1 ==============================
2 LLVM Language Reference Manual
3 ==============================
12 This document is a reference manual for the LLVM assembly language. LLVM
13 is a Static Single Assignment (SSA) based representation that provides
14 type safety, low-level operations, flexibility, and the capability of
15 representing 'all' high-level languages cleanly. It is the common code
16 representation used throughout all phases of the LLVM compilation
22 The LLVM code representation is designed to be used in three different
23 forms: as an in-memory compiler IR, as an on-disk bitcode representation
24 (suitable for fast loading by a Just-In-Time compiler), and as a human
25 readable assembly language representation. This allows LLVM to provide a
26 powerful intermediate representation for efficient compiler
27 transformations and analysis, while providing a natural means to debug
28 and visualize the transformations. The three different forms of LLVM are
29 all equivalent. This document describes the human readable
30 representation and notation.
32 The LLVM representation aims to be light-weight and low-level while
33 being expressive, typed, and extensible at the same time. It aims to be
34 a "universal IR" of sorts, by being at a low enough level that
35 high-level ideas may be cleanly mapped to it (similar to how
36 microprocessors are "universal IR's", allowing many source languages to
37 be mapped to them). By providing type information, LLVM can be used as
38 the target of optimizations: for example, through pointer analysis, it
39 can be proven that a C automatic variable is never accessed outside of
40 the current function, allowing it to be promoted to a simple SSA value
41 instead of a memory location.
48 It is important to note that this document describes 'well formed' LLVM
49 assembly language. There is a difference between what the parser accepts
50 and what is considered 'well formed'. For example, the following
51 instruction is syntactically okay, but not well formed:
57 because the definition of ``%x`` does not dominate all of its uses. The
58 LLVM infrastructure provides a verification pass that may be used to
59 verify that an LLVM module is well formed. This pass is automatically
60 run by the parser after parsing input assembly and by the optimizer
61 before it outputs bitcode. The violations pointed out by the verifier
62 pass indicate bugs in transformation passes or input to the parser.
69 LLVM identifiers come in two basic types: global and local. Global
70 identifiers (functions, global variables) begin with the ``'@'``
71 character. Local identifiers (register names, types) begin with the
72 ``'%'`` character. Additionally, there are three different formats for
73 identifiers, for different purposes:
75 #. Named values are represented as a string of characters with their
76 prefix. For example, ``%foo``, ``@DivisionByZero``,
77 ``%a.really.long.identifier``. The actual regular expression used is
78 '``[%@][-a-zA-Z$._][-a-zA-Z$._0-9]*``'. Identifiers that require other
79 characters in their names can be surrounded with quotes. Special
80 characters may be escaped using ``"\xx"`` where ``xx`` is the ASCII
81 code for the character in hexadecimal. In this way, any character can
82 be used in a name value, even quotes themselves. The ``"\01"`` prefix
83 can be used on global variables to suppress mangling.
84 #. Unnamed values are represented as an unsigned numeric value with
85 their prefix. For example, ``%12``, ``@2``, ``%44``.
86 #. Constants, which are described in the section Constants_ below.
88 LLVM requires that values start with a prefix for two reasons: Compilers
89 don't need to worry about name clashes with reserved words, and the set
90 of reserved words may be expanded in the future without penalty.
91 Additionally, unnamed identifiers allow a compiler to quickly come up
92 with a temporary variable without having to avoid symbol table
95 Reserved words in LLVM are very similar to reserved words in other
96 languages. There are keywords for different opcodes ('``add``',
97 '``bitcast``', '``ret``', etc...), for primitive type names ('``void``',
98 '``i32``', etc...), and others. These reserved words cannot conflict
99 with variable names, because none of them start with a prefix character
100 (``'%'`` or ``'@'``).
102 Here is an example of LLVM code to multiply the integer variable
109 %result = mul i32 %X, 8
111 After strength reduction:
115 %result = shl i32 %X, 3
121 %0 = add i32 %X, %X ; yields i32:%0
122 %1 = add i32 %0, %0 ; yields i32:%1
123 %result = add i32 %1, %1
125 This last way of multiplying ``%X`` by 8 illustrates several important
126 lexical features of LLVM:
128 #. Comments are delimited with a '``;``' and go until the end of line.
129 #. Unnamed temporaries are created when the result of a computation is
130 not assigned to a named value.
131 #. Unnamed temporaries are numbered sequentially (using a per-function
132 incrementing counter, starting with 0). Note that basic blocks and unnamed
133 function parameters are included in this numbering. For example, if the
134 entry basic block is not given a label name and all function parameters are
135 named, then it will get number 0.
137 It also shows a convention that we follow in this document. When
138 demonstrating instructions, we will follow an instruction with a comment
139 that defines the type and name of value produced.
147 LLVM programs are composed of ``Module``'s, each of which is a
148 translation unit of the input programs. Each module consists of
149 functions, global variables, and symbol table entries. Modules may be
150 combined together with the LLVM linker, which merges function (and
151 global variable) definitions, resolves forward declarations, and merges
152 symbol table entries. Here is an example of the "hello world" module:
156 ; Declare the string constant as a global constant.
157 @.str = private unnamed_addr constant [13 x i8] c"hello world\0A\00"
159 ; External declaration of the puts function
160 declare i32 @puts(i8* nocapture) nounwind
162 ; Definition of main function
163 define i32 @main() { ; i32()*
164 ; Convert [13 x i8]* to i8 *...
165 %cast210 = getelementptr [13 x i8], [13 x i8]* @.str, i64 0, i64 0
167 ; Call puts function to write out the string to stdout.
168 call i32 @puts(i8* %cast210)
173 !0 = !{i32 42, null, !"string"}
176 This example is made up of a :ref:`global variable <globalvars>` named
177 "``.str``", an external declaration of the "``puts``" function, a
178 :ref:`function definition <functionstructure>` for "``main``" and
179 :ref:`named metadata <namedmetadatastructure>` "``foo``".
181 In general, a module is made up of a list of global values (where both
182 functions and global variables are global values). Global values are
183 represented by a pointer to a memory location (in this case, a pointer
184 to an array of char, and a pointer to a function), and have one of the
185 following :ref:`linkage types <linkage>`.
192 All Global Variables and Functions have one of the following types of
196 Global values with "``private``" linkage are only directly
197 accessible by objects in the current module. In particular, linking
198 code into a module with an private global value may cause the
199 private to be renamed as necessary to avoid collisions. Because the
200 symbol is private to the module, all references can be updated. This
201 doesn't show up in any symbol table in the object file.
203 Similar to private, but the value shows as a local symbol
204 (``STB_LOCAL`` in the case of ELF) in the object file. This
205 corresponds to the notion of the '``static``' keyword in C.
206 ``available_externally``
207 Globals with "``available_externally``" linkage are never emitted
208 into the object file corresponding to the LLVM module. They exist to
209 allow inlining and other optimizations to take place given knowledge
210 of the definition of the global, which is known to be somewhere
211 outside the module. Globals with ``available_externally`` linkage
212 are allowed to be discarded at will, and are otherwise the same as
213 ``linkonce_odr``. This linkage type is only allowed on definitions,
216 Globals with "``linkonce``" linkage are merged with other globals of
217 the same name when linkage occurs. This can be used to implement
218 some forms of inline functions, templates, or other code which must
219 be generated in each translation unit that uses it, but where the
220 body may be overridden with a more definitive definition later.
221 Unreferenced ``linkonce`` globals are allowed to be discarded. Note
222 that ``linkonce`` linkage does not actually allow the optimizer to
223 inline the body of this function into callers because it doesn't
224 know if this definition of the function is the definitive definition
225 within the program or whether it will be overridden by a stronger
226 definition. To enable inlining and other optimizations, use
227 "``linkonce_odr``" linkage.
229 "``weak``" linkage has the same merging semantics as ``linkonce``
230 linkage, except that unreferenced globals with ``weak`` linkage may
231 not be discarded. This is used for globals that are declared "weak"
234 "``common``" linkage is most similar to "``weak``" linkage, but they
235 are used for tentative definitions in C, such as "``int X;``" at
236 global scope. Symbols with "``common``" linkage are merged in the
237 same way as ``weak symbols``, and they may not be deleted if
238 unreferenced. ``common`` symbols may not have an explicit section,
239 must have a zero initializer, and may not be marked
240 ':ref:`constant <globalvars>`'. Functions and aliases may not have
243 .. _linkage_appending:
246 "``appending``" linkage may only be applied to global variables of
247 pointer to array type. When two global variables with appending
248 linkage are linked together, the two global arrays are appended
249 together. This is the LLVM, typesafe, equivalent of having the
250 system linker append together "sections" with identical names when
253 The semantics of this linkage follow the ELF object file model: the
254 symbol is weak until linked, if not linked, the symbol becomes null
255 instead of being an undefined reference.
256 ``linkonce_odr``, ``weak_odr``
257 Some languages allow differing globals to be merged, such as two
258 functions with different semantics. Other languages, such as
259 ``C++``, ensure that only equivalent globals are ever merged (the
260 "one definition rule" --- "ODR"). Such languages can use the
261 ``linkonce_odr`` and ``weak_odr`` linkage types to indicate that the
262 global will only be merged with equivalent globals. These linkage
263 types are otherwise the same as their non-``odr`` versions.
265 If none of the above identifiers are used, the global is externally
266 visible, meaning that it participates in linkage and can be used to
267 resolve external symbol references.
269 It is illegal for a function *declaration* to have any linkage type
270 other than ``external`` or ``extern_weak``.
277 LLVM :ref:`functions <functionstructure>`, :ref:`calls <i_call>` and
278 :ref:`invokes <i_invoke>` can all have an optional calling convention
279 specified for the call. The calling convention of any pair of dynamic
280 caller/callee must match, or the behavior of the program is undefined.
281 The following calling conventions are supported by LLVM, and more may be
284 "``ccc``" - The C calling convention
285 This calling convention (the default if no other calling convention
286 is specified) matches the target C calling conventions. This calling
287 convention supports varargs function calls and tolerates some
288 mismatch in the declared prototype and implemented declaration of
289 the function (as does normal C).
290 "``fastcc``" - The fast calling convention
291 This calling convention attempts to make calls as fast as possible
292 (e.g. by passing things in registers). This calling convention
293 allows the target to use whatever tricks it wants to produce fast
294 code for the target, without having to conform to an externally
295 specified ABI (Application Binary Interface). `Tail calls can only
296 be optimized when this, the GHC or the HiPE convention is
297 used. <CodeGenerator.html#id80>`_ This calling convention does not
298 support varargs and requires the prototype of all callees to exactly
299 match the prototype of the function definition.
300 "``coldcc``" - The cold calling convention
301 This calling convention attempts to make code in the caller as
302 efficient as possible under the assumption that the call is not
303 commonly executed. As such, these calls often preserve all registers
304 so that the call does not break any live ranges in the caller side.
305 This calling convention does not support varargs and requires the
306 prototype of all callees to exactly match the prototype of the
307 function definition. Furthermore the inliner doesn't consider such function
309 "``cc 10``" - GHC convention
310 This calling convention has been implemented specifically for use by
311 the `Glasgow Haskell Compiler (GHC) <http://www.haskell.org/ghc>`_.
312 It passes everything in registers, going to extremes to achieve this
313 by disabling callee save registers. This calling convention should
314 not be used lightly but only for specific situations such as an
315 alternative to the *register pinning* performance technique often
316 used when implementing functional programming languages. At the
317 moment only X86 supports this convention and it has the following
320 - On *X86-32* only supports up to 4 bit type parameters. No
321 floating point types are supported.
322 - On *X86-64* only supports up to 10 bit type parameters and 6
323 floating point parameters.
325 This calling convention supports `tail call
326 optimization <CodeGenerator.html#id80>`_ but requires both the
327 caller and callee are using it.
328 "``cc 11``" - The HiPE calling convention
329 This calling convention has been implemented specifically for use by
330 the `High-Performance Erlang
331 (HiPE) <http://www.it.uu.se/research/group/hipe/>`_ compiler, *the*
332 native code compiler of the `Ericsson's Open Source Erlang/OTP
333 system <http://www.erlang.org/download.shtml>`_. It uses more
334 registers for argument passing than the ordinary C calling
335 convention and defines no callee-saved registers. The calling
336 convention properly supports `tail call
337 optimization <CodeGenerator.html#id80>`_ but requires that both the
338 caller and the callee use it. It uses a *register pinning*
339 mechanism, similar to GHC's convention, for keeping frequently
340 accessed runtime components pinned to specific hardware registers.
341 At the moment only X86 supports this convention (both 32 and 64
343 "``webkit_jscc``" - WebKit's JavaScript calling convention
344 This calling convention has been implemented for `WebKit FTL JIT
345 <https://trac.webkit.org/wiki/FTLJIT>`_. It passes arguments on the
346 stack right to left (as cdecl does), and returns a value in the
347 platform's customary return register.
348 "``anyregcc``" - Dynamic calling convention for code patching
349 This is a special convention that supports patching an arbitrary code
350 sequence in place of a call site. This convention forces the call
351 arguments into registers but allows them to be dynamically
352 allocated. This can currently only be used with calls to
353 llvm.experimental.patchpoint because only this intrinsic records
354 the location of its arguments in a side table. See :doc:`StackMaps`.
355 "``preserve_mostcc``" - The `PreserveMost` calling convention
356 This calling convention attempts to make the code in the caller as
357 unintrusive as possible. This convention behaves identically to the `C`
358 calling convention on how arguments and return values are passed, but it
359 uses a different set of caller/callee-saved registers. This alleviates the
360 burden of saving and recovering a large register set before and after the
361 call in the caller. If the arguments are passed in callee-saved registers,
362 then they will be preserved by the callee across the call. This doesn't
363 apply for values returned in callee-saved registers.
365 - On X86-64 the callee preserves all general purpose registers, except for
366 R11. R11 can be used as a scratch register. Floating-point registers
367 (XMMs/YMMs) are not preserved and need to be saved by the caller.
369 The idea behind this convention is to support calls to runtime functions
370 that have a hot path and a cold path. The hot path is usually a small piece
371 of code that doesn't use many registers. The cold path might need to call out to
372 another function and therefore only needs to preserve the caller-saved
373 registers, which haven't already been saved by the caller. The
374 `PreserveMost` calling convention is very similar to the `cold` calling
375 convention in terms of caller/callee-saved registers, but they are used for
376 different types of function calls. `coldcc` is for function calls that are
377 rarely executed, whereas `preserve_mostcc` function calls are intended to be
378 on the hot path and definitely executed a lot. Furthermore `preserve_mostcc`
379 doesn't prevent the inliner from inlining the function call.
381 This calling convention will be used by a future version of the ObjectiveC
382 runtime and should therefore still be considered experimental at this time.
383 Although this convention was created to optimize certain runtime calls to
384 the ObjectiveC runtime, it is not limited to this runtime and might be used
385 by other runtimes in the future too. The current implementation only
386 supports X86-64, but the intention is to support more architectures in the
388 "``preserve_allcc``" - The `PreserveAll` calling convention
389 This calling convention attempts to make the code in the caller even less
390 intrusive than the `PreserveMost` calling convention. This calling
391 convention also behaves identical to the `C` calling convention on how
392 arguments and return values are passed, but it uses a different set of
393 caller/callee-saved registers. This removes the burden of saving and
394 recovering a large register set before and after the call in the caller. If
395 the arguments are passed in callee-saved registers, then they will be
396 preserved by the callee across the call. This doesn't apply for values
397 returned in callee-saved registers.
399 - On X86-64 the callee preserves all general purpose registers, except for
400 R11. R11 can be used as a scratch register. Furthermore it also preserves
401 all floating-point registers (XMMs/YMMs).
403 The idea behind this convention is to support calls to runtime functions
404 that don't need to call out to any other functions.
406 This calling convention, like the `PreserveMost` calling convention, will be
407 used by a future version of the ObjectiveC runtime and should be considered
408 experimental at this time.
409 "``cc <n>``" - Numbered convention
410 Any calling convention may be specified by number, allowing
411 target-specific calling conventions to be used. Target specific
412 calling conventions start at 64.
414 More calling conventions can be added/defined on an as-needed basis, to
415 support Pascal conventions or any other well-known target-independent
418 .. _visibilitystyles:
423 All Global Variables and Functions have one of the following visibility
426 "``default``" - Default style
427 On targets that use the ELF object file format, default visibility
428 means that the declaration is visible to other modules and, in
429 shared libraries, means that the declared entity may be overridden.
430 On Darwin, default visibility means that the declaration is visible
431 to other modules. Default visibility corresponds to "external
432 linkage" in the language.
433 "``hidden``" - Hidden style
434 Two declarations of an object with hidden visibility refer to the
435 same object if they are in the same shared object. Usually, hidden
436 visibility indicates that the symbol will not be placed into the
437 dynamic symbol table, so no other module (executable or shared
438 library) can reference it directly.
439 "``protected``" - Protected style
440 On ELF, protected visibility indicates that the symbol will be
441 placed in the dynamic symbol table, but that references within the
442 defining module will bind to the local symbol. That is, the symbol
443 cannot be overridden by another module.
445 A symbol with ``internal`` or ``private`` linkage must have ``default``
453 All Global Variables, Functions and Aliases can have one of the following
457 "``dllimport``" causes the compiler to reference a function or variable via
458 a global pointer to a pointer that is set up by the DLL exporting the
459 symbol. On Microsoft Windows targets, the pointer name is formed by
460 combining ``__imp_`` and the function or variable name.
462 "``dllexport``" causes the compiler to provide a global pointer to a pointer
463 in a DLL, so that it can be referenced with the ``dllimport`` attribute. On
464 Microsoft Windows targets, the pointer name is formed by combining
465 ``__imp_`` and the function or variable name. Since this storage class
466 exists for defining a dll interface, the compiler, assembler and linker know
467 it is externally referenced and must refrain from deleting the symbol.
471 Thread Local Storage Models
472 ---------------------------
474 A variable may be defined as ``thread_local``, which means that it will
475 not be shared by threads (each thread will have a separated copy of the
476 variable). Not all targets support thread-local variables. Optionally, a
477 TLS model may be specified:
480 For variables that are only used within the current shared library.
482 For variables in modules that will not be loaded dynamically.
484 For variables defined in the executable and only used within it.
486 If no explicit model is given, the "general dynamic" model is used.
488 The models correspond to the ELF TLS models; see `ELF Handling For
489 Thread-Local Storage <http://people.redhat.com/drepper/tls.pdf>`_ for
490 more information on under which circumstances the different models may
491 be used. The target may choose a different TLS model if the specified
492 model is not supported, or if a better choice of model can be made.
494 A model can also be specified in a alias, but then it only governs how
495 the alias is accessed. It will not have any effect in the aliasee.
502 LLVM IR allows you to specify both "identified" and "literal" :ref:`structure
503 types <t_struct>`. Literal types are uniqued structurally, but identified types
504 are never uniqued. An :ref:`opaque structural type <t_opaque>` can also be used
505 to forward declare a type that is not yet available.
507 An example of a identified structure specification is:
511 %mytype = type { %mytype*, i32 }
513 Prior to the LLVM 3.0 release, identified types were structurally uniqued. Only
514 literal types are uniqued in recent versions of LLVM.
521 Global variables define regions of memory allocated at compilation time
524 Global variable definitions must be initialized.
526 Global variables in other translation units can also be declared, in which
527 case they don't have an initializer.
529 Either global variable definitions or declarations may have an explicit section
530 to be placed in and may have an optional explicit alignment specified.
532 A variable may be defined as a global ``constant``, which indicates that
533 the contents of the variable will **never** be modified (enabling better
534 optimization, allowing the global data to be placed in the read-only
535 section of an executable, etc). Note that variables that need runtime
536 initialization cannot be marked ``constant`` as there is a store to the
539 LLVM explicitly allows *declarations* of global variables to be marked
540 constant, even if the final definition of the global is not. This
541 capability can be used to enable slightly better optimization of the
542 program, but requires the language definition to guarantee that
543 optimizations based on the 'constantness' are valid for the translation
544 units that do not include the definition.
546 As SSA values, global variables define pointer values that are in scope
547 (i.e. they dominate) all basic blocks in the program. Global variables
548 always define a pointer to their "content" type because they describe a
549 region of memory, and all memory objects in LLVM are accessed through
552 Global variables can be marked with ``unnamed_addr`` which indicates
553 that the address is not significant, only the content. Constants marked
554 like this can be merged with other constants if they have the same
555 initializer. Note that a constant with significant address *can* be
556 merged with a ``unnamed_addr`` constant, the result being a constant
557 whose address is significant.
559 A global variable may be declared to reside in a target-specific
560 numbered address space. For targets that support them, address spaces
561 may affect how optimizations are performed and/or what target
562 instructions are used to access the variable. The default address space
563 is zero. The address space qualifier must precede any other attributes.
565 LLVM allows an explicit section to be specified for globals. If the
566 target supports it, it will emit globals to the section specified.
567 Additionally, the global can placed in a comdat if the target has the necessary
570 By default, global initializers are optimized by assuming that global
571 variables defined within the module are not modified from their
572 initial values before the start of the global initializer. This is
573 true even for variables potentially accessible from outside the
574 module, including those with external linkage or appearing in
575 ``@llvm.used`` or dllexported variables. This assumption may be suppressed
576 by marking the variable with ``externally_initialized``.
578 An explicit alignment may be specified for a global, which must be a
579 power of 2. If not present, or if the alignment is set to zero, the
580 alignment of the global is set by the target to whatever it feels
581 convenient. If an explicit alignment is specified, the global is forced
582 to have exactly that alignment. Targets and optimizers are not allowed
583 to over-align the global if the global has an assigned section. In this
584 case, the extra alignment could be observable: for example, code could
585 assume that the globals are densely packed in their section and try to
586 iterate over them as an array, alignment padding would break this
587 iteration. The maximum alignment is ``1 << 29``.
589 Globals can also have a :ref:`DLL storage class <dllstorageclass>`.
591 Variables and aliases can have a
592 :ref:`Thread Local Storage Model <tls_model>`.
596 [@<GlobalVarName> =] [Linkage] [Visibility] [DLLStorageClass] [ThreadLocal]
597 [unnamed_addr] [AddrSpace] [ExternallyInitialized]
598 <global | constant> <Type> [<InitializerConstant>]
599 [, section "name"] [, comdat [($name)]]
600 [, align <Alignment>]
602 For example, the following defines a global in a numbered address space
603 with an initializer, section, and alignment:
607 @G = addrspace(5) constant float 1.0, section "foo", align 4
609 The following example just declares a global variable
613 @G = external global i32
615 The following example defines a thread-local global with the
616 ``initialexec`` TLS model:
620 @G = thread_local(initialexec) global i32 0, align 4
622 .. _functionstructure:
627 LLVM function definitions consist of the "``define``" keyword, an
628 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
629 style <visibility>`, an optional :ref:`DLL storage class <dllstorageclass>`,
630 an optional :ref:`calling convention <callingconv>`,
631 an optional ``unnamed_addr`` attribute, a return type, an optional
632 :ref:`parameter attribute <paramattrs>` for the return type, a function
633 name, a (possibly empty) argument list (each with optional :ref:`parameter
634 attributes <paramattrs>`), optional :ref:`function attributes <fnattrs>`,
635 an optional section, an optional alignment,
636 an optional :ref:`comdat <langref_comdats>`,
637 an optional :ref:`garbage collector name <gc>`, an optional :ref:`prefix <prefixdata>`,
638 an optional :ref:`prologue <prologuedata>`, an opening
639 curly brace, a list of basic blocks, and a closing curly brace.
641 LLVM function declarations consist of the "``declare``" keyword, an
642 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
643 style <visibility>`, an optional :ref:`DLL storage class <dllstorageclass>`,
644 an optional :ref:`calling convention <callingconv>`,
645 an optional ``unnamed_addr`` attribute, a return type, an optional
646 :ref:`parameter attribute <paramattrs>` for the return type, a function
647 name, a possibly empty list of arguments, an optional alignment, an optional
648 :ref:`garbage collector name <gc>`, an optional :ref:`prefix <prefixdata>`,
649 and an optional :ref:`prologue <prologuedata>`.
651 A function definition contains a list of basic blocks, forming the CFG (Control
652 Flow Graph) for the function. Each basic block may optionally start with a label
653 (giving the basic block a symbol table entry), contains a list of instructions,
654 and ends with a :ref:`terminator <terminators>` instruction (such as a branch or
655 function return). If an explicit label is not provided, a block is assigned an
656 implicit numbered label, using the next value from the same counter as used for
657 unnamed temporaries (:ref:`see above<identifiers>`). For example, if a function
658 entry block does not have an explicit label, it will be assigned label "%0",
659 then the first unnamed temporary in that block will be "%1", etc.
661 The first basic block in a function is special in two ways: it is
662 immediately executed on entrance to the function, and it is not allowed
663 to have predecessor basic blocks (i.e. there can not be any branches to
664 the entry block of a function). Because the block can have no
665 predecessors, it also cannot have any :ref:`PHI nodes <i_phi>`.
667 LLVM allows an explicit section to be specified for functions. If the
668 target supports it, it will emit functions to the section specified.
669 Additionally, the function can be placed in a COMDAT.
671 An explicit alignment may be specified for a function. If not present,
672 or if the alignment is set to zero, the alignment of the function is set
673 by the target to whatever it feels convenient. If an explicit alignment
674 is specified, the function is forced to have at least that much
675 alignment. All alignments must be a power of 2.
677 If the ``unnamed_addr`` attribute is given, the address is known to not
678 be significant and two identical functions can be merged.
682 define [linkage] [visibility] [DLLStorageClass]
684 <ResultType> @<FunctionName> ([argument list])
685 [unnamed_addr] [fn Attrs] [section "name"] [comdat [($name)]]
686 [align N] [gc] [prefix Constant] [prologue Constant] { ... }
688 The argument list is a comma seperated sequence of arguments where each
689 argument is of the following form
693 <type> [parameter Attrs] [name]
701 Aliases, unlike function or variables, don't create any new data. They
702 are just a new symbol and metadata for an existing position.
704 Aliases have a name and an aliasee that is either a global value or a
707 Aliases may have an optional :ref:`linkage type <linkage>`, an optional
708 :ref:`visibility style <visibility>`, an optional :ref:`DLL storage class
709 <dllstorageclass>` and an optional :ref:`tls model <tls_model>`.
713 @<Name> = [Linkage] [Visibility] [DLLStorageClass] [ThreadLocal] [unnamed_addr] alias <AliaseeTy> @<Aliasee>
715 The linkage must be one of ``private``, ``internal``, ``linkonce``, ``weak``,
716 ``linkonce_odr``, ``weak_odr``, ``external``. Note that some system linkers
717 might not correctly handle dropping a weak symbol that is aliased.
719 Aliases that are not ``unnamed_addr`` are guaranteed to have the same address as
720 the aliasee expression. ``unnamed_addr`` ones are only guaranteed to point
723 Since aliases are only a second name, some restrictions apply, of which
724 some can only be checked when producing an object file:
726 * The expression defining the aliasee must be computable at assembly
727 time. Since it is just a name, no relocations can be used.
729 * No alias in the expression can be weak as the possibility of the
730 intermediate alias being overridden cannot be represented in an
733 * No global value in the expression can be a declaration, since that
734 would require a relocation, which is not possible.
741 Comdat IR provides access to COFF and ELF object file COMDAT functionality.
743 Comdats have a name which represents the COMDAT key. All global objects that
744 specify this key will only end up in the final object file if the linker chooses
745 that key over some other key. Aliases are placed in the same COMDAT that their
746 aliasee computes to, if any.
748 Comdats have a selection kind to provide input on how the linker should
749 choose between keys in two different object files.
753 $<Name> = comdat SelectionKind
755 The selection kind must be one of the following:
758 The linker may choose any COMDAT key, the choice is arbitrary.
760 The linker may choose any COMDAT key but the sections must contain the
763 The linker will choose the section containing the largest COMDAT key.
765 The linker requires that only section with this COMDAT key exist.
767 The linker may choose any COMDAT key but the sections must contain the
770 Note that the Mach-O platform doesn't support COMDATs and ELF only supports
771 ``any`` as a selection kind.
773 Here is an example of a COMDAT group where a function will only be selected if
774 the COMDAT key's section is the largest:
778 $foo = comdat largest
779 @foo = global i32 2, comdat($foo)
781 define void @bar() comdat($foo) {
785 As a syntactic sugar the ``$name`` can be omitted if the name is the same as
791 @foo = global i32 2, comdat
794 In a COFF object file, this will create a COMDAT section with selection kind
795 ``IMAGE_COMDAT_SELECT_LARGEST`` containing the contents of the ``@foo`` symbol
796 and another COMDAT section with selection kind
797 ``IMAGE_COMDAT_SELECT_ASSOCIATIVE`` which is associated with the first COMDAT
798 section and contains the contents of the ``@bar`` symbol.
800 There are some restrictions on the properties of the global object.
801 It, or an alias to it, must have the same name as the COMDAT group when
803 The contents and size of this object may be used during link-time to determine
804 which COMDAT groups get selected depending on the selection kind.
805 Because the name of the object must match the name of the COMDAT group, the
806 linkage of the global object must not be local; local symbols can get renamed
807 if a collision occurs in the symbol table.
809 The combined use of COMDATS and section attributes may yield surprising results.
816 @g1 = global i32 42, section "sec", comdat($foo)
817 @g2 = global i32 42, section "sec", comdat($bar)
819 From the object file perspective, this requires the creation of two sections
820 with the same name. This is necessary because both globals belong to different
821 COMDAT groups and COMDATs, at the object file level, are represented by
824 Note that certain IR constructs like global variables and functions may create
825 COMDATs in the object file in addition to any which are specified using COMDAT
826 IR. This arises, for example, when a global variable has linkonce_odr linkage.
828 .. _namedmetadatastructure:
833 Named metadata is a collection of metadata. :ref:`Metadata
834 nodes <metadata>` (but not metadata strings) are the only valid
835 operands for a named metadata.
837 #. Named metadata are represented as a string of characters with the
838 metadata prefix. The rules for metadata names are the same as for
839 identifiers, but quoted names are not allowed. ``"\xx"`` type escapes
840 are still valid, which allows any character to be part of a name.
844 ; Some unnamed metadata nodes, which are referenced by the named metadata.
849 !name = !{!0, !1, !2}
856 The return type and each parameter of a function type may have a set of
857 *parameter attributes* associated with them. Parameter attributes are
858 used to communicate additional information about the result or
859 parameters of a function. Parameter attributes are considered to be part
860 of the function, not of the function type, so functions with different
861 parameter attributes can have the same function type.
863 Parameter attributes are simple keywords that follow the type specified.
864 If multiple parameter attributes are needed, they are space separated.
869 declare i32 @printf(i8* noalias nocapture, ...)
870 declare i32 @atoi(i8 zeroext)
871 declare signext i8 @returns_signed_char()
873 Note that any attributes for the function result (``nounwind``,
874 ``readonly``) come immediately after the argument list.
876 Currently, only the following parameter attributes are defined:
879 This indicates to the code generator that the parameter or return
880 value should be zero-extended to the extent required by the target's
881 ABI (which is usually 32-bits, but is 8-bits for a i1 on x86-64) by
882 the caller (for a parameter) or the callee (for a return value).
884 This indicates to the code generator that the parameter or return
885 value should be sign-extended to the extent required by the target's
886 ABI (which is usually 32-bits) by the caller (for a parameter) or
887 the callee (for a return value).
889 This indicates that this parameter or return value should be treated
890 in a special target-dependent fashion during while emitting code for
891 a function call or return (usually, by putting it in a register as
892 opposed to memory, though some targets use it to distinguish between
893 two different kinds of registers). Use of this attribute is
896 This indicates that the pointer parameter should really be passed by
897 value to the function. The attribute implies that a hidden copy of
898 the pointee is made between the caller and the callee, so the callee
899 is unable to modify the value in the caller. This attribute is only
900 valid on LLVM pointer arguments. It is generally used to pass
901 structs and arrays by value, but is also valid on pointers to
902 scalars. The copy is considered to belong to the caller not the
903 callee (for example, ``readonly`` functions should not write to
904 ``byval`` parameters). This is not a valid attribute for return
907 The byval attribute also supports specifying an alignment with the
908 align attribute. It indicates the alignment of the stack slot to
909 form and the known alignment of the pointer specified to the call
910 site. If the alignment is not specified, then the code generator
911 makes a target-specific assumption.
917 The ``inalloca`` argument attribute allows the caller to take the
918 address of outgoing stack arguments. An ``inalloca`` argument must
919 be a pointer to stack memory produced by an ``alloca`` instruction.
920 The alloca, or argument allocation, must also be tagged with the
921 inalloca keyword. Only the last argument may have the ``inalloca``
922 attribute, and that argument is guaranteed to be passed in memory.
924 An argument allocation may be used by a call at most once because
925 the call may deallocate it. The ``inalloca`` attribute cannot be
926 used in conjunction with other attributes that affect argument
927 storage, like ``inreg``, ``nest``, ``sret``, or ``byval``. The
928 ``inalloca`` attribute also disables LLVM's implicit lowering of
929 large aggregate return values, which means that frontend authors
930 must lower them with ``sret`` pointers.
932 When the call site is reached, the argument allocation must have
933 been the most recent stack allocation that is still live, or the
934 results are undefined. It is possible to allocate additional stack
935 space after an argument allocation and before its call site, but it
936 must be cleared off with :ref:`llvm.stackrestore
939 See :doc:`InAlloca` for more information on how to use this
943 This indicates that the pointer parameter specifies the address of a
944 structure that is the return value of the function in the source
945 program. This pointer must be guaranteed by the caller to be valid:
946 loads and stores to the structure may be assumed by the callee
947 not to trap and to be properly aligned. This may only be applied to
948 the first parameter. This is not a valid attribute for return
952 This indicates that the pointer value may be assumed by the optimizer to
953 have the specified alignment.
955 Note that this attribute has additional semantics when combined with the
961 This indicates that objects accessed via pointer values
962 :ref:`based <pointeraliasing>` on the argument or return value are not also
963 accessed, during the execution of the function, via pointer values not
964 *based* on the argument or return value. The attribute on a return value
965 also has additional semantics described below. The caller shares the
966 responsibility with the callee for ensuring that these requirements are met.
967 For further details, please see the discussion of the NoAlias response in
968 :ref:`alias analysis <Must, May, or No>`.
970 Note that this definition of ``noalias`` is intentionally similar
971 to the definition of ``restrict`` in C99 for function arguments.
973 For function return values, C99's ``restrict`` is not meaningful,
974 while LLVM's ``noalias`` is. Furthermore, the semantics of the ``noalias``
975 attribute on return values are stronger than the semantics of the attribute
976 when used on function arguments. On function return values, the ``noalias``
977 attribute indicates that the function acts like a system memory allocation
978 function, returning a pointer to allocated storage disjoint from the
979 storage for any other object accessible to the caller.
982 This indicates that the callee does not make any copies of the
983 pointer that outlive the callee itself. This is not a valid
984 attribute for return values.
