ...)* @printf(i8* %msg, i32 12, i8 42) ; yields i32
6461 %X = tail call i32 @foo() ; yields i32
6462 %Y = tail call fastcc i32 @foo() ; yields i32
6463 call void %foo(i8 97 signext)
6465 %struct.A = type { i32, i8 }
6466 %r = call %struct.A @foo() ; yields { i32, i8 }
6467 %gr = extractvalue %struct.A %r, 0 ; yields i32
6468 %gr1 = extractvalue %struct.A %r, 1 ; yields i8
6469 %Z = call void @foo() noreturn ; indicates that %foo never returns normally
6470 %ZZ = call zeroext i32 @bar() ; Return value is %zero extended
6472 llvm treats calls to some functions with names and arguments that match
6473 the standard C99 library as being the C99 library functions, and may
6474 perform optimizations or generate code for them under that assumption.
6475 This is something we'd like to change in the future to provide better
6476 support for freestanding environments and non-C-based languages.
6480 '``va_arg``' Instruction
6481 ^^^^^^^^^^^^^^^^^^^^^^^^
6488 <resultval> = va_arg <va_list*> <arglist>, <argty>
6493 The '``va_arg``' instruction is used to access arguments passed through
6494 the "variable argument" area of a function call. It is used to implement
6495 the ``va_arg`` macro in C.
6500 This instruction takes a ``va_list*`` value and the type of the
6501 argument. It returns a value of the specified argument type and
6502 increments the ``va_list`` to point to the next argument. The actual
6503 type of ``va_list`` is target specific.
6508 The '``va_arg``' instruction loads an argument of the specified type
6509 from the specified ``va_list`` and causes the ``va_list`` to point to
6510 the next argument. For more information, see the variable argument
6511 handling :ref:`Intrinsic Functions <int_varargs>`.
6513 It is legal for this instruction to be called in a function which does
6514 not take a variable number of arguments, for example, the ``vfprintf``
6517 ``va_arg`` is an LLVM instruction instead of an :ref:`intrinsic
6518 function <intrinsics>` because it takes a type as an argument.
6523 See the :ref:`variable argument processing <int_varargs>` section.
6525 Note that the code generator does not yet fully support va\_arg on many
6526 targets. Also, it does not currently support va\_arg with aggregate
6527 types on any target.
6531 '``landingpad``' Instruction
6532 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6539 <resultval> = landingpad <resultty> personality <type> <pers_fn> <clause>+
6540 <resultval> = landingpad <resultty> personality <type> <pers_fn> cleanup <clause>*
6542 <clause> := catch <type> <value>
6543 <clause> := filter <array constant type> <array constant>
6548 The '``landingpad``' instruction is used by `LLVM's exception handling
6549 system <ExceptionHandling.html#overview>`_ to specify that a basic block
6550 is a landing pad --- one where the exception lands, and corresponds to the
6551 code found in the ``catch`` portion of a ``try``/``catch`` sequence. It
6552 defines values supplied by the personality function (``pers_fn``) upon
6553 re-entry to the function. The ``resultval`` has the type ``resultty``.
6558 This instruction takes a ``pers_fn`` value. This is the personality
6559 function associated with the unwinding mechanism. The optional
6560 ``cleanup`` flag indicates that the landing pad block is a cleanup.
6562 A ``clause`` begins with the clause type --- ``catch`` or ``filter`` --- and
6563 contains the global variable representing the "type" that may be caught
6564 or filtered respectively. Unlike the ``catch`` clause, the ``filter``
6565 clause takes an array constant as its argument. Use
6566 "``[0 x i8**] undef``" for a filter which cannot throw. The
6567 '``landingpad``' instruction must contain *at least* one ``clause`` or
6568 the ``cleanup`` flag.
6573 The '``landingpad``' instruction defines the values which are set by the
6574 personality function (``pers_fn``) upon re-entry to the function, and
6575 therefore the "result type" of the ``landingpad`` instruction. As with
6576 calling conventions, how the personality function results are
6577 represented in LLVM IR is target specific.
6579 The clauses are applied in order from top to bottom. If two
6580 ``landingpad`` instructions are merged together through inlining, the
6581 clauses from the calling function are appended to the list of clauses.
6582 When the call stack is being unwound due to an exception being thrown,
6583 the exception is compared against each ``clause`` in turn. If it doesn't
6584 match any of the clauses, and the ``cleanup`` flag is not set, then
6585 unwinding continues further up the call stack.
6587 The ``landingpad`` instruction has several restrictions:
6589 - A landing pad block is a basic block which is the unwind destination
6590 of an '``invoke``' instruction.
6591 - A landing pad block must have a '``landingpad``' instruction as its
6592 first non-PHI instruction.
6593 - There can be only one '``landingpad``' instruction within the landing
6595 - A basic block that is not a landing pad block may not include a
6596 '``landingpad``' instruction.
6597 - All '``landingpad``' instructions in a function must have the same
6598 personality function.
6603 .. code-block:: llvm
6605 ;; A landing pad which can catch an integer.
6606 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6608 ;; A landing pad that is a cleanup.
6609 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6611 ;; A landing pad which can catch an integer and can only throw a double.
6612 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6614 filter [1 x i8**] [@_ZTId]
6621 LLVM supports the notion of an "intrinsic function". These functions
6622 have well known names and semantics and are required to follow certain
6623 restrictions. Overall, these intrinsics represent an extension mechanism
6624 for the LLVM language that does not require changing all of the
6625 transformations in LLVM when adding to the language (or the bitcode
6626 reader/writer, the parser, etc...).
6628 Intrinsic function names must all start with an "``llvm.``" prefix. This
6629 prefix is reserved in LLVM for intrinsic names; thus, function names may
6630 not begin with this prefix. Intrinsic functions must always be external
6631 functions: you cannot define the body of intrinsic functions. Intrinsic
6632 functions may only be used in call or invoke instructions: it is illegal
6633 to take the address of an intrinsic function. Additionally, because
6634 intrinsic functions are part of the LLVM language, it is required if any
6635 are added that they be documented here.
6637 Some intrinsic functions can be overloaded, i.e., the intrinsic
6638 represents a family of functions that perform the same operation but on
6639 different data types. Because LLVM can represent over 8 million
6640 different integer types, overloading is used commonly to allow an
6641 intrinsic function to operate on any integer type. One or more of the
6642 argument types or the result type can be overloaded to accept any
6643 integer type. Argument types may also be defined as exactly matching a
6644 previous argument's type or the result type. This allows an intrinsic
6645 function which accepts multiple arguments, but needs all of them to be
6646 of the same type, to only be overloaded with respect to a single
6647 argument or the result.
6649 Overloaded intrinsics will have the names of its overloaded argument
6650 types encoded into its function name, each preceded by a period. Only
6651 those types which are overloaded result in a name suffix. Arguments
6652 whose type is matched against another type do not. For example, the
6653 ``llvm.ctpop`` function can take an integer of any width and returns an
6654 integer of exactly the same integer width. This leads to a family of
6655 functions such as ``i8 @llvm.ctpop.i8(i8 %val)`` and
6656 ``i29 @llvm.ctpop.i29(i29 %val)``. Only one type, the return type, is
6657 overloaded, and only one type suffix is required. Because the argument's
6658 type is matched against the return type, it does not require its own
6661 To learn how to add an intrinsic function, please see the `Extending
6662 LLVM Guide <ExtendingLLVM.html>`_.
6666 Variable Argument Handling Intrinsics
6667 -------------------------------------
6669 Variable argument support is defined in LLVM with the
6670 :ref:`va_arg <i_va_arg>` instruction and these three intrinsic
6671 functions. These functions are related to the similarly named macros
6672 defined in the ``<stdarg.h>`` header file.
6674 All of these functions operate on arguments that use a target-specific
6675 value type "``va_list``". The LLVM assembly language reference manual
6676 does not define what this type is, so all transformations should be
6677 prepared to handle these functions regardless of the type used.
6679 This example shows how the :ref:`va_arg <i_va_arg>` instruction and the
6680 variable argument handling intrinsic functions are used.
6682 .. code-block:: llvm
6684 define i32 @test(i32 %X, ...) {
6685 ; Initialize variable argument processing
6687 %ap2 = bitcast i8** %ap to i8*
6688 call void @llvm.va_start(i8* %ap2)
6690 ; Read a single integer argument
6691 %tmp = va_arg i8** %ap, i32
6693 ; Demonstrate usage of llvm.va_copy and llvm.va_end
6695 %aq2 = bitcast i8** %aq to i8*
6696 call void @llvm.va_copy(i8* %aq2, i8* %ap2)
6697 call void @llvm.va_end(i8* %aq2)
6699 ; Stop processing of arguments.
6700 call void @llvm.va_end(i8* %ap2)
6704 declare void @llvm.va_start(i8*)
6705 declare void @llvm.va_copy(i8*, i8*)
6706 declare void @llvm.va_end(i8*)
6710 '``llvm.va_start``' Intrinsic
6711 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6718 declare void @llvm.va_start(i8* <arglist>)
6723 The '``llvm.va_start``' intrinsic initializes ``*<arglist>`` for
6724 subsequent use by ``va_arg``.