989 This indicates that the pointer parameter can be excised using the
990 :ref:`trampoline intrinsics <int_trampoline>`. This is not a valid
991 attribute for return values and can only be applied to one parameter.
994 This indicates that the function always returns the argument as its return
995 value. This is an optimization hint to the code generator when generating
996 the caller, allowing tail call optimization and omission of register saves
997 and restores in some cases; it is not checked or enforced when generating
998 the callee. The parameter and the function return type must be valid
999 operands for the :ref:`bitcast instruction <i_bitcast>`. This is not a
1000 valid attribute for return values and can only be applied to one parameter.
1003 This indicates that the parameter or return pointer is not null. This
1004 attribute may only be applied to pointer typed parameters. This is not
1005 checked or enforced by LLVM, the caller must ensure that the pointer
1006 passed in is non-null, or the callee must ensure that the returned pointer
1009 ``dereferenceable(<n>)``
1010 This indicates that the parameter or return pointer is dereferenceable. This
1011 attribute may only be applied to pointer typed parameters. A pointer that
1012 is dereferenceable can be loaded from speculatively without a risk of
1013 trapping. The number of bytes known to be dereferenceable must be provided
1014 in parentheses. It is legal for the number of bytes to be less than the
1015 size of the pointee type. The ``nonnull`` attribute does not imply
1016 dereferenceability (consider a pointer to one element past the end of an
1017 array), however ``dereferenceable(<n>)`` does imply ``nonnull`` in
1018 ``addrspace(0)`` (which is the default address space).
1020 ``dereferenceable_or_null(<n>)``
1021 This indicates that the parameter or return value isn't both
1022 non-null and non-dereferenceable (up to ``<n>`` bytes) at the same
1023 time. All non-null pointers tagged with
1024 ``dereferenceable_or_null(<n>)`` are ``dereferenceable(<n>)``.
1025 For address space 0 ``dereferenceable_or_null(<n>)`` implies that
1026 a pointer is exactly one of ``dereferenceable(<n>)`` or ``null``,
1027 and in other address spaces ``dereferenceable_or_null(<n>)``
1028 implies that a pointer is at least one of ``dereferenceable(<n>)``
1029 or ``null`` (i.e. it may be both ``null`` and
1030 ``dereferenceable(<n>)``). This attribute may only be applied to
1031 pointer typed parameters.
1035 Garbage Collector Strategy Names
1036 --------------------------------
1038 Each function may specify a garbage collector strategy name, which is simply a
1041 .. code-block:: llvm
1043 define void @f() gc "name" { ... }
1045 The supported values of *name* includes those :ref:`built in to LLVM
1046 <builtin-gc-strategies>` and any provided by loaded plugins. Specifying a GC
1047 strategy will cause the compiler to alter its output in order to support the
1048 named garbage collection algorithm. Note that LLVM itself does not contain a
1049 garbage collector, this functionality is restricted to generating machine code
1050 which can interoperate with a collector provided externally.
1057 Prefix data is data associated with a function which the code
1058 generator will emit immediately before the function's entrypoint.
1059 The purpose of this feature is to allow frontends to associate
1060 language-specific runtime metadata with specific functions and make it
1061 available through the function pointer while still allowing the
1062 function pointer to be called.
1064 To access the data for a given function, a program may bitcast the
1065 function pointer to a pointer to the constant's type and dereference
1066 index -1. This implies that the IR symbol points just past the end of
1067 the prefix data. For instance, take the example of a function annotated
1068 with a single ``i32``,
1070 .. code-block:: llvm
1072 define void @f() prefix i32 123 { ... }
1074 The prefix data can be referenced as,
1076 .. code-block:: llvm
1078 %0 = bitcast void* () @f to i32*
1079 %a = getelementptr inbounds i32, i32* %0, i32 -1
1080 %b = load i32, i32* %a
1082 Prefix data is laid out as if it were an initializer for a global variable
1083 of the prefix data's type. The function will be placed such that the
1084 beginning of the prefix data is aligned. This means that if the size
1085 of the prefix data is not a multiple of the alignment size, the
1086 function's entrypoint will not be aligned. If alignment of the
1087 function's entrypoint is desired, padding must be added to the prefix
1090 A function may have prefix data but no body. This has similar semantics
1091 to the ``available_externally`` linkage in that the data may be used by the
1092 optimizers but will not be emitted in the object file.
1099 The ``prologue`` attribute allows arbitrary code (encoded as bytes) to
1100 be inserted prior to the function body. This can be used for enabling
1101 function hot-patching and instrumentation.
1103 To maintain the semantics of ordinary function calls, the prologue data must
1104 have a particular format. Specifically, it must begin with a sequence of
1105 bytes which decode to a sequence of machine instructions, valid for the
1106 module's target, which transfer control to the point immediately succeeding
1107 the prologue data, without performing any other visible action. This allows
1108 the inliner and other passes to reason about the semantics of the function
1109 definition without needing to reason about the prologue data. Obviously this
1110 makes the format of the prologue data highly target dependent.
1112 A trivial example of valid prologue data for the x86 architecture is ``i8 144``,
1113 which encodes the ``nop`` instruction:
1115 .. code-block:: llvm
1117 define void @f() prologue i8 144 { ... }
1119 Generally prologue data can be formed by encoding a relative branch instruction
1120 which skips the metadata, as in this example of valid prologue data for the
1121 x86_64 architecture, where the first two bytes encode ``jmp .+10``:
1123 .. code-block:: llvm
1125 %0 = type <{ i8, i8, i8* }>
1127 define void @f() prologue %0 <{ i8 235, i8 8, i8* @md}> { ... }
1129 A function may have prologue data but no body. This has similar semantics
1130 to the ``available_externally`` linkage in that the data may be used by the
1131 optimizers but will not be emitted in the object file.
1138 Attribute groups are groups of attributes that are referenced by objects within
1139 the IR. They are important for keeping ``.ll`` files readable, because a lot of
1140 functions will use the same set of attributes. In the degenerative case of a
1141 ``.ll`` file that corresponds to a single ``.c`` file, the single attribute
1142 group will capture the important command line flags used to build that file.
1144 An attribute group is a module-level object. To use an attribute group, an
1145 object references the attribute group's ID (e.g. ``#37``). An object may refer
1146 to more than one attribute group. In that situation, the attributes from the
1147 different groups are merged.
1149 Here is an example of attribute groups for a function that should always be
1150 inlined, has a stack alignment of 4, and which shouldn't use SSE instructions:
1152 .. code-block:: llvm
1154 ; Target-independent attributes:
1155 attributes #0 = { alwaysinline alignstack=4 }
1157 ; Target-dependent attributes:
1158 attributes #1 = { "no-sse" }
1160 ; Function @f has attributes: alwaysinline, alignstack=4, and "no-sse".
1161 define void @f() #0 #1 { ... }
1168 Function attributes are set to communicate additional information about
1169 a function. Function attributes are considered to be part of the
1170 function, not of the function type, so functions with different function
1171 attributes can have the same function type.
1173 Function attributes are simple keywords that follow the type specified.
1174 If multiple attributes are needed, they are space separated. For
1177 .. code-block:: llvm
1179 define void @f() noinline { ... }
1180 define void @f() alwaysinline { ... }
1181 define void @f() alwaysinline optsize { ... }
1182 define void @f() optsize { ... }
1185 This attribute indicates that, when emitting the prologue and
1186 epilogue, the backend should forcibly align the stack pointer.
1187 Specify the desired alignment, which must be a power of two, in
1190 This attribute indicates that the inliner should attempt to inline
1191 this function into callers whenever possible, ignoring any active
1192 inlining size threshold for this caller.
1194 This indicates that the callee function at a call site should be
1195 recognized as a built-in function, even though the function's declaration
1196 uses the ``nobuiltin`` attribute. This is only valid at call sites for
1197 direct calls to functions that are declared with the ``nobuiltin``
1200 This attribute indicates that this function is rarely called. When
1201 computing edge weights, basic blocks post-dominated by a cold
1202 function call are also considered to be cold; and, thus, given low
1205 This attribute indicates that the callee is dependent on a convergent
1206 thread execution pattern under certain parallel execution models.
1207 Transformations that are execution model agnostic may only move or
1208 tranform this call if the final location is control equivalent to its
1209 original position in the program, where control equivalence is defined as
1210 A dominates B and B post-dominates A, or vice versa.
1212 This attribute indicates that the source code contained a hint that
1213 inlining this function is desirable (such as the "inline" keyword in
1214 C/C++). It is just a hint; it imposes no requirements on the
1217 This attribute indicates that the function should be added to a
1218 jump-instruction table at code-generation time, and that all address-taken
1219 references to this function should be replaced with a reference to the
1220 appropriate jump-instruction-table function pointer. Note that this creates
1221 a new pointer for the original function, which means that code that depends
1222 on function-pointer identity can break. So, any function annotated with
1223 ``jumptable`` must also be ``unnamed_addr``.
1225 This attribute suggests that optimization passes and code generator
1226 passes make choices that keep the code size of this function as small
1227 as possible and perform optimizations that may sacrifice runtime
1228 performance in order to minimize the size of the generated code.
1230 This attribute disables prologue / epilogue emission for the
1231 function. This can have very system-specific consequences.
1233 This indicates that the callee function at a call site is not recognized as
1234 a built-in function. LLVM will retain the original call and not replace it
1235 with equivalent code based on the semantics of the built-in function, unless
1236 the call site uses the ``builtin`` attribute. This is valid at call sites
1237 and on function declarations and definitions.
1239 This attribute indicates that calls to the function cannot be
1240 duplicated. A call to a ``noduplicate`` function may be moved
1241 within its parent function, but may not be duplicated within
1242 its parent function.
1244 A function containing a ``noduplicate`` call may still
1245 be an inlining candidate, provided that the call is not
1246 duplicated by inlining. That implies that the function has
1247 internal linkage and only has one call site, so the original
1248 call is dead after inlining.
1250 This attributes disables implicit floating point instructions.
1252 This attribute indicates that the inliner should never inline this
1253 function in any situation. This attribute may not be used together
1254 with the ``alwaysinline`` attribute.
1256 This attribute suppresses lazy symbol binding for the function. This
1257 may make calls to the function faster, at the cost of extra program
1258 startup time if the function is not called during program startup.
1260 This attribute indicates that the code generator should not use a
1261 red zone, even if the target-specific ABI normally permits it.
1263 This function attribute indicates that the function never returns
1264 normally. This produces undefined behavior at runtime if the
1265 function ever does dynamically return.
1267 This function attribute indicates that the function never raises an
1268 exception. If the function does raise an exception, its runtime
1269 behavior is undefined. However, functions marked nounwind may still
1270 trap or generate asynchronous exceptions. Exception handling schemes
1271 that are recognized by LLVM to handle asynchronous exceptions, such
1272 as SEH, will still provide their implementation defined semantics.
1274 This function attribute indicates that the function is not optimized
1275 by any optimization or code generator passes with the
1276 exception of interprocedural optimization passes.
1277 This attribute cannot be used together with the ``alwaysinline``
1278 attribute; this attribute is also incompatible
1279 with the ``minsize`` attribute and the ``optsize`` attribute.
1281 This attribute requires the ``noinline`` attribute to be specified on
1282 the function as well, so the function is never inlined into any caller.
1283 Only functions with the ``alwaysinline`` attribute are valid
1284 candidates for inlining into the body of this function.
1286 This attribute suggests that optimization passes and code generator
1287 passes make choices that keep the code size of this function low,
1288 and otherwise do optimizations specifically to reduce code size as
1289 long as they do not significantly impact runtime performance.
1291 On a function, this attribute indicates that the function computes its
1292 result (or decides to unwind an exception) based strictly on its arguments,
1293 without dereferencing any pointer arguments or otherwise accessing
1294 any mutable state (e.g. memory, control registers, etc) visible to
1295 caller functions. It does not write through any pointer arguments
1296 (including ``byval`` arguments) and never changes any state visible
1297 to callers. This means that it cannot unwind exceptions by calling
1298 the ``C++`` exception throwing methods.
1300 On an argument, this attribute indicates that the function does not
1301 dereference that pointer argument, even though it may read or write the
1302 memory that the pointer points to if accessed through other pointers.
1304 On a function, this attribute indicates that the function does not write
1305 through any pointer arguments (including ``byval`` arguments) or otherwise
1306 modify any state (e.g. memory, control registers, etc) visible to
1307 caller functions. It may dereference pointer arguments and read
1308 state that may be set in the caller. A readonly function always
1309 returns the same value (or unwinds an exception identically) when
1310 called with the same set of arguments and global state. It cannot
1311 unwind an exception by calling the ``C++`` exception throwing
1314 On an argument, this attribute indicates that the function does not write
1315 through this pointer argument, even though it may write to the memory that
1316 the pointer points to.
1318 This attribute indicates that this function can return twice. The C
1319 ``setjmp`` is an example of such a function. The compiler disables
1320 some optimizations (like tail calls) in the caller of these
1323 This attribute indicates that
1324 `SafeStack <http://clang.llvm.org/docs/SafeStack.html>`_
1325 protection is enabled for this function.
1327 If a function that has a ``safestack`` attribute is inlined into a
1328 function that doesn't have a ``safestack`` attribute or which has an
1329 ``ssp``, ``sspstrong`` or ``sspreq`` attribute, then the resulting
1330 function will have a ``safestack`` attribute.
1331 ``sanitize_address``
1332 This attribute indicates that AddressSanitizer checks
1333 (dynamic address safety analysis) are enabled for this function.
1335 This attribute indicates that MemorySanitizer checks (dynamic detection
1336 of accesses to uninitialized memory) are enabled for this function.
1338 This attribute indicates that ThreadSanitizer checks
1339 (dynamic thread safety analysis) are enabled for this function.
1341 This attribute indicates that the function should emit a stack
1342 smashing protector. It is in the form of a "canary" --- a random value
1343 placed on the stack before the local variables that's checked upon
1344 return from the function to see if it has been overwritten. A
1345 heuristic is used to determine if a function needs stack protectors
1346 or not. The heuristic used will enable protectors for functions with:
1348 - Character arrays larger than ``ssp-buffer-size`` (default 8).
1349 - Aggregates containing character arrays larger than ``ssp-buffer-size``.
1350 - Calls to alloca() with variable sizes or constant sizes greater than
1351 ``ssp-buffer-size``.
1353 Variables that are identified as requiring a protector will be arranged
1354 on the stack such that they are adjacent to the stack protector guard.
1356 If a function that has an ``ssp`` attribute is inlined into a
1357 function that doesn't have an ``ssp`` attribute, then the resulting
1358 function will have an ``ssp`` attribute.
1360 This attribute indicates that the function should *always* emit a
1361 stack smashing protector. This overrides the ``ssp`` function
1364 Variables that are identified as requiring a protector will be arranged
1365 on the stack such that they are adjacent to the stack protector guard.
1366 The specific layout rules are:
1368 #. Large arrays and structures containing large arrays
1369 (``>= ssp-buffer-size``) are closest to the stack protector.
1370 #. Small arrays and structures containing small arrays
1371 (``< ssp-buffer-size``) are 2nd closest to the protector.
1372 #. Variables that have had their address taken are 3rd closest to the
1375 If a function that has an ``sspreq`` attribute is inlined into a
1376 function that doesn't have an ``sspreq`` attribute or which has an
1377 ``ssp`` or ``sspstrong`` attribute, then the resulting function will have
1378 an ``sspreq`` attribute.
1380 This attribute indicates that the function should emit a stack smashing
1381 protector. This attribute causes a strong heuristic to be used when
1382 determining if a function needs stack protectors. The strong heuristic
1383 will enable protectors for functions with:
1385 - Arrays of any size and type
1386 - Aggregates containing an array of any size and type.
1387 - Calls to alloca().
1388 - Local variables that have had their address taken.
1390 Variables that are identified as requiring a protector will be arranged
1391 on the stack such that they are adjacent to the stack protector guard.
1392 The specific layout rules are:
1394 #. Large arrays and structures containing large arrays
1395 (``>= ssp-buffer-size``) are closest to the stack protector.
1396 #. Small arrays and structures containing small arrays
1397 (``< ssp-buffer-size``) are 2nd closest to the protector.
1398 #. Variables that have had their address taken are 3rd closest to the
1401 This overrides the ``ssp`` function attribute.
1403 If a function that has an ``sspstrong`` attribute is inlined into a
1404 function that doesn't have an ``sspstrong`` attribute, then the
1405 resulting function will have an ``sspstrong`` attribute.
1407 This attribute indicates that the function will delegate to some other
1408 function with a tail call. The prototype of a thunk should not be used for
1409 optimization purposes. The caller is expected to cast the thunk prototype to
1410 match the thunk target prototype.
1412 This attribute indicates that the ABI being targeted requires that
1413 an unwind table entry be produce for this function even if we can
1414 show that no exceptions passes by it. This is normally the case for
1415 the ELF x86-64 abi, but it can be disabled for some compilation
1420 Module-Level Inline Assembly
1421 ----------------------------
1423 Modules may contain "module-level inline asm" blocks, which corresponds
1424 to the GCC "file scope inline asm" blocks. These blocks are internally
1425 concatenated by LLVM and treated as a single unit, but may be separated
1426 in the ``.ll`` file if desired. The syntax is very simple:
1428 .. code-block:: llvm
1430 module asm "inline asm code goes here"
1431 module asm "more can go here"
1433 The strings can contain any character by escaping non-printable
1434 characters. The escape sequence used is simply "\\xx" where "xx" is the
1435 two digit hex code for the number.
1437 The inline asm code is simply printed to the machine code .s file when
1438 assembly code is generated.
1440 .. _langref_datalayout:
1445 A module may specify a target specific data layout string that specifies
1446 how data is to be laid out in memory. The syntax for the data layout is
1449 .. code-block:: llvm
1451 target datalayout = "layout specification"
1453 The *layout specification* consists of a list of specifications
1454 separated by the minus sign character ('-'). Each specification starts
1455 with a letter and may include other information after the letter to
1456 define some aspect of the data layout. The specifications accepted are
1460 Specifies that the target lays out data in big-endian form. That is,
1461 the bits with the most significance have the lowest address
1464 Specifies that the target lays out data in little-endian form. That
1465 is, the bits with the least significance have the lowest address
1468 Specifies the natural alignment of the stack in bits. Alignment
1469 promotion of stack variables is limited to the natural stack
1470 alignment to avoid dynamic stack realignment. The stack alignment
1471 must be a multiple of 8-bits. If omitted, the natural stack
1472 alignment defaults to "unspecified", which does not prevent any
1473 alignment promotions.
1474 ``p[n]:<size>:<abi>:<pref>``
1475 This specifies the *size* of a pointer and its ``<abi>`` and
1476 ``<pref>``\erred alignments for address space ``n``. All sizes are in
1477 bits. The address space, ``n`` is optional, and if not specified,
1478 denotes the default address space 0. The value of ``n`` must be
1479 in the range [1,2^23).
1480 ``i<size>:<abi>:<pref>``
1481 This specifies the alignment for an integer type of a given bit
1482 ``<size>``. The value of ``<size>`` must be in the range [1,2^23).
1483 ``v<size>:<abi>:<pref>``
1484 This specifies the alignment for a vector type of a given bit
1486 ``f<size>:<abi>:<pref>``
1487 This specifies the alignment for a floating point type of a given bit
1488 ``<size>``. Only values of ``<size>`` that are supported by the target
1489 will work. 32 (float) and 64 (double) are supported on all targets; 80
1490 or 128 (different flavors of long double) are also supported on some
1493 This specifies the alignment for an object of aggregate type.
1495 If present, specifies that llvm names are mangled in the output. The
1498 * ``e``: ELF mangling: Private symbols get a ``.L`` prefix.
1499 * ``m``: Mips mangling: Private symbols get a ``$`` prefix.
1500 * ``o``: Mach-O mangling: Private symbols get ``L`` prefix. Other
1501 symbols get a ``_`` prefix.
1502 * ``w``: Windows COFF prefix: Similar to Mach-O, but stdcall and fastcall
1503 functions also get a suffix based on the frame size.
1504 ``n<size1>:<size2>:<size3>...``
1505 This specifies a set of native integer widths for the target CPU in
1506 bits. For example, it might contain ``n32`` for 32-bit PowerPC,
1507 ``n32:64`` for PowerPC 64, or ``n8:16:32:64`` for X86-64. Elements of
1508 this set are considered to support most general arithmetic operations
1511 On every specification that takes a ``<abi>:<pref>``, specifying the
1512 ``<pref>`` alignment is optional. If omitted, the preceding ``:``
1513 should be omitted too and ``<pref>`` will be equal to ``<abi>``.
1515 When constructing the data layout for a given target, LLVM starts with a
1516 default set of specifications which are then (possibly) overridden by
1517 the specifications in the ``datalayout`` keyword. The default
1518 specifications are given in this list:
1520 - ``E`` - big endian
1521 - ``p:64:64:64`` - 64-bit pointers with 64-bit alignment.
1522 - ``p[n]:64:64:64`` - Other address spaces are assumed to be the
1523 same as the default address space.
1524 - ``S0`` - natural stack alignment is unspecified
1525 - ``i1:8:8`` - i1 is 8-bit (byte) aligned
1526 - ``i8:8:8`` - i8 is 8-bit (byte) aligned
1527 - ``i16:16:16`` - i16 is 16-bit aligned
1528 - ``i32:32:32`` - i32 is 32-bit aligned
1529 - ``i64:32:64`` - i64 has ABI alignment of 32-bits but preferred
1530 alignment of 64-bits
1531 - ``f16:16:16`` - half is 16-bit aligned
1532 - ``f32:32:32`` - float is 32-bit aligned
1533 - ``f64:64:64`` - double is 64-bit aligned
1534 - ``f128:128:128`` - quad is 128-bit aligned
1535 - ``v64:64:64`` - 64-bit vector is 64-bit aligned
1536 - ``v128:128:128`` - 128-bit vector is 128-bit aligned
1537 - ``a:0:64`` - aggregates are 64-bit aligned
1539 When LLVM is determining the alignment for a given type, it uses the
1542 #. If the type sought is an exact match for one of the specifications,
1543 that specification is used.
1544 #. If no match is found, and the type sought is an integer type, then
1545 the smallest integer type that is larger than the bitwidth of the
1546 sought type is used. If none of the specifications are larger than
1547 the bitwidth then the largest integer type is used. For example,
1548 given the default specifications above, the i7 type will use the
1549 alignment of i8 (next largest) while both i65 and i256 will use the
1550 alignment of i64 (largest specified).
1551 #. If no match is found, and the type sought is a vector type, then the
1552 largest vector type that is smaller than the sought vector type will
1553 be used as a fall back. This happens because <128 x double> can be
1554 implemented in terms of 64 <2 x double>, for example.
1556 The function of the data layout string may not be what you expect.
1557 Notably, this is not a specification from the frontend of what alignment
1558 the code generator should use.
1560 Instead, if specified, the target data layout is required to match what
1561 the ultimate *code generator* expects. This string is used by the
1562 mid-level optimizers to improve code, and this only works if it matches
1563 what the ultimate code generator uses. There is no way to generate IR
1564 that does not embed this target-specific detail into the IR. If you
1565 don't specify the string, the default specifications will be used to
1566 generate a Data Layout and the optimization phases will operate
1567 accordingly and introduce target specificity into the IR with respect to
1568 these default specifications.
1575 A module may specify a target triple string that describes the target
1576 host. The syntax for the target triple is simply:
1578 .. code-block:: llvm
1580 target triple = "x86_64-apple-macosx10.7.0"
1582 The *target triple* string consists of a series of identifiers delimited
1583 by the minus sign character ('-'). The canonical forms are:
1587 ARCHITECTURE-VENDOR-OPERATING_SYSTEM
1588 ARCHITECTURE-VENDOR-OPERATING_SYSTEM-ENVIRONMENT
1590 This information is passed along to the backend so that it generates
1591 code for the proper architecture. It's possible to override this on the
1592 command line with the ``-mtriple`` command line option.
1594 .. _pointeraliasing:
1596 Pointer Aliasing Rules
1597 ----------------------
1599 Any memory access must be done through a pointer value associated with
1600 an address range of the memory access, otherwise the behavior is
1601 undefined. Pointer values are associated with address ranges according
1602 to the following rules:
1604 - A pointer value is associated with the addresses associated with any
1605 value it is *based* on.
1606 - An address of a global variable is associated with the address range
1607 of the variable's storage.
1608 - The result value of an allocation instruction is associated with the
1609 address range of the allocated storage.
1610 - A null pointer in the default address-space is associated with no
1612 - An integer constant other than zero or a pointer value returned from
1613 a function not defined within LLVM may be associated with address
1614 ranges allocated through mechanisms other than those provided by
1615 LLVM. Such ranges shall not overlap with any ranges of addresses
1616 allocated by mechanisms provided by LLVM.
1618 A pointer value is *based* on another pointer value according to the
1621 - A pointer value formed from a ``getelementptr`` operation is *based*
1622 on the first value operand of the ``getelementptr``.
1623 - The result value of a ``bitcast`` is *based* on the operand of the
1625 - A pointer value formed by an ``inttoptr`` is *based* on all pointer
1626 values that contribute (directly or indirectly) to the computation of
1627 the pointer's value.
1628 - The "*based* on" relationship is transitive.
1630 Note that this definition of *"based"* is intentionally similar to the
1631 definition of *"based"* in C99, though it is slightly weaker.
1633 LLVM IR does not associate types with memory. The result type of a
1634 ``load`` merely indicates the size and alignment of the memory from
1635 which to load, as well as the interpretation of the value. The first
1636 operand type of a ``store`` similarly only indicates the size and
1637 alignment of the store.
1639 Consequently, type-based alias analysis, aka TBAA, aka
1640 ``-fstrict-aliasing``, is not applicable to general unadorned LLVM IR.
1641 :ref:`Metadata <metadata>` may be used to encode additional information
1642 which specialized optimization passes may use to implement type-based
1647 Volatile Memory Accesses
1648 ------------------------
1650 Certain memory accesses, such as :ref:`load <i_load>`'s,
1651 :ref:`store <i_store>`'s, and :ref:`llvm.memcpy <int_memcpy>`'s may be
1652 marked ``volatile``. The optimizers must not change the number of
1653 volatile operations or change their order of execution relative to other
1654 volatile operations. The optimizers *may* change the order of volatile
1655 operations relative to non-volatile operations. This is not Java's
1656 "volatile" and has no cross-thread synchronization behavior.
1658 IR-level volatile loads and stores cannot safely be optimized into
1659 llvm.memcpy or llvm.memmove intrinsics even when those intrinsics are
1660 flagged volatile. Likewise, the backend should never split or merge
1661 target-legal volatile load/store instructions.
1663 .. admonition:: Rationale
1665 Platforms may rely on volatile loads and stores of natively supported
1666 data width to be executed as single instruction. For example, in C
1667 this holds for an l-value of volatile primitive type with native
1668 hardware support, but not necessarily for aggregate types. The
1669 frontend upholds these expectations, which are intentionally
1670 unspecified in the IR. The rules above ensure that IR transformation
1671 do not violate the frontend's contract with the language.
1675 Memory Model for Concurrent Operations
1676 --------------------------------------
1678 The LLVM IR does not define any way to start parallel threads of
1679 execution or to register signal handlers. Nonetheless, there are
1680 platform-specific ways to create them, and we define LLVM IR's behavior
1681 in their presence. This model is inspired by the C++0x memory model.
1683 For a more informal introduction to this model, see the :doc:`Atomics`.
1685 We define a *happens-before* partial order as the least partial order
1688 - Is a superset of single-thread program order, and
1689 - When a *synchronizes-with* ``b``, includes an edge from ``a`` to
1690 ``b``. *Synchronizes-with* pairs are introduced by platform-specific
1691 techniques, like pthread locks, thread creation, thread joining,
1692 etc., and by atomic instructions. (See also :ref:`Atomic Memory Ordering
1693 Constraints <ordering>`).
1695 Note that program order does not introduce *happens-before* edges
1696 between a thread and signals executing inside that thread.
1698 Every (defined) read operation (load instructions, memcpy, atomic
1699 loads/read-modify-writes, etc.) R reads a series of bytes written by
1700 (defined) write operations (store instructions, atomic
1701 stores/read-modify-writes, memcpy, etc.). For the purposes of this
1702 section, initialized globals are considered to have a write of the
1703 initializer which is atomic and happens before any other read or write
1704 of the memory in question. For each byte of a read R, R\ :sub:`byte`
1705 may see any write to the same byte, except:
1707 - If write\ :sub:`1` happens before write\ :sub:`2`, and
1708 write\ :sub:`2` happens before R\ :sub:`byte`, then
1709 R\ :sub:`byte` does not see write\ :sub:`1`.
1710 - If R\ :sub:`byte` happens before write\ :sub:`3`, then
1711 R\ :sub:`byte` does not see write\ :sub:`3`.
1713 Given that definition, R\ :sub:`byte` is defined as follows:
1715 - If R is volatile, the result is target-dependent. (Volatile is
1716 supposed to give guarantees which can support ``sig_atomic_t`` in
1717 C/C++, and may be used for accesses to addresses that do not behave
1718 like normal memory. It does not generally provide cross-thread
1720 - Otherwise, if there is no write to the same byte that happens before
1721 R\ :sub:`byte`, R\ :sub:`byte` returns ``undef`` for that byte.
1722 - Otherwise, if R\ :sub:`byte` may see exactly one write,
1723 R\ :sub:`byte` returns the value written by that write.
1724 - Otherwise, if R is atomic, and all the writes R\ :sub:`byte` may
1725 see are atomic, it chooses one of the values written. See the :ref:`Atomic
1726 Memory Ordering Constraints <ordering>` section for additional
1727 constraints on how the choice is made.
1728 - Otherwise R\ :sub:`byte` returns ``undef``.
1730 R returns the value composed of the series of bytes it read. This
1731 implies that some bytes within the value may be ``undef`` **without**
1732 the entire value being ``undef``. Note that this only defines the
1733 semantics of the operation; it doesn't mean that targets will emit more
1734 than one instruction to read the series of bytes.
1736 Note that in cases where none of the atomic intrinsics are used, this
1737 model places only one restriction on IR transformations on top of what
1738 is required for single-threaded execution: introducing a store to a byte
1739 which might not otherwise be stored is not allowed in general.
1740 (Specifically, in the case where another thread might write to and read
1741 from an address, introducing a store can change a load that may see
1742 exactly one write into a load that may see multiple writes.)
1746 Atomic Memory Ordering Constraints
1747 ----------------------------------
1749 Atomic instructions (:ref:`cmpxchg <i_cmpxchg>`,
1750 :ref:`atomicrmw <i_atomicrmw>`, :ref:`fence <i_fence>`,
1751 :ref:`atomic load <i_load>`, and :ref:`atomic store <i_store>`) take
1752 ordering parameters that determine which other atomic instructions on
1753 the same address they *synchronize with*. These semantics are borrowed
1754 from Java and C++0x, but are somewhat more colloquial. If these
1755 descriptions aren't precise enough, check those specs (see spec
1756 references in the :doc:`atomics guide <Atomics>`).
1757 :ref:`fence <i_fence>` instructions treat these orderings somewhat
1758 differently since they don't take an address. See that instruction's
1759 documentation for details.
1761 For a simpler introduction to the ordering constraints, see the
1765 The set of values that can be read is governed by the happens-before
1766 partial order. A value cannot be read unless some operation wrote
1767 it. This is intended to provide a guarantee strong enough to model
1768 Java's non-volatile shared variables. This ordering cannot be
1769 specified for read-modify-write operations; it is not strong enough
1770 to make them atomic in any interesting way.
1772 In addition to the guarantees of ``unordered``, there is a single
1773 total order for modifications by ``monotonic`` operations on each
1774 address. All modification orders must be compatible with the
1775 happens-before order. There is no guarantee that the modification
1776 orders can be combined to a global total order for the whole program
1777 (and this often will not be possible). The read in an atomic
1778 read-modify-write operation (:ref:`cmpxchg <i_cmpxchg>` and
1779 :ref:`atomicrmw <i_atomicrmw>`) reads the value in the modification
1780 order immediately before the value it writes. If one atomic read
1781 happens before another atomic read of the same address, the later
1782 read must see the same value or a later value in the address's
1783 modification order. This disallows reordering of ``monotonic`` (or
1784 stronger) operations on the same address. If an address is written
1785 ``monotonic``-ally by one thread, and other threads ``monotonic``-ally
1786 read that address repeatedly, the other threads must eventually see
1787 the write. This corresponds to the C++0x/C1x
1788 ``memory_order_relaxed``.
1790 In addition to the guarantees of ``monotonic``, a
1791 *synchronizes-with* edge may be formed with a ``release`` operation.
1792 This is intended to model C++'s ``memory_order_acquire``.
1794 In addition to the guarantees of ``monotonic``, if this operation
1795 writes a value which is subsequently read by an ``acquire``
1796 operation, it *synchronizes-with* that operation. (This isn't a
1797 complete description; see the C++0x definition of a release
1798 sequence.) This corresponds to the C++0x/C1x
1799 ``memory_order_release``.
1800 ``acq_rel`` (acquire+release)
1801 Acts as both an ``acquire`` and ``release`` operation on its
1802 address. This corresponds to the C++0x/C1x ``memory_order_acq_rel``.
1803 ``seq_cst`` (sequentially consistent)
1804 In addition to the guarantees of ``acq_rel`` (``acquire`` for an
1805 operation that only reads, ``release`` for an operation that only
1806 writes), there is a global total order on all
1807 sequentially-consistent operations on all addresses, which is
1808 consistent with the *happens-before* partial order and with the
1809 modification orders of all the affected addresses. Each
1810 sequentially-consistent read sees the last preceding write to the
1811 same address in this global order. This corresponds to the C++0x/C1x
1812 ``memory_order_seq_cst`` and Java volatile.
1816 If an atomic operation is marked ``singlethread``, it only *synchronizes
1817 with* or participates in modification and seq\_cst total orderings with
1818 other operations running in the same thread (for example, in signal
1826 LLVM IR floating-point binary ops (:ref:`fadd <i_fadd>`,
1827 :ref:`fsub <i_fsub>`, :ref:`fmul <i_fmul>`, :ref:`fdiv <i_fdiv>`,
1828 :ref:`frem <i_frem>`) have the following flags that can be set to enable
1829 otherwise unsafe floating point operations
1832 No NaNs - Allow optimizations to assume the arguments and result are not
1833 NaN. Such optimizations are required to retain defined behavior over
1834 NaNs, but the value of the result is undefined.
1837 No Infs - Allow optimizations to assume the arguments and result are not
1838 +/-Inf. Such optimizations are required to retain defined behavior over
1839 +/-Inf, but the value of the result is undefined.
1842 No Signed Zeros - Allow optimizations to treat the sign of a zero
1843 argument or result as insignificant.
1846 Allow Reciprocal - Allow optimizations to use the reciprocal of an
1847 argument rather than perform division.
1850 Fast - Allow algebraically equivalent transformations that may
1851 dramatically change results in floating point (e.g. reassociate). This
1852 flag implies all the others.
1856 Use-list Order Directives
1857 -------------------------
1859 Use-list directives encode the in-memory order of each use-list, allowing the
1860 order to be recreated. ``<order-indexes>`` is a comma-separated list of
1861 indexes that are assigned to the referenced value's uses. The referenced
1862 value's use-list is immediately sorted by these indexes.
1864 Use-list directives may appear at function scope or global scope. They are not
1865 instructions, and have no effect on the semantics of the IR. When they're at
1866 function scope, they must appear after the terminator of the final basic block.
1868 If basic blocks have their address taken via ``blockaddress()`` expressions,
1869 ``uselistorder_bb`` can be used to reorder their use-lists from outside their
1876 uselistorder <ty> <value>, { <order-indexes> }
1877 uselistorder_bb @function, %block { <order-indexes> }
1883 define void @foo(i32 %arg1, i32 %arg2) {
1885 ; ... instructions ...
1887 ; ... instructions ...
1889 ; At function scope.
1890 uselistorder i32 %arg1, { 1, 0, 2 }
1891 uselistorder label %bb, { 1, 0 }
1895 uselistorder i32* @global, { 1, 2, 0 }
1896 uselistorder i32 7, { 1, 0 }
1897 uselistorder i32 (i32) @bar, { 1, 0 }
1898 uselistorder_bb @foo, %bb, { 5, 1, 3, 2, 0, 4 }
1905 The LLVM type system is one of the most important features of the
1906 intermediate representation. Being typed enables a number of
1907 optimizations to be performed on the intermediate representation
1908 directly, without having to do extra analyses on the side before the
1909 transformation. A strong type system makes it easier to read the
1910 generated code and enables novel analyses and transformations that are
1911 not feasible to perform on normal three address code representations.
1921 The void type does not represent any value and has no size.
1939 The function type can be thought of as a function signature. It consists of a
1940 return type and a list of formal parameter types. The return type of a function
1941 type is a void type or first class type --- except for :ref:`label <t_label>`
1942 and :ref:`metadata <t_metadata>` types.
1948 <returntype> (<parameter list>)
1950 ...where '``<parameter list>``' is a comma-separated list of type
1951 specifiers. Optionally, the parameter list may include a type ``...``, which
1952 indicates that the function takes a variable number of arguments. Variable
1953 argument functions can access their arguments with the :ref:`variable argument
1954 handling intrinsic <int_varargs>` functions. '``<returntype>``' is any type
1955 except :ref:`label <t_label>` and :ref:`metadata <t_metadata>`.
1959 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1960 | ``i32 (i32)`` | function taking an ``i32``, returning an ``i32`` |
1961 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1962 | ``float (i16, i32 *) *`` | :ref:`Pointer <t_pointer>` to a function that takes an ``i16`` and a :ref:`pointer <t_pointer>` to ``i32``, returning ``float``. |
1963 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1964 | ``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. |
1965 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1966 | ``{i32, i32} (i32)`` | A function taking an ``i32``, returning a :ref:`structure <t_struct>` containing two ``i32`` values |
1967 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1974 The :ref:`first class <t_firstclass>` types are perhaps the most important.
1975 Values of these types are the only ones which can be produced by
1983 These are the types that are valid in registers from CodeGen's perspective.
1992 The integer type is a very simple type that simply specifies an
1993 arbitrary bit width for the integer type desired. Any bit width from 1
1994 bit to 2\ :sup:`23`\ -1 (about 8 million) can be specified.
2002 The number of bits the integer will occupy is specified by the ``N``
2008 +----------------+------------------------------------------------+
2009 | ``i1`` | a single-bit integer. |
2010 +----------------+------------------------------------------------+
2011 | ``i32`` | a 32-bit integer. |
2012 +----------------+------------------------------------------------+
2013 | ``i1942652`` | a really big integer of over 1 million bits. |
2014 +----------------+------------------------------------------------+
2018 Floating Point Types
2019 """"""""""""""""""""
2028 - 16-bit floating point value
2031 - 32-bit floating point value
2034 - 64-bit floating point value
2037 - 128-bit floating point value (112-bit mantissa)
2040 - 80-bit floating point value (X87)
2043 - 128-bit floating point value (two 64-bits)
2050 The x86_mmx type represents a value held in an MMX register on an x86
2051 machine. The operations allowed on it are quite limited: parameters and
2052 return values, load and store, and bitcast. User-specified MMX
2053 instructions are represented as intrinsic or asm calls with arguments
2054 and/or results of this type. There are no arrays, vectors or constants
2071 The pointer type is used to specify memory locations. Pointers are
2072 commonly used to reference objects in memory.
2074 Pointer types may have an optional address space attribute defining the
2075 numbered address space where the pointed-to object resides. The default
2076 address space is number zero. The semantics of non-zero address spaces
2077 are target-specific.
2079 Note that LLVM does not permit pointers to void (``void*``) nor does it
2080 permit pointers to labels (``label*``). Use ``i8*`` instead.
2090 +-------------------------+--------------------------------------------------------------------------------------------------------------+
2091 | ``[4 x i32]*`` | A :ref:`pointer <t_pointer>` to :ref:`array <t_array>` of four ``i32`` values. |
2092 +-------------------------+--------------------------------------------------------------------------------------------------------------+
2093 | ``i32 (i32*) *`` | A :ref:`pointer <t_pointer>` to a :ref:`function <t_function>` that takes an ``i32*``, returning an ``i32``. |
2094 +-------------------------+--------------------------------------------------------------------------------------------------------------+
2095 | ``i32 addrspace(5)*`` | A :ref:`pointer <t_pointer>` to an ``i32`` value that resides in address space #5. |
2096 +-------------------------+--------------------------------------------------------------------------------------------------------------+
2105 A vector type is a simple derived type that represents a vector of
2106 elements. Vector types are used when multiple primitive data are
2107 operated in parallel using a single instruction (SIMD). A vector type
2108 requires a size (number of elements) and an underlying primitive data
2109 type. Vector types are considered :ref:`first class <t_firstclass>`.
2115 < <# elements> x <elementtype> >
2117 The number of elements is a constant integer value larger than 0;
2118 elementtype may be any integer, floating point or pointer type. Vectors
2119 of size zero are not allowed.
2123 +-------------------+--------------------------------------------------+
2124 | ``<4 x i32>`` | Vector of 4 32-bit integer values. |
2125 +-------------------+--------------------------------------------------+
2126 | ``<8 x float>`` | Vector of 8 32-bit floating-point values. |
2127 +-------------------+--------------------------------------------------+
2128 | ``<2 x i64>`` | Vector of 2 64-bit integer values. |
2129 +-------------------+--------------------------------------------------+
2130 | ``<4 x i64*>`` | Vector of 4 pointers to 64-bit integer values. |
2131 +-------------------+--------------------------------------------------+
2140 The label type represents code labels.
2155 The metadata type represents embedded metadata. No derived types may be
2156 created from metadata except for :ref:`function <t_function>` arguments.
2169 Aggregate Types are a subset of derived types that can contain multiple
2170 member types. :ref:`Arrays <t_array>` and :ref:`structs <t_struct>` are
2171 aggregate types. :ref:`Vectors <t_vector>` are not considered to be
2181 The array type is a very simple derived type that arranges elements
2182 sequentially in memory. The array type requires a size (number of
2183 elements) and an underlying data type.
2189 [<# elements> x <elementtype>]
2191 The number of elements is a constant integer value; ``elementtype`` may
2192 be any type with a size.
2196 +------------------+--------------------------------------+
2197 | ``[40 x i32]`` | Array of 40 32-bit integer values. |
2198 +------------------+--------------------------------------+
2199 | ``[41 x i32]`` | Array of 41 32-bit integer values. |
2200 +------------------+--------------------------------------+
2201 | ``[4 x i8]`` | Array of 4 8-bit integer values. |
2202 +------------------+--------------------------------------+
2204 Here are some examples of multidimensional arrays:
2206 +-----------------------------+----------------------------------------------------------+
2207 | ``[3 x [4 x i32]]`` | 3x4 array of 32-bit integer values. |
2208 +-----------------------------+----------------------------------------------------------+
2209 | ``[12 x [10 x float]]`` | 12x10 array of single precision floating point values. |
2210 +-----------------------------+----------------------------------------------------------+
2211 | ``[2 x [3 x [4 x i16]]]`` | 2x3x4 array of 16-bit integer values. |
2212 +-----------------------------+----------------------------------------------------------+
2214 There is no restriction on indexing beyond the end of the array implied
2215 by a static type (though there are restrictions on indexing beyond the
2216 bounds of an allocated object in some cases). This means that
2217 single-dimension 'variable sized array' addressing can be implemented in
2218 LLVM with a zero length array type. An implementation of 'pascal style
2219 arrays' in LLVM could use the type "``{ i32, [0 x float]}``", for
2229 The structure type is used to represent a collection of data members
2230 together in memory. The elements of a structure may be any type that has
2233 Structures in memory are accessed using '``load``' and '``store``' by
2234 getting a pointer to a field with the '``getelementptr``' instruction.
2235 Structures in registers are accessed using the '``extractvalue``' and
2236 '``insertvalue``' instructions.
2238 Structures may optionally be "packed" structures, which indicate that
2239 the alignment of the struct is one byte, and that there is no padding
2240 between the elements. In non-packed structs, padding between field types
2241 is inserted as defined by the DataLayout string in the module, which is
2242 required to match what the underlying code generator expects.
2244 Structures can either be "literal" or "identified". A literal structure
2245 is defined inline with other types (e.g. ``{i32, i32}*``) whereas
2246 identified types are always defined at the top level with a name.
2247 Literal types are uniqued by their contents and can never be recursive
2248 or opaque since there is no way to write one. Identified types can be
2249 recursive, can be opaqued, and are never uniqued.
2255 %T1 = type { <type list> } ; Identified normal struct type
2256 %T2 = type <{ <type list> }> ; Identified packed struct type
2260 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2261 | ``{ i32, i32, i32 }`` | A triple of three ``i32`` values |
2262 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2263 | ``{ 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``. |
2264 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2265 | ``<{ i8, i32 }>`` | A packed struct known to be 5 bytes in size. |
2266 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2270 Opaque Structure Types
2271 """"""""""""""""""""""
2275 Opaque structure types are used to represent named structure types that
2276 do not have a body specified. This corresponds (for example) to the C
2277 notion of a forward declared structure.
2288 +--------------+-------------------+
2289 | ``opaque`` | An opaque type. |
2290 +--------------+-------------------+
2297 LLVM has several different basic types of constants. This section
2298 describes them all and their syntax.
2303 **Boolean constants**
2304 The two strings '``true``' and '``false``' are both valid constants
2306 **Integer constants**
2307 Standard integers (such as '4') are constants of the
2308 :ref:`integer <t_integer>` type. Negative numbers may be used with
2310 **Floating point constants**
2311 Floating point constants use standard decimal notation (e.g.
2312 123.421), exponential notation (e.g. 1.23421e+2), or a more precise
2313 hexadecimal notation (see below). The assembler requires the exact
2314 decimal value of a floating-point constant. For example, the
2315 assembler accepts 1.25 but rejects 1.3 because 1.3 is a repeating
2316 decimal in binary. Floating point constants must have a :ref:`floating
2317 point <t_floating>` type.
2318 **Null pointer constants**
2319 The identifier '``null``' is recognized as a null pointer constant
2320 and must be of :ref:`pointer type <t_pointer>`.
2322 The one non-intuitive notation for constants is the hexadecimal form of
2323 floating point constants. For example, the form
2324 '``double 0x432ff973cafa8000``' is equivalent to (but harder to read
2325 than) '``double 4.5e+15``'. The only time hexadecimal floating point
2326 constants are required (and the only time that they are generated by the
2327 disassembler) is when a floating point constant must be emitted but it
2328 cannot be represented as a decimal floating point number in a reasonable
2329 number of digits. For example, NaN's, infinities, and other special
2330 values are represented in their IEEE hexadecimal format so that assembly
2331 and disassembly do not cause any bits to change in the constants.
2333 When using the hexadecimal form, constants of types half, float, and
2334 double are represented using the 16-digit form shown above (which
2335 matches the IEEE754 representation for double); half and float values
2336 must, however, be exactly representable as IEEE 754 half and single
2337 precision, respectively. Hexadecimal format is always used for long
2338 double, and there are three forms of long double. The 80-bit format used
2339 by x86 is represented as ``0xK`` followed by 20 hexadecimal digits. The
2340 128-bit format used by PowerPC (two adjacent doubles) is represented by
2341 ``0xM`` followed by 32 hexadecimal digits. The IEEE 128-bit format is
2342 represented by ``0xL`` followed by 32 hexadecimal digits. Long doubles
2343 will only work if they match the long double format on your target.
2344 The IEEE 16-bit format (half precision) is represented by ``0xH``
2345 followed by 4 hexadecimal digits. All hexadecimal formats are big-endian
2346 (sign bit at the left).
2348 There are no constants of type x86_mmx.
2350 .. _complexconstants:
2355 Complex constants are a (potentially recursive) combination of simple
2356 constants and smaller complex constants.
2358 **Structure constants**
2359 Structure constants are represented with notation similar to
2360 structure type definitions (a comma separated list of elements,
2361 surrounded by braces (``{}``)). For example:
2362 "``{ i32 4, float 17.0, i32* @G }``", where "``@G``" is declared as
2363 "``@G = external global i32``". Structure constants must have
2364 :ref:`structure type <t_struct>`, and the number and types of elements
2365 must match those specified by the type.
2367 Array constants are represented with notation similar to array type
2368 definitions (a comma separated list of elements, surrounded by
2369 square brackets (``[]``)). For example:
2370 "``[ i32 42, i32 11, i32 74 ]``". Array constants must have
2371 :ref:`array type <t_array>`, and the number and types of elements must
2372 match those specified by the type. As a special case, character array
2373 constants may also be represented as a double-quoted string using the ``c``
2374 prefix. For example: "``c"Hello World\0A\00"``".
2375 **Vector constants**
2376 Vector constants are represented with notation similar to vector
2377 type definitions (a comma separated list of elements, surrounded by
2378 less-than/greater-than's (``<>``)). For example:
2379 "``< i32 42, i32 11, i32 74, i32 100 >``". Vector constants
2380 must have :ref:`vector type <t_vector>`, and the number and types of
2381 elements must match those specified by the type.
2382 **Zero initialization**
2383 The string '``zeroinitializer``' can be used to zero initialize a
2384 value to zero of *any* type, including scalar and
2385 :ref:`aggregate <t_aggregate>` types. This is often used to avoid
2386 having to print large zero initializers (e.g. for large arrays) and
2387 is always exactly equivalent to using explicit zero initializers.
2389 A metadata node is a constant tuple without types. For example:
2390 "``!{!0, !{!2, !0}, !"test"}``". Metadata can reference constant values,
2391 for example: "``!{!0, i32 0, i8* @global, i64 (i64)* @function, !"str"}``".
2392 Unlike other typed constants that are meant to be interpreted as part of
2393 the instruction stream, metadata is a place to attach additional
2394 information such as debug info.
2396 Global Variable and Function Addresses
2397 --------------------------------------
2399 The addresses of :ref:`global variables <globalvars>` and
2400 :ref:`functions <functionstructure>` are always implicitly valid
2401 (link-time) constants. These constants are explicitly referenced when
2402 the :ref:`identifier for the global <identifiers>` is used and always have
2403 :ref:`pointer <t_pointer>` type. For example, the following is a legal LLVM
2406 .. code-block:: llvm
2410 @Z = global [2 x i32*] [ i32* @X, i32* @Y ]
2417 The string '``undef``' can be used anywhere a constant is expected, and
2418 indicates that the user of the value may receive an unspecified
2419 bit-pattern. Undefined values may be of any type (other than '``label``'
2420 or '``void``') and be used anywhere a constant is permitted.
2422 Undefined values are useful because they indicate to the compiler that
2423 the program is well defined no matter what value is used. This gives the
2424 compiler more freedom to optimize. Here are some examples of
2425 (potentially surprising) transformations that are valid (in pseudo IR):
2427 .. code-block:: llvm
2437 This is safe because all of the output bits are affected by the undef
2438 bits. Any output bit can have a zero or one depending on the input bits.
2440 .. code-block:: llvm
2451 These logical operations have bits that are not always affected by the
2452 input. For example, if ``%X`` has a zero bit, then the output of the
2453 '``and``' operation will always be a zero for that bit, no matter what
2454 the corresponding bit from the '``undef``' is. As such, it is unsafe to
2455 optimize or assume that the result of the '``and``' is '``undef``'.
2456 However, it is safe to assume that all bits of the '``undef``' could be
2457 0, and optimize the '``and``' to 0. Likewise, it is safe to assume that
2458 all the bits of the '``undef``' operand to the '``or``' could be set,
2459 allowing the '``or``' to be folded to -1.
2461 .. code-block:: llvm
2463 %A = select undef, %X, %Y
2464 %B = select undef, 42, %Y
2465 %C = select %X, %Y, undef
2475 This set of examples shows that undefined '``select``' (and conditional
2476 branch) conditions can go *either way*, but they have to come from one
2477 of the two operands. In the ``%A`` example, if ``%X`` and ``%Y`` were
2478 both known to have a clear low bit, then ``%A`` would have to have a
2479 cleared low bit. However, in the ``%C`` example, the optimizer is
2480 allowed to assume that the '``undef``' operand could be the same as
2481 ``%Y``, allowing the whole '``select``' to be eliminated.
2483 .. code-block:: llvm
2485 %A = xor undef, undef
2502 This example points out that two '``undef``' operands are not
2503 necessarily the same. This can be surprising to people (and also matches
2504 C semantics) where they assume that "``X^X``" is always zero, even if
2505 ``X`` is undefined. This isn't true for a number of reasons, but the
2506 short answer is that an '``undef``' "variable" can arbitrarily change
2507 its value over its "live range". This is true because the variable
2508 doesn't actually *have a live range*. Instead, the value is logically
2509 read from arbitrary registers that happen to be around when needed, so
2510 the value is not necessarily consistent over time. In fact, ``%A`` and
2511 ``%C`` need to have the same semantics or the core LLVM "replace all
2512 uses with" concept would not hold.
2514 .. code-block:: llvm
2522 These examples show the crucial difference between an *undefined value*
2523 and *undefined behavior*. An undefined value (like '``undef``') is
2524 allowed to have an arbitrary bit-pattern. This means that the ``%A``
2525 operation can be constant folded to '``undef``', because the '``undef``'
2526 could be an SNaN, and ``fdiv`` is not (currently) defined on SNaN's.
2527 However, in the second example, we can make a more aggressive
2528 assumption: because the ``undef`` is allowed to be an arbitrary value,
2529 we are allowed to assume that it could be zero. Since a divide by zero
2530 has *undefined behavior*, we are allowed to assume that the operation
2531 does not execute at all. This allows us to delete the divide and all
2532 code after it. Because the undefined operation "can't happen", the
2533 optimizer can assume that it occurs in dead code.
2535 .. code-block:: llvm
2537 a: store undef -> %X
2538 b: store %X -> undef
2543 These examples reiterate the ``fdiv`` example: a store *of* an undefined
2544 value can be assumed to not have any effect; we can assume that the
2545 value is overwritten with bits that happen to match what was already
2546 there. However, a store *to* an undefined location could clobber
2547 arbitrary memory, therefore, it has undefined behavior.
2554 Poison values are similar to :ref:`undef values <undefvalues>`, however
2555 they also represent the fact that an instruction or constant expression
2556 that cannot evoke side effects has nevertheless detected a condition
2557 that results in undefined behavior.
2559 There is currently no way of representing a poison value in the IR; they
2560 only exist when produced by operations such as :ref:`add <i_add>` with
2563 Poison value behavior is defined in terms of value *dependence*:
2565 - Values other than :ref:`phi <i_phi>` nodes depend on their operands.
2566 - :ref:`Phi <i_phi>` nodes depend on the operand corresponding to
2567 their dynamic predecessor basic block.
2568 - Function arguments depend on the corresponding actual argument values
2569 in the dynamic callers of their functions.
2570 - :ref:`Call <i_call>` instructions depend on the :ref:`ret <i_ret>`
2571 instructions that dynamically transfer control back to them.
2572 - :ref:`Invoke <i_invoke>` instructions depend on the
2573 :ref:`ret <i_ret>`, :ref:`resume <i_resume>`, or exception-throwing
2574 call instructions that dynamically transfer control back to them.
2575 - Non-volatile loads and stores depend on the most recent stores to all
2576 of the referenced memory addresses, following the order in the IR
2577 (including loads and stores implied by intrinsics such as
2578 :ref:`@llvm.memcpy <int_memcpy>`.)
2579 - An instruction with externally visible side effects depends on the
2580 most recent preceding instruction with externally visible side
2581 effects, following the order in the IR. (This includes :ref:`volatile
2582 operations <volatile>`.)
2583 - An instruction *control-depends* on a :ref:`terminator
2584 instruction <terminators>` if the terminator instruction has
2585 multiple successors and the instruction is always executed when
2586 control transfers to one of the successors, and may not be executed
2587 when control is transferred to another.
2588 - Additionally, an instruction also *control-depends* on a terminator
2589 instruction if the set of instructions it otherwise depends on would
2590 be different if the terminator had transferred control to a different
2592 - Dependence is transitive.
2594 Poison values have the same behavior as :ref:`undef values <undefvalues>`,
2595 with the additional effect that any instruction that has a *dependence*
2596 on a poison value has undefined behavior.
2598 Here are some examples:
2600 .. code-block:: llvm
2603 %poison = sub nuw i32 0, 1 ; Results in a poison value.
2604 %still_poison = and i32 %poison, 0 ; 0, but also poison.
2605 %poison_yet_again = getelementptr i32, i32* @h, i32 %still_poison
2606 store i32 0, i32* %poison_yet_again ; memory at @h[0] is poisoned
2608 store i32 %poison, i32* @g ; Poison value stored to memory.
2609 %poison2 = load i32, i32* @g ; Poison value loaded back from memory.
2611 store volatile i32 %poison, i32* @g ; External observation; undefined behavior.
2613 %narrowaddr = bitcast i32* @g to i16*
2614 %wideaddr = bitcast i32* @g to i64*
2615 %poison3 = load i16, i16* %narrowaddr ; Returns a poison value.
2616 %poison4 = load i64, i64* %wideaddr ; Returns a poison value.
2618 %cmp = icmp slt i32 %poison, 0 ; Returns a poison value.
2619 br i1 %cmp, label %true, label %end ; Branch to either destination.
2622 store volatile i32 0, i32* @g ; This is control-dependent on %cmp, so
2623 ; it has undefined behavior.
2627 %p = phi i32 [ 0, %entry ], [ 1, %true ]
2628 ; Both edges into this PHI are
2629 ; control-dependent on %cmp, so this
2630 ; always results in a poison value.
2632 store volatile i32 0, i32* @g ; This would depend on the store in %true
2633 ; if %cmp is true, or the store in %entry
2634 ; otherwise, so this is undefined behavior.
2636 br i1 %cmp, label %second_true, label %second_end
2637 ; The same branch again, but this time the
2638 ; true block doesn't have side effects.
2645 store volatile i32 0, i32* @g ; This time, the instruction always depends
2646 ; on the store in %end. Also, it is
2647 ; control-equivalent to %end, so this is
2648 ; well-defined (ignoring earlier undefined
2649 ; behavior in this example).
2653 Addresses of Basic Blocks
2654 -------------------------
2656 ``blockaddress(@function, %block)``
2658 The '``blockaddress``' constant computes the address of the specified
2659 basic block in the specified function, and always has an ``i8*`` type.
2660 Taking the address of the entry block is illegal.
2662 This value only has defined behavior when used as an operand to the
2663 ':ref:`indirectbr <i_indirectbr>`' instruction, or for comparisons
2664 against null. Pointer equality tests between labels addresses results in
2665 undefined behavior --- though, again, comparison against null is ok, and
2666 no label is equal to the null pointer. This may be passed around as an
2667 opaque pointer sized value as long as the bits are not inspected. This
2668 allows ``ptrtoint`` and arithmetic to be performed on these values so
2669 long as the original value is reconstituted before the ``indirectbr``
2672 Finally, some targets may provide defined semantics when using the value
2673 as the operand to an inline assembly, but that is target specific.
2677 Constant Expressions
2678 --------------------
2680 Constant expressions are used to allow expressions involving other
2681 constants to be used as constants. Constant expressions may be of any
2682 :ref:`first class <t_firstclass>` type and may involve any LLVM operation
2683 that does not have side effects (e.g. load and call are not supported).
2684 The following is the syntax for constant expressions:
2686 ``trunc (CST to TYPE)``
2687 Truncate a constant to another type. The bit size of CST must be
2688 larger than the bit size of TYPE. Both types must be integers.
2689 ``zext (CST to TYPE)``
2690 Zero extend a constant to another type. The bit size of CST must be
2691 smaller than the bit size of TYPE. Both types must be integers.
2692 ``sext (CST to TYPE)``
2693 Sign extend a constant to another type. The bit size of CST must be
2694 smaller than the bit size of TYPE. Both types must be integers.
2695 ``fptrunc (CST to TYPE)``
2696 Truncate a floating point constant to another floating point type.
2697 The size of CST must be larger than the size of TYPE. Both types
2698 must be floating point.
2699 ``fpext (CST to TYPE)``
2700 Floating point extend a constant to another type. The size of CST
2701 must be smaller or equal to the size of TYPE. Both types must be
2703 ``fptoui (CST to TYPE)``
2704 Convert a floating point constant to the corresponding unsigned
2705 integer constant. TYPE must be a scalar or vector integer type. CST
2706 must be of scalar or vector floating point type. Both CST and TYPE
2707 must be scalars, or vectors of the same number of elements. If the
2708 value won't fit in the integer type, the results are undefined.
2709 ``fptosi (CST to TYPE)``
2710 Convert a floating point constant to the corresponding signed
2711 integer constant. TYPE must be a scalar or vector integer type. CST
2712 must be of scalar or vector floating point type. Both CST and TYPE
2713 must be scalars, or vectors of the same number of elements. If the
2714 value won't fit in the integer type, the results are undefined.
2715 ``uitofp (CST to TYPE)``
2716 Convert an unsigned integer constant to the corresponding floating
2717 point constant. TYPE must be a scalar or vector floating point type.
2718 CST must be of scalar or vector integer type. Both CST and TYPE must
2719 be scalars, or vectors of the same number of elements. If the value
2720 won't fit in the floating point type, the results are undefined.
2721 ``sitofp (CST to TYPE)``
2722 Convert a signed integer constant to the corresponding floating
2723 point constant. TYPE must be a scalar or vector floating point type.
2724 CST must be of scalar or vector integer type. Both CST and TYPE must
2725 be scalars, or vectors of the same number of elements. If the value
2726 won't fit in the floating point type, the results are undefined.
2727 ``ptrtoint (CST to TYPE)``
2728 Convert a pointer typed constant to the corresponding integer
2729 constant. ``TYPE`` must be an integer type. ``CST`` must be of
2730 pointer type. The ``CST`` value is zero extended, truncated, or
2731 unchanged to make it fit in ``TYPE``.
2732 ``inttoptr (CST to TYPE)``
2733 Convert an integer constant to a pointer constant. TYPE must be a
2734 pointer type. CST must be of integer type. The CST value is zero
2735 extended, truncated, or unchanged to make it fit in a pointer size.
2736 This one is *really* dangerous!
2737 ``bitcast (CST to TYPE)``
2738 Convert a constant, CST, to another TYPE. The constraints of the
2739 operands are the same as those for the :ref:`bitcast
2740 instruction <i_bitcast>`.
2741 ``addrspacecast (CST to TYPE)``
2742 Convert a constant pointer or constant vector of pointer, CST, to another
2743 TYPE in a different address space. The constraints of the operands are the
2744 same as those for the :ref:`addrspacecast instruction <i_addrspacecast>`.
2745 ``getelementptr (TY, CSTPTR, IDX0, IDX1, ...)``, ``getelementptr inbounds (TY, CSTPTR, IDX0, IDX1, ...)``
2746 Perform the :ref:`getelementptr operation <i_getelementptr>` on
2747 constants. As with the :ref:`getelementptr <i_getelementptr>`
2748 instruction, the index list may have zero or more indexes, which are
2749 required to make sense for the type of "pointer to TY".
2750 ``select (COND, VAL1, VAL2)``
2751 Perform the :ref:`select operation <i_select>` on constants.
2752 ``icmp COND (VAL1, VAL2)``
2753 Performs the :ref:`icmp operation <i_icmp>` on constants.
2754 ``fcmp COND (VAL1, VAL2)``
2755 Performs the :ref:`fcmp operation <i_fcmp>` on constants.
2756 ``extractelement (VAL, IDX)``
2757 Perform the :ref:`extractelement operation <i_extractelement>` on
2759 ``insertelement (VAL, ELT, IDX)``
2760 Perform the :ref:`insertelement operation <i_insertelement>` on
2762 ``shufflevector (VEC1, VEC2, IDXMASK)``
2763 Perform the :ref:`shufflevector operation <i_shufflevector>` on
2765 ``extractvalue (VAL, IDX0, IDX1, ...)``
2766 Perform the :ref:`extractvalue operation <i_extractvalue>` on
2767 constants. The index list is interpreted in a similar manner as
2768 indices in a ':ref:`getelementptr <i_getelementptr>`' operation. At
2769 least one index value must be specified.
2770 ``insertvalue (VAL, ELT, IDX0, IDX1, ...)``
2771 Perform the :ref:`insertvalue operation <i_insertvalue>` on constants.
2772 The index list is interpreted in a similar manner as indices in a
2773 ':ref:`getelementptr <i_getelementptr>`' operation. At least one index
2774 value must be specified.
2775 ``OPCODE (LHS, RHS)``
2776 Perform the specified operation of the LHS and RHS constants. OPCODE
2777 may be any of the :ref:`binary <binaryops>` or :ref:`bitwise
2778 binary <bitwiseops>` operations. The constraints on operands are
2779 the same as those for the corresponding instruction (e.g. no bitwise
2780 operations on floating point values are allowed).
2787 Inline Assembler Expressions
2788 ----------------------------
2790 LLVM supports inline assembler expressions (as opposed to :ref:`Module-Level
2791 Inline Assembly <moduleasm>`) through the use of a special value. This
2792 value represents the inline assembler as a string (containing the
2793 instructions to emit), a list of operand constraints (stored as a
2794 string), a flag that indicates whether or not the inline asm expression
2795 has side effects, and a flag indicating whether the function containing
2796 the asm needs to align its stack conservatively. An example inline
2797 assembler expression is:
2799 .. code-block:: llvm
2801 i32 (i32) asm "bswap $0", "=r,r"
2803 Inline assembler expressions may **only** be used as the callee operand
2804 of a :ref:`call <i_call>` or an :ref:`invoke <i_invoke>` instruction.
2805 Thus, typically we have:
2807 .. code-block:: llvm
2809 %X = call i32 asm "bswap $0", "=r,r"(i32 %Y)
2811 Inline asms with side effects not visible in the constraint list must be
2812 marked as having side effects. This is done through the use of the
2813 '``sideeffect``' keyword, like so:
2815 .. code-block:: llvm
2817 call void asm sideeffect "eieio", ""()
2819 In some cases inline asms will contain code that will not work unless
2820 the stack is aligned in some way, such as calls or SSE instructions on
2821 x86, yet will not contain code that does that alignment within the asm.
2822 The compiler should make conservative assumptions about what the asm
2823 might contain and should generate its usual stack alignment code in the
2824 prologue if the '``alignstack``' keyword is present:
2826 .. code-block:: llvm
2828 call void asm alignstack "eieio", ""()
2830 Inline asms also support using non-standard assembly dialects. The
2831 assumed dialect is ATT. When the '``inteldialect``' keyword is present,
2832 the inline asm is using the Intel dialect. Currently, ATT and Intel are
2833 the only supported dialects. An example is:
2835 .. code-block:: llvm
2837 call void asm inteldialect "eieio", ""()
2839 If multiple keywords appear the '``sideeffect``' keyword must come
2840 first, the '``alignstack``' keyword second and the '``inteldialect``'
2846 The call instructions that wrap inline asm nodes may have a
2847 "``!srcloc``" MDNode attached to it that contains a list of constant
2848 integers. If present, the code generator will use the integer as the
2849 location cookie value when report errors through the ``LLVMContext``
2850 error reporting mechanisms. This allows a front-end to correlate backend
2851 errors that occur with inline asm back to the source code that produced
2854 .. code-block:: llvm
2856 call void asm sideeffect "something bad", ""(), !srcloc !42
2858 !42 = !{ i32 1234567 }
2860 It is up to the front-end to make sense of the magic numbers it places
2861 in the IR. If the MDNode contains multiple constants, the code generator
2862 will use the one that corresponds to the line of the asm that the error
2870 LLVM IR allows metadata to be attached to instructions in the program
2871 that can convey extra information about the code to the optimizers and
2872 code generator. One example application of metadata is source-level
2873 debug information. There are two metadata primitives: strings and nodes.
2875 Metadata does not have a type, and is not a value. If referenced from a
2876 ``call`` instruction, it uses the ``metadata`` type.
2878 All metadata are identified in syntax by a exclamation point ('``!``').
2880 .. _metadata-string:
2882 Metadata Nodes and Metadata Strings
2883 -----------------------------------
2885 A metadata string is a string surrounded by double quotes. It can
2886 contain any character by escaping non-printable characters with
2887 "``\xx``" where "``xx``" is the two digit hex code. For example:
2890 Metadata nodes are represented with notation similar to structure
2891 constants (a comma separated list of elements, surrounded by braces and
2892 preceded by an exclamation point). Metadata nodes can have any values as
2893 their operand. For example:
2895 .. code-block:: llvm
2897 !{ !"test\00", i32 10}
2899 Metadata nodes that aren't uniqued use the ``distinct`` keyword. For example:
2901 .. code-block:: llvm
2903 !0 = distinct !{!"test\00", i32 10}
2905 ``distinct`` nodes are useful when nodes shouldn't be merged based on their
2906 content. They can also occur when transformations cause uniquing collisions
2907 when metadata operands change.
2909 A :ref:`named metadata <namedmetadatastructure>` is a collection of
2910 metadata nodes, which can be looked up in the module symbol table. For
2913 .. code-block:: llvm
2917 Metadata can be used as function arguments. Here ``llvm.dbg.value``
2918 function is using two metadata arguments:
2920 .. code-block:: llvm
2922 call void @llvm.dbg.value(metadata !24, i64 0, metadata !25)
2924 Metadata can be attached with an instruction. Here metadata ``!21`` is
2925 attached to the ``add`` instruction using the ``!dbg`` identifier:
2927 .. code-block:: llvm
2929 %indvar.next = add i64 %indvar, 1, !dbg !21
2931 More information about specific metadata nodes recognized by the
2932 optimizers and code generator is found below.