6729 The argument is a pointer to a ``va_list`` element to initialize.
6734 The '``llvm.va_start``' intrinsic works just like the ``va_start`` macro
6735 available in C. In a target-dependent way, it initializes the
6736 ``va_list`` element to which the argument points, so that the next call
6737 to ``va_arg`` will produce the first variable argument passed to the
6738 function. Unlike the C ``va_start`` macro, this intrinsic does not need
6739 to know the last argument of the function as the compiler can figure
6742 '``llvm.va_end``' Intrinsic
6743 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6750 declare void @llvm.va_end(i8* <arglist>)
6755 The '``llvm.va_end``' intrinsic destroys ``*<arglist>``, which has been
6756 initialized previously with ``llvm.va_start`` or ``llvm.va_copy``.
6761 The argument is a pointer to a ``va_list`` to destroy.
6766 The '``llvm.va_end``' intrinsic works just like the ``va_end`` macro
6767 available in C. In a target-dependent way, it destroys the ``va_list``
6768 element to which the argument points. Calls to
6769 :ref:`llvm.va_start <int_va_start>` and
6770 :ref:`llvm.va_copy <int_va_copy>` must be matched exactly with calls to
6775 '``llvm.va_copy``' Intrinsic
6776 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6783 declare void @llvm.va_copy(i8* <destarglist>, i8* <srcarglist>)
6788 The '``llvm.va_copy``' intrinsic copies the current argument position
6789 from the source argument list to the destination argument list.
6794 The first argument is a pointer to a ``va_list`` element to initialize.
6795 The second argument is a pointer to a ``va_list`` element to copy from.
6800 The '``llvm.va_copy``' intrinsic works just like the ``va_copy`` macro
6801 available in C. In a target-dependent way, it copies the source
6802 ``va_list`` element into the destination ``va_list`` element. This
6803 intrinsic is necessary because the `` llvm.va_start`` intrinsic may be
6804 arbitrarily complex and require, for example, memory allocation.
6806 Accurate Garbage Collection Intrinsics
6807 --------------------------------------
6809 LLVM support for `Accurate Garbage Collection <GarbageCollection.html>`_
6810 (GC) requires the implementation and generation of these intrinsics.
6811 These intrinsics allow identification of :ref:`GC roots on the
6812 stack <int_gcroot>`, as well as garbage collector implementations that
6813 require :ref:`read <int_gcread>` and :ref:`write <int_gcwrite>` barriers.
6814 Front-ends for type-safe garbage collected languages should generate
6815 these intrinsics to make use of the LLVM garbage collectors. For more
6816 details, see `Accurate Garbage Collection with
6817 LLVM <GarbageCollection.html>`_.
6819 The garbage collection intrinsics only operate on objects in the generic
6820 address space (address space zero).
6824 '``llvm.gcroot``' Intrinsic
6825 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6832 declare void @llvm.gcroot(i8** %ptrloc, i8* %metadata)
6837 The '``llvm.gcroot``' intrinsic declares the existence of a GC root to
6838 the code generator, and allows some metadata to be associated with it.
6843 The first argument specifies the address of a stack object that contains
6844 the root pointer. The second pointer (which must be either a constant or
6845 a global value address) contains the meta-data to be associated with the
6851 At runtime, a call to this intrinsic stores a null pointer into the
6852 "ptrloc" location. At compile-time, the code generator generates
6853 information to allow the runtime to find the pointer at GC safe points.
6854 The '``llvm.gcroot``' intrinsic may only be used in a function which
6855 :ref:`specifies a GC algorithm <gc>`.
6859 '``llvm.gcread``' Intrinsic
6860 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6867 declare i8* @llvm.gcread(i8* %ObjPtr, i8** %Ptr)
6872 The '``llvm.gcread``' intrinsic identifies reads of references from heap
6873 locations, allowing garbage collector implementations that require read
6879 The second argument is the address to read from, which should be an
6880 address allocated from the garbage collector. The first object is a
6881 pointer to the start of the referenced object, if needed by the language
6882 runtime (otherwise null).
6887 The '``llvm.gcread``' intrinsic has the same semantics as a load
6888 instruction, but may be replaced with substantially more complex code by
6889 the garbage collector runtime, as needed. The '``llvm.gcread``'
6890 intrinsic may only be used in a function which :ref:`specifies a GC
6895 '``llvm.gcwrite``' Intrinsic
6896 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6903 declare void @llvm.gcwrite(i8* %P1, i8* %Obj, i8** %P2)
6908 The '``llvm.gcwrite``' intrinsic identifies writes of references to heap
6909 locations, allowing garbage collector implementations that require write
6910 barriers (such as generational or reference counting collectors).
6915 The first argument is the reference to store, the second is the start of
6916 the object to store it to, and the third is the address of the field of
6917 Obj to store to. If the runtime does not require a pointer to the
6918 object, Obj may be null.
6923 The '``llvm.gcwrite``' intrinsic has the same semantics as a store
6924 instruction, but may be replaced with substantially more complex code by
6925 the garbage collector runtime, as needed. The '``llvm.gcwrite``'
6926 intrinsic may only be used in a function which :ref:`specifies a GC
6929 Code Generator Intrinsics
6930 -------------------------
6932 These intrinsics are provided by LLVM to expose special features that
6933 may only be implemented with code generator support.
6935 '``llvm.returnaddress``' Intrinsic
6936 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6943 declare i8 *@llvm.returnaddress(i32 <level>)
6948 The '``llvm.returnaddress``' intrinsic attempts to compute a
6949 target-specific value indicating the return address of the current
6950 function or one of its callers.
6955 The argument to this intrinsic indicates which function to return the
6956 address for. Zero indicates the calling function, one indicates its
6957 caller, etc. The argument is **required** to be a constant integer
6963 The '``llvm.returnaddress``' intrinsic either returns a pointer
6964 indicating the return address of the specified call frame, or zero if it
6965 cannot be identified. The value returned by this intrinsic is likely to
6966 be incorrect or 0 for arguments other than zero, so it should only be
6967 used for debugging purposes.
6969 Note that calling this intrinsic does not prevent function inlining or
6970 other aggressive transformations, so the value returned may not be that
6971 of the obvious source-language caller.
6973 '``llvm.frameaddress``' Intrinsic
6974 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6981 declare i8* @llvm.frameaddress(i32 <level>)
6986 The '``llvm.frameaddress``' intrinsic attempts to return the
6987 target-specific frame pointer value for the specified stack frame.
6992 The argument to this intrinsic indicates which function to return the
6993 frame pointer for. Zero indicates the calling function, one indicates
6994 its caller, etc. The argument is **required** to be a constant integer
7000 The '``llvm.frameaddress``' intrinsic either returns a pointer
7001 indicating the frame address of the specified call frame, or zero if it
7002 cannot be identified. The value returned by this intrinsic is likely to
7003 be incorrect or 0 for arguments other than zero, so it should only be
7004 used for debugging purposes.
7006 Note that calling this intrinsic does not prevent function inlining or
7007 other aggressive transformations, so the value returned may not be that
7008 of the obvious source-language caller.
7010 .. _int_read_register:
7011 .. _int_write_register:
7013 '``llvm.read_register``' and '``llvm.write_register``' Intrinsics
7014 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7021 declare i32 @llvm.read_register.i32(metadata)
7022 declare i64 @llvm.read_register.i64(metadata)
7023 declare void @llvm.write_register.i32(metadata, i32 @value)
7024 declare void @llvm.write_register.i64(metadata, i64 @value)
7025 !0 = metadata !{metadata !"sp\00"}
7030 The '``llvm.read_register``' and '``llvm.write_register``' intrinsics
7031 provides access to the named register. The register must be valid on
7032 the architecture being compiled to. The type needs to be compatible
7033 with the register being read.
7038 The '``llvm.read_register``' intrinsic returns the current value of the
7039 register, where possible. The '``llvm.write_register``' intrinsic sets
7040 the current value of the register, where possible.
7042 This is useful to implement named register global variables that need
7043 to always be mapped to a specific register, as is common practice on
7044 bare-metal programs including OS kernels.
7046 The compiler doesn't check for register availability or use of the used
7047 register in surrounding code, including inline assembly. Because of that,
7048 allocatable registers are not supported.
7050 Warning: So far it only works with the stack pointer on selected
7051 architectures (ARM, AArch64, PowerPC and x86_64). Significant amount of
7052 work is needed to support other registers and even more so, allocatable
7057 '``llvm.stacksave``' Intrinsic
7058 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7065 declare i8* @llvm.stacksave()
7070 The '``llvm.stacksave``' intrinsic is used to remember the current state
7071 of the function stack, for use with
7072 :ref:`llvm.stackrestore <int_stackrestore>`. This is useful for
7073 implementing language features like scoped automatic variable sized
7079 This intrinsic returns a opaque pointer value that can be passed to
7080 :ref:`llvm.stackrestore <int_stackrestore>`. When an
7081 ``llvm.stackrestore`` intrinsic is executed with a value saved from
7082 ``llvm.stacksave``, it effectively restores the state of the stack to
7083 the state it was in when the ``llvm.stacksave`` intrinsic executed. In
7084 practice, this pops any :ref:`alloca <i_alloca>` blocks from the stack that
7085 were allocated after the ``llvm.stacksave`` was executed.
7087 .. _int_stackrestore:
7089 '``llvm.stackrestore``' Intrinsic
7090 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7097 declare void @llvm.stackrestore(i8* %ptr)
7102 The '``llvm.stackrestore``' intrinsic is used to restore the state of
7103 the function stack to the state it was in when the corresponding
7104 :ref:`llvm.stacksave <int_stacksave>` intrinsic executed. This is
7105 useful for implementing language features like scoped automatic variable
7106 sized arrays in C99.