2934 .. _specialized-metadata:
2936 Specialized Metadata Nodes
2937 ^^^^^^^^^^^^^^^^^^^^^^^^^^
2939 Specialized metadata nodes are custom data structures in metadata (as opposed
2940 to generic tuples). Their fields are labelled, and can be specified in any
2943 These aren't inherently debug info centric, but currently all the specialized
2944 metadata nodes are related to debug info.
2951 ``DICompileUnit`` nodes represent a compile unit. The ``enums:``,
2952 ``retainedTypes:``, ``subprograms:``, ``globals:`` and ``imports:`` fields are
2953 tuples containing the debug info to be emitted along with the compile unit,
2954 regardless of code optimizations (some nodes are only emitted if there are
2955 references to them from instructions).
2957 .. code-block:: llvm
2959 !0 = !DICompileUnit(language: DW_LANG_C99, file: !1, producer: "clang",
2960 isOptimized: true, flags: "-O2", runtimeVersion: 2,
2961 splitDebugFilename: "abc.debug", emissionKind: 1,
2962 enums: !2, retainedTypes: !3, subprograms: !4,
2963 globals: !5, imports: !6)
2965 Compile unit descriptors provide the root scope for objects declared in a
2966 specific compilation unit. File descriptors are defined using this scope.
2967 These descriptors are collected by a named metadata ``!llvm.dbg.cu``. They
2968 keep track of subprograms, global variables, type information, and imported
2969 entities (declarations and namespaces).
2976 ``DIFile`` nodes represent files. The ``filename:`` can include slashes.
2978 .. code-block:: llvm
2980 !0 = !DIFile(filename: "path/to/file", directory: "/path/to/dir")
2982 Files are sometimes used in ``scope:`` fields, and are the only valid target
2983 for ``file:`` fields.
2990 ``DIBasicType`` nodes represent primitive types, such as ``int``, ``bool`` and
2991 ``float``. ``tag:`` defaults to ``DW_TAG_base_type``.
2993 .. code-block:: llvm
2995 !0 = !DIBasicType(name: "unsigned char", size: 8, align: 8,
2996 encoding: DW_ATE_unsigned_char)
2997 !1 = !DIBasicType(tag: DW_TAG_unspecified_type, name: "decltype(nullptr)")
2999 The ``encoding:`` describes the details of the type. Usually it's one of the
3002 .. code-block:: llvm
3008 DW_ATE_signed_char = 6
3010 DW_ATE_unsigned_char = 8
3012 .. _DISubroutineType:
3017 ``DISubroutineType`` nodes represent subroutine types. Their ``types:`` field
3018 refers to a tuple; the first operand is the return type, while the rest are the
3019 types of the formal arguments in order. If the first operand is ``null``, that
3020 represents a function with no return value (such as ``void foo() {}`` in C++).
3022 .. code-block:: llvm
3024 !0 = !BasicType(name: "int", size: 32, align: 32, DW_ATE_signed)
3025 !1 = !BasicType(name: "char", size: 8, align: 8, DW_ATE_signed_char)
3026 !2 = !DISubroutineType(types: !{null, !0, !1}) ; void (int, char)
3033 ``DIDerivedType`` nodes represent types derived from other types, such as
3036 .. code-block:: llvm
3038 !0 = !DIBasicType(name: "unsigned char", size: 8, align: 8,
3039 encoding: DW_ATE_unsigned_char)
3040 !1 = !DIDerivedType(tag: DW_TAG_pointer_type, baseType: !0, size: 32,
3043 The following ``tag:`` values are valid:
3045 .. code-block:: llvm
3047 DW_TAG_formal_parameter = 5
3049 DW_TAG_pointer_type = 15
3050 DW_TAG_reference_type = 16
3052 DW_TAG_ptr_to_member_type = 31
3053 DW_TAG_const_type = 38
3054 DW_TAG_volatile_type = 53
3055 DW_TAG_restrict_type = 55
3057 ``DW_TAG_member`` is used to define a member of a :ref:`composite type
3058 <DICompositeType>` or :ref:`subprogram <DISubprogram>`. The type of the member
3059 is the ``baseType:``. The ``offset:`` is the member's bit offset.
3060 ``DW_TAG_formal_parameter`` is used to define a member which is a formal
3061 argument of a subprogram.
3063 ``DW_TAG_typedef`` is used to provide a name for the ``baseType:``.
3065 ``DW_TAG_pointer_type``, ``DW_TAG_reference_type``, ``DW_TAG_const_type``,
3066 ``DW_TAG_volatile_type`` and ``DW_TAG_restrict_type`` are used to qualify the
3069 Note that the ``void *`` type is expressed as a type derived from NULL.
3071 .. _DICompositeType:
3076 ``DICompositeType`` nodes represent types composed of other types, like
3077 structures and unions. ``elements:`` points to a tuple of the composed types.
3079 If the source language supports ODR, the ``identifier:`` field gives the unique
3080 identifier used for type merging between modules. When specified, other types
3081 can refer to composite types indirectly via a :ref:`metadata string
3082 <metadata-string>` that matches their identifier.
3084 .. code-block:: llvm
3086 !0 = !DIEnumerator(name: "SixKind", value: 7)
3087 !1 = !DIEnumerator(name: "SevenKind", value: 7)
3088 !2 = !DIEnumerator(name: "NegEightKind", value: -8)
3089 !3 = !DICompositeType(tag: DW_TAG_enumeration_type, name: "Enum", file: !12,
3090 line: 2, size: 32, align: 32, identifier: "_M4Enum",
3091 elements: !{!0, !1, !2})
3093 The following ``tag:`` values are valid:
3095 .. code-block:: llvm
3097 DW_TAG_array_type = 1
3098 DW_TAG_class_type = 2
3099 DW_TAG_enumeration_type = 4
3100 DW_TAG_structure_type = 19
3101 DW_TAG_union_type = 23
3102 DW_TAG_subroutine_type = 21
3103 DW_TAG_inheritance = 28
3106 For ``DW_TAG_array_type``, the ``elements:`` should be :ref:`subrange
3107 descriptors <DISubrange>`, each representing the range of subscripts at that
3108 level of indexing. The ``DIFlagVector`` flag to ``flags:`` indicates that an
3109 array type is a native packed vector.
3111 For ``DW_TAG_enumeration_type``, the ``elements:`` should be :ref:`enumerator
3112 descriptors <DIEnumerator>`, each representing the definition of an enumeration
3113 value for the set. All enumeration type descriptors are collected in the
3114 ``enums:`` field of the :ref:`compile unit <DICompileUnit>`.
3116 For ``DW_TAG_structure_type``, ``DW_TAG_class_type``, and
3117 ``DW_TAG_union_type``, the ``elements:`` should be :ref:`derived types
3118 <DIDerivedType>` with ``tag: DW_TAG_member`` or ``tag: DW_TAG_inheritance``.
3125 ``DISubrange`` nodes are the elements for ``DW_TAG_array_type`` variants of
3126 :ref:`DICompositeType`. ``count: -1`` indicates an empty array.
3128 .. code-block:: llvm
3130 !0 = !DISubrange(count: 5, lowerBound: 0) ; array counting from 0
3131 !1 = !DISubrange(count: 5, lowerBound: 1) ; array counting from 1
3132 !2 = !DISubrange(count: -1) ; empty array.
3139 ``DIEnumerator`` nodes are the elements for ``DW_TAG_enumeration_type``
3140 variants of :ref:`DICompositeType`.
3142 .. code-block:: llvm
3144 !0 = !DIEnumerator(name: "SixKind", value: 7)
3145 !1 = !DIEnumerator(name: "SevenKind", value: 7)
3146 !2 = !DIEnumerator(name: "NegEightKind", value: -8)
3148 DITemplateTypeParameter
3149 """""""""""""""""""""""
3151 ``DITemplateTypeParameter`` nodes represent type parameters to generic source
3152 language constructs. They are used (optionally) in :ref:`DICompositeType` and
3153 :ref:`DISubprogram` ``templateParams:`` fields.
3155 .. code-block:: llvm
3157 !0 = !DITemplateTypeParameter(name: "Ty", type: !1)
3159 DITemplateValueParameter
3160 """"""""""""""""""""""""
3162 ``DITemplateValueParameter`` nodes represent value parameters to generic source
3163 language constructs. ``tag:`` defaults to ``DW_TAG_template_value_parameter``,
3164 but if specified can also be set to ``DW_TAG_GNU_template_template_param`` or
3165 ``DW_TAG_GNU_template_param_pack``. They are used (optionally) in
3166 :ref:`DICompositeType` and :ref:`DISubprogram` ``templateParams:`` fields.
3168 .. code-block:: llvm
3170 !0 = !DITemplateValueParameter(name: "Ty", type: !1, value: i32 7)
3175 ``DINamespace`` nodes represent namespaces in the source language.
3177 .. code-block:: llvm
3179 !0 = !DINamespace(name: "myawesomeproject", scope: !1, file: !2, line: 7)
3184 ``DIGlobalVariable`` nodes represent global variables in the source language.
3186 .. code-block:: llvm
3188 !0 = !DIGlobalVariable(name: "foo", linkageName: "foo", scope: !1,
3189 file: !2, line: 7, type: !3, isLocal: true,
3190 isDefinition: false, variable: i32* @foo,
3193 All global variables should be referenced by the `globals:` field of a
3194 :ref:`compile unit <DICompileUnit>`.
3201 ``DISubprogram`` nodes represent functions from the source language. The
3202 ``variables:`` field points at :ref:`variables <DILocalVariable>` that must be
3203 retained, even if their IR counterparts are optimized out of the IR. The
3204 ``type:`` field must point at an :ref:`DISubroutineType`.
3206 .. code-block:: llvm
3208 !0 = !DISubprogram(name: "foo", linkageName: "_Zfoov", scope: !1,
3209 file: !2, line: 7, type: !3, isLocal: true,
3210 isDefinition: false, scopeLine: 8, containingType: !4,
3211 virtuality: DW_VIRTUALITY_pure_virtual, virtualIndex: 10,
3212 flags: DIFlagPrototyped, isOptimized: true,
3213 function: void ()* @_Z3foov,
3214 templateParams: !5, declaration: !6, variables: !7)
3221 ``DILexicalBlock`` nodes describe nested blocks within a :ref:`subprogram
3222 <DISubprogram>`. The line number and column numbers are used to dinstinguish
3223 two lexical blocks at same depth. They are valid targets for ``scope:``
3226 .. code-block:: llvm
3228 !0 = distinct !DILexicalBlock(scope: !1, file: !2, line: 7, column: 35)
3230 Usually lexical blocks are ``distinct`` to prevent node merging based on
3233 .. _DILexicalBlockFile:
3238 ``DILexicalBlockFile`` nodes are used to discriminate between sections of a
3239 :ref:`lexical block <DILexicalBlock>`. The ``file:`` field can be changed to
3240 indicate textual inclusion, or the ``discriminator:`` field can be used to
3241 discriminate between control flow within a single block in the source language.
3243 .. code-block:: llvm
3245 !0 = !DILexicalBlock(scope: !3, file: !4, line: 7, column: 35)
3246 !1 = !DILexicalBlockFile(scope: !0, file: !4, discriminator: 0)
3247 !2 = !DILexicalBlockFile(scope: !0, file: !4, discriminator: 1)
3254 ``DILocation`` nodes represent source debug locations. The ``scope:`` field is
3255 mandatory, and points at an :ref:`DILexicalBlockFile`, an
3256 :ref:`DILexicalBlock`, or an :ref:`DISubprogram`.
3258 .. code-block:: llvm
3260 !0 = !DILocation(line: 2900, column: 42, scope: !1, inlinedAt: !2)
3262 .. _DILocalVariable:
3267 ``DILocalVariable`` nodes represent local variables in the source language.
3268 Instead of ``DW_TAG_variable``, they use LLVM-specific fake tags to
3269 discriminate between local variables (``DW_TAG_auto_variable``) and subprogram
3270 arguments (``DW_TAG_arg_variable``). In the latter case, the ``arg:`` field
3271 specifies the argument position, and this variable will be included in the
3272 ``variables:`` field of its :ref:`DISubprogram`.
3274 .. code-block:: llvm
3276 !0 = !DILocalVariable(tag: DW_TAG_arg_variable, name: "this", arg: 0,
3277 scope: !3, file: !2, line: 7, type: !3,
3278 flags: DIFlagArtificial)
3279 !1 = !DILocalVariable(tag: DW_TAG_arg_variable, name: "x", arg: 1,
3280 scope: !4, file: !2, line: 7, type: !3)
3281 !1 = !DILocalVariable(tag: DW_TAG_auto_variable, name: "y",
3282 scope: !5, file: !2, line: 7, type: !3)
3287 ``DIExpression`` nodes represent DWARF expression sequences. They are used in
3288 :ref:`debug intrinsics<dbg_intrinsics>` (such as ``llvm.dbg.declare``) to
3289 describe how the referenced LLVM variable relates to the source language
3292 The current supported vocabulary is limited:
3294 - ``DW_OP_deref`` dereferences the working expression.
3295 - ``DW_OP_plus, 93`` adds ``93`` to the working expression.
3296 - ``DW_OP_bit_piece, 16, 8`` specifies the offset and size (``16`` and ``8``
3297 here, respectively) of the variable piece from the working expression.
3299 .. code-block:: llvm
3301 !0 = !DIExpression(DW_OP_deref)
3302 !1 = !DIExpression(DW_OP_plus, 3)
3303 !2 = !DIExpression(DW_OP_bit_piece, 3, 7)
3304 !3 = !DIExpression(DW_OP_deref, DW_OP_plus, 3, DW_OP_bit_piece, 3, 7)
3309 ``DIObjCProperty`` nodes represent Objective-C property nodes.
3311 .. code-block:: llvm
3313 !3 = !DIObjCProperty(name: "foo", file: !1, line: 7, setter: "setFoo",
3314 getter: "getFoo", attributes: 7, type: !2)
3319 ``DIImportedEntity`` nodes represent entities (such as modules) imported into a
3322 .. code-block:: llvm
3324 !2 = !DIImportedEntity(tag: DW_TAG_imported_module, name: "foo", scope: !0,
3325 entity: !1, line: 7)
3330 In LLVM IR, memory does not have types, so LLVM's own type system is not
3331 suitable for doing TBAA. Instead, metadata is added to the IR to
3332 describe a type system of a higher level language. This can be used to
3333 implement typical C/C++ TBAA, but it can also be used to implement
3334 custom alias analysis behavior for other languages.
3336 The current metadata format is very simple. TBAA metadata nodes have up
3337 to three fields, e.g.:
3339 .. code-block:: llvm
3341 !0 = !{ !"an example type tree" }
3342 !1 = !{ !"int", !0 }
3343 !2 = !{ !"float", !0 }
3344 !3 = !{ !"const float", !2, i64 1 }
3346 The first field is an identity field. It can be any value, usually a
3347 metadata string, which uniquely identifies the type. The most important
3348 name in the tree is the name of the root node. Two trees with different
3349 root node names are entirely disjoint, even if they have leaves with
3352 The second field identifies the type's parent node in the tree, or is
3353 null or omitted for a root node. A type is considered to alias all of
3354 its descendants and all of its ancestors in the tree. Also, a type is
3355 considered to alias all types in other trees, so that bitcode produced
3356 from multiple front-ends is handled conservatively.
3358 If the third field is present, it's an integer which if equal to 1
3359 indicates that the type is "constant" (meaning
3360 ``pointsToConstantMemory`` should return true; see `other useful
3361 AliasAnalysis methods <AliasAnalysis.html#OtherItfs>`_).
3363 '``tbaa.struct``' Metadata
3364 ^^^^^^^^^^^^^^^^^^^^^^^^^^
3366 The :ref:`llvm.memcpy <int_memcpy>` is often used to implement
3367 aggregate assignment operations in C and similar languages, however it
3368 is defined to copy a contiguous region of memory, which is more than
3369 strictly necessary for aggregate types which contain holes due to
3370 padding. Also, it doesn't contain any TBAA information about the fields
3373 ``!tbaa.struct`` metadata can describe which memory subregions in a
3374 memcpy are padding and what the TBAA tags of the struct are.
3376 The current metadata format is very simple. ``!tbaa.struct`` metadata
3377 nodes are a list of operands which are in conceptual groups of three.
3378 For each group of three, the first operand gives the byte offset of a
3379 field in bytes, the second gives its size in bytes, and the third gives
3382 .. code-block:: llvm
3384 !4 = !{ i64 0, i64 4, !1, i64 8, i64 4, !2 }
3386 This describes a struct with two fields. The first is at offset 0 bytes
3387 with size 4 bytes, and has tbaa tag !1. The second is at offset 8 bytes
3388 and has size 4 bytes and has tbaa tag !2.
3390 Note that the fields need not be contiguous. In this example, there is a
3391 4 byte gap between the two fields. This gap represents padding which
3392 does not carry useful data and need not be preserved.
3394 '``noalias``' and '``alias.scope``' Metadata
3395 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3397 ``noalias`` and ``alias.scope`` metadata provide the ability to specify generic
3398 noalias memory-access sets. This means that some collection of memory access
3399 instructions (loads, stores, memory-accessing calls, etc.) that carry
3400 ``noalias`` metadata can specifically be specified not to alias with some other
3401 collection of memory access instructions that carry ``alias.scope`` metadata.
3402 Each type of metadata specifies a list of scopes where each scope has an id and
3403 a domain. When evaluating an aliasing query, if for some domain, the set
3404 of scopes with that domain in one instruction's ``alias.scope`` list is a
3405 subset of (or equal to) the set of scopes for that domain in another
3406 instruction's ``noalias`` list, then the two memory accesses are assumed not to
3409 The metadata identifying each domain is itself a list containing one or two
3410 entries. The first entry is the name of the domain. Note that if the name is a
3411 string then it can be combined accross functions and translation units. A
3412 self-reference can be used to create globally unique domain names. A
3413 descriptive string may optionally be provided as a second list entry.
3415 The metadata identifying each scope is also itself a list containing two or
3416 three entries. The first entry is the name of the scope. Note that if the name
3417 is a string then it can be combined accross functions and translation units. A
3418 self-reference can be used to create globally unique scope names. A metadata
3419 reference to the scope's domain is the second entry. A descriptive string may
3420 optionally be provided as a third list entry.
3424 .. code-block:: llvm
3426 ; Two scope domains:
3430 ; Some scopes in these domains:
3436 !5 = !{!4} ; A list containing only scope !4
3440 ; These two instructions don't alias:
3441 %0 = load float, float* %c, align 4, !alias.scope !5
3442 store float %0, float* %arrayidx.i, align 4, !noalias !5
3444 ; These two instructions also don't alias (for domain !1, the set of scopes
3445 ; in the !alias.scope equals that in the !noalias list):
3446 %2 = load float, float* %c, align 4, !alias.scope !5
3447 store float %2, float* %arrayidx.i2, align 4, !noalias !6
3449 ; These two instructions may alias (for domain !0, the set of scopes in
3450 ; the !noalias list is not a superset of, or equal to, the scopes in the
3451 ; !alias.scope list):
3452 %2 = load float, float* %c, align 4, !alias.scope !6
3453 store float %0, float* %arrayidx.i, align 4, !noalias !7
3455 '``fpmath``' Metadata
3456 ^^^^^^^^^^^^^^^^^^^^^
3458 ``fpmath`` metadata may be attached to any instruction of floating point
3459 type. It can be used to express the maximum acceptable error in the
3460 result of that instruction, in ULPs, thus potentially allowing the
3461 compiler to use a more efficient but less accurate method of computing
3462 it. ULP is defined as follows:
3464 If ``x`` is a real number that lies between two finite consecutive
3465 floating-point numbers ``a`` and ``b``, without being equal to one
3466 of them, then ``ulp(x) = |b - a|``, otherwise ``ulp(x)`` is the
3467 distance between the two non-equal finite floating-point numbers
3468 nearest ``x``. Moreover, ``ulp(NaN)`` is ``NaN``.
3470 The metadata node shall consist of a single positive floating point
3471 number representing the maximum relative error, for example:
3473 .. code-block:: llvm
3475 !0 = !{ float 2.5 } ; maximum acceptable inaccuracy is 2.5 ULPs
3479 '``range``' Metadata
3480 ^^^^^^^^^^^^^^^^^^^^
3482 ``range`` metadata may be attached only to ``load``, ``call`` and ``invoke`` of
3483 integer types. It expresses the possible ranges the loaded value or the value
3484 returned by the called function at this call site is in. The ranges are
3485 represented with a flattened list of integers. The loaded value or the value
3486 returned is known to be in the union of the ranges defined by each consecutive
3487 pair. Each pair has the following properties:
3489 - The type must match the type loaded by the instruction.
3490 - The pair ``a,b`` represents the range ``[a,b)``.
3491 - Both ``a`` and ``b`` are constants.
3492 - The range is allowed to wrap.
3493 - The range should not represent the full or empty set. That is,
3496 In addition, the pairs must be in signed order of the lower bound and
3497 they must be non-contiguous.
3501 .. code-block:: llvm
3503 %a = load i8, i8* %x, align 1, !range !0 ; Can only be 0 or 1
3504 %b = load i8, i8* %y, align 1, !range !1 ; Can only be 255 (-1), 0 or 1
3505 %c = call i8 @foo(), !range !2 ; Can only be 0, 1, 3, 4 or 5
3506 %d = invoke i8 @bar() to label %cont
3507 unwind label %lpad, !range !3 ; Can only be -2, -1, 3, 4 or 5
3509 !0 = !{ i8 0, i8 2 }
3510 !1 = !{ i8 255, i8 2 }
3511 !2 = !{ i8 0, i8 2, i8 3, i8 6 }
3512 !3 = !{ i8 -2, i8 0, i8 3, i8 6 }
3517 It is sometimes useful to attach information to loop constructs. Currently,
3518 loop metadata is implemented as metadata attached to the branch instruction
3519 in the loop latch block. This type of metadata refer to a metadata node that is
3520 guaranteed to be separate for each loop. The loop identifier metadata is
3521 specified with the name ``llvm.loop``.
3523 The loop identifier metadata is implemented using a metadata that refers to
3524 itself to avoid merging it with any other identifier metadata, e.g.,
3525 during module linkage or function inlining. That is, each loop should refer
3526 to their own identification metadata even if they reside in separate functions.
3527 The following example contains loop identifier metadata for two separate loop
3530 .. code-block:: llvm
3535 The loop identifier metadata can be used to specify additional
3536 per-loop metadata. Any operands after the first operand can be treated
3537 as user-defined metadata. For example the ``llvm.loop.unroll.count``
3538 suggests an unroll factor to the loop unroller:
3540 .. code-block:: llvm
3542 br i1 %exitcond, label %._crit_edge, label %.lr.ph, !llvm.loop !0
3545 !1 = !{!"llvm.loop.unroll.count", i32 4}
3547 '``llvm.loop.vectorize``' and '``llvm.loop.interleave``'
3548 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3550 Metadata prefixed with ``llvm.loop.vectorize`` or ``llvm.loop.interleave`` are
3551 used to control per-loop vectorization and interleaving parameters such as
3552 vectorization width and interleave count. These metadata should be used in
3553 conjunction with ``llvm.loop`` loop identification metadata. The
3554 ``llvm.loop.vectorize`` and ``llvm.loop.interleave`` metadata are only
3555 optimization hints and the optimizer will only interleave and vectorize loops if
3556 it believes it is safe to do so. The ``llvm.mem.parallel_loop_access`` metadata
3557 which contains information about loop-carried memory dependencies can be helpful
3558 in determining the safety of these transformations.
3560 '``llvm.loop.interleave.count``' Metadata
3561 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3563 This metadata suggests an interleave count to the loop interleaver.
3564 The first operand is the string ``llvm.loop.interleave.count`` and the
3565 second operand is an integer specifying the interleave count. For
3568 .. code-block:: llvm
3570 !0 = !{!"llvm.loop.interleave.count", i32 4}
3572 Note that setting ``llvm.loop.interleave.count`` to 1 disables interleaving
3573 multiple iterations of the loop. If ``llvm.loop.interleave.count`` is set to 0
3574 then the interleave count will be determined automatically.
3576 '``llvm.loop.vectorize.enable``' Metadata
3577 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3579 This metadata selectively enables or disables vectorization for the loop. The
3580 first operand is the string ``llvm.loop.vectorize.enable`` and the second operand
3581 is a bit. If the bit operand value is 1 vectorization is enabled. A value of
3582 0 disables vectorization:
3584 .. code-block:: llvm
3586 !0 = !{!"llvm.loop.vectorize.enable", i1 0}
3587 !1 = !{!"llvm.loop.vectorize.enable", i1 1}
3589 '``llvm.loop.vectorize.width``' Metadata
3590 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3592 This metadata sets the target width of the vectorizer. The first
3593 operand is the string ``llvm.loop.vectorize.width`` and the second
3594 operand is an integer specifying the width. For example:
3596 .. code-block:: llvm
3598 !0 = !{!"llvm.loop.vectorize.width", i32 4}
3600 Note that setting ``llvm.loop.vectorize.width`` to 1 disables
3601 vectorization of the loop. If ``llvm.loop.vectorize.width`` is set to
3602 0 or if the loop does not have this metadata the width will be
3603 determined automatically.
3605 '``llvm.loop.unroll``'
3606 ^^^^^^^^^^^^^^^^^^^^^^
3608 Metadata prefixed with ``llvm.loop.unroll`` are loop unrolling
3609 optimization hints such as the unroll factor. ``llvm.loop.unroll``
3610 metadata should be used in conjunction with ``llvm.loop`` loop
3611 identification metadata. The ``llvm.loop.unroll`` metadata are only
3612 optimization hints and the unrolling will only be performed if the
3613 optimizer believes it is safe to do so.
3615 '``llvm.loop.unroll.count``' Metadata
3616 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3618 This metadata suggests an unroll factor to the loop unroller. The
3619 first operand is the string ``llvm.loop.unroll.count`` and the second
3620 operand is a positive integer specifying the unroll factor. For
3623 .. code-block:: llvm
3625 !0 = !{!"llvm.loop.unroll.count", i32 4}
3627 If the trip count of the loop is less than the unroll count the loop
3628 will be partially unrolled.
3630 '``llvm.loop.unroll.disable``' Metadata
3631 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3633 This metadata either disables loop unrolling. The metadata has a single operand
3634 which is the string ``llvm.loop.unroll.disable``. For example:
3636 .. code-block:: llvm
3638 !0 = !{!"llvm.loop.unroll.disable"}
3640 '``llvm.loop.unroll.runtime.disable``' Metadata
3641 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3643 This metadata either disables runtime loop unrolling. The metadata has a single
3644 operand which is the string ``llvm.loop.unroll.runtime.disable``. For example:
3646 .. code-block:: llvm
3648 !0 = !{!"llvm.loop.unroll.runtime.disable"}
3650 '``llvm.loop.unroll.full``' Metadata
3651 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3653 This metadata either suggests that the loop should be unrolled fully. The
3654 metadata has a single operand which is the string ``llvm.loop.unroll.disable``.
3657 .. code-block:: llvm
3659 !0 = !{!"llvm.loop.unroll.full"}
3664 Metadata types used to annotate memory accesses with information helpful
3665 for optimizations are prefixed with ``llvm.mem``.
3667 '``llvm.mem.parallel_loop_access``' Metadata
3668 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3670 The ``llvm.mem.parallel_loop_access`` metadata refers to a loop identifier,
3671 or metadata containing a list of loop identifiers for nested loops.
3672 The metadata is attached to memory accessing instructions and denotes that
3673 no loop carried memory dependence exist between it and other instructions denoted
3674 with the same loop identifier.
3676 Precisely, given two instructions ``m1`` and ``m2`` that both have the
3677 ``llvm.mem.parallel_loop_access`` metadata, with ``L1`` and ``L2`` being the
3678 set of loops associated with that metadata, respectively, then there is no loop
3679 carried dependence between ``m1`` and ``m2`` for loops in both ``L1`` and
3682 As a special case, if all memory accessing instructions in a loop have
3683 ``llvm.mem.parallel_loop_access`` metadata that refers to that loop, then the
3684 loop has no loop carried memory dependences and is considered to be a parallel
3687 Note that if not all memory access instructions have such metadata referring to
3688 the loop, then the loop is considered not being trivially parallel. Additional
3689 memory dependence analysis is required to make that determination. As a fail
3690 safe mechanism, this causes loops that were originally parallel to be considered
3691 sequential (if optimization passes that are unaware of the parallel semantics
3692 insert new memory instructions into the loop body).
3694 Example of a loop that is considered parallel due to its correct use of
3695 both ``llvm.loop`` and ``llvm.mem.parallel_loop_access``
3696 metadata types that refer to the same loop identifier metadata.
3698 .. code-block:: llvm
3702 %val0 = load i32, i32* %arrayidx, !llvm.mem.parallel_loop_access !0
3704 store i32 %val0, i32* %arrayidx1, !llvm.mem.parallel_loop_access !0
3706 br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !0
3712 It is also possible to have nested parallel loops. In that case the
3713 memory accesses refer to a list of loop identifier metadata nodes instead of
3714 the loop identifier metadata node directly:
3716 .. code-block:: llvm
3720 %val1 = load i32, i32* %arrayidx3, !llvm.mem.parallel_loop_access !2
3722 br label %inner.for.body
3726 %val0 = load i32, i32* %arrayidx1, !llvm.mem.parallel_loop_access !0
3728 store i32 %val0, i32* %arrayidx2, !llvm.mem.parallel_loop_access !0
3730 br i1 %exitcond, label %inner.for.end, label %inner.for.body, !llvm.loop !1
3734 store i32 %val1, i32* %arrayidx4, !llvm.mem.parallel_loop_access !2
3736 br i1 %exitcond, label %outer.for.end, label %outer.for.body, !llvm.loop !2
3738 outer.for.end: ; preds = %for.body
3740 !0 = !{!1, !2} ; a list of loop identifiers
3741 !1 = !{!1} ; an identifier for the inner loop
3742 !2 = !{!2} ; an identifier for the outer loop
3747 The ``llvm.bitsets`` global metadata is used to implement
3748 :doc:`bitsets <BitSets>`.
3750 Module Flags Metadata
3751 =====================
3753 Information about the module as a whole is difficult to convey to LLVM's
3754 subsystems. The LLVM IR isn't sufficient to transmit this information.
3755 The ``llvm.module.flags`` named metadata exists in order to facilitate
3756 this. These flags are in the form of key / value pairs --- much like a
3757 dictionary --- making it easy for any subsystem who cares about a flag to
3760 The ``llvm.module.flags`` metadata contains a list of metadata triplets.
3761 Each triplet has the following form:
3763 - The first element is a *behavior* flag, which specifies the behavior
3764 when two (or more) modules are merged together, and it encounters two
3765 (or more) metadata with the same ID. The supported behaviors are
3767 - The second element is a metadata string that is a unique ID for the
3768 metadata. Each module may only have one flag entry for each unique ID (not
3769 including entries with the **Require** behavior).
3770 - The third element is the value of the flag.
3772 When two (or more) modules are merged together, the resulting
3773 ``llvm.module.flags`` metadata is the union of the modules' flags. That is, for
3774 each unique metadata ID string, there will be exactly one entry in the merged
3775 modules ``llvm.module.flags`` metadata table, and the value for that entry will
3776 be determined by the merge behavior flag, as described below. The only exception
3777 is that entries with the *Require* behavior are always preserved.
3779 The following behaviors are supported:
3790 Emits an error if two values disagree, otherwise the resulting value
3791 is that of the operands.
3795 Emits a warning if two values disagree. The result value will be the
3796 operand for the flag from the first module being linked.
3800 Adds a requirement that another module flag be present and have a
3801 specified value after linking is performed. The value must be a
3802 metadata pair, where the first element of the pair is the ID of the
3803 module flag to be restricted, and the second element of the pair is
3804 the value the module flag should be restricted to. This behavior can
3805 be used to restrict the allowable results (via triggering of an
3806 error) of linking IDs with the **Override** behavior.
3810 Uses the specified value, regardless of the behavior or value of the
3811 other module. If both modules specify **Override**, but the values
3812 differ, an error will be emitted.
3816 Appends the two values, which are required to be metadata nodes.
3820 Appends the two values, which are required to be metadata
3821 nodes. However, duplicate entries in the second list are dropped
3822 during the append operation.
3824 It is an error for a particular unique flag ID to have multiple behaviors,
3825 except in the case of **Require** (which adds restrictions on another metadata
3826 value) or **Override**.
3828 An example of module flags:
3830 .. code-block:: llvm
3832 !0 = !{ i32 1, !"foo", i32 1 }
3833 !1 = !{ i32 4, !"bar", i32 37 }
3834 !2 = !{ i32 2, !"qux", i32 42 }
3835 !3 = !{ i32 3, !"qux",
3840 !llvm.module.flags = !{ !0, !1, !2, !3 }
3842 - Metadata ``!0`` has the ID ``!"foo"`` and the value '1'. The behavior
3843 if two or more ``!"foo"`` flags are seen is to emit an error if their
3844 values are not equal.
3846 - Metadata ``!1`` has the ID ``!"bar"`` and the value '37'. The
3847 behavior if two or more ``!"bar"`` flags are seen is to use the value
3850 - Metadata ``!2`` has the ID ``!"qux"`` and the value '42'. The
3851 behavior if two or more ``!"qux"`` flags are seen is to emit a
3852 warning if their values are not equal.
3854 - Metadata ``!3`` has the ID ``!"qux"`` and the value:
3860 The behavior is to emit an error if the ``llvm.module.flags`` does not
3861 contain a flag with the ID ``!"foo"`` that has the value '1' after linking is
3864 Objective-C Garbage Collection Module Flags Metadata
3865 ----------------------------------------------------
3867 On the Mach-O platform, Objective-C stores metadata about garbage
3868 collection in a special section called "image info". The metadata
3869 consists of a version number and a bitmask specifying what types of
3870 garbage collection are supported (if any) by the file. If two or more
3871 modules are linked together their garbage collection metadata needs to
3872 be merged rather than appended together.
3874 The Objective-C garbage collection module flags metadata consists of the
3875 following key-value pairs:
3884 * - ``Objective-C Version``
3885 - **[Required]** --- The Objective-C ABI version. Valid values are 1 and 2.
3887 * - ``Objective-C Image Info Version``
3888 - **[Required]** --- The version of the image info section. Currently
3891 * - ``Objective-C Image Info Section``
3892 - **[Required]** --- The section to place the metadata. Valid values are
3893 ``"__OBJC, __image_info, regular"`` for Objective-C ABI version 1, and
3894 ``"__DATA,__objc_imageinfo, regular, no_dead_strip"`` for
3895 Objective-C ABI version 2.
3897 * - ``Objective-C Garbage Collection``
3898 - **[Required]** --- Specifies whether garbage collection is supported or
3899 not. Valid values are 0, for no garbage collection, and 2, for garbage
3900 collection supported.
3902 * - ``Objective-C GC Only``
3903 - **[Optional]** --- Specifies that only garbage collection is supported.
3904 If present, its value must be 6. This flag requires that the
3905 ``Objective-C Garbage Collection`` flag have the value 2.
3907 Some important flag interactions:
3909 - If a module with ``Objective-C Garbage Collection`` set to 0 is
3910 merged with a module with ``Objective-C Garbage Collection`` set to
3911 2, then the resulting module has the
3912 ``Objective-C Garbage Collection`` flag set to 0.
3913 - A module with ``Objective-C Garbage Collection`` set to 0 cannot be
3914 merged with a module with ``Objective-C GC Only`` set to 6.
3916 Automatic Linker Flags Module Flags Metadata
3917 --------------------------------------------
3919 Some targets support embedding flags to the linker inside individual object
3920 files. Typically this is used in conjunction with language extensions which
3921 allow source files to explicitly declare the libraries they depend on, and have
3922 these automatically be transmitted to the linker via object files.
3924 These flags are encoded in the IR using metadata in the module flags section,
3925 using the ``Linker Options`` key. The merge behavior for this flag is required
3926 to be ``AppendUnique``, and the value for the key is expected to be a metadata
3927 node which should be a list of other metadata nodes, each of which should be a
3928 list of metadata strings defining linker options.
3930 For example, the following metadata section specifies two separate sets of
3931 linker options, presumably to link against ``libz`` and the ``Cocoa``
3934 !0 = !{ i32 6, !"Linker Options",
3937 !{ !"-framework", !"Cocoa" } } }
3938 !llvm.module.flags = !{ !0 }
3940 The metadata encoding as lists of lists of options, as opposed to a collapsed
3941 list of options, is chosen so that the IR encoding can use multiple option
3942 strings to specify e.g., a single library, while still having that specifier be
3943 preserved as an atomic element that can be recognized by a target specific
3944 assembly writer or object file emitter.
3946 Each individual option is required to be either a valid option for the target's
3947 linker, or an option that is reserved by the target specific assembly writer or
3948 object file emitter. No other aspect of these options is defined by the IR.
3950 C type width Module Flags Metadata
3951 ----------------------------------
3953 The ARM backend emits a section into each generated object file describing the
3954 options that it was compiled with (in a compiler-independent way) to prevent
3955 linking incompatible objects, and to allow automatic library selection. Some
3956 of these options are not visible at the IR level, namely wchar_t width and enum
3959 To pass this information to the backend, these options are encoded in module
3960 flags metadata, using the following key-value pairs:
3970 - * 0 --- sizeof(wchar_t) == 4
3971 * 1 --- sizeof(wchar_t) == 2
3974 - * 0 --- Enums are at least as large as an ``int``.
3975 * 1 --- Enums are stored in the smallest integer type which can
3976 represent all of its values.
3978 For example, the following metadata section specifies that the module was
3979 compiled with a ``wchar_t`` width of 4 bytes, and the underlying type of an
3980 enum is the smallest type which can represent all of its values::
3982 !llvm.module.flags = !{!0, !1}
3983 !0 = !{i32 1, !"short_wchar", i32 1}
3984 !1 = !{i32 1, !"short_enum", i32 0}
3986 .. _intrinsicglobalvariables:
3988 Intrinsic Global Variables
3989 ==========================
3991 LLVM has a number of "magic" global variables that contain data that
3992 affect code generation or other IR semantics. These are documented here.
3993 All globals of this sort should have a section specified as
3994 "``llvm.metadata``". This section and all globals that start with
3995 "``llvm.``" are reserved for use by LLVM.
3999 The '``llvm.used``' Global Variable
4000 -----------------------------------
4002 The ``@llvm.used`` global is an array which has
4003 :ref:`appending linkage <linkage_appending>`. This array contains a list of
4004 pointers to named global variables, functions and aliases which may optionally
4005 have a pointer cast formed of bitcast or getelementptr. For example, a legal
4008 .. code-block:: llvm
4013 @llvm.used = appending global [2 x i8*] [
4015 i8* bitcast (i32* @Y to i8*)
4016 ], section "llvm.metadata"
4018 If a symbol appears in the ``@llvm.used`` list, then the compiler, assembler,
4019 and linker are required to treat the symbol as if there is a reference to the
4020 symbol that it cannot see (which is why they have to be named). For example, if
4021 a variable has internal linkage and no references other than that from the
4022 ``@llvm.used`` list, it cannot be deleted. This is commonly used to represent
4023 references from inline asms and other things the compiler cannot "see", and
4024 corresponds to "``attribute((used))``" in GNU C.
4026 On some targets, the code generator must emit a directive to the
4027 assembler or object file to prevent the assembler and linker from
4028 molesting the symbol.
4030 .. _gv_llvmcompilerused:
4032 The '``llvm.compiler.used``' Global Variable
4033 --------------------------------------------
4035 The ``@llvm.compiler.used`` directive is the same as the ``@llvm.used``
4036 directive, except that it only prevents the compiler from touching the
4037 symbol. On targets that support it, this allows an intelligent linker to
4038 optimize references to the symbol without being impeded as it would be
4041 This is a rare construct that should only be used in rare circumstances,
4042 and should not be exposed to source languages.
4044 .. _gv_llvmglobalctors:
4046 The '``llvm.global_ctors``' Global Variable
4047 -------------------------------------------
4049 .. code-block:: llvm
4051 %0 = type { i32, void ()*, i8* }
4052 @llvm.global_ctors = appending global [1 x %0] [%0 { i32 65535, void ()* @ctor, i8* @data }]
4054 The ``@llvm.global_ctors`` array contains a list of constructor
4055 functions, priorities, and an optional associated global or function.
4056 The functions referenced by this array will be called in ascending order
4057 of priority (i.e. lowest first) when the module is loaded. The order of
4058 functions with the same priority is not defined.
4060 If the third field is present, non-null, and points to a global variable
4061 or function, the initializer function will only run if the associated
4062 data from the current module is not discarded.
4064 .. _llvmglobaldtors:
4066 The '``llvm.global_dtors``' Global Variable
4067 -------------------------------------------
4069 .. code-block:: llvm
4071 %0 = type { i32, void ()*, i8* }
4072 @llvm.global_dtors = appending global [1 x %0] [%0 { i32 65535, void ()* @dtor, i8* @data }]
4074 The ``@llvm.global_dtors`` array contains a list of destructor
4075 functions, priorities, and an optional associated global or function.
4076 The functions referenced by this array will be called in descending
4077 order of priority (i.e. highest first) when the module is unloaded. The
4078 order of functions with the same priority is not defined.
4080 If the third field is present, non-null, and points to a global variable
4081 or function, the destructor function will only run if the associated
4082 data from the current module is not discarded.
4084 Instruction Reference
4085 =====================
4087 The LLVM instruction set consists of several different classifications
4088 of instructions: :ref:`terminator instructions <terminators>`, :ref:`binary
4089 instructions <binaryops>`, :ref:`bitwise binary
4090 instructions <bitwiseops>`, :ref:`memory instructions <memoryops>`, and
4091 :ref:`other instructions <otherops>`.
4095 Terminator Instructions
4096 -----------------------
4098 As mentioned :ref:`previously <functionstructure>`, every basic block in a
4099 program ends with a "Terminator" instruction, which indicates which
4100 block should be executed after the current block is finished. These
4101 terminator instructions typically yield a '``void``' value: they produce
4102 control flow, not values (the one exception being the
4103 ':ref:`invoke <i_invoke>`' instruction).
4105 The terminator instructions are: ':ref:`ret <i_ret>`',
4106 ':ref:`br <i_br>`', ':ref:`switch <i_switch>`',
4107 ':ref:`indirectbr <i_indirectbr>`', ':ref:`invoke <i_invoke>`',
4108 ':ref:`resume <i_resume>`', and ':ref:`unreachable <i_unreachable>`'.
4112 '``ret``' Instruction
4113 ^^^^^^^^^^^^^^^^^^^^^
4120 ret <type> <value> ; Return a value from a non-void function
4121 ret void ; Return from void function
4126 The '``ret``' instruction is used to return control flow (and optionally
4127 a value) from a function back to the caller.
4129 There are two forms of the '``ret``' instruction: one that returns a
4130 value and then causes control flow, and one that just causes control
4136 The '``ret``' instruction optionally accepts a single argument, the
4137 return value. The type of the return value must be a ':ref:`first
4138 class <t_firstclass>`' type.
4140 A function is not :ref:`well formed <wellformed>` if it it has a non-void
4141 return type and contains a '``ret``' instruction with no return value or
4142 a return value with a type that does not match its type, or if it has a
4143 void return type and contains a '``ret``' instruction with a return
4149 When the '``ret``' instruction is executed, control flow returns back to
4150 the calling function's context. If the caller is a
4151 ":ref:`call <i_call>`" instruction, execution continues at the
4152 instruction after the call. If the caller was an
4153 ":ref:`invoke <i_invoke>`" instruction, execution continues at the
4154 beginning of the "normal" destination block. If the instruction returns
4155 a value, that value shall set the call or invoke instruction's return
4161 .. code-block:: llvm
4163 ret i32 5 ; Return an integer value of 5
4164 ret void ; Return from a void function
4165 ret { i32, i8 } { i32 4, i8 2 } ; Return a struct of values 4 and 2
4169 '``br``' Instruction
4170 ^^^^^^^^^^^^^^^^^^^^
4177 br i1 <cond>, label <iftrue>, label <iffalse>
4178 br label <dest> ; Unconditional branch
4183 The '``br``' instruction is used to cause control flow to transfer to a
4184 different basic block in the current function. There are two forms of
4185 this instruction, corresponding to a conditional branch and an
4186 unconditional branch.
4191 The conditional branch form of the '``br``' instruction takes a single
4192 '``i1``' value and two '``label``' values. The unconditional form of the
4193 '``br``' instruction takes a single '``label``' value as a target.
4198 Upon execution of a conditional '``br``' instruction, the '``i1``'
4199 argument is evaluated. If the value is ``true``, control flows to the
4200 '``iftrue``' ``label`` argument. If "cond" is ``false``, control flows
4201 to the '``iffalse``' ``label`` argument.
4206 .. code-block:: llvm
4209 %cond = icmp eq i32 %a, %b
4210 br i1 %cond, label %IfEqual, label %IfUnequal
4218 '``switch``' Instruction
4219 ^^^^^^^^^^^^^^^^^^^^^^^^
4226 switch <intty> <value>, label <defaultdest> [ <intty> <val>, label <dest> ... ]
4231 The '``switch``' instruction is used to transfer control flow to one of
4232 several different places. It is a generalization of the '``br``'
4233 instruction, allowing a branch to occur to one of many possible
4239 The '``switch``' instruction uses three parameters: an integer
4240 comparison value '``value``', a default '``label``' destination, and an
4241 array of pairs of comparison value constants and '``label``'s. The table
4242 is not allowed to contain duplicate constant entries.
4247 The ``switch`` instruction specifies a table of values and destinations.
4248 When the '``switch``' instruction is executed, this table is searched
4249 for the given value. If the value is found, control flow is transferred
4250 to the corresponding destination; otherwise, control flow is transferred
4251 to the default destination.
4256 Depending on properties of the target machine and the particular
4257 ``switch`` instruction, this instruction may be code generated in
4258 different ways. For example, it could be generated as a series of
4259 chained conditional branches or with a lookup table.
4264 .. code-block:: llvm
4266 ; Emulate a conditional br instruction
4267 %Val = zext i1 %value to i32
4268 switch i32 %Val, label %truedest [ i32 0, label %falsedest ]
4270 ; Emulate an unconditional br instruction
4271 switch i32 0, label %dest [ ]
4273 ; Implement a jump table:
4274 switch i32 %val, label %otherwise [ i32 0, label %onzero
4276 i32 2, label %ontwo ]
4280 '``indirectbr``' Instruction
4281 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4288 indirectbr <somety>* <address>, [ label <dest1>, label <dest2>, ... ]
4293 The '``indirectbr``' instruction implements an indirect branch to a
4294 label within the current function, whose address is specified by
4295 "``address``". Address must be derived from a
4296 :ref:`blockaddress <blockaddress>` constant.
4301 The '``address``' argument is the address of the label to jump to. The
4302 rest of the arguments indicate the full set of possible destinations
4303 that the address may point to. Blocks are allowed to occur multiple
4304 times in the destination list, though this isn't particularly useful.
4306 This destination list is required so that dataflow analysis has an
4307 accurate understanding of the CFG.
4312 Control transfers to the block specified in the address argument. All
4313 possible destination blocks must be listed in the label list, otherwise
4314 this instruction has undefined behavior. This implies that jumps to
4315 labels defined in other functions have undefined behavior as well.
4320 This is typically implemented with a jump through a register.
4325 .. code-block:: llvm
4327 indirectbr i8* %Addr, [ label %bb1, label %bb2, label %bb3 ]
4331 '``invoke``' Instruction
4332 ^^^^^^^^^^^^^^^^^^^^^^^^
4339 <result> = invoke [cconv] [ret attrs] <ptr to function ty> <function ptr val>(<function args>) [fn attrs]
4340 to label <normal label> unwind label <exception label>
4345 The '``invoke``' instruction causes control to transfer to a specified
4346 function, with the possibility of control flow transfer to either the
4347 '``normal``' label or the '``exception``' label. If the callee function
4348 returns with the "``ret``" instruction, control flow will return to the
4349 "normal" label. If the callee (or any indirect callees) returns via the
4350 ":ref:`resume <i_resume>`" instruction or other exception handling
4351 mechanism, control is interrupted and continued at the dynamically
4352 nearest "exception" label.
4354 The '``exception``' label is a `landing
4355 pad <ExceptionHandling.html#overview>`_ for the exception. As such,
4356 '``exception``' label is required to have the
4357 ":ref:`landingpad <i_landingpad>`" instruction, which contains the
4358 information about the behavior of the program after unwinding happens,
4359 as its first non-PHI instruction. The restrictions on the
4360 "``landingpad``" instruction's tightly couples it to the "``invoke``"
4361 instruction, so that the important information contained within the
4362 "``landingpad``" instruction can't be lost through normal code motion.
4367 This instruction requires several arguments:
4369 #. The optional "cconv" marker indicates which :ref:`calling
4370 convention <callingconv>` the call should use. If none is
4371 specified, the call defaults to using C calling conventions.
4372 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
4373 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
4375 #. '``ptr to function ty``': shall be the signature of the pointer to
4376 function value being invoked. In most cases, this is a direct
4377 function invocation, but indirect ``invoke``'s are just as possible,
4378 branching off an arbitrary pointer to function value.
4379 #. '``function ptr val``': An LLVM value containing a pointer to a
4380 function to be invoked.
4381 #. '``function args``': argument list whose types match the function
4382 signature argument types and parameter attributes. All arguments must
4383 be of :ref:`first class <t_firstclass>` type. If the function signature
4384 indicates the function accepts a variable number of arguments, the
4385 extra arguments can be specified.
4386 #. '``normal label``': the label reached when the called function
4387 executes a '``ret``' instruction.
4388 #. '``exception label``': the label reached when a callee returns via
4389 the :ref:`resume <i_resume>` instruction or other exception handling
4391 #. The optional :ref:`function attributes <fnattrs>` list. Only
4392 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
4393 attributes are valid here.
4398 This instruction is designed to operate as a standard '``call``'
4399 instruction in most regards. The primary difference is that it
4400 establishes an association with a label, which is used by the runtime
4401 library to unwind the stack.
4403 This instruction is used in languages with destructors to ensure that
4404 proper cleanup is performed in the case of either a ``longjmp`` or a
4405 thrown exception. Additionally, this is important for implementation of
4406 '``catch``' clauses in high-level languages that support them.
4408 For the purposes of the SSA form, the definition of the value returned
4409 by the '``invoke``' instruction is deemed to occur on the edge from the
4410 current block to the "normal" label. If the callee unwinds then no
4411 return value is available.
4416 .. code-block:: llvm
4418 %retval = invoke i32 @Test(i32 15) to label %Continue
4419 unwind label %TestCleanup ; i32:retval set
4420 %retval = invoke coldcc i32 %Testfnptr(i32 15) to label %Continue
4421 unwind label %TestCleanup ; i32:retval set
4425 '``resume``' Instruction
4426 ^^^^^^^^^^^^^^^^^^^^^^^^
4433 resume <type> <value>
4438 The '``resume``' instruction is a terminator instruction that has no
4444 The '``resume``' instruction requires one argument, which must have the
4445 same type as the result of any '``landingpad``' instruction in the same
4451 The '``resume``' instruction resumes propagation of an existing
4452 (in-flight) exception whose unwinding was interrupted with a
4453 :ref:`landingpad <i_landingpad>` instruction.
4458 .. code-block:: llvm
4460 resume { i8*, i32 } %exn
4464 '``unreachable``' Instruction
4465 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4477 The '``unreachable``' instruction has no defined semantics. This
4478 instruction is used to inform the optimizer that a particular portion of
4479 the code is not reachable. This can be used to indicate that the code
4480 after a no-return function cannot be reached, and other facts.
4485 The '``unreachable``' instruction has no defined semantics.
4492 Binary operators are used to do most of the computation in a program.
4493 They require two operands of the same type, execute an operation on
4494 them, and produce a single value. The operands might represent multiple
4495 data, as is the case with the :ref:`vector <t_vector>` data type. The
4496 result value has the same type as its operands.
4498 There are several different binary operators:
4502 '``add``' Instruction
4503 ^^^^^^^^^^^^^^^^^^^^^
4510 <result> = add <ty> <op1>, <op2> ; yields ty:result
4511 <result> = add nuw <ty> <op1>, <op2> ; yields ty:result
4512 <result> = add nsw <ty> <op1>, <op2> ; yields ty:result
4513 <result> = add nuw nsw <ty> <op1>, <op2> ; yields ty:result
4518 The '``add``' instruction returns the sum of its two operands.
4523 The two arguments to the '``add``' instruction must be
4524 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4525 arguments must have identical types.
4530 The value produced is the integer sum of the two operands.
4532 If the sum has unsigned overflow, the result returned is the
4533 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
4536 Because LLVM integers use a two's complement representation, this
4537 instruction is appropriate for both signed and unsigned integers.
4539 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
4540 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
4541 result value of the ``add`` is a :ref:`poison value <poisonvalues>` if
4542 unsigned and/or signed overflow, respectively, occurs.
4547 .. code-block:: llvm
4549 <result> = add i32 4, %var ; yields i32:result = 4 + %var
4553 '``fadd``' Instruction
4554 ^^^^^^^^^^^^^^^^^^^^^^
4561 <result> = fadd [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
4566 The '``fadd``' instruction returns the sum of its two operands.
4571 The two arguments to the '``fadd``' instruction must be :ref:`floating
4572 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
4573 Both arguments must have identical types.
4578 The value produced is the floating point sum of the two operands. This
4579 instruction can also take any number of :ref:`fast-math flags <fastmath>`,
4580 which are optimization hints to enable otherwise unsafe floating point
4586 .. code-block:: llvm
4588 <result> = fadd float 4.0, %var ; yields float:result = 4.0 + %var
4590 '``sub``' Instruction
4591 ^^^^^^^^^^^^^^^^^^^^^
4598 <result> = sub <ty> <op1>, <op2> ; yields ty:result
4599 <result> = sub nuw <ty> <op1>, <op2> ; yields ty:result
4600 <result> = sub nsw <ty> <op1>, <op2> ; yields ty:result
4601 <result> = sub nuw nsw <ty> <op1>, <op2> ; yields ty:result
4606 The '``sub``' instruction returns the difference of its two operands.
4608 Note that the '``sub``' instruction is used to represent the '``neg``'
4609 instruction present in most other intermediate representations.
4614 The two arguments to the '``sub``' instruction must be
4615 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4616 arguments must have identical types.
4621 The value produced is the integer difference of the two operands.
4623 If the difference has unsigned overflow, the result returned is the
4624 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
4627 Because LLVM integers use a two's complement representation, this
4628 instruction is appropriate for both signed and unsigned integers.
4630 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
4631 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
4632 result value of the ``sub`` is a :ref:`poison value <poisonvalues>` if
4633 unsigned and/or signed overflow, respectively, occurs.
4638 .. code-block:: llvm
4640 <result> = sub i32 4, %var ; yields i32:result = 4 - %var
4641 <result> = sub i32 0, %val ; yields i32:result = -%var
4645 '``fsub``' Instruction
4646 ^^^^^^^^^^^^^^^^^^^^^^
4653 <result> = fsub [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
4658 The '``fsub``' instruction returns the difference of its two operands.
4660 Note that the '``fsub``' instruction is used to represent the '``fneg``'
4661 instruction present in most other intermediate representations.
4666 The two arguments to the '``fsub``' instruction must be :ref:`floating
4667 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
4668 Both arguments must have identical types.
4673 The value produced is the floating point difference of the two operands.
4674 This instruction can also take any number of :ref:`fast-math
4675 flags <fastmath>`, which are optimization hints to enable otherwise
4676 unsafe floating point optimizations:
4681 .. code-block:: llvm
4683 <result> = fsub float 4.0, %var ; yields float:result = 4.0 - %var
4684 <result> = fsub float -0.0, %val ; yields float:result = -%var
4686 '``mul``' Instruction
4687 ^^^^^^^^^^^^^^^^^^^^^
4694 <result> = mul <ty> <op1>, <op2> ; yields ty:result
4695 <result> = mul nuw <ty> <op1>, <op2> ; yields ty:result
4696 <result> = mul nsw <ty> <op1>, <op2> ; yields ty:result
4697 <result> = mul nuw nsw <ty> <op1>, <op2> ; yields ty:result
4702 The '``mul``' instruction returns the product of its two operands.
4707 The two arguments to the '``mul``' instruction must be
4708 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4709 arguments must have identical types.
4714 The value produced is the integer product of the two operands.
4716 If the result of the multiplication has unsigned overflow, the result
4717 returned is the mathematical result modulo 2\ :sup:`n`\ , where n is the
4718 bit width of the result.
4720 Because LLVM integers use a two's complement representation, and the
4721 result is the same width as the operands, this instruction returns the
4722 correct result for both signed and unsigned integers. If a full product
4723 (e.g. ``i32`` * ``i32`` -> ``i64``) is needed, the operands should be
4724 sign-extended or zero-extended as appropriate to the width of the full
4727 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
4728 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
4729 result value of the ``mul`` is a :ref:`poison value <poisonvalues>` if
4730 unsigned and/or signed overflow, respectively, occurs.
4735 .. code-block:: llvm
4737 <result> = mul i32 4, %var ; yields i32:result = 4 * %var
4741 '``fmul``' Instruction
4742 ^^^^^^^^^^^^^^^^^^^^^^
4749 <result> = fmul [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
4754 The '``fmul``' instruction returns the product of its two operands.
4759 The two arguments to the '``fmul``' instruction must be :ref:`floating
4760 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
4761 Both arguments must have identical types.
4766 The value produced is the floating point product of the two operands.
4767 This instruction can also take any number of :ref:`fast-math
4768 flags <fastmath>`, which are optimization hints to enable otherwise
4769 unsafe floating point optimizations:
4774 .. code-block:: llvm
4776 <result> = fmul float 4.0, %var ; yields float:result = 4.0 * %var
4778 '``udiv``' Instruction
4779 ^^^^^^^^^^^^^^^^^^^^^^
4786 <result> = udiv <ty> <op1>, <op2> ; yields ty:result
4787 <result> = udiv exact <ty> <op1>, <op2> ; yields ty:result
4792 The '``udiv``' instruction returns the quotient of its two operands.
4797 The two arguments to the '``udiv``' instruction must be
4798 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4799 arguments must have identical types.
4804 The value produced is the unsigned integer quotient of the two operands.
4806 Note that unsigned integer division and signed integer division are
4807 distinct operations; for signed integer division, use '``sdiv``'.
4809 Division by zero leads to undefined behavior.
4811 If the ``exact`` keyword is present, the result value of the ``udiv`` is
4812 a :ref:`poison value <poisonvalues>` if %op1 is not a multiple of %op2 (as
4813 such, "((a udiv exact b) mul b) == a").
4818 .. code-block:: llvm
4820 <result> = udiv i32 4, %var ; yields i32:result = 4 / %var
4822 '``sdiv``' Instruction
4823 ^^^^^^^^^^^^^^^^^^^^^^
4830 <result> = sdiv <ty> <op1>, <op2> ; yields ty:result
4831 <result> = sdiv exact <ty> <op1>, <op2> ; yields ty:result
4836 The '``sdiv``' instruction returns the quotient of its two operands.
4841 The two arguments to the '``sdiv``' instruction must be
4842 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4843 arguments must have identical types.
4848 The value produced is the signed integer quotient of the two operands
4849 rounded towards zero.
4851 Note that signed integer division and unsigned integer division are
4852 distinct operations; for unsigned integer division, use '``udiv``'.
4854 Division by zero leads to undefined behavior. Overflow also leads to
4855 undefined behavior; this is a rare case, but can occur, for example, by
4856 doing a 32-bit division of -2147483648 by -1.
4858 If the ``exact`` keyword is present, the result value of the ``sdiv`` is
4859 a :ref:`poison value <poisonvalues>` if the result would be rounded.
4864 .. code-block:: llvm
4866 <result> = sdiv i32 4, %var ; yields i32:result = 4 / %var
4870 '``fdiv``' Instruction
4871 ^^^^^^^^^^^^^^^^^^^^^^
4878 <result> = fdiv [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
4883 The '``fdiv``' instruction returns the quotient of its two operands.
4888 The two arguments to the '``fdiv``' instruction must be :ref:`floating
4889 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
4890 Both arguments must have identical types.
4895 The value produced is the floating point quotient of the two operands.
4896 This instruction can also take any number of :ref:`fast-math
4897 flags <fastmath>`, which are optimization hints to enable otherwise
4898 unsafe floating point optimizations:
4903 .. code-block:: llvm
4905 <result> = fdiv float 4.0, %var ; yields float:result = 4.0 / %var
4907 '``urem``' Instruction
4908 ^^^^^^^^^^^^^^^^^^^^^^
4915 <result> = urem <ty> <op1>, <op2> ; yields ty:result
4920 The '``urem``' instruction returns the remainder from the unsigned
4921 division of its two arguments.
4926 The two arguments to the '``urem``' instruction must be
4927 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4928 arguments must have identical types.
4933 This instruction returns the unsigned integer *remainder* of a division.
4934 This instruction always performs an unsigned division to get the
4937 Note that unsigned integer remainder and signed integer remainder are
4938 distinct operations; for signed integer remainder, use '``srem``'.
4940 Taking the remainder of a division by zero leads to undefined behavior.
4945 .. code-block:: llvm
4947 <result> = urem i32 4, %var ; yields i32:result = 4 % %var
4949 '``srem``' Instruction
4950 ^^^^^^^^^^^^^^^^^^^^^^
4957 <result> = srem <ty> <op1>, <op2> ; yields ty:result
4962 The '``srem``' instruction returns the remainder from the signed
4963 division of its two operands. This instruction can also take
4964 :ref:`vector <t_vector>` versions of the values in which case the elements
4970 The two arguments to the '``srem``' instruction must be
4971 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4972 arguments must have identical types.
4977 This instruction returns the *remainder* of a division (where the result
4978 is either zero or has the same sign as the dividend, ``op1``), not the
4979 *modulo* operator (where the result is either zero or has the same sign
4980 as the divisor, ``op2``) of a value. For more information about the
4981 difference, see `The Math
4982 Forum <http://mathforum.org/dr.math/problems/anne.4.28.99.html>`_. For a
4983 table of how this is implemented in various languages, please see
4985 operation <http://en.wikipedia.org/wiki/Modulo_operation>`_.
4987 Note that signed integer remainder and unsigned integer remainder are
4988 distinct operations; for unsigned integer remainder, use '``urem``'.
4990 Taking the remainder of a division by zero leads to undefined behavior.
4991 Overflow also leads to undefined behavior; this is a rare case, but can
4992 occur, for example, by taking the remainder of a 32-bit division of
4993 -2147483648 by -1. (The remainder doesn't actually overflow, but this
4994 rule lets srem be implemented using instructions that return both the
4995 result of the division and the remainder.)
5000 .. code-block:: llvm
5002 <result> = srem i32 4, %var ; yields i32:result = 4 % %var
5006 '``frem``' Instruction
5007 ^^^^^^^^^^^^^^^^^^^^^^
5014 <result> = frem [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
5019 The '``frem``' instruction returns the remainder from the division of
5025 The two arguments to the '``frem``' instruction must be :ref:`floating
5026 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
5027 Both arguments must have identical types.
5032 This instruction returns the *remainder* of a division. The remainder
5033 has the same sign as the dividend. This instruction can also take any
5034 number of :ref:`fast-math flags <fastmath>`, which are optimization hints
5035 to enable otherwise unsafe floating point optimizations:
5040 .. code-block:: llvm
5042 <result> = frem float 4.0, %var ; yields float:result = 4.0 % %var
5046 Bitwise Binary Operations
5047 -------------------------
5049 Bitwise binary operators are used to do various forms of bit-twiddling
5050 in a program. They are generally very efficient instructions and can
5051 commonly be strength reduced from other instructions. They require two
5052 operands of the same type, execute an operation on them, and produce a
5053 single value. The resulting value is the same type as its operands.
5055 '``shl``' Instruction
5056 ^^^^^^^^^^^^^^^^^^^^^
5063 <result> = shl <ty> <op1>, <op2> ; yields ty:result
5064 <result> = shl nuw <ty> <op1>, <op2> ; yields ty:result
5065 <result> = shl nsw <ty> <op1>, <op2> ; yields ty:result
5066 <result> = shl nuw nsw <ty> <op1>, <op2> ; yields ty:result
5071 The '``shl``' instruction returns the first operand shifted to the left
5072 a specified number of bits.
5077 Both arguments to the '``shl``' instruction must be the same
5078 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
5079 '``op2``' is treated as an unsigned value.
5084 The value produced is ``op1`` \* 2\ :sup:`op2` mod 2\ :sup:`n`,
5085 where ``n`` is the width of the result. If ``op2`` is (statically or
5086 dynamically) equal to or larger than the number of bits in
5087 ``op1``, the result is undefined. If the arguments are vectors, each
5088 vector element of ``op1`` is shifted by the corresponding shift amount
5091 If the ``nuw`` keyword is present, then the shift produces a :ref:`poison
5092 value <poisonvalues>` if it shifts out any non-zero bits. If the
5093 ``nsw`` keyword is present, then the shift produces a :ref:`poison
5094 value <poisonvalues>` if it shifts out any bits that disagree with the
5095 resultant sign bit. As such, NUW/NSW have the same semantics as they
5096 would if the shift were expressed as a mul instruction with the same
5097 nsw/nuw bits in (mul %op1, (shl 1, %op2)).
5102 .. code-block:: llvm
5104 <result> = shl i32 4, %var ; yields i32: 4 << %var
5105 <result> = shl i32 4, 2 ; yields i32: 16
5106 <result> = shl i32 1, 10 ; yields i32: 1024
5107 <result> = shl i32 1, 32 ; undefined
5108 <result> = shl <2 x i32> < i32 1, i32 1>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 2, i32 4>
5110 '``lshr``' Instruction
5111 ^^^^^^^^^^^^^^^^^^^^^^
5118 <result> = lshr <ty> <op1>, <op2> ; yields ty:result
5119 <result> = lshr exact <ty> <op1>, <op2> ; yields ty:result
5124 The '``lshr``' instruction (logical shift right) returns the first
5125 operand shifted to the right a specified number of bits with zero fill.
5130 Both arguments to the '``lshr``' instruction must be the same
5131 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
5132 '``op2``' is treated as an unsigned value.
5137 This instruction always performs a logical shift right operation. The
5138 most significant bits of the result will be filled with zero bits after
5139 the shift. If ``op2`` is (statically or dynamically) equal to or larger
5140 than the number of bits in ``op1``, the result is undefined. If the
5141 arguments are vectors, each vector element of ``op1`` is shifted by the
5142 corresponding shift amount in ``op2``.
5144 If the ``exact`` keyword is present, the result value of the ``lshr`` is
5145 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
5151 .. code-block:: llvm
5153 <result> = lshr i32 4, 1 ; yields i32:result = 2
5154 <result> = lshr i32 4, 2 ; yields i32:result = 1
5155 <result> = lshr i8 4, 3 ; yields i8:result = 0
5156 <result> = lshr i8 -2, 1 ; yields i8:result = 0x7F
5157 <result> = lshr i32 1, 32 ; undefined
5158 <result> = lshr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 0x7FFFFFFF, i32 1>
5160 '``ashr``' Instruction
5161 ^^^^^^^^^^^^^^^^^^^^^^
5168 <result> = ashr <ty> <op1>, <op2> ; yields ty:result
5169 <result> = ashr exact <ty> <op1>, <op2> ; yields ty:result
5174 The '``ashr``' instruction (arithmetic shift right) returns the first
5175 operand shifted to the right a specified number of bits with sign
5181 Both arguments to the '``ashr``' instruction must be the same
5182 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
5183 '``op2``' is treated as an unsigned value.
5188 This instruction always performs an arithmetic shift right operation,
5189 The most significant bits of the result will be filled with the sign bit
5190 of ``op1``. If ``op2`` is (statically or dynamically) equal to or larger
5191 than the number of bits in ``op1``, the result is undefined. If the
5192 arguments are vectors, each vector element of ``op1`` is shifted by the
5193 corresponding shift amount in ``op2``.
5195 If the ``exact`` keyword is present, the result value of the ``ashr`` is
5196 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
5202 .. code-block:: llvm
5204 <result> = ashr i32 4, 1 ; yields i32:result = 2
5205 <result> = ashr i32 4, 2 ; yields i32:result = 1
5206 <result> = ashr i8 4, 3 ; yields i8:result = 0
5207 <result> = ashr i8 -2, 1 ; yields i8:result = -1
5208 <result> = ashr i32 1, 32 ; undefined
5209 <result> = ashr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 3> ; yields: result=<2 x i32> < i32 -1, i32 0>
5211 '``and``' Instruction
5212 ^^^^^^^^^^^^^^^^^^^^^
5219 <result> = and <ty> <op1>, <op2> ; yields ty:result
5224 The '``and``' instruction returns the bitwise logical and of its two
5230 The two arguments to the '``and``' instruction must be
5231 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
5232 arguments must have identical types.
5237 The truth table used for the '``and``' instruction is:
5254 .. code-block:: llvm
5256 <result> = and i32 4, %var ; yields i32:result = 4 & %var
5257 <result> = and i32 15, 40 ; yields i32:result = 8
5258 <result> = and i32 4, 8 ; yields i32:result = 0
5260 '``or``' Instruction
5261 ^^^^^^^^^^^^^^^^^^^^
5268 <result> = or <ty> <op1>, <op2> ; yields ty:result
5273 The '``or``' instruction returns the bitwise logical inclusive or of its
5279 The two arguments to the '``or``' instruction must be
5280 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
5281 arguments must have identical types.