7111 See the description for :ref:`llvm.stacksave <int_stacksave>`.
7113 '``llvm.prefetch``' Intrinsic
7114 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7121 declare void @llvm.prefetch(i8* <address>, i32 <rw>, i32 <locality>, i32 <cache type>)
7126 The '``llvm.prefetch``' intrinsic is a hint to the code generator to
7127 insert a prefetch instruction if supported; otherwise, it is a noop.
7128 Prefetches have no effect on the behavior of the program but can change
7129 its performance characteristics.
7134 ``address`` is the address to be prefetched, ``rw`` is the specifier
7135 determining if the fetch should be for a read (0) or write (1), and
7136 ``locality`` is a temporal locality specifier ranging from (0) - no
7137 locality, to (3) - extremely local keep in cache. The ``cache type``
7138 specifies whether the prefetch is performed on the data (1) or
7139 instruction (0) cache. The ``rw``, ``locality`` and ``cache type``
7140 arguments must be constant integers.
7145 This intrinsic does not modify the behavior of the program. In
7146 particular, prefetches cannot trap and do not produce a value. On
7147 targets that support this intrinsic, the prefetch can provide hints to
7148 the processor cache for better performance.
7150 '``llvm.pcmarker``' Intrinsic
7151 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7158 declare void @llvm.pcmarker(i32 <id>)
7163 The '``llvm.pcmarker``' intrinsic is a method to export a Program
7164 Counter (PC) in a region of code to simulators and other tools. The
7165 method is target specific, but it is expected that the marker will use
7166 exported symbols to transmit the PC of the marker. The marker makes no
7167 guarantees that it will remain with any specific instruction after
7168 optimizations. It is possible that the presence of a marker will inhibit
7169 optimizations. The intended use is to be inserted after optimizations to
7170 allow correlations of simulation runs.
7175 ``id`` is a numerical id identifying the marker.
7180 This intrinsic does not modify the behavior of the program. Backends
7181 that do not support this intrinsic may ignore it.
7183 '``llvm.readcyclecounter``' Intrinsic
7184 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7191 declare i64 @llvm.readcyclecounter()
7196 The '``llvm.readcyclecounter``' intrinsic provides access to the cycle
7197 counter register (or similar low latency, high accuracy clocks) on those
7198 targets that support it. On X86, it should map to RDTSC. On Alpha, it
7199 should map to RPCC. As the backing counters overflow quickly (on the
7200 order of 9 seconds on alpha), this should only be used for small
7206 When directly supported, reading the cycle counter should not modify any
7207 memory. Implementations are allowed to either return a application
7208 specific value or a system wide value. On backends without support, this
7209 is lowered to a constant 0.
7211 Note that runtime support may be conditional on the privilege-level code is
7212 running at and the host platform.
7214 '``llvm.clear_cache``' Intrinsic
7215 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7222 declare void @llvm.clear_cache(i8*, i8*)
7227 The '``llvm.clear_cache``' intrinsic ensures visibility of modifications
7228 in the specified range to the execution unit of the processor. On
7229 targets with non-unified instruction and data cache, the implementation
7230 flushes the instruction cache.
7235 On platforms with coherent instruction and data caches (e.g. x86), this
7236 intrinsic is a nop. On platforms with non-coherent instruction and data
7237 cache (e.g. ARM, MIPS), the intrinsic is lowered either to appropriate
7238 instructions or a system call, if cache flushing requires special
7241 The default behavior is to emit a call to ``__clear_cache`` from the run
7244 This instrinsic does *not* empty the instruction pipeline. Modifications
7245 of the current function are outside the scope of the intrinsic.
7247 Standard C Library Intrinsics
7248 -----------------------------
7250 LLVM provides intrinsics for a few important standard C library
7251 functions. These intrinsics allow source-language front-ends to pass
7252 information about the alignment of the pointer arguments to the code
7253 generator, providing opportunity for more efficient code generation.
7257 '``llvm.memcpy``' Intrinsic
7258 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7263 This is an overloaded intrinsic. You can use ``llvm.memcpy`` on any
7264 integer bit width and for different address spaces. Not all targets
7265 support all bit widths however.
7269 declare void @llvm.memcpy.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
7270 i32 <len>, i32 <align>, i1 <isvolatile>)
7271 declare void @llvm.memcpy.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
7272 i64 <len>, i32 <align>, i1 <isvolatile>)
7277 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
7278 source location to the destination location.
7280 Note that, unlike the standard libc function, the ``llvm.memcpy.*``
7281 intrinsics do not return a value, takes extra alignment/isvolatile
7282 arguments and the pointers can be in specified address spaces.
7287 The first argument is a pointer to the destination, the second is a
7288 pointer to the source. The third argument is an integer argument
7289 specifying the number of bytes to copy, the fourth argument is the
7290 alignment of the source and destination locations, and the fifth is a
7291 boolean indicating a volatile access.
7293 If the call to this intrinsic has an alignment value that is not 0 or 1,
7294 then the caller guarantees that both the source and destination pointers
7295 are aligned to that boundary.
7297 If the ``isvolatile`` parameter is ``true``, the ``llvm.memcpy`` call is
7298 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
7299 very cleanly specified and it is unwise to depend on it.
7304 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
7305 source location to the destination location, which are not allowed to
7306 overlap. It copies "len" bytes of memory over. If the argument is known
7307 to be aligned to some boundary, this can be specified as the fourth
7308 argument, otherwise it should be set to 0 or 1 (both meaning no alignment).
7310 '``llvm.memmove``' Intrinsic
7311 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7316 This is an overloaded intrinsic. You can use llvm.memmove on any integer
7317 bit width and for different address space. Not all targets support all
7322 declare void @llvm.memmove.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
7323 i32 <len>, i32 <align>, i1 <isvolatile>)
7324 declare void @llvm.memmove.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
7325 i64 <len>, i32 <align>, i1 <isvolatile>)
7330 The '``llvm.memmove.*``' intrinsics move a block of memory from the
7331 source location to the destination location. It is similar to the
7332 '``llvm.memcpy``' intrinsic but allows the two memory locations to
7335 Note that, unlike the standard libc function, the ``llvm.memmove.*``
7336 intrinsics do not return a value, takes extra alignment/isvolatile
7337 arguments and the pointers can be in specified address spaces.
7342 The first argument is a pointer to the destination, the second is a
7343 pointer to the source. The third argument is an integer argument
7344 specifying the number of bytes to copy, the fourth argument is the
7345 alignment of the source and destination locations, and the fifth is a
7346 boolean indicating a volatile access.
7348 If the call to this intrinsic has an alignment value that is not 0 or 1,
7349 then the caller guarantees that the source and destination pointers are
7350 aligned to that boundary.
7352 If the ``isvolatile`` parameter is ``true``, the ``llvm.memmove`` call
7353 is a :ref:`volatile operation <volatile>`. The detailed access behavior is
7354 not very cleanly specified and it is unwise to depend on it.
7359 The '``llvm.memmove.*``' intrinsics copy a block of memory from the
7360 source location to the destination location, which may overlap. It
7361 copies "len" bytes of memory over. If the argument is known to be
7362 aligned to some boundary, this can be specified as the fourth argument,
7363 otherwise it should be set to 0 or 1 (both meaning no alignment).
7365 '``llvm.memset.*``' Intrinsics
7366 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7371 This is an overloaded intrinsic. You can use llvm.memset on any integer
7372 bit width and for different address spaces. However, not all targets
7373 support all bit widths.
7377 declare void @llvm.memset.p0i8.i32(i8* <dest>, i8 <val>,
7378 i32 <len>, i32 <align>, i1 <isvolatile>)
7379 declare void @llvm.memset.p0i8.i64(i8* <dest>, i8 <val>,
7380 i64 <len>, i32 <align>, i1 <isvolatile>)
7385 The '``llvm.memset.*``' intrinsics fill a block of memory with a
7386 particular byte value.
7388 Note that, unlike the standard libc function, the ``llvm.memset``
7389 intrinsic does not return a value and takes extra alignment/volatile
7390 arguments. Also, the destination can be in an arbitrary address space.
7395 The first argument is a pointer to the destination to fill, the second
7396 is the byte value with which to fill it, the third argument is an
7397 integer argument specifying the number of bytes to fill, and the fourth
7398 argument is the known alignment of the destination location.
7400 If the call to this intrinsic has an alignment value that is not 0 or 1,
7401 then the caller guarantees that the destination pointer is aligned to
7404 If the ``isvolatile`` parameter is ``true``, the ``llvm.memset`` call is
7405 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
7406 very cleanly specified and it is unwise to depend on it.
7411 The '``llvm.memset.*``' intrinsics fill "len" bytes of memory starting
7412 at the destination location. If the argument is known to be aligned to
7413 some boundary, this can be specified as the fourth argument, otherwise
7414 it should be set to 0 or 1 (both meaning no alignment).
7416 '``llvm.sqrt.*``' Intrinsic
7417 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7422 This is an overloaded intrinsic. You can use ``llvm.sqrt`` on any
7423 floating point or vector of floating point type. Not all targets support
7428 declare float @llvm.sqrt.f32(float %Val)
7429 declare double @llvm.sqrt.f64(double %Val)
7430 declare x86_fp80 @llvm.sqrt.f80(x86_fp80 %Val)
7431 declare fp128 @llvm.sqrt.f128(fp128 %Val)
7432 declare ppc_fp128 @llvm.sqrt.ppcf128(ppc_fp128 %Val)
7437 The '``llvm.sqrt``' intrinsics return the sqrt of the specified operand,
7438 returning the same value as the libm '``sqrt``' functions would. Unlike
7439 ``sqrt`` in libm, however, ``llvm.sqrt`` has undefined behavior for
7440 negative numbers other than -0.0 (which allows for better optimization,
7441 because there is no need to worry about errno being set).