5286 The truth table used for the '``or``' instruction is:
5305 <result> = or i32 4, %var ; yields i32:result = 4 | %var
5306 <result> = or i32 15, 40 ; yields i32:result = 47
5307 <result> = or i32 4, 8 ; yields i32:result = 12
5309 '``xor``' Instruction
5310 ^^^^^^^^^^^^^^^^^^^^^
5317 <result> = xor <ty> <op1>, <op2> ; yields ty:result
5322 The '``xor``' instruction returns the bitwise logical exclusive or of
5323 its two operands. The ``xor`` is used to implement the "one's
5324 complement" operation, which is the "~" operator in C.
5329 The two arguments to the '``xor``' instruction must be
5330 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
5331 arguments must have identical types.
5336 The truth table used for the '``xor``' instruction is:
5353 .. code-block:: llvm
5355 <result> = xor i32 4, %var ; yields i32:result = 4 ^ %var
5356 <result> = xor i32 15, 40 ; yields i32:result = 39
5357 <result> = xor i32 4, 8 ; yields i32:result = 12
5358 <result> = xor i32 %V, -1 ; yields i32:result = ~%V
5363 LLVM supports several instructions to represent vector operations in a
5364 target-independent manner. These instructions cover the element-access
5365 and vector-specific operations needed to process vectors effectively.
5366 While LLVM does directly support these vector operations, many
5367 sophisticated algorithms will want to use target-specific intrinsics to
5368 take full advantage of a specific target.
5370 .. _i_extractelement:
5372 '``extractelement``' Instruction
5373 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5380 <result> = extractelement <n x <ty>> <val>, <ty2> <idx> ; yields <ty>
5385 The '``extractelement``' instruction extracts a single scalar element
5386 from a vector at a specified index.
5391 The first operand of an '``extractelement``' instruction is a value of
5392 :ref:`vector <t_vector>` type. The second operand is an index indicating
5393 the position from which to extract the element. The index may be a
5394 variable of any integer type.
5399 The result is a scalar of the same type as the element type of ``val``.
5400 Its value is the value at position ``idx`` of ``val``. If ``idx``
5401 exceeds the length of ``val``, the results are undefined.
5406 .. code-block:: llvm
5408 <result> = extractelement <4 x i32> %vec, i32 0 ; yields i32
5410 .. _i_insertelement:
5412 '``insertelement``' Instruction
5413 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5420 <result> = insertelement <n x <ty>> <val>, <ty> <elt>, <ty2> <idx> ; yields <n x <ty>>
5425 The '``insertelement``' instruction inserts a scalar element into a
5426 vector at a specified index.
5431 The first operand of an '``insertelement``' instruction is a value of
5432 :ref:`vector <t_vector>` type. The second operand is a scalar value whose
5433 type must equal the element type of the first operand. The third operand
5434 is an index indicating the position at which to insert the value. The
5435 index may be a variable of any integer type.
5440 The result is a vector of the same type as ``val``. Its element values
5441 are those of ``val`` except at position ``idx``, where it gets the value
5442 ``elt``. If ``idx`` exceeds the length of ``val``, the results are
5448 .. code-block:: llvm
5450 <result> = insertelement <4 x i32> %vec, i32 1, i32 0 ; yields <4 x i32>
5452 .. _i_shufflevector:
5454 '``shufflevector``' Instruction
5455 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5462 <result> = shufflevector <n x <ty>> <v1>, <n x <ty>> <v2>, <m x i32> <mask> ; yields <m x <ty>>
5467 The '``shufflevector``' instruction constructs a permutation of elements
5468 from two input vectors, returning a vector with the same element type as
5469 the input and length that is the same as the shuffle mask.
5474 The first two operands of a '``shufflevector``' instruction are vectors
5475 with the same type. The third argument is a shuffle mask whose element
5476 type is always 'i32'. The result of the instruction is a vector whose
5477 length is the same as the shuffle mask and whose element type is the
5478 same as the element type of the first two operands.
5480 The shuffle mask operand is required to be a constant vector with either
5481 constant integer or undef values.
5486 The elements of the two input vectors are numbered from left to right
5487 across both of the vectors. The shuffle mask operand specifies, for each
5488 element of the result vector, which element of the two input vectors the
5489 result element gets. The element selector may be undef (meaning "don't
5490 care") and the second operand may be undef if performing a shuffle from
5496 .. code-block:: llvm
5498 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
5499 <4 x i32> <i32 0, i32 4, i32 1, i32 5> ; yields <4 x i32>
5500 <result> = shufflevector <4 x i32> %v1, <4 x i32> undef,
5501 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32> - Identity shuffle.
5502 <result> = shufflevector <8 x i32> %v1, <8 x i32> undef,
5503 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32>
5504 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
5505 <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7 > ; yields <8 x i32>
5507 Aggregate Operations
5508 --------------------
5510 LLVM supports several instructions for working with
5511 :ref:`aggregate <t_aggregate>` values.
5515 '``extractvalue``' Instruction
5516 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5523 <result> = extractvalue <aggregate type> <val>, <idx>{, <idx>}*
5528 The '``extractvalue``' instruction extracts the value of a member field
5529 from an :ref:`aggregate <t_aggregate>` value.
5534 The first operand of an '``extractvalue``' instruction is a value of
5535 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The operands are
5536 constant indices to specify which value to extract in a similar manner
5537 as indices in a '``getelementptr``' instruction.
5539 The major differences to ``getelementptr`` indexing are:
5541 - Since the value being indexed is not a pointer, the first index is
5542 omitted and assumed to be zero.
5543 - At least one index must be specified.
5544 - Not only struct indices but also array indices must be in bounds.
5549 The result is the value at the position in the aggregate specified by
5555 .. code-block:: llvm
5557 <result> = extractvalue {i32, float} %agg, 0 ; yields i32
5561 '``insertvalue``' Instruction
5562 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5569 <result> = insertvalue <aggregate type> <val>, <ty> <elt>, <idx>{, <idx>}* ; yields <aggregate type>
5574 The '``insertvalue``' instruction inserts a value into a member field in
5575 an :ref:`aggregate <t_aggregate>` value.
5580 The first operand of an '``insertvalue``' instruction is a value of
5581 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The second operand is
5582 a first-class value to insert. The following operands are constant
5583 indices indicating the position at which to insert the value in a
5584 similar manner as indices in a '``extractvalue``' instruction. The value
5585 to insert must have the same type as the value identified by the
5591 The result is an aggregate of the same type as ``val``. Its value is
5592 that of ``val`` except that the value at the position specified by the
5593 indices is that of ``elt``.
5598 .. code-block:: llvm
5600 %agg1 = insertvalue {i32, float} undef, i32 1, 0 ; yields {i32 1, float undef}
5601 %agg2 = insertvalue {i32, float} %agg1, float %val, 1 ; yields {i32 1, float %val}
5602 %agg3 = insertvalue {i32, {float}} undef, float %val, 1, 0 ; yields {i32 undef, {float %val}}
5606 Memory Access and Addressing Operations
5607 ---------------------------------------
5609 A key design point of an SSA-based representation is how it represents
5610 memory. In LLVM, no memory locations are in SSA form, which makes things
5611 very simple. This section describes how to read, write, and allocate
5616 '``alloca``' Instruction
5617 ^^^^^^^^^^^^^^^^^^^^^^^^
5624 <result> = alloca [inalloca] <type> [, <ty> <NumElements>] [, align <alignment>] ; yields type*:result
5629 The '``alloca``' instruction allocates memory on the stack frame of the
5630 currently executing function, to be automatically released when this
5631 function returns to its caller. The object is always allocated in the
5632 generic address space (address space zero).
5637 The '``alloca``' instruction allocates ``sizeof(<type>)*NumElements``
5638 bytes of memory on the runtime stack, returning a pointer of the
5639 appropriate type to the program. If "NumElements" is specified, it is
5640 the number of elements allocated, otherwise "NumElements" is defaulted
5641 to be one. If a constant alignment is specified, the value result of the
5642 allocation is guaranteed to be aligned to at least that boundary. The
5643 alignment may not be greater than ``1 << 29``. If not specified, or if
5644 zero, the target can choose to align the allocation on any convenient
5645 boundary compatible with the type.
5647 '``type``' may be any sized type.
5652 Memory is allocated; a pointer is returned. The operation is undefined
5653 if there is insufficient stack space for the allocation. '``alloca``'d
5654 memory is automatically released when the function returns. The
5655 '``alloca``' instruction is commonly used to represent automatic
5656 variables that must have an address available. When the function returns
5657 (either with the ``ret`` or ``resume`` instructions), the memory is
5658 reclaimed. Allocating zero bytes is legal, but the result is undefined.
5659 The order in which memory is allocated (ie., which way the stack grows)
5665 .. code-block:: llvm
5667 %ptr = alloca i32 ; yields i32*:ptr
5668 %ptr = alloca i32, i32 4 ; yields i32*:ptr
5669 %ptr = alloca i32, i32 4, align 1024 ; yields i32*:ptr
5670 %ptr = alloca i32, align 1024 ; yields i32*:ptr
5674 '``load``' Instruction
5675 ^^^^^^^^^^^^^^^^^^^^^^
5682 <result> = load [volatile] <ty>, <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>][, !invariant.load !<index>][, !nonnull !<index>][, !dereferenceable !<index>][, !dereferenceable_or_null !<index>]
5683 <result> = load atomic [volatile] <ty>* <pointer> [singlethread] <ordering>, align <alignment>
5684 !<index> = !{ i32 1 }
5689 The '``load``' instruction is used to read from memory.
5694 The argument to the ``load`` instruction specifies the memory address
5695 from which to load. The type specified must be a :ref:`first
5696 class <t_firstclass>` type. If the ``load`` is marked as ``volatile``,
5697 then the optimizer is not allowed to modify the number or order of
5698 execution of this ``load`` with other :ref:`volatile
5699 operations <volatile>`.
5701 If the ``load`` is marked as ``atomic``, it takes an extra
5702 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
5703 ``release`` and ``acq_rel`` orderings are not valid on ``load``
5704 instructions. Atomic loads produce :ref:`defined <memmodel>` results
5705 when they may see multiple atomic stores. The type of the pointee must
5706 be an integer type whose bit width is a power of two greater than or
5707 equal to eight and less than or equal to a target-specific size limit.
5708 ``align`` must be explicitly specified on atomic loads, and the load has
5709 undefined behavior if the alignment is not set to a value which is at
5710 least the size in bytes of the pointee. ``!nontemporal`` does not have
5711 any defined semantics for atomic loads.
5713 The optional constant ``align`` argument specifies the alignment of the
5714 operation (that is, the alignment of the memory address). A value of 0
5715 or an omitted ``align`` argument means that the operation has the ABI
5716 alignment for the target. It is the responsibility of the code emitter
5717 to ensure that the alignment information is correct. Overestimating the
5718 alignment results in undefined behavior. Underestimating the alignment
5719 may produce less efficient code. An alignment of 1 is always safe. The
5720 maximum possible alignment is ``1 << 29``.
5722 The optional ``!nontemporal`` metadata must reference a single
5723 metadata name ``<index>`` corresponding to a metadata node with one
5724 ``i32`` entry of value 1. The existence of the ``!nontemporal``
5725 metadata on the instruction tells the optimizer and code generator
5726 that this load is not expected to be reused in the cache. The code
5727 generator may select special instructions to save cache bandwidth, such
5728 as the ``MOVNT`` instruction on x86.
5730 The optional ``!invariant.load`` metadata must reference a single
5731 metadata name ``<index>`` corresponding to a metadata node with no
5732 entries. The existence of the ``!invariant.load`` metadata on the
5733 instruction tells the optimizer and code generator that the address
5734 operand to this load points to memory which can be assumed unchanged.
5735 Being invariant does not imply that a location is dereferenceable,
5736 but it does imply that once the location is known dereferenceable
5737 its value is henceforth unchanging.
5739 The optional ``!nonnull`` metadata must reference a single
5740 metadata name ``<index>`` corresponding to a metadata node with no
5741 entries. The existence of the ``!nonnull`` metadata on the
5742 instruction tells the optimizer that the value loaded is known to
5743 never be null. This is analogous to the ''nonnull'' attribute
5744 on parameters and return values. This metadata can only be applied
5745 to loads of a pointer type.
5747 The optional ``!dereferenceable`` metadata must reference a single
5748 metadata name ``<index>`` corresponding to a metadata node with one ``i64``
5749 entry. The existence of the ``!dereferenceable`` metadata on the instruction
5750 tells the optimizer that the value loaded is known to be dereferenceable.
5751 The number of bytes known to be dereferenceable is specified by the integer
5752 value in the metadata node. This is analogous to the ''dereferenceable''
5753 attribute on parameters and return values. This metadata can only be applied
5754 to loads of a pointer type.
5756 The optional ``!dereferenceable_or_null`` metadata must reference a single
5757 metadata name ``<index>`` corresponding to a metadata node with one ``i64``
5758 entry. The existence of the ``!dereferenceable_or_null`` metadata on the
5759 instruction tells the optimizer that the value loaded is known to be either
5760 dereferenceable or null.
5761 The number of bytes known to be dereferenceable is specified by the integer
5762 value in the metadata node. This is analogous to the ''dereferenceable_or_null''
5763 attribute on parameters and return values. This metadata can only be applied
5764 to loads of a pointer type.
5769 The location of memory pointed to is loaded. If the value being loaded
5770 is of scalar type then the number of bytes read does not exceed the
5771 minimum number of bytes needed to hold all bits of the type. For
5772 example, loading an ``i24`` reads at most three bytes. When loading a
5773 value of a type like ``i20`` with a size that is not an integral number
5774 of bytes, the result is undefined if the value was not originally
5775 written using a store of the same type.
5780 .. code-block:: llvm
5782 %ptr = alloca i32 ; yields i32*:ptr
5783 store i32 3, i32* %ptr ; yields void
5784 %val = load i32, i32* %ptr ; yields i32:val = i32 3
5788 '``store``' Instruction
5789 ^^^^^^^^^^^^^^^^^^^^^^^
5796 store [volatile] <ty> <value>, <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>] ; yields void
5797 store atomic [volatile] <ty> <value>, <ty>* <pointer> [singlethread] <ordering>, align <alignment> ; yields void
5802 The '``store``' instruction is used to write to memory.
5807 There are two arguments to the ``store`` instruction: a value to store
5808 and an address at which to store it. The type of the ``<pointer>``
5809 operand must be a pointer to the :ref:`first class <t_firstclass>` type of
5810 the ``<value>`` operand. If the ``store`` is marked as ``volatile``,
5811 then the optimizer is not allowed to modify the number or order of
5812 execution of this ``store`` with other :ref:`volatile
5813 operations <volatile>`.
5815 If the ``store`` is marked as ``atomic``, it takes an extra
5816 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
5817 ``acquire`` and ``acq_rel`` orderings aren't valid on ``store``
5818 instructions. Atomic loads produce :ref:`defined <memmodel>` results
5819 when they may see multiple atomic stores. The type of the pointee must
5820 be an integer type whose bit width is a power of two greater than or
5821 equal to eight and less than or equal to a target-specific size limit.
5822 ``align`` must be explicitly specified on atomic stores, and the store
5823 has undefined behavior if the alignment is not set to a value which is
5824 at least the size in bytes of the pointee. ``!nontemporal`` does not
5825 have any defined semantics for atomic stores.
5827 The optional constant ``align`` argument specifies the alignment of the
5828 operation (that is, the alignment of the memory address). A value of 0
5829 or an omitted ``align`` argument means that the operation has the ABI
5830 alignment for the target. It is the responsibility of the code emitter
5831 to ensure that the alignment information is correct. Overestimating the
5832 alignment results in undefined behavior. Underestimating the
5833 alignment may produce less efficient code. An alignment of 1 is always
5834 safe. The maximum possible alignment is ``1 << 29``.
5836 The optional ``!nontemporal`` metadata must reference a single metadata
5837 name ``<index>`` corresponding to a metadata node with one ``i32`` entry of
5838 value 1. The existence of the ``!nontemporal`` metadata on the instruction
5839 tells the optimizer and code generator that this load is not expected to
5840 be reused in the cache. The code generator may select special
5841 instructions to save cache bandwidth, such as the MOVNT instruction on
5847 The contents of memory are updated to contain ``<value>`` at the
5848 location specified by the ``<pointer>`` operand. If ``<value>`` is
5849 of scalar type then the number of bytes written does not exceed the
5850 minimum number of bytes needed to hold all bits of the type. For
5851 example, storing an ``i24`` writes at most three bytes. When writing a
5852 value of a type like ``i20`` with a size that is not an integral number
5853 of bytes, it is unspecified what happens to the extra bits that do not
5854 belong to the type, but they will typically be overwritten.
5859 .. code-block:: llvm
5861 %ptr = alloca i32 ; yields i32*:ptr
5862 store i32 3, i32* %ptr ; yields void
5863 %val = load i32* %ptr ; yields i32:val = i32 3
5867 '``fence``' Instruction
5868 ^^^^^^^^^^^^^^^^^^^^^^^
5875 fence [singlethread] <ordering> ; yields void
5880 The '``fence``' instruction is used to introduce happens-before edges
5886 '``fence``' instructions take an :ref:`ordering <ordering>` argument which
5887 defines what *synchronizes-with* edges they add. They can only be given
5888 ``acquire``, ``release``, ``acq_rel``, and ``seq_cst`` orderings.
5893 A fence A which has (at least) ``release`` ordering semantics
5894 *synchronizes with* a fence B with (at least) ``acquire`` ordering
5895 semantics if and only if there exist atomic operations X and Y, both
5896 operating on some atomic object M, such that A is sequenced before X, X
5897 modifies M (either directly or through some side effect of a sequence
5898 headed by X), Y is sequenced before B, and Y observes M. This provides a
5899 *happens-before* dependency between A and B. Rather than an explicit
5900 ``fence``, one (but not both) of the atomic operations X or Y might
5901 provide a ``release`` or ``acquire`` (resp.) ordering constraint and
5902 still *synchronize-with* the explicit ``fence`` and establish the
5903 *happens-before* edge.
5905 A ``fence`` which has ``seq_cst`` ordering, in addition to having both
5906 ``acquire`` and ``release`` semantics specified above, participates in
5907 the global program order of other ``seq_cst`` operations and/or fences.
5909 The optional ":ref:`singlethread <singlethread>`" argument specifies
5910 that the fence only synchronizes with other fences in the same thread.
5911 (This is useful for interacting with signal handlers.)
5916 .. code-block:: llvm
5918 fence acquire ; yields void
5919 fence singlethread seq_cst ; yields void
5923 '``cmpxchg``' Instruction
5924 ^^^^^^^^^^^^^^^^^^^^^^^^^
5931 cmpxchg [weak] [volatile] <ty>* <pointer>, <ty> <cmp>, <ty> <new> [singlethread] <success ordering> <failure ordering> ; yields { ty, i1 }
5936 The '``cmpxchg``' instruction is used to atomically modify memory. It
5937 loads a value in memory and compares it to a given value. If they are
5938 equal, it tries to store a new value into the memory.
5943 There are three arguments to the '``cmpxchg``' instruction: an address
5944 to operate on, a value to compare to the value currently be at that
5945 address, and a new value to place at that address if the compared values
5946 are equal. The type of '<cmp>' must be an integer type whose bit width
5947 is a power of two greater than or equal to eight and less than or equal
5948 to a target-specific size limit. '<cmp>' and '<new>' must have the same
5949 type, and the type of '<pointer>' must be a pointer to that type. If the
5950 ``cmpxchg`` is marked as ``volatile``, then the optimizer is not allowed
5951 to modify the number or order of execution of this ``cmpxchg`` with
5952 other :ref:`volatile operations <volatile>`.
5954 The success and failure :ref:`ordering <ordering>` arguments specify how this
5955 ``cmpxchg`` synchronizes with other atomic operations. Both ordering parameters
5956 must be at least ``monotonic``, the ordering constraint on failure must be no
5957 stronger than that on success, and the failure ordering cannot be either
5958 ``release`` or ``acq_rel``.
5960 The optional "``singlethread``" argument declares that the ``cmpxchg``
5961 is only atomic with respect to code (usually signal handlers) running in
5962 the same thread as the ``cmpxchg``. Otherwise the cmpxchg is atomic with
5963 respect to all other code in the system.
5965 The pointer passed into cmpxchg must have alignment greater than or
5966 equal to the size in memory of the operand.
5971 The contents of memory at the location specified by the '``<pointer>``' operand
5972 is read and compared to '``<cmp>``'; if the read value is the equal, the
5973 '``<new>``' is written. The original value at the location is returned, together
5974 with a flag indicating success (true) or failure (false).
5976 If the cmpxchg operation is marked as ``weak`` then a spurious failure is
5977 permitted: the operation may not write ``<new>`` even if the comparison
5980 If the cmpxchg operation is strong (the default), the i1 value is 1 if and only
5981 if the value loaded equals ``cmp``.
5983 A successful ``cmpxchg`` is a read-modify-write instruction for the purpose of
5984 identifying release sequences. A failed ``cmpxchg`` is equivalent to an atomic
5985 load with an ordering parameter determined the second ordering parameter.
5990 .. code-block:: llvm
5993 %orig = atomic load i32, i32* %ptr unordered ; yields i32
5997 %cmp = phi i32 [ %orig, %entry ], [%old, %loop]
5998 %squared = mul i32 %cmp, %cmp
5999 %val_success = cmpxchg i32* %ptr, i32 %cmp, i32 %squared acq_rel monotonic ; yields { i32, i1 }
6000 %value_loaded = extractvalue { i32, i1 } %val_success, 0
6001 %success = extractvalue { i32, i1 } %val_success, 1
6002 br i1 %success, label %done, label %loop
6009 '``atomicrmw``' Instruction
6010 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6017 atomicrmw [volatile] <operation> <ty>* <pointer>, <ty> <value> [singlethread] <ordering> ; yields ty
6022 The '``atomicrmw``' instruction is used to atomically modify memory.
6027 There are three arguments to the '``atomicrmw``' instruction: an
6028 operation to apply, an address whose value to modify, an argument to the
6029 operation. The operation must be one of the following keywords:
6043 The type of '<value>' must be an integer type whose bit width is a power
6044 of two greater than or equal to eight and less than or equal to a
6045 target-specific size limit. The type of the '``<pointer>``' operand must
6046 be a pointer to that type. If the ``atomicrmw`` is marked as
6047 ``volatile``, then the optimizer is not allowed to modify the number or
6048 order of execution of this ``atomicrmw`` with other :ref:`volatile
6049 operations <volatile>`.
6054 The contents of memory at the location specified by the '``<pointer>``'
6055 operand are atomically read, modified, and written back. The original
6056 value at the location is returned. The modification is specified by the
6059 - xchg: ``*ptr = val``
6060 - add: ``*ptr = *ptr + val``
6061 - sub: ``*ptr = *ptr - val``
6062 - and: ``*ptr = *ptr & val``
6063 - nand: ``*ptr = ~(*ptr & val)``
6064 - or: ``*ptr = *ptr | val``
6065 - xor: ``*ptr = *ptr ^ val``
6066 - max: ``*ptr = *ptr > val ? *ptr : val`` (using a signed comparison)
6067 - min: ``*ptr = *ptr < val ? *ptr : val`` (using a signed comparison)
6068 - umax: ``*ptr = *ptr > val ? *ptr : val`` (using an unsigned
6070 - umin: ``*ptr = *ptr < val ? *ptr : val`` (using an unsigned
6076 .. code-block:: llvm
6078 %old = atomicrmw add i32* %ptr, i32 1 acquire ; yields i32
6080 .. _i_getelementptr:
6082 '``getelementptr``' Instruction
6083 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6090 <result> = getelementptr <ty>, <ty>* <ptrval>{, <ty> <idx>}*
6091 <result> = getelementptr inbounds <ty>, <ty>* <ptrval>{, <ty> <idx>}*
6092 <result> = getelementptr <ty>, <ptr vector> <ptrval>, <vector index type> <idx>
6097 The '``getelementptr``' instruction is used to get the address of a
6098 subelement of an :ref:`aggregate <t_aggregate>` data structure. It performs
6099 address calculation only and does not access memory.
6104 The first argument is always a type used as the basis for the calculations.
6105 The second argument is always a pointer or a vector of pointers, and is the
6106 base address to start from. The remaining arguments are indices
6107 that indicate which of the elements of the aggregate object are indexed.
6108 The interpretation of each index is dependent on the type being indexed
6109 into. The first index always indexes the pointer value given as the
6110 first argument, the second index indexes a value of the type pointed to
6111 (not necessarily the value directly pointed to, since the first index
6112 can be non-zero), etc. The first type indexed into must be a pointer
6113 value, subsequent types can be arrays, vectors, and structs. Note that
6114 subsequent types being indexed into can never be pointers, since that
6115 would require loading the pointer before continuing calculation.
6117 The type of each index argument depends on the type it is indexing into.
6118 When indexing into a (optionally packed) structure, only ``i32`` integer
6119 **constants** are allowed (when using a vector of indices they must all
6120 be the **same** ``i32`` integer constant). When indexing into an array,
6121 pointer or vector, integers of any width are allowed, and they are not
6122 required to be constant. These integers are treated as signed values
6125 For example, let's consider a C code fragment and how it gets compiled
6141 int *foo(struct ST *s) {
6142 return &s[1].Z.B[5][13];
6145 The LLVM code generated by Clang is:
6147 .. code-block:: llvm
6149 %struct.RT = type { i8, [10 x [20 x i32]], i8 }
6150 %struct.ST = type { i32, double, %struct.RT }
6152 define i32* @foo(%struct.ST* %s) nounwind uwtable readnone optsize ssp {
6154 %arrayidx = getelementptr inbounds %struct.ST, %struct.ST* %s, i64 1, i32 2, i32 1, i64 5, i64 13
6161 In the example above, the first index is indexing into the
6162 '``%struct.ST*``' type, which is a pointer, yielding a '``%struct.ST``'
6163 = '``{ i32, double, %struct.RT }``' type, a structure. The second index
6164 indexes into the third element of the structure, yielding a
6165 '``%struct.RT``' = '``{ i8 , [10 x [20 x i32]], i8 }``' type, another
6166 structure. The third index indexes into the second element of the
6167 structure, yielding a '``[10 x [20 x i32]]``' type, an array. The two
6168 dimensions of the array are subscripted into, yielding an '``i32``'
6169 type. The '``getelementptr``' instruction returns a pointer to this
6170 element, thus computing a value of '``i32*``' type.
6172 Note that it is perfectly legal to index partially through a structure,
6173 returning a pointer to an inner element. Because of this, the LLVM code
6174 for the given testcase is equivalent to:
6176 .. code-block:: llvm
6178 define i32* @foo(%struct.ST* %s) {
6179 %t1 = getelementptr %struct.ST, %struct.ST* %s, i32 1 ; yields %struct.ST*:%t1
6180 %t2 = getelementptr %struct.ST, %struct.ST* %t1, i32 0, i32 2 ; yields %struct.RT*:%t2
6181 %t3 = getelementptr %struct.RT, %struct.RT* %t2, i32 0, i32 1 ; yields [10 x [20 x i32]]*:%t3
6182 %t4 = getelementptr [10 x [20 x i32]], [10 x [20 x i32]]* %t3, i32 0, i32 5 ; yields [20 x i32]*:%t4
6183 %t5 = getelementptr [20 x i32], [20 x i32]* %t4, i32 0, i32 13 ; yields i32*:%t5
6187 If the ``inbounds`` keyword is present, the result value of the
6188 ``getelementptr`` is a :ref:`poison value <poisonvalues>` if the base
6189 pointer is not an *in bounds* address of an allocated object, or if any
6190 of the addresses that would be formed by successive addition of the
6191 offsets implied by the indices to the base address with infinitely
6192 precise signed arithmetic are not an *in bounds* address of that
6193 allocated object. The *in bounds* addresses for an allocated object are
6194 all the addresses that point into the object, plus the address one byte
6195 past the end. In cases where the base is a vector of pointers the
6196 ``inbounds`` keyword applies to each of the computations element-wise.
6198 If the ``inbounds`` keyword is not present, the offsets are added to the
6199 base address with silently-wrapping two's complement arithmetic. If the
6200 offsets have a different width from the pointer, they are sign-extended
6201 or truncated to the width of the pointer. The result value of the
6202 ``getelementptr`` may be outside the object pointed to by the base
6203 pointer. The result value may not necessarily be used to access memory
6204 though, even if it happens to point into allocated storage. See the
6205 :ref:`Pointer Aliasing Rules <pointeraliasing>` section for more
6208 The getelementptr instruction is often confusing. For some more insight
6209 into how it works, see :doc:`the getelementptr FAQ <GetElementPtr>`.
6214 .. code-block:: llvm
6216 ; yields [12 x i8]*:aptr
6217 %aptr = getelementptr {i32, [12 x i8]}, {i32, [12 x i8]}* %saptr, i64 0, i32 1
6219 %vptr = getelementptr {i32, <2 x i8>}, {i32, <2 x i8>}* %svptr, i64 0, i32 1, i32 1
6221 %eptr = getelementptr [12 x i8], [12 x i8]* %aptr, i64 0, i32 1
6223 %iptr = getelementptr [10 x i32], [10 x i32]* @arr, i16 0, i16 0
6225 In cases where the pointer argument is a vector of pointers, each index
6226 must be a vector with the same number of elements. For example:
6228 .. code-block:: llvm
6230 %A = getelementptr i8, <4 x i8*> %ptrs, <4 x i64> %offsets,
6232 Conversion Operations
6233 ---------------------
6235 The instructions in this category are the conversion instructions
6236 (casting) which all take a single operand and a type. They perform
6237 various bit conversions on the operand.
6239 '``trunc .. to``' Instruction
6240 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6247 <result> = trunc <ty> <value> to <ty2> ; yields ty2
6252 The '``trunc``' instruction truncates its operand to the type ``ty2``.
6257 The '``trunc``' instruction takes a value to trunc, and a type to trunc
6258 it to. Both types must be of :ref:`integer <t_integer>` types, or vectors
6259 of the same number of integers. The bit size of the ``value`` must be
6260 larger than the bit size of the destination type, ``ty2``. Equal sized
6261 types are not allowed.
6266 The '``trunc``' instruction truncates the high order bits in ``value``
6267 and converts the remaining bits to ``ty2``. Since the source size must
6268 be larger than the destination size, ``trunc`` cannot be a *no-op cast*.
6269 It will always truncate bits.
6274 .. code-block:: llvm
6276 %X = trunc i32 257 to i8 ; yields i8:1
6277 %Y = trunc i32 123 to i1 ; yields i1:true
6278 %Z = trunc i32 122 to i1 ; yields i1:false
6279 %W = trunc <2 x i16> <i16 8, i16 7> to <2 x i8> ; yields <i8 8, i8 7>
6281 '``zext .. to``' Instruction
6282 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6289 <result> = zext <ty> <value> to <ty2> ; yields ty2
6294 The '``zext``' instruction zero extends its operand to type ``ty2``.
6299 The '``zext``' instruction takes a value to cast, and a type to cast it
6300 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
6301 the same number of integers. The bit size of the ``value`` must be
6302 smaller than the bit size of the destination type, ``ty2``.
6307 The ``zext`` fills the high order bits of the ``value`` with zero bits
6308 until it reaches the size of the destination type, ``ty2``.
6310 When zero extending from i1, the result will always be either 0 or 1.
6315 .. code-block:: llvm
6317 %X = zext i32 257 to i64 ; yields i64:257
6318 %Y = zext i1 true to i32 ; yields i32:1
6319 %Z = zext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
6321 '``sext .. to``' Instruction
6322 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6329 <result> = sext <ty> <value> to <ty2> ; yields ty2
6334 The '``sext``' sign extends ``value`` to the type ``ty2``.
6339 The '``sext``' instruction takes a value to cast, and a type to cast it
6340 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
6341 the same number of integers. The bit size of the ``value`` must be
6342 smaller than the bit size of the destination type, ``ty2``.
6347 The '``sext``' instruction performs a sign extension by copying the sign
6348 bit (highest order bit) of the ``value`` until it reaches the bit size
6349 of the type ``ty2``.
6351 When sign extending from i1, the extension always results in -1 or 0.
6356 .. code-block:: llvm
6358 %X = sext i8 -1 to i16 ; yields i16 :65535
6359 %Y = sext i1 true to i32 ; yields i32:-1
6360 %Z = sext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
6362 '``fptrunc .. to``' Instruction
6363 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6370 <result> = fptrunc <ty> <value> to <ty2> ; yields ty2
6375 The '``fptrunc``' instruction truncates ``value`` to type ``ty2``.
6380 The '``fptrunc``' instruction takes a :ref:`floating point <t_floating>`
6381 value to cast and a :ref:`floating point <t_floating>` type to cast it to.
6382 The size of ``value`` must be larger than the size of ``ty2``. This
6383 implies that ``fptrunc`` cannot be used to make a *no-op cast*.
6388 The '``fptrunc``' instruction truncates a ``value`` from a larger
6389 :ref:`floating point <t_floating>` type to a smaller :ref:`floating
6390 point <t_floating>` type. If the value cannot fit within the
6391 destination type, ``ty2``, then the results are undefined.
6396 .. code-block:: llvm
6398 %X = fptrunc double 123.0 to float ; yields float:123.0
6399 %Y = fptrunc double 1.0E+300 to float ; yields undefined
6401 '``fpext .. to``' Instruction
6402 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6409 <result> = fpext <ty> <value> to <ty2> ; yields ty2
6414 The '``fpext``' extends a floating point ``value`` to a larger floating
6420 The '``fpext``' instruction takes a :ref:`floating point <t_floating>`
6421 ``value`` to cast, and a :ref:`floating point <t_floating>` type to cast it
6422 to. The source type must be smaller than the destination type.
6427 The '``fpext``' instruction extends the ``value`` from a smaller
6428 :ref:`floating point <t_floating>` type to a larger :ref:`floating
6429 point <t_floating>` type. The ``fpext`` cannot be used to make a
6430 *no-op cast* because it always changes bits. Use ``bitcast`` to make a
6431 *no-op cast* for a floating point cast.
6436 .. code-block:: llvm
6438 %X = fpext float 3.125 to double ; yields double:3.125000e+00
6439 %Y = fpext double %X to fp128 ; yields fp128:0xL00000000000000004000900000000000
6441 '``fptoui .. to``' Instruction
6442 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6449 <result> = fptoui <ty> <value> to <ty2> ; yields ty2
6454 The '``fptoui``' converts a floating point ``value`` to its unsigned
6455 integer equivalent of type ``ty2``.
6460 The '``fptoui``' instruction takes a value to cast, which must be a
6461 scalar or vector :ref:`floating point <t_floating>` value, and a type to
6462 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
6463 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
6464 type with the same number of elements as ``ty``
6469 The '``fptoui``' instruction converts its :ref:`floating
6470 point <t_floating>` operand into the nearest (rounding towards zero)
6471 unsigned integer value. If the value cannot fit in ``ty2``, the results
6477 .. code-block:: llvm
6479 %X = fptoui double 123.0 to i32 ; yields i32:123
6480 %Y = fptoui float 1.0E+300 to i1 ; yields undefined:1
6481 %Z = fptoui float 1.04E+17 to i8 ; yields undefined:1
6483 '``fptosi .. to``' Instruction
6484 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6491 <result> = fptosi <ty> <value> to <ty2> ; yields ty2
6496 The '``fptosi``' instruction converts :ref:`floating point <t_floating>`
6497 ``value`` to type ``ty2``.
6502 The '``fptosi``' instruction takes a value to cast, which must be a
6503 scalar or vector :ref:`floating point <t_floating>` value, and a type to
6504 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
6505 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
6506 type with the same number of elements as ``ty``
6511 The '``fptosi``' instruction converts its :ref:`floating
6512 point <t_floating>` operand into the nearest (rounding towards zero)
6513 signed integer value. If the value cannot fit in ``ty2``, the results
6519 .. code-block:: llvm
6521 %X = fptosi double -123.0 to i32 ; yields i32:-123
6522 %Y = fptosi float 1.0E-247 to i1 ; yields undefined:1
6523 %Z = fptosi float 1.04E+17 to i8 ; yields undefined:1
6525 '``uitofp .. to``' Instruction
6526 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6533 <result> = uitofp <ty> <value> to <ty2> ; yields ty2
6538 The '``uitofp``' instruction regards ``value`` as an unsigned integer
6539 and converts that value to the ``ty2`` type.
6544 The '``uitofp``' instruction takes a value to cast, which must be a
6545 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
6546 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
6547 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
6548 type with the same number of elements as ``ty``
6553 The '``uitofp``' instruction interprets its operand as an unsigned
6554 integer quantity and converts it to the corresponding floating point
6555 value. If the value cannot fit in the floating point value, the results
6561 .. code-block:: llvm
6563 %X = uitofp i32 257 to float ; yields float:257.0
6564 %Y = uitofp i8 -1 to double ; yields double:255.0
6566 '``sitofp .. to``' Instruction
6567 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6574 <result> = sitofp <ty> <value> to <ty2> ; yields ty2
6579 The '``sitofp``' instruction regards ``value`` as a signed integer and
6580 converts that value to the ``ty2`` type.
6585 The '``sitofp``' instruction takes a value to cast, which must be a
6586 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
6587 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
6588 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
6589 type with the same number of elements as ``ty``
6594 The '``sitofp``' instruction interprets its operand as a signed integer
6595 quantity and converts it to the corresponding floating point value. If
6596 the value cannot fit in the floating point value, the results are
6602 .. code-block:: llvm
6604 %X = sitofp i32 257 to float ; yields float:257.0
6605 %Y = sitofp i8 -1 to double ; yields double:-1.0
6609 '``ptrtoint .. to``' Instruction
6610 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6617 <result> = ptrtoint <ty> <value> to <ty2> ; yields ty2
6622 The '``ptrtoint``' instruction converts the pointer or a vector of
6623 pointers ``value`` to the integer (or vector of integers) type ``ty2``.
6628 The '``ptrtoint``' instruction takes a ``value`` to cast, which must be
6629 a value of type :ref:`pointer <t_pointer>` or a vector of pointers, and a
6630 type to cast it to ``ty2``, which must be an :ref:`integer <t_integer>` or
6631 a vector of integers type.
6636 The '``ptrtoint``' instruction converts ``value`` to integer type
6637 ``ty2`` by interpreting the pointer value as an integer and either
6638 truncating or zero extending that value to the size of the integer type.
6639 If ``value`` is smaller than ``ty2`` then a zero extension is done. If
6640 ``value`` is larger than ``ty2`` then a truncation is done. If they are
6641 the same size, then nothing is done (*no-op cast*) other than a type
6647 .. code-block:: llvm
6649 %X = ptrtoint i32* %P to i8 ; yields truncation on 32-bit architecture
6650 %Y = ptrtoint i32* %P to i64 ; yields zero extension on 32-bit architecture
6651 %Z = ptrtoint <4 x i32*> %P to <4 x i64>; yields vector zero extension for a vector of addresses on 32-bit architecture
6655 '``inttoptr .. to``' Instruction
6656 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6663 <result> = inttoptr <ty> <value> to <ty2> ; yields ty2
6668 The '``inttoptr``' instruction converts an integer ``value`` to a
6669 pointer type, ``ty2``.
6674 The '``inttoptr``' instruction takes an :ref:`integer <t_integer>` value to
6675 cast, and a type to cast it to, which must be a :ref:`pointer <t_pointer>`
6681 The '``inttoptr``' instruction converts ``value`` to type ``ty2`` by
6682 applying either a zero extension or a truncation depending on the size
6683 of the integer ``value``. If ``value`` is larger than the size of a
6684 pointer then a truncation is done. If ``value`` is smaller than the size
6685 of a pointer then a zero extension is done. If they are the same size,
6686 nothing is done (*no-op cast*).
6691 .. code-block:: llvm
6693 %X = inttoptr i32 255 to i32* ; yields zero extension on 64-bit architecture
6694 %Y = inttoptr i32 255 to i32* ; yields no-op on 32-bit architecture
6695 %Z = inttoptr i64 0 to i32* ; yields truncation on 32-bit architecture
6696 %Z = inttoptr <4 x i32> %G to <4 x i8*>; yields truncation of vector G to four pointers
6700 '``bitcast .. to``' Instruction
6701 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6708 <result> = bitcast <ty> <value> to <ty2> ; yields ty2
6713 The '``bitcast``' instruction converts ``value`` to type ``ty2`` without
6719 The '``bitcast``' instruction takes a value to cast, which must be a
6720 non-aggregate first class value, and a type to cast it to, which must
6721 also be a non-aggregate :ref:`first class <t_firstclass>` type. The
6722 bit sizes of ``value`` and the destination type, ``ty2``, must be
6723 identical. If the source type is a pointer, the destination type must
6724 also be a pointer of the same size. This instruction supports bitwise
6725 conversion of vectors to integers and to vectors of other types (as
6726 long as they have the same size).
6731 The '``bitcast``' instruction converts ``value`` to type ``ty2``. It
6732 is always a *no-op cast* because no bits change with this
6733 conversion. The conversion is done as if the ``value`` had been stored
6734 to memory and read back as type ``ty2``. Pointer (or vector of
6735 pointers) types may only be converted to other pointer (or vector of
6736 pointers) types with the same address space through this instruction.
6737 To convert pointers to other types, use the :ref:`inttoptr <i_inttoptr>`
6738 or :ref:`ptrtoint <i_ptrtoint>` instructions first.
6743 .. code-block:: llvm
6745 %X = bitcast i8 255 to i8 ; yields i8 :-1
6746 %Y = bitcast i32* %x to sint* ; yields sint*:%x
6747 %Z = bitcast <2 x int> %V to i64; ; yields i64: %V
6748 %Z = bitcast <2 x i32*> %V to <2 x i64*> ; yields <2 x i64*>
6750 .. _i_addrspacecast:
6752 '``addrspacecast .. to``' Instruction
6753 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6760 <result> = addrspacecast <pty> <ptrval> to <pty2> ; yields pty2
6765 The '``addrspacecast``' instruction converts ``ptrval`` from ``pty`` in
6766 address space ``n`` to type ``pty2`` in address space ``m``.
6771 The '``addrspacecast``' instruction takes a pointer or vector of pointer value
6772 to cast and a pointer type to cast it to, which must have a different
6778 The '``addrspacecast``' instruction converts the pointer value
6779 ``ptrval`` to type ``pty2``. It can be a *no-op cast* or a complex
6780 value modification, depending on the target and the address space
6781 pair. Pointer conversions within the same address space must be
6782 performed with the ``bitcast`` instruction. Note that if the address space
6783 conversion is legal then both result and operand refer to the same memory
6789 .. code-block:: llvm
6791 %X = addrspacecast i32* %x to i32 addrspace(1)* ; yields i32 addrspace(1)*:%x
6792 %Y = addrspacecast i32 addrspace(1)* %y to i64 addrspace(2)* ; yields i64 addrspace(2)*:%y
6793 %Z = addrspacecast <4 x i32*> %z to <4 x float addrspace(3)*> ; yields <4 x float addrspace(3)*>:%z
6800 The instructions in this category are the "miscellaneous" instructions,
6801 which defy better classification.
6805 '``icmp``' Instruction
6806 ^^^^^^^^^^^^^^^^^^^^^^
6813 <result> = icmp <cond> <ty> <op1>, <op2> ; yields i1 or <N x i1>:result
6818 The '``icmp``' instruction returns a boolean value or a vector of
6819 boolean values based on comparison of its two integer, integer vector,
6820 pointer, or pointer vector operands.
6825 The '``icmp``' instruction takes three operands. The first operand is
6826 the condition code indicating the kind of comparison to perform. It is
6827 not a value, just a keyword. The possible condition code are:
6830 #. ``ne``: not equal
6831 #. ``ugt``: unsigned greater than
6832 #. ``uge``: unsigned greater or equal
6833 #. ``ult``: unsigned less than
6834 #. ``ule``: unsigned less or equal
6835 #. ``sgt``: signed greater than
6836 #. ``sge``: signed greater or equal
6837 #. ``slt``: signed less than
6838 #. ``sle``: signed less or equal
6840 The remaining two arguments must be :ref:`integer <t_integer>` or
6841 :ref:`pointer <t_pointer>` or integer :ref:`vector <t_vector>` typed. They
6842 must also be identical types.
6847 The '``icmp``' compares ``op1`` and ``op2`` according to the condition
6848 code given as ``cond``. The comparison performed always yields either an
6849 :ref:`i1 <t_integer>` or vector of ``i1`` result, as follows:
6851 #. ``eq``: yields ``true`` if the operands are equal, ``false``
6852 otherwise. No sign interpretation is necessary or performed.
6853 #. ``ne``: yields ``true`` if the operands are unequal, ``false``
6854 otherwise. No sign interpretation is necessary or performed.
6855 #. ``ugt``: interprets the operands as unsigned values and yields
6856 ``true`` if ``op1`` is greater than ``op2``.
6857 #. ``uge``: interprets the operands as unsigned values and yields
6858 ``true`` if ``op1`` is greater than or equal to ``op2``.
6859 #. ``ult``: interprets the operands as unsigned values and yields
6860 ``true`` if ``op1`` is less than ``op2``.
6861 #. ``ule``: interprets the operands as unsigned values and yields
6862 ``true`` if ``op1`` is less than or equal to ``op2``.
6863 #. ``sgt``: interprets the operands as signed values and yields ``true``
6864 if ``op1`` is greater than ``op2``.
6865 #. ``sge``: interprets the operands as signed values and yields ``true``
6866 if ``op1`` is greater than or equal to ``op2``.
6867 #. ``slt``: interprets the operands as signed values and yields ``true``
6868 if ``op1`` is less than ``op2``.
6869 #. ``sle``: interprets the operands as signed values and yields ``true``
6870 if ``op1`` is less than or equal to ``op2``.
6872 If the operands are :ref:`pointer <t_pointer>` typed, the pointer values
6873 are compared as if they were integers.
6875 If the operands are integer vectors, then they are compared element by
6876 element. The result is an ``i1`` vector with the same number of elements
6877 as the values being compared. Otherwise, the result is an ``i1``.
6882 .. code-block:: llvm
6884 <result> = icmp eq i32 4, 5 ; yields: result=false
6885 <result> = icmp ne float* %X, %X ; yields: result=false
6886 <result> = icmp ult i16 4, 5 ; yields: result=true
6887 <result> = icmp sgt i16 4, 5 ; yields: result=false
6888 <result> = icmp ule i16 -4, 5 ; yields: result=false
6889 <result> = icmp sge i16 4, 5 ; yields: result=false
6891 Note that the code generator does not yet support vector types with the
6892 ``icmp`` instruction.
6896 '``fcmp``' Instruction
6897 ^^^^^^^^^^^^^^^^^^^^^^
6904 <result> = fcmp <cond> <ty> <op1>, <op2> ; yields i1 or <N x i1>:result
6909 The '``fcmp``' instruction returns a boolean value or vector of boolean
6910 values based on comparison of its operands.
6912 If the operands are floating point scalars, then the result type is a
6913 boolean (:ref:`i1 <t_integer>`).
6915 If the operands are floating point vectors, then the result type is a
6916 vector of boolean with the same number of elements as the operands being
6922 The '``fcmp``' instruction takes three operands. The first operand is
6923 the condition code indicating the kind of comparison to perform. It is
6924 not a value, just a keyword. The possible condition code are:
6926 #. ``false``: no comparison, always returns false
6927 #. ``oeq``: ordered and equal
6928 #. ``ogt``: ordered and greater than
6929 #. ``oge``: ordered and greater than or equal
6930 #. ``olt``: ordered and less than
6931 #. ``ole``: ordered and less than or equal
6932 #. ``one``: ordered and not equal
6933 #. ``ord``: ordered (no nans)
6934 #. ``ueq``: unordered or equal
6935 #. ``ugt``: unordered or greater than
6936 #. ``uge``: unordered or greater than or equal
6937 #. ``ult``: unordered or less than
6938 #. ``ule``: unordered or less than or equal
6939 #. ``une``: unordered or not equal
6940 #. ``uno``: unordered (either nans)
6941 #. ``true``: no comparison, always returns true
6943 *Ordered* means that neither operand is a QNAN while *unordered* means
6944 that either operand may be a QNAN.
6946 Each of ``val1`` and ``val2`` arguments must be either a :ref:`floating
6947 point <t_floating>` type or a :ref:`vector <t_vector>` of floating point
6948 type. They must have identical types.
6953 The '``fcmp``' instruction compares ``op1`` and ``op2`` according to the
6954 condition code given as ``cond``. If the operands are vectors, then the
6955 vectors are compared element by element. Each comparison performed
6956 always yields an :ref:`i1 <t_integer>` result, as follows:
6958 #. ``false``: always yields ``false``, regardless of operands.
6959 #. ``oeq``: yields ``true`` if both operands are not a QNAN and ``op1``
6960 is equal to ``op2``.
6961 #. ``ogt``: yields ``true`` if both operands are not a QNAN and ``op1``
6962 is greater than ``op2``.
6963 #. ``oge``: yields ``true`` if both operands are not a QNAN and ``op1``
6964 is greater than or equal to ``op2``.
6965 #. ``olt``: yields ``true`` if both operands are not a QNAN and ``op1``
6966 is less than ``op2``.
6967 #. ``ole``: yields ``true`` if both operands are not a QNAN and ``op1``
6968 is less than or equal to ``op2``.
6969 #. ``one``: yields ``true`` if both operands are not a QNAN and ``op1``
6970 is not equal to ``op2``.
6971 #. ``ord``: yields ``true`` if both operands are not a QNAN.
6972 #. ``ueq``: yields ``true`` if either operand is a QNAN or ``op1`` is
6974 #. ``ugt``: yields ``true`` if either operand is a QNAN or ``op1`` is
6975 greater than ``op2``.
6976 #. ``uge``: yields ``true`` if either operand is a QNAN or ``op1`` is
6977 greater than or equal to ``op2``.
6978 #. ``ult``: yields ``true`` if either operand is a QNAN or ``op1`` is
6980 #. ``ule``: yields ``true`` if either operand is a QNAN or ``op1`` is
6981 less than or equal to ``op2``.
6982 #. ``une``: yields ``true`` if either operand is a QNAN or ``op1`` is
6983 not equal to ``op2``.
6984 #. ``uno``: yields ``true`` if either operand is a QNAN.
6985 #. ``true``: always yields ``true``, regardless of operands.
6990 .. code-block:: llvm
6992 <result> = fcmp oeq float 4.0, 5.0 ; yields: result=false
6993 <result> = fcmp one float 4.0, 5.0 ; yields: result=true
6994 <result> = fcmp olt float 4.0, 5.0 ; yields: result=true
6995 <result> = fcmp ueq double 1.0, 2.0 ; yields: result=false
6997 Note that the code generator does not yet support vector types with the
6998 ``fcmp`` instruction.
7002 '``phi``' Instruction
7003 ^^^^^^^^^^^^^^^^^^^^^
7010 <result> = phi <ty> [ <val0>, <label0>], ...
7015 The '``phi``' instruction is used to implement the φ node in the SSA
7016 graph representing the function.
7021 The type of the incoming values is specified with the first type field.
7022 After this, the '``phi``' instruction takes a list of pairs as
7023 arguments, with one pair for each predecessor basic block of the current
7024 block. Only values of :ref:`first class <t_firstclass>` type may be used as
7025 the value arguments to the PHI node. Only labels may be used as the
7028 There must be no non-phi instructions between the start of a basic block
7029 and the PHI instructions: i.e. PHI instructions must be first in a basic
7032 For the purposes of the SSA form, the use of each incoming value is
7033 deemed to occur on the edge from the corresponding predecessor block to
7034 the current block (but after any definition of an '``invoke``'
7035 instruction's return value on the same edge).
7040 At runtime, the '``phi``' instruction logically takes on the value
7041 specified by the pair corresponding to the predecessor basic block that
7042 executed just prior to the current block.
7047 .. code-block:: llvm
7049 Loop: ; Infinite loop that counts from 0 on up...
7050 %indvar = phi i32 [ 0, %LoopHeader ], [ %nextindvar, %Loop ]
7051 %nextindvar = add i32 %indvar, 1
7056 '``select``' Instruction
7057 ^^^^^^^^^^^^^^^^^^^^^^^^
7064 <result> = select selty <cond>, <ty> <val1>, <ty> <val2> ; yields ty
7066 selty is either i1 or {<N x i1>}
7071 The '``select``' instruction is used to choose one value based on a
7072 condition, without IR-level branching.
7077 The '``select``' instruction requires an 'i1' value or a vector of 'i1'
7078 values indicating the condition, and two values of the same :ref:`first
7079 class <t_firstclass>` type.
7084 If the condition is an i1 and it evaluates to 1, the instruction returns
7085 the first value argument; otherwise, it returns the second value
7088 If the condition is a vector of i1, then the value arguments must be
7089 vectors of the same size, and the selection is done element by element.
7091 If the condition is an i1 and the value arguments are vectors of the
7092 same size, then an entire vector is selected.
7097 .. code-block:: llvm
7099 %X = select i1 true, i8 17, i8 42 ; yields i8:17
7103 '``call``' Instruction
7104 ^^^^^^^^^^^^^^^^^^^^^^
7111 <result> = [tail | musttail] call [cconv] [ret attrs] <ty> [<fnty>*] <fnptrval>(<function args>) [fn attrs]
7116 The '``call``' instruction represents a simple function call.
7121 This instruction requires several arguments:
7123 #. The optional ``tail`` and ``musttail`` markers indicate that the optimizers
7124 should perform tail call optimization. The ``tail`` marker is a hint that
7125 `can be ignored <CodeGenerator.html#sibcallopt>`_. The ``musttail`` marker
7126 means that the call must be tail call optimized in order for the program to
7127 be correct. The ``musttail`` marker provides these guarantees:
7129 #. The call will not cause unbounded stack growth if it is part of a
7130 recursive cycle in the call graph.
7131 #. Arguments with the :ref:`inalloca <attr_inalloca>` attribute are
7134 Both markers imply that the callee does not access allocas or varargs from
7135 the caller. Calls marked ``musttail`` must obey the following additional
7138 - The call must immediately precede a :ref:`ret <i_ret>` instruction,
7139 or a pointer bitcast followed by a ret instruction.
7140 - The ret instruction must return the (possibly bitcasted) value
7141 produced by the call or void.
7142 - The caller and callee prototypes must match. Pointer types of
7143 parameters or return types may differ in pointee type, but not
7145 - The calling conventions of the caller and callee must match.
7146 - All ABI-impacting function attributes, such as sret, byval, inreg,
7147 returned, and inalloca, must match.
7148 - The callee must be varargs iff the caller is varargs. Bitcasting a
7149 non-varargs function to the appropriate varargs type is legal so
7150 long as the non-varargs prefixes obey the other rules.
7152 Tail call optimization for calls marked ``tail`` is guaranteed to occur if
7153 the following conditions are met:
7155 - Caller and callee both have the calling convention ``fastcc``.
7156 - The call is in tail position (ret immediately follows call and ret
7157 uses value of call or is void).
7158 - Option ``-tailcallopt`` is enabled, or
7159 ``llvm::GuaranteedTailCallOpt`` is ``true``.
7160 - `Platform-specific constraints are
7161 met. <CodeGenerator.html#tailcallopt>`_
7163 #. The optional "cconv" marker indicates which :ref:`calling
7164 convention <callingconv>` the call should use. If none is
7165 specified, the call defaults to using C calling conventions. The
7166 calling convention of the call must match the calling convention of
7167 the target function, or else the behavior is undefined.
7168 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
7169 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
7171 #. '``ty``': the type of the call instruction itself which is also the
7172 type of the return value. Functions that return no value are marked
7174 #. '``fnty``': shall be the signature of the pointer to function value
7175 being invoked. The argument types must match the types implied by
7176 this signature. This type can be omitted if the function is not
7177 varargs and if the function type does not return a pointer to a
7179 #. '``fnptrval``': An LLVM value containing a pointer to a function to
7180 be invoked. In most cases, this is a direct function invocation, but
7181 indirect ``call``'s are just as possible, calling an arbitrary pointer
7183 #. '``function args``': argument list whose types match the function
7184 signature argument types and parameter attributes. All arguments must
7185 be of :ref:`first class <t_firstclass>` type. If the function signature
7186 indicates the function accepts a variable number of arguments, the
7187 extra arguments can be specified.
7188 #. The optional :ref:`function attributes <fnattrs>` list. Only
7189 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
7190 attributes are valid here.
7195 The '``call``' instruction is used to cause control flow to transfer to
7196 a specified function, with its incoming arguments bound to the specified
7197 values. Upon a '``ret``' instruction in the called function, control
7198 flow continues with the instruction after the function call, and the
7199 return value of the function is bound to the result argument.
7204 .. code-block:: llvm
7206 %retval = call i32 @test(i32 %argc)
7207 call i32 (i8*, ...)* @printf(i8* %msg, i32 12, i8 42) ; yields i32
7208 %X = tail call i32 @foo() ; yields i32
7209 %Y = tail call fastcc i32 @foo() ; yields i32
7210 call void %foo(i8 97 signext)
7212 %struct.A = type { i32, i8 }
7213 %r = call %struct.A @foo() ; yields { i32, i8 }
7214 %gr = extractvalue %struct.A %r, 0 ; yields i32
7215 %gr1 = extractvalue %struct.A %r, 1 ; yields i8
7216 %Z = call void @foo() noreturn ; indicates that %foo never returns normally
7217 %ZZ = call zeroext i32 @bar() ; Return value is %zero extended
7219 llvm treats calls to some functions with names and arguments that match
7220 the standard C99 library as being the C99 library functions, and may
7221 perform optimizations or generate code for them under that assumption.
7222 This is something we'd like to change in the future to provide better
7223 support for freestanding environments and non-C-based languages.
7227 '``va_arg``' Instruction
7228 ^^^^^^^^^^^^^^^^^^^^^^^^
7235 <resultval> = va_arg <va_list*> <arglist>, <argty>
7240 The '``va_arg``' instruction is used to access arguments passed through
7241 the "variable argument" area of a function call. It is used to implement
7242 the ``va_arg`` macro in C.
7247 This instruction takes a ``va_list*`` value and the type of the
7248 argument. It returns a value of the specified argument type and
7249 increments the ``va_list`` to point to the next argument. The actual
7250 type of ``va_list`` is target specific.
7255 The '``va_arg``' instruction loads an argument of the specified type
7256 from the specified ``va_list`` and causes the ``va_list`` to point to
7257 the next argument. For more information, see the variable argument
7258 handling :ref:`Intrinsic Functions <int_varargs>`.
7260 It is legal for this instruction to be called in a function which does
7261 not take a variable number of arguments, for example, the ``vfprintf``
7264 ``va_arg`` is an LLVM instruction instead of an :ref:`intrinsic
7265 function <intrinsics>` because it takes a type as an argument.
7270 See the :ref:`variable argument processing <int_varargs>` section.
7272 Note that the code generator does not yet fully support va\_arg on many
7273 targets. Also, it does not currently support va\_arg with aggregate
7274 types on any target.
7278 '``landingpad``' Instruction
7279 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7286 <resultval> = landingpad <resultty> personality <type> <pers_fn> <clause>+
7287 <resultval> = landingpad <resultty> personality <type> <pers_fn> cleanup <clause>*
7289 <clause> := catch <type> <value>
7290 <clause> := filter <array constant type> <array constant>
7295 The '``landingpad``' instruction is used by `LLVM's exception handling
7296 system <ExceptionHandling.html#overview>`_ to specify that a basic block
7297 is a landing pad --- one where the exception lands, and corresponds to the
7298 code found in the ``catch`` portion of a ``try``/``catch`` sequence. It
7299 defines values supplied by the personality function (``pers_fn``) upon
7300 re-entry to the function. The ``resultval`` has the type ``resultty``.
7305 This instruction takes a ``pers_fn`` value. This is the personality
7306 function associated with the unwinding mechanism. The optional
7307 ``cleanup`` flag indicates that the landing pad block is a cleanup.
7309 A ``clause`` begins with the clause type --- ``catch`` or ``filter`` --- and
7310 contains the global variable representing the "type" that may be caught
7311 or filtered respectively. Unlike the ``catch`` clause, the ``filter``
7312 clause takes an array constant as its argument. Use
7313 "``[0 x i8**] undef``" for a filter which cannot throw. The
7314 '``landingpad``' instruction must contain *at least* one ``clause`` or
7315 the ``cleanup`` flag.
7320 The '``landingpad``' instruction defines the values which are set by the
7321 personality function (``pers_fn``) upon re-entry to the function, and
7322 therefore the "result type" of the ``landingpad`` instruction. As with
7323 calling conventions, how the personality function results are
7324 represented in LLVM IR is target specific.
7326 The clauses are applied in order from top to bottom. If two
7327 ``landingpad`` instructions are merged together through inlining, the
7328 clauses from the calling function are appended to the list of clauses.
7329 When the call stack is being unwound due to an exception being thrown,
7330 the exception is compared against each ``clause`` in turn. If it doesn't
7331 match any of the clauses, and the ``cleanup`` flag is not set, then
7332 unwinding continues further up the call stack.
7334 The ``landingpad`` instruction has several restrictions:
7336 - A landing pad block is a basic block which is the unwind destination
7337 of an '``invoke``' instruction.
7338 - A landing pad block must have a '``landingpad``' instruction as its
7339 first non-PHI instruction.
7340 - There can be only one '``landingpad``' instruction within the landing
7342 - A basic block that is not a landing pad block may not include a
7343 '``landingpad``' instruction.
7344 - All '``landingpad``' instructions in a function must have the same
7345 personality function.
7350 .. code-block:: llvm
7352 ;; A landing pad which can catch an integer.
7353 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
7355 ;; A landing pad that is a cleanup.
7356 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
7358 ;; A landing pad which can catch an integer and can only throw a double.
7359 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
7361 filter [1 x i8**] [@_ZTId]
7368 LLVM supports the notion of an "intrinsic function". These functions
7369 have well known names and semantics and are required to follow certain
7370 restrictions. Overall, these intrinsics represent an extension mechanism
7371 for the LLVM language that does not require changing all of the
7372 transformations in LLVM when adding to the language (or the bitcode
7373 reader/writer, the parser, etc...).
7375 Intrinsic function names must all start with an "``llvm.``" prefix. This
7376 prefix is reserved in LLVM for intrinsic names; thus, function names may
7377 not begin with this prefix. Intrinsic functions must always be external
7378 functions: you cannot define the body of intrinsic functions. Intrinsic
7379 functions may only be used in call or invoke instructions: it is illegal
7380 to take the address of an intrinsic function. Additionally, because
7381 intrinsic functions are part of the LLVM language, it is required if any
7382 are added that they be documented here.
7384 Some intrinsic functions can be overloaded, i.e., the intrinsic
7385 represents a family of functions that perform the same operation but on
7386 different data types. Because LLVM can represent over 8 million
7387 different integer types, overloading is used commonly to allow an
7388 intrinsic function to operate on any integer type. One or more of the
7389 argument types or the result type can be overloaded to accept any
7390 integer type. Argument types may also be defined as exactly matching a
7391 previous argument's type or the result type. This allows an intrinsic
7392 function which accepts multiple arguments, but needs all of them to be
7393 of the same type, to only be overloaded with respect to a single
7394 argument or the result.
7396 Overloaded intrinsics will have the names of its overloaded argument
7397 types encoded into its function name, each preceded by a period. Only
7398 those types which are overloaded result in a name suffix. Arguments
7399 whose type is matched against another type do not. For example, the
7400 ``llvm.ctpop`` function can take an integer of any width and returns an
7401 integer of exactly the same integer width. This leads to a family of
7402 functions such as ``i8 @llvm.ctpop.i8(i8 %val)`` and
7403 ``i29 @llvm.ctpop.i29(i29 %val)``. Only one type, the return type, is
7404 overloaded, and only one type suffix is required. Because the argument's
7405 type is matched against the return type, it does not require its own
7408 To learn how to add an intrinsic function, please see the `Extending
7409 LLVM Guide <ExtendingLLVM.html>`_.
7413 Variable Argument Handling Intrinsics
7414 -------------------------------------
7416 Variable argument support is defined in LLVM with the
7417 :ref:`va_arg <i_va_arg>` instruction and these three intrinsic
7418 functions. These functions are related to the similarly named macros
7419 defined in the ``<stdarg.h>`` header file.
7421 All of these functions operate on arguments that use a target-specific
7422 value type "``va_list``". The LLVM assembly language reference manual
7423 does not define what this type is, so all transformations should be
7424 prepared to handle these functions regardless of the type used.
7426 This example shows how the :ref:`va_arg <i_va_arg>` instruction and the
7427 variable argument handling intrinsic functions are used.
7429 .. code-block:: llvm
7431 ; This struct is different for every platform. For most platforms,
7432 ; it is merely an i8*.
7433 %struct.va_list = type { i8* }
7435 ; For Unix x86_64 platforms, va_list is the following struct:
7436 ; %struct.va_list = type { i32, i32, i8*, i8* }
7438 define i32 @test(i32 %X, ...) {
7439 ; Initialize variable argument processing
7440 %ap = alloca %struct.va_list
7441 %ap2 = bitcast %struct.va_list* %ap to i8*
7442 call void @llvm.va_start(i8* %ap2)
7444 ; Read a single integer argument
7445 %tmp = va_arg i8* %ap2, i32
7447 ; Demonstrate usage of llvm.va_copy and llvm.va_end
7449 %aq2 = bitcast i8** %aq to i8*
7450 call void @llvm.va_copy(i8* %aq2, i8* %ap2)
7451 call void @llvm.va_end(i8* %aq2)
7453 ; Stop processing of arguments.
7454 call void @llvm.va_end(i8* %ap2)
7458 declare void @llvm.va_start(i8*)
7459 declare void @llvm.va_copy(i8*, i8*)
7460 declare void @llvm.va_end(i8*)
7464 '``llvm.va_start``' Intrinsic
7465 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7472 declare void @llvm.va_start(i8* <arglist>)
7477 The '``llvm.va_start``' intrinsic initializes ``*<arglist>`` for
7478 subsequent use by ``va_arg``.
7483 The argument is a pointer to a ``va_list`` element to initialize.
7488 The '``llvm.va_start``' intrinsic works just like the ``va_start`` macro
7489 available in C. In a target-dependent way, it initializes the
7490 ``va_list`` element to which the argument points, so that the next call
7491 to ``va_arg`` will produce the first variable argument passed to the
7492 function. Unlike the C ``va_start`` macro, this intrinsic does not need
7493 to know the last argument of the function as the compiler can figure
7496 '``llvm.va_end``' Intrinsic
7497 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7504 declare void @llvm.va_end(i8* <arglist>)
7509 The '``llvm.va_end``' intrinsic destroys ``*<arglist>``, which has been
7510 initialized previously with ``llvm.va_start`` or ``llvm.va_copy``.
7515 The argument is a pointer to a ``va_list`` to destroy.
7520 The '``llvm.va_end``' intrinsic works just like the ``va_end`` macro
7521 available in C. In a target-dependent way, it destroys the ``va_list``
7522 element to which the argument points. Calls to
7523 :ref:`llvm.va_start <int_va_start>` and
7524 :ref:`llvm.va_copy <int_va_copy>` must be matched exactly with calls to
7529 '``llvm.va_copy``' Intrinsic
7530 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7537 declare void @llvm.va_copy(i8* <destarglist>, i8* <srcarglist>)
7542 The '``llvm.va_copy``' intrinsic copies the current argument position
7543 from the source argument list to the destination argument list.
7548 The first argument is a pointer to a ``va_list`` element to initialize.
7549 The second argument is a pointer to a ``va_list`` element to copy from.
7554 The '``llvm.va_copy``' intrinsic works just like the ``va_copy`` macro
7555 available in C. In a target-dependent way, it copies the source
7556 ``va_list`` element into the destination ``va_list`` element. This
7557 intrinsic is necessary because the `` llvm.va_start`` intrinsic may be
7558 arbitrarily complex and require, for example, memory allocation.
7560 Accurate Garbage Collection Intrinsics
7561 --------------------------------------
7563 LLVM's support for `Accurate Garbage Collection <GarbageCollection.html>`_
7564 (GC) requires the frontend to generate code containing appropriate intrinsic
7565 calls and select an appropriate GC strategy which knows how to lower these
7566 intrinsics in a manner which is appropriate for the target collector.
7568 These intrinsics allow identification of :ref:`GC roots on the
7569 stack <int_gcroot>`, as well as garbage collector implementations that
7570 require :ref:`read <int_gcread>` and :ref:`write <int_gcwrite>` barriers.
7571 Frontends for type-safe garbage collected languages should generate
7572 these intrinsics to make use of the LLVM garbage collectors. For more
7573 details, see `Garbage Collection with LLVM <GarbageCollection.html>`_.
7575 Experimental Statepoint Intrinsics
7576 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7578 LLVM provides an second experimental set of intrinsics for describing garbage
7579 collection safepoints in compiled code. These intrinsics are an alternative
7580 to the ``llvm.gcroot`` intrinsics, but are compatible with the ones for
7581 :ref:`read <int_gcread>` and :ref:`write <int_gcwrite>` barriers. The
7582 differences in approach are covered in the `Garbage Collection with LLVM
7583 <GarbageCollection.html>`_ documentation. The intrinsics themselves are
7584 described in :doc:`Statepoints`.
7588 '``llvm.gcroot``' Intrinsic
7589 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7596 declare void @llvm.gcroot(i8** %ptrloc, i8* %metadata)
7601 The '``llvm.gcroot``' intrinsic declares the existence of a GC root to
7602 the code generator, and allows some metadata to be associated with it.
7607 The first argument specifies the address of a stack object that contains
7608 the root pointer. The second pointer (which must be either a constant or
7609 a global value address) contains the meta-data to be associated with the
7615 At runtime, a call to this intrinsic stores a null pointer into the
7616 "ptrloc" location. At compile-time, the code generator generates
7617 information to allow the runtime to find the pointer at GC safe points.