7442 ``llvm.sqrt(-0.0)`` is defined to return -0.0 like IEEE sqrt.
7447 The argument and return value are floating point numbers of the same
7453 This function returns the sqrt of the specified operand if it is a
7454 nonnegative floating point number.
7456 '``llvm.powi.*``' Intrinsic
7457 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7462 This is an overloaded intrinsic. You can use ``llvm.powi`` on any
7463 floating point or vector of floating point type. Not all targets support
7468 declare float @llvm.powi.f32(float %Val, i32 %power)
7469 declare double @llvm.powi.f64(double %Val, i32 %power)
7470 declare x86_fp80 @llvm.powi.f80(x86_fp80 %Val, i32 %power)
7471 declare fp128 @llvm.powi.f128(fp128 %Val, i32 %power)
7472 declare ppc_fp128 @llvm.powi.ppcf128(ppc_fp128 %Val, i32 %power)
7477 The '``llvm.powi.*``' intrinsics return the first operand raised to the
7478 specified (positive or negative) power. The order of evaluation of
7479 multiplications is not defined. When a vector of floating point type is
7480 used, the second argument remains a scalar integer value.
7485 The second argument is an integer power, and the first is a value to
7486 raise to that power.
7491 This function returns the first value raised to the second power with an
7492 unspecified sequence of rounding operations.
7494 '``llvm.sin.*``' Intrinsic
7495 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7500 This is an overloaded intrinsic. You can use ``llvm.sin`` on any
7501 floating point or vector of floating point type. Not all targets support
7506 declare float @llvm.sin.f32(float %Val)
7507 declare double @llvm.sin.f64(double %Val)
7508 declare x86_fp80 @llvm.sin.f80(x86_fp80 %Val)
7509 declare fp128 @llvm.sin.f128(fp128 %Val)
7510 declare ppc_fp128 @llvm.sin.ppcf128(ppc_fp128 %Val)
7515 The '``llvm.sin.*``' intrinsics return the sine of the operand.
7520 The argument and return value are floating point numbers of the same
7526 This function returns the sine of the specified operand, returning the
7527 same values as the libm ``sin`` functions would, and handles error
7528 conditions in the same way.
7530 '``llvm.cos.*``' Intrinsic
7531 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7536 This is an overloaded intrinsic. You can use ``llvm.cos`` on any
7537 floating point or vector of floating point type. Not all targets support
7542 declare float @llvm.cos.f32(float %Val)
7543 declare double @llvm.cos.f64(double %Val)
7544 declare x86_fp80 @llvm.cos.f80(x86_fp80 %Val)
7545 declare fp128 @llvm.cos.f128(fp128 %Val)
7546 declare ppc_fp128 @llvm.cos.ppcf128(ppc_fp128 %Val)
7551 The '``llvm.cos.*``' intrinsics return the cosine of the operand.
7556 The argument and return value are floating point numbers of the same
7562 This function returns the cosine of the specified operand, returning the
7563 same values as the libm ``cos`` functions would, and handles error
7564 conditions in the same way.
7566 '``llvm.pow.*``' Intrinsic
7567 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7572 This is an overloaded intrinsic. You can use ``llvm.pow`` on any
7573 floating point or vector of floating point type. Not all targets support
7578 declare float @llvm.pow.f32(float %Val, float %Power)
7579 declare double @llvm.pow.f64(double %Val, double %Power)
7580 declare x86_fp80 @llvm.pow.f80(x86_fp80 %Val, x86_fp80 %Power)
7581 declare fp128 @llvm.pow.f128(fp128 %Val, fp128 %Power)
7582 declare ppc_fp128 @llvm.pow.ppcf128(ppc_fp128 %Val, ppc_fp128 Power)
7587 The '``llvm.pow.*``' intrinsics return the first operand raised to the
7588 specified (positive or negative) power.
7593 The second argument is a floating point power, and the first is a value
7594 to raise to that power.
7599 This function returns the first value raised to the second power,
7600 returning the same values as the libm ``pow`` functions would, and
7601 handles error conditions in the same way.
7603 '``llvm.exp.*``' Intrinsic
7604 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7609 This is an overloaded intrinsic. You can use ``llvm.exp`` on any
7610 floating point or vector of floating point type. Not all targets support
7615 declare float @llvm.exp.f32(float %Val)
7616 declare double @llvm.exp.f64(double %Val)
7617 declare x86_fp80 @llvm.exp.f80(x86_fp80 %Val)
7618 declare fp128 @llvm.exp.f128(fp128 %Val)
7619 declare ppc_fp128 @llvm.exp.ppcf128(ppc_fp128 %Val)
7624 The '``llvm.exp.*``' intrinsics perform the exp function.
7629 The argument and return value are floating point numbers of the same
7635 This function returns the same values as the libm ``exp`` functions
7636 would, and handles error conditions in the same way.
7638 '``llvm.exp2.*``' Intrinsic
7639 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7644 This is an overloaded intrinsic. You can use ``llvm.exp2`` on any
7645 floating point or vector of floating point type. Not all targets support
7650 declare float @llvm.exp2.f32(float %Val)
7651 declare double @llvm.exp2.f64(double %Val)
7652 declare x86_fp80 @llvm.exp2.f80(x86_fp80 %Val)
7653 declare fp128 @llvm.exp2.f128(fp128 %Val)
7654 declare ppc_fp128 @llvm.exp2.ppcf128(ppc_fp128 %Val)
7659 The '``llvm.exp2.*``' intrinsics perform the exp2 function.
7664 The argument and return value are floating point numbers of the same
7670 This function returns the same values as the libm ``exp2`` functions
7671 would, and handles error conditions in the same way.
7673 '``llvm.log.*``' Intrinsic
7674 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7679 This is an overloaded intrinsic. You can use ``llvm.log`` on any
7680 floating point or vector of floating point type. Not all targets support
7685 declare float @llvm.log.f32(float %Val)
7686 declare double @llvm.log.f64(double %Val)
7687 declare x86_fp80 @llvm.log.f80(x86_fp80 %Val)
7688 declare fp128 @llvm.log.f128(fp128 %Val)
7689 declare ppc_fp128 @llvm.log.ppcf128(ppc_fp128 %Val)
7694 The '``llvm.log.*``' intrinsics perform the log function.
7699 The argument and return value are floating point numbers of the same
7705 This function returns the same values as the libm ``log`` functions
7706 would, and handles error conditions in the same way.
7708 '``llvm.log10.*``' Intrinsic
7709 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7714 This is an overloaded intrinsic. You can use ``llvm.log10`` on any
7715 floating point or vector of floating point type. Not all targets support
7720 declare float @llvm.log10.f32(float %Val)
7721 declare double @llvm.log10.f64(double %Val)
7722 declare x86_fp80 @llvm.log10.f80(x86_fp80 %Val)
7723 declare fp128 @llvm.log10.f128(fp128 %Val)
7724 declare ppc_fp128 @llvm.log10.ppcf128(ppc_fp128 %Val)
7729 The '``llvm.log10.*``' intrinsics perform the log10 function.
7734 The argument and return value are floating point numbers of the same
7740 This function returns the same values as the libm ``log10`` functions
7741 would, and handles error conditions in the same way.
7743 '``llvm.log2.*``' Intrinsic
7744 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7749 This is an overloaded intrinsic. You can use ``llvm.log2`` on any
7750 floating point or vector of floating point type. Not all targets support
7755 declare float @llvm.log2.f32(float %Val)
7756 declare double @llvm.log2.f64(double %Val)
7757 declare x86_fp80 @llvm.log2.f80(x86_fp80 %Val)
7758 declare fp128 @llvm.log2.f128(fp128 %Val)
7759 declare ppc_fp128 @llvm.log2.ppcf128(ppc_fp128 %Val)
7764 The '``llvm.log2.*``' intrinsics perform the log2 function.
7769 The argument and return value are floating point numbers of the same
7775 This function returns the same values as the libm ``log2`` functions
7776 would, and handles error conditions in the same way.
7778 '``llvm.fma.*``' Intrinsic
7779 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7784 This is an overloaded intrinsic. You can use ``llvm.fma`` on any
7785 floating point or vector of floating point type. Not all targets support
7790 declare float @llvm.fma.f32(float %a, float %b, float %c)
7791 declare double @llvm.fma.f64(double %a, double %b, double %c)
7792 declare x86_fp80 @llvm.fma.f80(x86_fp80 %a, x86_fp80 %b, x86_fp80 %c)
7793 declare fp128 @llvm.fma.f128(fp128 %a, fp128 %b, fp128 %c)
7794 declare ppc_fp128 @llvm.fma.ppcf128(ppc_fp128 %a, ppc_fp128 %b, ppc_fp128 %c)
7799 The '``llvm.fma.*``' intrinsics perform the fused multiply-add
7805 The argument and return value are floating point numbers of the same
7811 This function returns the same values as the libm ``fma`` functions
7812 would, and does not set errno.
7814 '``llvm.fabs.*``' Intrinsic
7815 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7820 This is an overloaded intrinsic. You can use ``llvm.fabs`` on any
7821 floating point or vector of floating point type. Not all targets support
7826 declare float @llvm.fabs.f32(float %Val)
7827 declare double @llvm.fabs.f64(double %Val)
7828 declare x86_fp80 @llvm.fabs.f80(x86_fp80 %Val)
7829 declare fp128 @llvm.fabs.f128(fp128 %Val)
7830 declare ppc_fp128 @llvm.fabs.ppcf128(ppc_fp128 %Val)
7835 The '``llvm.fabs.*``' intrinsics return the absolute value of the
7841 The argument and return value are floating point numbers of the same
7847 This function returns the same values as the libm ``fabs`` functions
7848 would, and handles error conditions in the same way.
7850 '``llvm.copysign.*``' Intrinsic
7851 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7856 This is an overloaded intrinsic. You can use ``llvm.copysign`` on any
7857 floating point or vector of floating point type. Not all targets support
7862 declare float @llvm.copysign.f32(float %Mag, float %Sgn)
7863 declare double @llvm.copysign.f64(double %Mag, double %Sgn)
7864 declare x86_fp80 @llvm.copysign.f80(x86_fp80 %Mag, x86_fp80 %Sgn)
7865 declare fp128 @llvm.copysign.f128(fp128 %Mag, fp128 %Sgn)
7866 declare ppc_fp128 @llvm.copysign.ppcf128(ppc_fp128 %Mag, ppc_fp128 %Sgn)
7871 The '``llvm.copysign.*``' intrinsics return a value with the magnitude of the
7872 first operand and the sign of the second operand.
7877 The arguments and return value are floating point numbers of the same
7883 This function returns the same values as the libm ``copysign``
7884 functions would, and handles error conditions in the same way.
7886 '``llvm.floor.*``' Intrinsic
7887 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7892 This is an overloaded intrinsic. You can use ``llvm.floor`` on any
7893 floating point or vector of floating point type. Not all targets support
7898 declare float @llvm.floor.f32(float %Val)
7899 declare double @llvm.floor.f64(double %Val)
7900 declare x86_fp80 @llvm.floor.f80(x86_fp80 %Val)
7901 declare fp128 @llvm.floor.f128(fp128 %Val)
7902 declare ppc_fp128 @llvm.floor.ppcf128(ppc_fp128 %Val)
7907 The '``llvm.floor.*``' intrinsics return the floor of the operand.
7912 The argument and return value are floating point numbers of the same
7918 This function returns the same values as the libm ``floor`` functions
7919 would, and handles error conditions in the same way.
7921 '``llvm.ceil.*``' Intrinsic
7922 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7927 This is an overloaded intrinsic. You can use ``llvm.ceil`` on any
7928 floating point or vector of floating point type. Not all targets support
7933 declare float @llvm.ceil.f32(float %Val)
7934 declare double @llvm.ceil.f64(double %Val)
7935 declare x86_fp80 @llvm.ceil.f80(x86_fp80 %Val)
7936 declare fp128 @llvm.ceil.f128(fp128 %Val)
7937 declare ppc_fp128 @llvm.ceil.ppcf128(ppc_fp128 %Val)
7942 The '``llvm.ceil.*``' intrinsics return the ceiling of the operand.
7947 The argument and return value are floating point numbers of the same
7953 This function returns the same values as the libm ``ceil`` functions
7954 would, and handles error conditions in the same way.
7956 '``llvm.trunc.*``' Intrinsic
7957 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7962 This is an overloaded intrinsic. You can use ``llvm.trunc`` on any
7963 floating point or vector of floating point type. Not all targets support
7968 declare float @llvm.trunc.f32(float %Val)
7969 declare double @llvm.trunc.f64(double %Val)
7970 declare x86_fp80 @llvm.trunc.f80(x86_fp80 %Val)
7971 declare fp128 @llvm.trunc.f128(fp128 %Val)
7972 declare ppc_fp128 @llvm.trunc.ppcf128(ppc_fp128 %Val)
7977 The '``llvm.trunc.*``' intrinsics returns the operand rounded to the
7978 nearest integer not larger in magnitude than the operand.
7983 The argument and return value are floating point numbers of the same
7989 This function returns the same values as the libm ``trunc`` functions
7990 would, and handles error conditions in the same way.
7992 '``llvm.rint.*``' Intrinsic
7993 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7998 This is an overloaded intrinsic. You can use ``llvm.rint`` on any
7999 floating point or vector of floating point type. Not all targets support
8004 declare float @llvm.rint.f32(float %Val)
8005 declare double @llvm.rint.f64(double %Val)
8006 declare x86_fp80 @llvm.rint.f80(x86_fp80 %Val)
8007 declare fp128 @llvm.rint.f128(fp128 %Val)
8008 declare ppc_fp128 @llvm.rint.ppcf128(ppc_fp128 %Val)
8013 The '``llvm.rint.*``' intrinsics returns the operand rounded to the
8014 nearest integer. It may raise an inexact floating-point exception if the
8015 operand isn't an integer.
8020 The argument and return value are floating point numbers of the same
8026 This function returns the same values as the libm ``rint`` functions
8027 would, and handles error conditions in the same way.
8029 '``llvm.nearbyint.*``' Intrinsic
8030 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8035 This is an overloaded intrinsic. You can use ``llvm.nearbyint`` on any
8036 floating point or vector of floating point type. Not all targets support
8041 declare float @llvm.nearbyint.f32(float %Val)
8042 declare double @llvm.nearbyint.f64(double %Val)
8043 declare x86_fp80 @llvm.nearbyint.f80(x86_fp80 %Val)
8044 declare fp128 @llvm.nearbyint.f128(fp128 %Val)
8045 declare ppc_fp128 @llvm.nearbyint.ppcf128(ppc_fp128 %Val)
8050 The '``llvm.nearbyint.*``' intrinsics returns the operand rounded to the
8056 The argument and return value are floating point numbers of the same
8062 This function returns the same values as the libm ``nearbyint``
8063 functions would, and handles error conditions in the same way.
8065 '``llvm.round.*``' Intrinsic
8066 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8071 This is an overloaded intrinsic. You can use ``llvm.round`` on any
8072 floating point or vector of floating point type. Not all targets support
8077 declare float @llvm.round.f32(float %Val)
8078 declare double @llvm.round.f64(double %Val)
8079 declare x86_fp80 @llvm.round.f80(x86_fp80 %Val)
8080 declare fp128 @llvm.round.f128(fp128 %Val)
8081 declare ppc_fp128 @llvm.round.ppcf128(ppc_fp128 %Val)
8086 The '``llvm.round.*``' intrinsics returns the operand rounded to the
8092 The argument and return value are floating point numbers of the same
8098 This function returns the same values as the libm ``round``
8099 functions would, and handles error conditions in the same way.
8101 Bit Manipulation Intrinsics
8102 ---------------------------
8104 LLVM provides intrinsics for a few important bit manipulation
8105 operations. These allow efficient code generation for some algorithms.
8107 '``llvm.bswap.*``' Intrinsics
8108 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8113 This is an overloaded intrinsic function. You can use bswap on any
8114 integer type that is an even number of bytes (i.e. BitWidth % 16 == 0).
8118 declare i16 @llvm.bswap.i16(i16 <id>)
8119 declare i32 @llvm.bswap.i32(i32 <id>)
8120 declare i64 @llvm.bswap.i64(i64 <id>)
8125 The '``llvm.bswap``' family of intrinsics is used to byte swap integer
8126 values with an even number of bytes (positive multiple of 16 bits).
8127 These are useful for performing operations on data that is not in the
8128 target's native byte order.
8133 The ``llvm.bswap.i16`` intrinsic returns an i16 value that has the high
8134 and low byte of the input i16 swapped. Similarly, the ``llvm.bswap.i32``
8135 intrinsic returns an i32 value that has the four bytes of the input i32
8136 swapped, so that if the input bytes are numbered 0, 1, 2, 3 then the
8137 returned i32 will have its bytes in 3, 2, 1, 0 order. The
8138 ``llvm.bswap.i48``, ``llvm.bswap.i64`` and other intrinsics extend this
8139 concept to additional even-byte lengths (6 bytes, 8 bytes and more,
8142 '``llvm.ctpop.*``' Intrinsic
8143 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8148 This is an overloaded intrinsic. You can use llvm.ctpop on any integer
8149 bit width, or on any vector with integer elements. Not all targets
8150 support all bit widths or vector types, however.
8154 declare i8 @llvm.ctpop.i8(i8 <src>)
8155 declare i16 @llvm.ctpop.i16(i16 <src>)
8156 declare i32 @llvm.ctpop.i32(i32 <src>)
8157 declare i64 @llvm.ctpop.i64(i64 <src>)
8158 declare i256 @llvm.ctpop.i256(i256 <src>)
8159 declare <2 x i32> @llvm.ctpop.v2i32(<2 x i32> <src>)
8164 The '``llvm.ctpop``' family of intrinsics counts the number of bits set
8170 The only argument is the value to be counted. The argument may be of any
8171 integer type, or a vector with integer elements. The return type must
8172 match the argument type.
8177 The '``llvm.ctpop``' intrinsic counts the 1's in a variable, or within
8178 each element of a vector.
8180 '``llvm.ctlz.*``' Intrinsic
8181 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8186 This is an overloaded intrinsic. You can use ``llvm.ctlz`` on any
8187 integer bit width, or any vector whose elements are integers. Not all
8188 targets support all bit widths or vector types, however.
8192 declare i8 @llvm.ctlz.i8 (i8 <src>, i1 <is_zero_undef>)
8193 declare i16 @llvm.ctlz.i16 (i16 <src>, i1 <is_zero_undef>)
8194 declare i32 @llvm.ctlz.i32 (i32 <src>, i1 <is_zero_undef>)
8195 declare i64 @llvm.ctlz.i64 (i64 <src>, i1 <is_zero_undef>)
8196 declare i256 @llvm.ctlz.i256(i256 <src>, i1 <is_zero_undef>)
8197 declase <2 x i32> @llvm.ctlz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
8202 The '``llvm.ctlz``' family of intrinsic functions counts the number of
8203 leading zeros in a variable.
8208 The first argument is the value to be counted. This argument may be of
8209 any integer type, or a vectory with integer element type. The return
8210 type must match the first argument type.
8212 The second argument must be a constant and is a flag to indicate whether
8213 the intrinsic should ensure that a zero as the first argument produces a
8214 defined result. Historically some architectures did not provide a
8215 defined result for zero values as efficiently, and many algorithms are
8216 now predicated on avoiding zero-value inputs.
8221 The '``llvm.ctlz``' intrinsic counts the leading (most significant)
8222 zeros in a variable, or within each element of the vector. If
8223 ``src == 0`` then the result is the size in bits of the type of ``src``
8224 if ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
8225 ``llvm.ctlz(i32 2) = 30``.
8227 '``llvm.cttz.*``' Intrinsic
8228 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8233 This is an overloaded intrinsic. You can use ``llvm.cttz`` on any
8234 integer bit width, or any vector of integer elements. Not all targets
8235 support all bit widths or vector types, however.
8239 declare i8 @llvm.cttz.i8 (i8 <src>, i1 <is_zero_undef>)
8240 declare i16 @llvm.cttz.i16 (i16 <src>, i1 <is_zero_undef>)
8241 declare i32 @llvm.cttz.i32 (i32 <src>, i1 <is_zero_undef>)
8242 declare i64 @llvm.cttz.i64 (i64 <src>, i1 <is_zero_undef>)
8243 declare i256 @llvm.cttz.i256(i256 <src>, i1 <is_zero_undef>)
8244 declase <2 x i32> @llvm.cttz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
8249 The '``llvm.cttz``' family of intrinsic functions counts the number of
8255 The first argument is the value to be counted. This argument may be of
8256 any integer type, or a vectory with integer element type. The return
8257 type must match the first argument type.
8259 The second argument must be a constant and is a flag to indicate whether
8260 the intrinsic should ensure that a zero as the first argument produces a
8261 defined result. Historically some architectures did not provide a
8262 defined result for zero values as efficiently, and many algorithms are
8263 now predicated on avoiding zero-value inputs.
8268 The '``llvm.cttz``' intrinsic counts the trailing (least significant)
8269 zeros in a variable, or within each element of a vector. If ``src == 0``
8270 then the result is the size in bits of the type of ``src`` if
8271 ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
8272 ``llvm.cttz(2) = 1``.
8274 Arithmetic with Overflow Intrinsics
8275 -----------------------------------
8277 LLVM provides intrinsics for some arithmetic with overflow operations.
8279 '``llvm.sadd.with.overflow.*``' Intrinsics
8280 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8285 This is an overloaded intrinsic. You can use ``llvm.sadd.with.overflow``
8286 on any integer bit width.
8290 declare {i16, i1} @llvm.sadd.with.overflow.i16(i16 %a, i16 %b)
8291 declare {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
8292 declare {i64, i1} @llvm.sadd.with.overflow.i64(i64 %a, i64 %b)
8297 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
8298 a signed addition of the two arguments, and indicate whether an overflow
8299 occurred during the signed summation.
8304 The arguments (%a and %b) and the first element of the result structure
8305 may be of integer types of any bit width, but they must have the same
8306 bit width. The second element of the result structure must be of type
8307 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
8313 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
8314 a signed addition of the two variables. They return a structure --- the
8315 first element of which is the signed summation, and the second element
8316 of which is a bit specifying if the signed summation resulted in an
8322 .. code-block:: llvm
8324 %res = call {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
8325 %sum = extractvalue {i32, i1} %res, 0
8326 %obit = extractvalue {i32, i1} %res, 1
8327 br i1 %obit, label %overflow, label %normal
8329 '``llvm.uadd.with.overflow.*``' Intrinsics
8330 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8335 This is an overloaded intrinsic. You can use ``llvm.uadd.with.overflow``
8336 on any integer bit width.
8340 declare {i16, i1} @llvm.uadd.with.overflow.i16(i16 %a, i16 %b)
8341 declare {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
8342 declare {i64, i1} @llvm.uadd.with.overflow.i64(i64 %a, i64 %b)
8347 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
8348 an unsigned addition of the two arguments, and indicate whether a carry
8349 occurred during the unsigned summation.
8354 The arguments (%a and %b) and the first element of the result structure
8355 may be of integer types of any bit width, but they must have the same
8356 bit width. The second element of the result structure must be of type
8357 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
8363 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
8364 an unsigned addition of the two arguments. They return a structure --- the
8365 first element of which is the sum, and the second element of which is a
8366 bit specifying if the unsigned summation resulted in a carry.
8371 .. code-block:: llvm
8373 %res = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
8374 %sum = extractvalue {i32, i1} %res, 0
8375 %obit = extractvalue {i32, i1} %res, 1
8376 br i1 %obit, label %carry, label %normal
8378 '``llvm.ssub.with.overflow.*``' Intrinsics
8379 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8384 This is an overloaded intrinsic. You can use ``llvm.ssub.with.overflow``
8385 on any integer bit width.
8389 declare {i16, i1} @llvm.ssub.with.overflow.i16(i16 %a, i16 %b)
8390 declare {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
8391 declare {i64, i1} @llvm.ssub.with.overflow.i64(i64 %a, i64 %b)
8396 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
8397 a signed subtraction of the two arguments, and indicate whether an
8398 overflow occurred during the signed subtraction.
8403 The arguments (%a and %b) and the first element of the result structure
8404 may be of integer types of any bit width, but they must have the same
8405 bit width. The second element of the result structure must be of type
8406 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
8412 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
8413 a signed subtraction of the two arguments. They return a structure --- the
8414 first element of which is the subtraction, and the second element of
8415 which is a bit specifying if the signed subtraction resulted in an
8421 .. code-block:: llvm
8423 %res = call {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
8424 %sum = extractvalue {i32, i1} %res, 0
8425 %obit = extractvalue {i32, i1} %res, 1
8426 br i1 %obit, label %overflow, label %normal
8428 '``llvm.usub.with.overflow.*``' Intrinsics
8429 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8434 This is an overloaded intrinsic. You can use ``llvm.usub.with.overflow``
8435 on any integer bit width.
8439 declare {i16, i1} @llvm.usub.with.overflow.i16(i16 %a, i16 %b)
8440 declare {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
8441 declare {i64, i1} @llvm.usub.with.overflow.i64(i64 %a, i64 %b)
8446 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
8447 an unsigned subtraction of the two arguments, and indicate whether an
8448 overflow occurred during the unsigned subtraction.
8453 The arguments (%a and %b) and the first element of the result structure
8454 may be of integer types of any bit width, but they must have the same
8455 bit width. The second element of the result structure must be of type
8456 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
8462 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
8463 an unsigned subtraction of the two arguments. They return a structure ---
8464 the first element of which is the subtraction, and the second element of
8465 which is a bit specifying if the unsigned subtraction resulted in an
8471 .. code-block:: llvm
8473 %res = call {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
8474 %sum = extractvalue {i32, i1} %res, 0
8475 %obit = extractvalue {i32, i1} %res, 1
8476 br i1 %obit, label %overflow, label %normal
8478 '``llvm.smul.with.overflow.*``' Intrinsics
8479 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8484 This is an overloaded intrinsic. You can use ``llvm.smul.with.overflow``
8485 on any integer bit width.
8489 declare {i16, i1} @llvm.smul.with.overflow.i16(i16 %a, i16 %b)
8490 declare {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
8491 declare {i64, i1} @llvm.smul.with.overflow.i64(i64 %a, i64 %b)
8496 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
8497 a signed multiplication of the two arguments, and indicate whether an
8498 overflow occurred during the signed multiplication.
8503 The arguments (%a and %b) and the first element of the result structure
8504 may be of integer types of any bit width, but they must have the same
8505 bit width. The second element of the result structure must be of type
8506 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
8512 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
8513 a signed multiplication of the two arguments. They return a structure ---
8514 the first element of which is the multiplication, and the second element
8515 of which is a bit specifying if the signed multiplication resulted in an
8521 .. code-block:: llvm
8523 %res = call {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
8524 %sum = extractvalue {i32, i1} %res, 0
8525 %obit = extractvalue {i32, i1} %res, 1
8526 br i1 %obit, label %overflow, label %normal
8528 '``llvm.umul.with.overflow.*``' Intrinsics
8529 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8534 This is an overloaded intrinsic. You can use ``llvm.umul.with.overflow``
8535 on any integer bit width.
8539 declare {i16, i1} @llvm.umul.with.overflow.i16(i16 %a, i16 %b)
8540 declare {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
8541 declare {i64, i1} @llvm.umul.with.overflow.i64(i64 %a, i64 %b)
8546 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
8547 a unsigned multiplication of the two arguments, and indicate whether an
8548 overflow occurred during the unsigned multiplication.
8553 The arguments (%a and %b) and the first element of the result structure
8554 may be of integer types of any bit width, but they must have the same
8555 bit width. The second element of the result structure must be of type
8556 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
8562 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
8563 an unsigned multiplication of the two arguments. They return a structure ---
8564 the first element of which is the multiplication, and the second
8565 element of which is a bit specifying if the unsigned multiplication
8566 resulted in an overflow.
8571 .. code-block:: llvm
8573 %res = call {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
8574 %sum = extractvalue {i32, i1} %res, 0
8575 %obit = extractvalue {i32, i1} %res, 1
8576 br i1 %obit, label %overflow, label %normal
8578 Specialised Arithmetic Intrinsics
8579 ---------------------------------
8581 '``llvm.fmuladd.*``' Intrinsic
8582 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8589 declare float @llvm.fmuladd.f32(float %a, float %b, float %c)
8590 declare double @llvm.fmuladd.f64(double %a, double %b, double %c)
8595 The '``llvm.fmuladd.*``' intrinsic functions represent multiply-add
8596 expressions that can be fused if the code generator determines that (a) the
8597 target instruction set has support for a fused operation, and (b) that the
8598 fused operation is more efficient than the equivalent, separate pair of mul
8599 and add instructions.
8604 The '``llvm.fmuladd.*``' intrinsics each take three arguments: two
8605 multiplicands, a and b, and an addend c.
8614 %0 = call float @llvm.fmuladd.f32(%a, %b, %c)
8616 is equivalent to the expression a \* b + c, except that rounding will
8617 not be performed between the multiplication and addition steps if the
8618 code generator fuses the operations. Fusion is not guaranteed, even if
8619 the target platform supports it. If a fused multiply-add is required the
8620 corresponding llvm.fma.\* intrinsic function should be used
8621 instead. This never sets errno, just as '``llvm.fma.*``'.
8626 .. code-block:: llvm
8628 %r2 = call float @llvm.fmuladd.f32(float %a, float %b, float %c) ; yields float:r2 = (a * b) + c
8630 Half Precision Floating Point Intrinsics
8631 ----------------------------------------
8633 For most target platforms, half precision floating point is a
8634 storage-only format. This means that it is a dense encoding (in memory)
8635 but does not support computation in the format.
8637 This means that code must first load the half-precision floating point
8638 value as an i16, then convert it to float with
8639 :ref:`llvm.convert.from.fp16 <int_convert_from_fp16>`. Computation can
8640 then be performed on the float value (including extending to double
8641 etc). To store the value back to memory, it is first converted to float
8642 if needed, then converted to i16 with
8643 :ref:`llvm.convert.to.fp16 <int_convert_to_fp16>`, then storing as an
8646 .. _int_convert_to_fp16:
8648 '``llvm.convert.to.fp16``' Intrinsic
8649 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8656 declare i16 @llvm.convert.to.fp16.f32(float %a)
8657 declare i16 @llvm.convert.to.fp16.f64(double %a)
8662 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion from a
8663 conventional floating point type to half precision floating point format.
8668 The intrinsic function contains single argument - the value to be
8674 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion from a
8675 conventional floating point format to half precision floating point format. The
8676 return value is an ``i16`` which contains the converted number.
8681 .. code-block:: llvm
8683 %res = call i16 @llvm.convert.to.fp16.f32(float %a)
8684 store i16 %res, i16* @x, align 2
8686 .. _int_convert_from_fp16:
8688 '``llvm.convert.from.fp16``' Intrinsic
8689 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8696 declare float @llvm.convert.from.fp16.f32(i16 %a)
8697 declare double @llvm.convert.from.fp16.f64(i16 %a)
8702 The '``llvm.convert.from.fp16``' intrinsic function performs a
8703 conversion from half precision floating point format to single precision
8704 floating point format.
8709 The intrinsic function contains single argument - the value to be
8715 The '``llvm.convert.from.fp16``' intrinsic function performs a
8716 conversion from half single precision floating point format to single
8717 precision floating point format. The input half-float value is
8718 represented by an ``i16`` value.
8723 .. code-block:: llvm
8725 %a = load i16* @x, align 2
8726 %res = call float @llvm.convert.from.fp16(i16 %a)
8731 The LLVM debugger intrinsics (which all start with ``llvm.dbg.``
8732 prefix), are described in the `LLVM Source Level
8733 Debugging <SourceLevelDebugging.html#format_common_intrinsics>`_
8736 Exception Handling Intrinsics
8737 -----------------------------
8739 The LLVM exception handling intrinsics (which all start with
8740 ``llvm.eh.`` prefix), are described in the `LLVM Exception
8741 Handling <ExceptionHandling.html#format_common_intrinsics>`_ document.
8745 Trampoline Intrinsics
8746 ---------------------
8748 These intrinsics make it possible to excise one parameter, marked with
8749 the :ref:`nest <nest>` attribute, from a function. The result is a
8750 callable function pointer lacking the nest parameter - the caller does
8751 not need to provide a value for it. Instead, the value to use is stored
8752 in advance in a "trampoline", a block of memory usually allocated on the
8753 stack, which also contains code to splice the nest value into the
8754 argument list. This is used to implement the GCC nested function address
8757 For example, if the function is ``i32 f(i8* nest %c, i32 %x, i32 %y)``
8758 then the resulting function pointer has signature ``i32 (i32, i32)*``.
8759 It can be created as follows:
8761 .. code-block:: llvm
8763 %tramp = alloca [10 x i8], align 4 ; size and alignment only correct for X86
8764 %tramp1 = getelementptr [10 x i8]* %tramp, i32 0, i32 0
8765 call i8* @llvm.init.trampoline(i8* %tramp1, i8* bitcast (i32 (i8*, i32, i32)* @f to i8*), i8* %nval)
8766 %p = call i8* @llvm.adjust.trampoline(i8* %tramp1)
8767 %fp = bitcast i8* %p to i32 (i32, i32)*
8769 The call ``%val = call i32 %fp(i32 %x, i32 %y)`` is then equivalent to
8770 ``%val = call i32 %f(i8* %nval, i32 %x, i32 %y)``.
8774 '``llvm.init.trampoline``' Intrinsic
8775 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8782 declare void @llvm.init.trampoline(i8* <tramp>, i8* <func>, i8* <nval>)
8787 This fills the memory pointed to by ``tramp`` with executable code,
8788 turning it into a trampoline.
8793 The ``llvm.init.trampoline`` intrinsic takes three arguments, all
8794 pointers. The ``tramp`` argument must point to a sufficiently large and
8795 sufficiently aligned block of memory; this memory is written to by the
8796 intrinsic. Note that the size and the alignment are target-specific -
8797 LLVM currently provides no portable way of determining them, so a
8798 front-end that generates this intrinsic needs to have some
8799 target-specific knowledge. The ``func`` argument must hold a function
8800 bitcast to an ``i8*``.
8805 The block of memory pointed to by ``tramp`` is filled with target
8806 dependent code, turning it into a function. Then ``tramp`` needs to be
8807 passed to :ref:`llvm.adjust.trampoline <int_at>` to get a pointer which can
8808 be :ref:`bitcast (to a new function) and called <int_trampoline>`. The new
8809 function's signature is the same as that of ``func`` with any arguments
8810 marked with the ``nest`` attribute removed. At most one such ``nest``
8811 argument is allowed, and it must be of pointer type. Calling the new
8812 function is equivalent to calling ``func`` with the same argument list,
8813 but with ``nval`` used for the missing ``nest`` argument. If, after
8814 calling ``llvm.init.trampoline``, the memory pointed to by ``tramp`` is
8815 modified, then the effect of any later call to the returned function
8816 pointer is undefined.
8820 '``llvm.adjust.trampoline``' Intrinsic
8821 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8828 declare i8* @llvm.adjust.trampoline(i8* <tramp>)
8833 This performs any required machine-specific adjustment to the address of
8834 a trampoline (passed as ``tramp``).
8839 ``tramp`` must point to a block of memory which already has trampoline
8840 code filled in by a previous call to
8841 :ref:`llvm.init.trampoline <int_it>`.
8846 On some architectures the address of the code to be executed needs to be
8847 different than the address where the trampoline is actually stored. This
8848 intrinsic returns the executable address corresponding to ``tramp``
8849 after performing the required machine specific adjustments. The pointer
8850 returned can then be :ref:`bitcast and executed <int_trampoline>`.
8855 This class of intrinsics provides information about the lifetime of
8856 memory objects and ranges where variables are immutable.
8860 '``llvm.lifetime.start``' Intrinsic
8861 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8868 declare void @llvm.lifetime.start(i64 <size>, i8* nocapture <ptr>)
8873 The '``llvm.lifetime.start``' intrinsic specifies the start of a memory
8879 The first argument is a constant integer representing the size of the
8880 object, or -1 if it is variable sized. The second argument is a pointer
8886 This intrinsic indicates that before this point in the code, the value
8887 of the memory pointed to by ``ptr`` is dead. This means that it is known
8888 to never be used and has an undefined value. A load from the pointer
8889 that precedes this intrinsic can be replaced with ``'undef'``.
8893 '``llvm.lifetime.end``' Intrinsic
8894 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8901 declare void @llvm.lifetime.end(i64 <size>, i8* nocapture <ptr>)
8906 The '``llvm.lifetime.end``' intrinsic specifies the end of a memory
8912 The first argument is a constant integer representing the size of the
8913 object, or -1 if it is variable sized. The second argument is a pointer
8919 This intrinsic indicates that after this point in the code, the value of
8920 the memory pointed to by ``ptr`` is dead. This means that it is known to
8921 never be used and has an undefined value. Any stores into the memory
8922 object following this intrinsic may be removed as dead.
8924 '``llvm.invariant.start``' Intrinsic
8925 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8932 declare {}* @llvm.invariant.start(i64 <size>, i8* nocapture <ptr>)
8937 The '``llvm.invariant.start``' intrinsic specifies that the contents of
8938 a memory object will not change.
8943 The first argument is a constant integer representing the size of the
8944 object, or -1 if it is variable sized. The second argument is a pointer
8950 This intrinsic indicates that until an ``llvm.invariant.end`` that uses
8951 the return value, the referenced memory location is constant and
8954 '``llvm.invariant.end``' Intrinsic
8955 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8962 declare void @llvm.invariant.end({}* <start>, i64 <size>, i8* nocapture <ptr>)
8967 The '``llvm.invariant.end``' intrinsic specifies that the contents of a
8968 memory object are mutable.
8973 The first argument is the matching ``llvm.invariant.start`` intrinsic.
8974 The second argument is a constant integer representing the size of the
8975 object, or -1 if it is variable sized and the third argument is a
8976 pointer to the object.
8981 This intrinsic indicates that the memory is mutable again.
8986 This class of intrinsics is designed to be generic and has no specific
8989 '``llvm.var.annotation``' Intrinsic
8990 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8997 declare void @llvm.var.annotation(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
9002 The '``llvm.var.annotation``' intrinsic.
9007 The first argument is a pointer to a value, the second is a pointer to a
9008 global string, the third is a pointer to a global string which is the
9009 source file name, and the last argument is the line number.
9014 This intrinsic allows annotation of local variables with arbitrary
9015 strings. This can be useful for special purpose optimizations that want
9016 to look for these annotations. These have no other defined use; they are
9017 ignored by code generation and optimization.
9019 '``llvm.ptr.annotation.*``' Intrinsic
9020 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9025 This is an overloaded intrinsic. You can use '``llvm.ptr.annotation``' on a
9026 pointer to an integer of any width. *NOTE* you must specify an address space for
9027 the pointer. The identifier for the default address space is the integer
9032 declare i8* @llvm.ptr.annotation.p<address space>i8(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
9033 declare i16* @llvm.ptr.annotation.p<address space>i16(i16* <val>, i8* <str>, i8* <str>, i32 <int>)
9034 declare i32* @llvm.ptr.annotation.p<address space>i32(i32* <val>, i8* <str>, i8* <str>, i32 <int>)
9035 declare i64* @llvm.ptr.annotation.p<address space>i64(i64* <val>, i8* <str>, i8* <str>, i32 <int>)
9036 declare i256* @llvm.ptr.annotation.p<address space>i256(i256* <val>, i8* <str>, i8* <str>, i32 <int>)
9041 The '``llvm.ptr.annotation``' intrinsic.
9046 The first argument is a pointer to an integer value of arbitrary bitwidth
9047 (result of some expression), the second is a pointer to a global string, the
9048 third is a pointer to a global string which is the source file name, and the
9049 last argument is the line number. It returns the value of the first argument.
9054 This intrinsic allows annotation of a pointer to an integer with arbitrary
9055 strings. This can be useful for special purpose optimizations that want to look
9056 for these annotations. These have no other defined use; they are ignored by code
9057 generation and optimization.
9059 '``llvm.annotation.*``' Intrinsic
9060 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9065 This is an overloaded intrinsic. You can use '``llvm.annotation``' on
9066 any integer bit width.
9070 declare i8 @llvm.annotation.i8(i8 <val>, i8* <str>, i8* <str>, i32 <int>)
9071 declare i16 @llvm.annotation.i16(i16 <val>, i8* <str>, i8* <str>, i32 <int>)
9072 declare i32 @llvm.annotation.i32(i32 <val>, i8* <str>, i8* <str>, i32 <int>)
9073 declare i64 @llvm.annotation.i64(i64 <val>, i8* <str>, i8* <str>, i32 <int>)
9074 declare i256 @llvm.annotation.i256(i256 <val>, i8* <str>, i8* <str>, i32 <int>)
9079 The '``llvm.annotation``' intrinsic.
9084 The first argument is an integer value (result of some expression), the
9085 second is a pointer to a global string, the third is a pointer to a
9086 global string which is the source file name, and the last argument is
9087 the line number. It returns the value of the first argument.
9092 This intrinsic allows annotations to be put on arbitrary expressions
9093 with arbitrary strings. This can be useful for special purpose
9094 optimizations that want to look for these annotations. These have no
9095 other defined use; they are ignored by code generation and optimization.
9097 '``llvm.trap``' Intrinsic
9098 ^^^^^^^^^^^^^^^^^^^^^^^^^
9105 declare void @llvm.trap() noreturn nounwind
9110 The '``llvm.trap``' intrinsic.
9120 This intrinsic is lowered to the target dependent trap instruction. If
9121 the target does not have a trap instruction, this intrinsic will be
9122 lowered to a call of the ``abort()`` function.
9124 '``llvm.debugtrap``' Intrinsic
9125 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9132 declare void @llvm.debugtrap() nounwind
9137 The '``llvm.debugtrap``' intrinsic.
9147 This intrinsic is lowered to code which is intended to cause an
9148 execution trap with the intention of requesting the attention of a
9151 '``llvm.stackprotector``' Intrinsic
9152 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9159 declare void @llvm.stackprotector(i8* <guard>, i8** <slot>)
9164 The ``llvm.stackprotector`` intrinsic takes the ``guard`` and stores it
9165 onto the stack at ``slot``. The stack slot is adjusted to ensure that it
9166 is placed on the stack before local variables.
9171 The ``llvm.stackprotector`` intrinsic requires two pointer arguments.
9172 The first argument is the value loaded from the stack guard
9173 ``@__stack_chk_guard``. The second variable is an ``alloca`` that has
9174 enough space to hold the value of the guard.
9179 This intrinsic causes the prologue/epilogue inserter to force the position of
9180 the ``AllocaInst`` stack slot to be before local variables on the stack. This is
9181 to ensure that if a local variable on the stack is overwritten, it will destroy
9182 the value of the guard. When the function exits, the guard on the stack is
9183 checked against the original guard by ``llvm.stackprotectorcheck``. If they are
9184 different, then ``llvm.stackprotectorcheck`` causes the program to abort by
9185 calling the ``__stack_chk_fail()`` function.
9187 '``llvm.stackprotectorcheck``' Intrinsic
9188 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9195 declare void @llvm.stackprotectorcheck(i8** <guard>)
9200 The ``llvm.stackprotectorcheck`` intrinsic compares ``guard`` against an already
9201 created stack protector and if they are not equal calls the
9202 ``__stack_chk_fail()`` function.
9207 The ``llvm.stackprotectorcheck`` intrinsic requires one pointer argument, the
9208 the variable ``@__stack_chk_guard``.
9213 This intrinsic is provided to perform the stack protector check by comparing
9214 ``guard`` with the stack slot created by ``llvm.stackprotector`` and if the
9215 values do not match call the ``__stack_chk_fail()`` function.
9217 The reason to provide this as an IR level intrinsic instead of implementing it
9218 via other IR operations is that in order to perform this operation at the IR
9219 level without an intrinsic, one would need to create additional basic blocks to
9220 handle the success/failure cases. This makes it difficult to stop the stack
9221 protector check from disrupting sibling tail calls in Codegen. With this
9222 intrinsic, we are able to generate the stack protector basic blocks late in
9223 codegen after the tail call decision has occurred.
9225 '``llvm.objectsize``' Intrinsic
9226 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9233 declare i32 @llvm.objectsize.i32(i8* <object>, i1 <min>)
9234 declare i64 @llvm.objectsize.i64(i8* <object>, i1 <min>)
9239 The ``llvm.objectsize`` intrinsic is designed to provide information to
9240 the optimizers to determine at compile time whether a) an operation
9241 (like memcpy) will overflow a buffer that corresponds to an object, or
9242 b) that a runtime check for overflow isn't necessary. An object in this
9243 context means an allocation of a specific class, structure, array, or
9249 The ``llvm.objectsize`` intrinsic takes two arguments. The first
9250 argument is a pointer to or into the ``object``. The second argument is
9251 a boolean and determines whether ``llvm.objectsize`` returns 0 (if true)
9252 or -1 (if false) when the object size is unknown. The second argument
9253 only accepts constants.
9258 The ``llvm.objectsize`` intrinsic is lowered to a constant representing
9259 the size of the object concerned. If the size cannot be determined at
9260 compile time, ``llvm.objectsize`` returns ``i32/i64 -1 or 0`` (depending
9261 on the ``min`` argument).
9263 '``llvm.expect``' Intrinsic
9264 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9269 This is an overloaded intrinsic. You can use ``llvm.expect`` on any
9274 declare i1 @llvm.expect.i1(i1 <val>, i1 <expected_val>)
9275 declare i32 @llvm.expect.i32(i32 <val>, i32 <expected_val>)
9276 declare i64 @llvm.expect.i64(i64 <val>, i64 <expected_val>)
9281 The ``llvm.expect`` intrinsic provides information about expected (the
9282 most probable) value of ``val``, which can be used by optimizers.
9287 The ``llvm.expect`` intrinsic takes two arguments. The first argument is
9288 a value. The second argument is an expected value, this needs to be a
9289 constant value, variables are not allowed.
9294 This intrinsic is lowered to the ``val``.
9296 '``llvm.donothing``' Intrinsic
9297 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9304 declare void @llvm.donothing() nounwind readnone
9309 The ``llvm.donothing`` intrinsic doesn't perform any operation. It's the
9310 only intrinsic that can be called with an invoke instruction.
9320 This intrinsic does nothing, and it's removed by optimizers and ignored
9323 Stack Map Intrinsics
9324 --------------------
9326 LLVM provides experimental intrinsics to support runtime patching
9327 mechanisms commonly desired in dynamic language JITs. These intrinsics
9328 are described in :doc:`StackMaps`.