7618 The '``llvm.gcroot``' intrinsic may only be used in a function which
7619 :ref:`specifies a GC algorithm <gc>`.
7623 '``llvm.gcread``' Intrinsic
7624 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7631 declare i8* @llvm.gcread(i8* %ObjPtr, i8** %Ptr)
7636 The '``llvm.gcread``' intrinsic identifies reads of references from heap
7637 locations, allowing garbage collector implementations that require read
7643 The second argument is the address to read from, which should be an
7644 address allocated from the garbage collector. The first object is a
7645 pointer to the start of the referenced object, if needed by the language
7646 runtime (otherwise null).
7651 The '``llvm.gcread``' intrinsic has the same semantics as a load
7652 instruction, but may be replaced with substantially more complex code by
7653 the garbage collector runtime, as needed. The '``llvm.gcread``'
7654 intrinsic may only be used in a function which :ref:`specifies a GC
7659 '``llvm.gcwrite``' Intrinsic
7660 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7667 declare void @llvm.gcwrite(i8* %P1, i8* %Obj, i8** %P2)
7672 The '``llvm.gcwrite``' intrinsic identifies writes of references to heap
7673 locations, allowing garbage collector implementations that require write
7674 barriers (such as generational or reference counting collectors).
7679 The first argument is the reference to store, the second is the start of
7680 the object to store it to, and the third is the address of the field of
7681 Obj to store to. If the runtime does not require a pointer to the
7682 object, Obj may be null.
7687 The '``llvm.gcwrite``' intrinsic has the same semantics as a store
7688 instruction, but may be replaced with substantially more complex code by
7689 the garbage collector runtime, as needed. The '``llvm.gcwrite``'
7690 intrinsic may only be used in a function which :ref:`specifies a GC
7693 Code Generator Intrinsics
7694 -------------------------
7696 These intrinsics are provided by LLVM to expose special features that
7697 may only be implemented with code generator support.
7699 '``llvm.returnaddress``' Intrinsic
7700 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7707 declare i8 *@llvm.returnaddress(i32 <level>)
7712 The '``llvm.returnaddress``' intrinsic attempts to compute a
7713 target-specific value indicating the return address of the current
7714 function or one of its callers.
7719 The argument to this intrinsic indicates which function to return the
7720 address for. Zero indicates the calling function, one indicates its
7721 caller, etc. The argument is **required** to be a constant integer
7727 The '``llvm.returnaddress``' intrinsic either returns a pointer
7728 indicating the return address of the specified call frame, or zero if it
7729 cannot be identified. The value returned by this intrinsic is likely to
7730 be incorrect or 0 for arguments other than zero, so it should only be
7731 used for debugging purposes.
7733 Note that calling this intrinsic does not prevent function inlining or
7734 other aggressive transformations, so the value returned may not be that
7735 of the obvious source-language caller.
7737 '``llvm.frameaddress``' Intrinsic
7738 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7745 declare i8* @llvm.frameaddress(i32 <level>)
7750 The '``llvm.frameaddress``' intrinsic attempts to return the
7751 target-specific frame pointer value for the specified stack frame.
7756 The argument to this intrinsic indicates which function to return the
7757 frame pointer for. Zero indicates the calling function, one indicates
7758 its caller, etc. The argument is **required** to be a constant integer
7764 The '``llvm.frameaddress``' intrinsic either returns a pointer
7765 indicating the frame address of the specified call frame, or zero if it
7766 cannot be identified. The value returned by this intrinsic is likely to
7767 be incorrect or 0 for arguments other than zero, so it should only be
7768 used for debugging purposes.
7770 Note that calling this intrinsic does not prevent function inlining or
7771 other aggressive transformations, so the value returned may not be that
7772 of the obvious source-language caller.
7774 '``llvm.frameescape``' and '``llvm.framerecover``' Intrinsics
7775 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7782 declare void @llvm.frameescape(...)
7783 declare i8* @llvm.framerecover(i8* %func, i8* %fp, i32 %idx)
7788 The '``llvm.frameescape``' intrinsic escapes offsets of a collection of static
7789 allocas, and the '``llvm.framerecover``' intrinsic applies those offsets to a
7790 live frame pointer to recover the address of the allocation. The offset is
7791 computed during frame layout of the caller of ``llvm.frameescape``.
7796 All arguments to '``llvm.frameescape``' must be pointers to static allocas or
7797 casts of static allocas. Each function can only call '``llvm.frameescape``'
7798 once, and it can only do so from the entry block.
7800 The ``func`` argument to '``llvm.framerecover``' must be a constant
7801 bitcasted pointer to a function defined in the current module. The code
7802 generator cannot determine the frame allocation offset of functions defined in
7805 The ``fp`` argument to '``llvm.framerecover``' must be a frame
7806 pointer of a call frame that is currently live. The return value of
7807 '``llvm.frameaddress``' is one way to produce such a value, but most platforms
7808 also expose the frame pointer through stack unwinding mechanisms.
7810 The ``idx`` argument to '``llvm.framerecover``' indicates which alloca passed to
7811 '``llvm.frameescape``' to recover. It is zero-indexed.
7816 These intrinsics allow a group of functions to access one stack memory
7817 allocation in an ancestor stack frame. The memory returned from
7818 '``llvm.frameallocate``' may be allocated prior to stack realignment, so the
7819 memory is only aligned to the ABI-required stack alignment. Each function may
7820 only call '``llvm.frameallocate``' one or zero times from the function entry
7821 block. The frame allocation intrinsic inhibits inlining, as any frame
7822 allocations in the inlined function frame are likely to be at a different
7823 offset from the one used by '``llvm.framerecover``' called with the
7826 .. _int_read_register:
7827 .. _int_write_register:
7829 '``llvm.read_register``' and '``llvm.write_register``' Intrinsics
7830 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7837 declare i32 @llvm.read_register.i32(metadata)
7838 declare i64 @llvm.read_register.i64(metadata)
7839 declare void @llvm.write_register.i32(metadata, i32 @value)
7840 declare void @llvm.write_register.i64(metadata, i64 @value)
7846 The '``llvm.read_register``' and '``llvm.write_register``' intrinsics
7847 provides access to the named register. The register must be valid on
7848 the architecture being compiled to. The type needs to be compatible
7849 with the register being read.
7854 The '``llvm.read_register``' intrinsic returns the current value of the
7855 register, where possible. The '``llvm.write_register``' intrinsic sets
7856 the current value of the register, where possible.
7858 This is useful to implement named register global variables that need
7859 to always be mapped to a specific register, as is common practice on
7860 bare-metal programs including OS kernels.
7862 The compiler doesn't check for register availability or use of the used
7863 register in surrounding code, including inline assembly. Because of that,
7864 allocatable registers are not supported.
7866 Warning: So far it only works with the stack pointer on selected
7867 architectures (ARM, AArch64, PowerPC and x86_64). Significant amount of
7868 work is needed to support other registers and even more so, allocatable
7873 '``llvm.stacksave``' Intrinsic
7874 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7881 declare i8* @llvm.stacksave()
7886 The '``llvm.stacksave``' intrinsic is used to remember the current state
7887 of the function stack, for use with
7888 :ref:`llvm.stackrestore <int_stackrestore>`. This is useful for
7889 implementing language features like scoped automatic variable sized
7895 This intrinsic returns a opaque pointer value that can be passed to
7896 :ref:`llvm.stackrestore <int_stackrestore>`. When an
7897 ``llvm.stackrestore`` intrinsic is executed with a value saved from
7898 ``llvm.stacksave``, it effectively restores the state of the stack to
7899 the state it was in when the ``llvm.stacksave`` intrinsic executed. In
7900 practice, this pops any :ref:`alloca <i_alloca>` blocks from the stack that
7901 were allocated after the ``llvm.stacksave`` was executed.
7903 .. _int_stackrestore:
7905 '``llvm.stackrestore``' Intrinsic
7906 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7913 declare void @llvm.stackrestore(i8* %ptr)
7918 The '``llvm.stackrestore``' intrinsic is used to restore the state of
7919 the function stack to the state it was in when the corresponding
7920 :ref:`llvm.stacksave <int_stacksave>` intrinsic executed. This is
7921 useful for implementing language features like scoped automatic variable
7922 sized arrays in C99.
7927 See the description for :ref:`llvm.stacksave <int_stacksave>`.
7929 '``llvm.prefetch``' Intrinsic
7930 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7937 declare void @llvm.prefetch(i8* <address>, i32 <rw>, i32 <locality>, i32 <cache type>)
7942 The '``llvm.prefetch``' intrinsic is a hint to the code generator to
7943 insert a prefetch instruction if supported; otherwise, it is a noop.
7944 Prefetches have no effect on the behavior of the program but can change
7945 its performance characteristics.
7950 ``address`` is the address to be prefetched, ``rw`` is the specifier
7951 determining if the fetch should be for a read (0) or write (1), and
7952 ``locality`` is a temporal locality specifier ranging from (0) - no
7953 locality, to (3) - extremely local keep in cache. The ``cache type``
7954 specifies whether the prefetch is performed on the data (1) or
7955 instruction (0) cache. The ``rw``, ``locality`` and ``cache type``
7956 arguments must be constant integers.
7961 This intrinsic does not modify the behavior of the program. In
7962 particular, prefetches cannot trap and do not produce a value. On
7963 targets that support this intrinsic, the prefetch can provide hints to
7964 the processor cache for better performance.
7966 '``llvm.pcmarker``' Intrinsic
7967 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7974 declare void @llvm.pcmarker(i32 <id>)
7979 The '``llvm.pcmarker``' intrinsic is a method to export a Program
7980 Counter (PC) in a region of code to simulators and other tools. The
7981 method is target specific, but it is expected that the marker will use
7982 exported symbols to transmit the PC of the marker. The marker makes no
7983 guarantees that it will remain with any specific instruction after
7984 optimizations. It is possible that the presence of a marker will inhibit
7985 optimizations. The intended use is to be inserted after optimizations to
7986 allow correlations of simulation runs.
7991 ``id`` is a numerical id identifying the marker.
7996 This intrinsic does not modify the behavior of the program. Backends
7997 that do not support this intrinsic may ignore it.
7999 '``llvm.readcyclecounter``' Intrinsic
8000 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8007 declare i64 @llvm.readcyclecounter()
8012 The '``llvm.readcyclecounter``' intrinsic provides access to the cycle
8013 counter register (or similar low latency, high accuracy clocks) on those
8014 targets that support it. On X86, it should map to RDTSC. On Alpha, it
8015 should map to RPCC. As the backing counters overflow quickly (on the
8016 order of 9 seconds on alpha), this should only be used for small
8022 When directly supported, reading the cycle counter should not modify any
8023 memory. Implementations are allowed to either return a application
8024 specific value or a system wide value. On backends without support, this
8025 is lowered to a constant 0.
8027 Note that runtime support may be conditional on the privilege-level code is
8028 running at and the host platform.
8030 '``llvm.clear_cache``' Intrinsic
8031 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8038 declare void @llvm.clear_cache(i8*, i8*)
8043 The '``llvm.clear_cache``' intrinsic ensures visibility of modifications
8044 in the specified range to the execution unit of the processor. On
8045 targets with non-unified instruction and data cache, the implementation
8046 flushes the instruction cache.
8051 On platforms with coherent instruction and data caches (e.g. x86), this
8052 intrinsic is a nop. On platforms with non-coherent instruction and data
8053 cache (e.g. ARM, MIPS), the intrinsic is lowered either to appropriate
8054 instructions or a system call, if cache flushing requires special
8057 The default behavior is to emit a call to ``__clear_cache`` from the run
8060 This instrinsic does *not* empty the instruction pipeline. Modifications
8061 of the current function are outside the scope of the intrinsic.
8063 '``llvm.instrprof_increment``' Intrinsic
8064 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8071 declare void @llvm.instrprof_increment(i8* <name>, i64 <hash>,
8072 i32 <num-counters>, i32 <index>)
8077 The '``llvm.instrprof_increment``' intrinsic can be emitted by a
8078 frontend for use with instrumentation based profiling. These will be
8079 lowered by the ``-instrprof`` pass to generate execution counts of a
8085 The first argument is a pointer to a global variable containing the
8086 name of the entity being instrumented. This should generally be the
8087 (mangled) function name for a set of counters.
8089 The second argument is a hash value that can be used by the consumer
8090 of the profile data to detect changes to the instrumented source, and
8091 the third is the number of counters associated with ``name``. It is an
8092 error if ``hash`` or ``num-counters`` differ between two instances of
8093 ``instrprof_increment`` that refer to the same name.
8095 The last argument refers to which of the counters for ``name`` should
8096 be incremented. It should be a value between 0 and ``num-counters``.
8101 This intrinsic represents an increment of a profiling counter. It will
8102 cause the ``-instrprof`` pass to generate the appropriate data
8103 structures and the code to increment the appropriate value, in a
8104 format that can be written out by a compiler runtime and consumed via
8105 the ``llvm-profdata`` tool.
8107 Standard C Library Intrinsics
8108 -----------------------------
8110 LLVM provides intrinsics for a few important standard C library
8111 functions. These intrinsics allow source-language front-ends to pass
8112 information about the alignment of the pointer arguments to the code
8113 generator, providing opportunity for more efficient code generation.
8117 '``llvm.memcpy``' Intrinsic
8118 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8123 This is an overloaded intrinsic. You can use ``llvm.memcpy`` on any
8124 integer bit width and for different address spaces. Not all targets
8125 support all bit widths however.
8129 declare void @llvm.memcpy.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
8130 i32 <len>, i32 <align>, i1 <isvolatile>)
8131 declare void @llvm.memcpy.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
8132 i64 <len>, i32 <align>, i1 <isvolatile>)
8137 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
8138 source location to the destination location.
8140 Note that, unlike the standard libc function, the ``llvm.memcpy.*``
8141 intrinsics do not return a value, takes extra alignment/isvolatile
8142 arguments and the pointers can be in specified address spaces.
8147 The first argument is a pointer to the destination, the second is a
8148 pointer to the source. The third argument is an integer argument
8149 specifying the number of bytes to copy, the fourth argument is the
8150 alignment of the source and destination locations, and the fifth is a
8151 boolean indicating a volatile access.
8153 If the call to this intrinsic has an alignment value that is not 0 or 1,
8154 then the caller guarantees that both the source and destination pointers
8155 are aligned to that boundary.
8157 If the ``isvolatile`` parameter is ``true``, the ``llvm.memcpy`` call is
8158 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
8159 very cleanly specified and it is unwise to depend on it.
8164 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
8165 source location to the destination location, which are not allowed to
8166 overlap. It copies "len" bytes of memory over. If the argument is known
8167 to be aligned to some boundary, this can be specified as the fourth
8168 argument, otherwise it should be set to 0 or 1 (both meaning no alignment).
8170 '``llvm.memmove``' Intrinsic
8171 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8176 This is an overloaded intrinsic. You can use llvm.memmove on any integer
8177 bit width and for different address space. Not all targets support all
8182 declare void @llvm.memmove.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
8183 i32 <len>, i32 <align>, i1 <isvolatile>)
8184 declare void @llvm.memmove.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
8185 i64 <len>, i32 <align>, i1 <isvolatile>)
8190 The '``llvm.memmove.*``' intrinsics move a block of memory from the
8191 source location to the destination location. It is similar to the
8192 '``llvm.memcpy``' intrinsic but allows the two memory locations to
8195 Note that, unlike the standard libc function, the ``llvm.memmove.*``
8196 intrinsics do not return a value, takes extra alignment/isvolatile
8197 arguments and the pointers can be in specified address spaces.
8202 The first argument is a pointer to the destination, the second is a
8203 pointer to the source. The third argument is an integer argument
8204 specifying the number of bytes to copy, the fourth argument is the
8205 alignment of the source and destination locations, and the fifth is a
8206 boolean indicating a volatile access.
8208 If the call to this intrinsic has an alignment value that is not 0 or 1,
8209 then the caller guarantees that the source and destination pointers are
8210 aligned to that boundary.
8212 If the ``isvolatile`` parameter is ``true``, the ``llvm.memmove`` call
8213 is a :ref:`volatile operation <volatile>`. The detailed access behavior is
8214 not very cleanly specified and it is unwise to depend on it.
8219 The '``llvm.memmove.*``' intrinsics copy a block of memory from the
8220 source location to the destination location, which may overlap. It
8221 copies "len" bytes of memory over. If the argument is known to be
8222 aligned to some boundary, this can be specified as the fourth argument,
8223 otherwise it should be set to 0 or 1 (both meaning no alignment).
8225 '``llvm.memset.*``' Intrinsics
8226 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8231 This is an overloaded intrinsic. You can use llvm.memset on any integer
8232 bit width and for different address spaces. However, not all targets
8233 support all bit widths.
8237 declare void @llvm.memset.p0i8.i32(i8* <dest>, i8 <val>,
8238 i32 <len>, i32 <align>, i1 <isvolatile>)
8239 declare void @llvm.memset.p0i8.i64(i8* <dest>, i8 <val>,
8240 i64 <len>, i32 <align>, i1 <isvolatile>)
8245 The '``llvm.memset.*``' intrinsics fill a block of memory with a
8246 particular byte value.
8248 Note that, unlike the standard libc function, the ``llvm.memset``
8249 intrinsic does not return a value and takes extra alignment/volatile
8250 arguments. Also, the destination can be in an arbitrary address space.
8255 The first argument is a pointer to the destination to fill, the second
8256 is the byte value with which to fill it, the third argument is an
8257 integer argument specifying the number of bytes to fill, and the fourth
8258 argument is the known alignment of the destination location.
8260 If the call to this intrinsic has an alignment value that is not 0 or 1,
8261 then the caller guarantees that the destination pointer is aligned to
8264 If the ``isvolatile`` parameter is ``true``, the ``llvm.memset`` call is
8265 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
8266 very cleanly specified and it is unwise to depend on it.
8271 The '``llvm.memset.*``' intrinsics fill "len" bytes of memory starting
8272 at the destination location. If the argument is known to be aligned to
8273 some boundary, this can be specified as the fourth argument, otherwise
8274 it should be set to 0 or 1 (both meaning no alignment).
8276 '``llvm.sqrt.*``' Intrinsic
8277 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8282 This is an overloaded intrinsic. You can use ``llvm.sqrt`` on any
8283 floating point or vector of floating point type. Not all targets support
8288 declare float @llvm.sqrt.f32(float %Val)
8289 declare double @llvm.sqrt.f64(double %Val)
8290 declare x86_fp80 @llvm.sqrt.f80(x86_fp80 %Val)
8291 declare fp128 @llvm.sqrt.f128(fp128 %Val)
8292 declare ppc_fp128 @llvm.sqrt.ppcf128(ppc_fp128 %Val)
8297 The '``llvm.sqrt``' intrinsics return the sqrt of the specified operand,
8298 returning the same value as the libm '``sqrt``' functions would. Unlike
8299 ``sqrt`` in libm, however, ``llvm.sqrt`` has undefined behavior for
8300 negative numbers other than -0.0 (which allows for better optimization,
8301 because there is no need to worry about errno being set).
8302 ``llvm.sqrt(-0.0)`` is defined to return -0.0 like IEEE sqrt.
8307 The argument and return value are floating point numbers of the same
8313 This function returns the sqrt of the specified operand if it is a
8314 nonnegative floating point number.
8316 '``llvm.powi.*``' Intrinsic
8317 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8322 This is an overloaded intrinsic. You can use ``llvm.powi`` on any
8323 floating point or vector of floating point type. Not all targets support
8328 declare float @llvm.powi.f32(float %Val, i32 %power)
8329 declare double @llvm.powi.f64(double %Val, i32 %power)
8330 declare x86_fp80 @llvm.powi.f80(x86_fp80 %Val, i32 %power)
8331 declare fp128 @llvm.powi.f128(fp128 %Val, i32 %power)
8332 declare ppc_fp128 @llvm.powi.ppcf128(ppc_fp128 %Val, i32 %power)
8337 The '``llvm.powi.*``' intrinsics return the first operand raised to the
8338 specified (positive or negative) power. The order of evaluation of
8339 multiplications is not defined. When a vector of floating point type is
8340 used, the second argument remains a scalar integer value.
8345 The second argument is an integer power, and the first is a value to
8346 raise to that power.
8351 This function returns the first value raised to the second power with an
8352 unspecified sequence of rounding operations.
8354 '``llvm.sin.*``' Intrinsic
8355 ^^^^^^^^^^^^^^^^^^^^^^^^^^
8360 This is an overloaded intrinsic. You can use ``llvm.sin`` on any
8361 floating point or vector of floating point type. Not all targets support
8366 declare float @llvm.sin.f32(float %Val)
8367 declare double @llvm.sin.f64(double %Val)
8368 declare x86_fp80 @llvm.sin.f80(x86_fp80 %Val)
8369 declare fp128 @llvm.sin.f128(fp128 %Val)
8370 declare ppc_fp128 @llvm.sin.ppcf128(ppc_fp128 %Val)
8375 The '``llvm.sin.*``' intrinsics return the sine of the operand.
8380 The argument and return value are floating point numbers of the same
8386 This function returns the sine of the specified operand, returning the
8387 same values as the libm ``sin`` functions would, and handles error
8388 conditions in the same way.
8390 '``llvm.cos.*``' Intrinsic
8391 ^^^^^^^^^^^^^^^^^^^^^^^^^^
8396 This is an overloaded intrinsic. You can use ``llvm.cos`` on any
8397 floating point or vector of floating point type. Not all targets support
8402 declare float @llvm.cos.f32(float %Val)
8403 declare double @llvm.cos.f64(double %Val)
8404 declare x86_fp80 @llvm.cos.f80(x86_fp80 %Val)
8405 declare fp128 @llvm.cos.f128(fp128 %Val)
8406 declare ppc_fp128 @llvm.cos.ppcf128(ppc_fp128 %Val)
8411 The '``llvm.cos.*``' intrinsics return the cosine of the operand.
8416 The argument and return value are floating point numbers of the same
8422 This function returns the cosine of the specified operand, returning the
8423 same values as the libm ``cos`` functions would, and handles error
8424 conditions in the same way.
8426 '``llvm.pow.*``' Intrinsic
8427 ^^^^^^^^^^^^^^^^^^^^^^^^^^
8432 This is an overloaded intrinsic. You can use ``llvm.pow`` on any
8433 floating point or vector of floating point type. Not all targets support
8438 declare float @llvm.pow.f32(float %Val, float %Power)
8439 declare double @llvm.pow.f64(double %Val, double %Power)
8440 declare x86_fp80 @llvm.pow.f80(x86_fp80 %Val, x86_fp80 %Power)
8441 declare fp128 @llvm.pow.f128(fp128 %Val, fp128 %Power)
8442 declare ppc_fp128 @llvm.pow.ppcf128(ppc_fp128 %Val, ppc_fp128 Power)
8447 The '``llvm.pow.*``' intrinsics return the first operand raised to the
8448 specified (positive or negative) power.
8453 The second argument is a floating point power, and the first is a value
8454 to raise to that power.
8459 This function returns the first value raised to the second power,
8460 returning the same values as the libm ``pow`` functions would, and
8461 handles error conditions in the same way.
8463 '``llvm.exp.*``' Intrinsic
8464 ^^^^^^^^^^^^^^^^^^^^^^^^^^
8469 This is an overloaded intrinsic. You can use ``llvm.exp`` on any
8470 floating point or vector of floating point type. Not all targets support
8475 declare float @llvm.exp.f32(float %Val)
8476 declare double @llvm.exp.f64(double %Val)
8477 declare x86_fp80 @llvm.exp.f80(x86_fp80 %Val)
8478 declare fp128 @llvm.exp.f128(fp128 %Val)
8479 declare ppc_fp128 @llvm.exp.ppcf128(ppc_fp128 %Val)
8484 The '``llvm.exp.*``' intrinsics perform the exp function.
8489 The argument and return value are floating point numbers of the same
8495 This function returns the same values as the libm ``exp`` functions
8496 would, and handles error conditions in the same way.
8498 '``llvm.exp2.*``' Intrinsic
8499 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8504 This is an overloaded intrinsic. You can use ``llvm.exp2`` on any
8505 floating point or vector of floating point type. Not all targets support
8510 declare float @llvm.exp2.f32(float %Val)
8511 declare double @llvm.exp2.f64(double %Val)
8512 declare x86_fp80 @llvm.exp2.f80(x86_fp80 %Val)
8513 declare fp128 @llvm.exp2.f128(fp128 %Val)
8514 declare ppc_fp128 @llvm.exp2.ppcf128(ppc_fp128 %Val)
8519 The '``llvm.exp2.*``' intrinsics perform the exp2 function.
8524 The argument and return value are floating point numbers of the same
8530 This function returns the same values as the libm ``exp2`` functions
8531 would, and handles error conditions in the same way.
8533 '``llvm.log.*``' Intrinsic
8534 ^^^^^^^^^^^^^^^^^^^^^^^^^^
8539 This is an overloaded intrinsic. You can use ``llvm.log`` on any
8540 floating point or vector of floating point type. Not all targets support
8545 declare float @llvm.log.f32(float %Val)
8546 declare double @llvm.log.f64(double %Val)
8547 declare x86_fp80 @llvm.log.f80(x86_fp80 %Val)
8548 declare fp128 @llvm.log.f128(fp128 %Val)
8549 declare ppc_fp128 @llvm.log.ppcf128(ppc_fp128 %Val)
8554 The '``llvm.log.*``' intrinsics perform the log function.
8559 The argument and return value are floating point numbers of the same
8565 This function returns the same values as the libm ``log`` functions
8566 would, and handles error conditions in the same way.
8568 '``llvm.log10.*``' Intrinsic
8569 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8574 This is an overloaded intrinsic. You can use ``llvm.log10`` on any
8575 floating point or vector of floating point type. Not all targets support
8580 declare float @llvm.log10.f32(float %Val)
8581 declare double @llvm.log10.f64(double %Val)
8582 declare x86_fp80 @llvm.log10.f80(x86_fp80 %Val)
8583 declare fp128 @llvm.log10.f128(fp128 %Val)
8584 declare ppc_fp128 @llvm.log10.ppcf128(ppc_fp128 %Val)
8589 The '``llvm.log10.*``' intrinsics perform the log10 function.
8594 The argument and return value are floating point numbers of the same
8600 This function returns the same values as the libm ``log10`` functions
8601 would, and handles error conditions in the same way.
8603 '``llvm.log2.*``' Intrinsic
8604 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8609 This is an overloaded intrinsic. You can use ``llvm.log2`` on any
8610 floating point or vector of floating point type. Not all targets support
8615 declare float @llvm.log2.f32(float %Val)
8616 declare double @llvm.log2.f64(double %Val)
8617 declare x86_fp80 @llvm.log2.f80(x86_fp80 %Val)
8618 declare fp128 @llvm.log2.f128(fp128 %Val)
8619 declare ppc_fp128 @llvm.log2.ppcf128(ppc_fp128 %Val)
8624 The '``llvm.log2.*``' intrinsics perform the log2 function.
8629 The argument and return value are floating point numbers of the same
8635 This function returns the same values as the libm ``log2`` functions
8636 would, and handles error conditions in the same way.
8638 '``llvm.fma.*``' Intrinsic
8639 ^^^^^^^^^^^^^^^^^^^^^^^^^^
8644 This is an overloaded intrinsic. You can use ``llvm.fma`` on any
8645 floating point or vector of floating point type. Not all targets support
8650 declare float @llvm.fma.f32(float %a, float %b, float %c)
8651 declare double @llvm.fma.f64(double %a, double %b, double %c)
8652 declare x86_fp80 @llvm.fma.f80(x86_fp80 %a, x86_fp80 %b, x86_fp80 %c)
8653 declare fp128 @llvm.fma.f128(fp128 %a, fp128 %b, fp128 %c)
8654 declare ppc_fp128 @llvm.fma.ppcf128(ppc_fp128 %a, ppc_fp128 %b, ppc_fp128 %c)
8659 The '``llvm.fma.*``' intrinsics perform the fused multiply-add
8665 The argument and return value are floating point numbers of the same
8671 This function returns the same values as the libm ``fma`` functions
8672 would, and does not set errno.
8674 '``llvm.fabs.*``' Intrinsic
8675 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8680 This is an overloaded intrinsic. You can use ``llvm.fabs`` on any
8681 floating point or vector of floating point type. Not all targets support
8686 declare float @llvm.fabs.f32(float %Val)
8687 declare double @llvm.fabs.f64(double %Val)
8688 declare x86_fp80 @llvm.fabs.f80(x86_fp80 %Val)
8689 declare fp128 @llvm.fabs.f128(fp128 %Val)
8690 declare ppc_fp128 @llvm.fabs.ppcf128(ppc_fp128 %Val)
8695 The '``llvm.fabs.*``' intrinsics return the absolute value of the
8701 The argument and return value are floating point numbers of the same
8707 This function returns the same values as the libm ``fabs`` functions
8708 would, and handles error conditions in the same way.
8710 '``llvm.minnum.*``' Intrinsic
8711 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8716 This is an overloaded intrinsic. You can use ``llvm.minnum`` on any
8717 floating point or vector of floating point type. Not all targets support
8722 declare float @llvm.minnum.f32(float %Val0, float %Val1)
8723 declare double @llvm.minnum.f64(double %Val0, double %Val1)
8724 declare x86_fp80 @llvm.minnum.f80(x86_fp80 %Val0, x86_fp80 %Val1)
8725 declare fp128 @llvm.minnum.f128(fp128 %Val0, fp128 %Val1)
8726 declare ppc_fp128 @llvm.minnum.ppcf128(ppc_fp128 %Val0, ppc_fp128 %Val1)
8731 The '``llvm.minnum.*``' intrinsics return the minimum of the two
8738 The arguments and return value are floating point numbers of the same
8744 Follows the IEEE-754 semantics for minNum, which also match for libm's
8747 If either operand is a NaN, returns the other non-NaN operand. Returns
8748 NaN only if both operands are NaN. If the operands compare equal,
8749 returns a value that compares equal to both operands. This means that
8750 fmin(+/-0.0, +/-0.0) could return either -0.0 or 0.0.
8752 '``llvm.maxnum.*``' Intrinsic
8753 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8758 This is an overloaded intrinsic. You can use ``llvm.maxnum`` on any
8759 floating point or vector of floating point type. Not all targets support
8764 declare float @llvm.maxnum.f32(float %Val0, float %Val1l)
8765 declare double @llvm.maxnum.f64(double %Val0, double %Val1)
8766 declare x86_fp80 @llvm.maxnum.f80(x86_fp80 %Val0, x86_fp80 %Val1)
8767 declare fp128 @llvm.maxnum.f128(fp128 %Val0, fp128 %Val1)
8768 declare ppc_fp128 @llvm.maxnum.ppcf128(ppc_fp128 %Val0, ppc_fp128 %Val1)
8773 The '``llvm.maxnum.*``' intrinsics return the maximum of the two
8780 The arguments and return value are floating point numbers of the same
8785 Follows the IEEE-754 semantics for maxNum, which also match for libm's
8788 If either operand is a NaN, returns the other non-NaN operand. Returns
8789 NaN only if both operands are NaN. If the operands compare equal,
8790 returns a value that compares equal to both operands. This means that
8791 fmax(+/-0.0, +/-0.0) could return either -0.0 or 0.0.
8793 '``llvm.copysign.*``' Intrinsic
8794 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8799 This is an overloaded intrinsic. You can use ``llvm.copysign`` on any
8800 floating point or vector of floating point type. Not all targets support
8805 declare float @llvm.copysign.f32(float %Mag, float %Sgn)
8806 declare double @llvm.copysign.f64(double %Mag, double %Sgn)
8807 declare x86_fp80 @llvm.copysign.f80(x86_fp80 %Mag, x86_fp80 %Sgn)
8808 declare fp128 @llvm.copysign.f128(fp128 %Mag, fp128 %Sgn)
8809 declare ppc_fp128 @llvm.copysign.ppcf128(ppc_fp128 %Mag, ppc_fp128 %Sgn)
8814 The '``llvm.copysign.*``' intrinsics return a value with the magnitude of the
8815 first operand and the sign of the second operand.
8820 The arguments and return value are floating point numbers of the same
8826 This function returns the same values as the libm ``copysign``
8827 functions would, and handles error conditions in the same way.
8829 '``llvm.floor.*``' Intrinsic
8830 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8835 This is an overloaded intrinsic. You can use ``llvm.floor`` on any
8836 floating point or vector of floating point type. Not all targets support
8841 declare float @llvm.floor.f32(float %Val)
8842 declare double @llvm.floor.f64(double %Val)
8843 declare x86_fp80 @llvm.floor.f80(x86_fp80 %Val)
8844 declare fp128 @llvm.floor.f128(fp128 %Val)
8845 declare ppc_fp128 @llvm.floor.ppcf128(ppc_fp128 %Val)
8850 The '``llvm.floor.*``' intrinsics return the floor of the operand.
8855 The argument and return value are floating point numbers of the same
8861 This function returns the same values as the libm ``floor`` functions
8862 would, and handles error conditions in the same way.
8864 '``llvm.ceil.*``' Intrinsic
8865 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8870 This is an overloaded intrinsic. You can use ``llvm.ceil`` on any
8871 floating point or vector of floating point type. Not all targets support
8876 declare float @llvm.ceil.f32(float %Val)
8877 declare double @llvm.ceil.f64(double %Val)
8878 declare x86_fp80 @llvm.ceil.f80(x86_fp80 %Val)
8879 declare fp128 @llvm.ceil.f128(fp128 %Val)
8880 declare ppc_fp128 @llvm.ceil.ppcf128(ppc_fp128 %Val)
8885 The '``llvm.ceil.*``' intrinsics return the ceiling of the operand.
8890 The argument and return value are floating point numbers of the same
8896 This function returns the same values as the libm ``ceil`` functions
8897 would, and handles error conditions in the same way.
8899 '``llvm.trunc.*``' Intrinsic
8900 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8905 This is an overloaded intrinsic. You can use ``llvm.trunc`` on any
8906 floating point or vector of floating point type. Not all targets support
8911 declare float @llvm.trunc.f32(float %Val)
8912 declare double @llvm.trunc.f64(double %Val)
8913 declare x86_fp80 @llvm.trunc.f80(x86_fp80 %Val)
8914 declare fp128 @llvm.trunc.f128(fp128 %Val)
8915 declare ppc_fp128 @llvm.trunc.ppcf128(ppc_fp128 %Val)
8920 The '``llvm.trunc.*``' intrinsics returns the operand rounded to the
8921 nearest integer not larger in magnitude than the operand.
8926 The argument and return value are floating point numbers of the same
8932 This function returns the same values as the libm ``trunc`` functions
8933 would, and handles error conditions in the same way.
8935 '``llvm.rint.*``' Intrinsic
8936 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8941 This is an overloaded intrinsic. You can use ``llvm.rint`` on any
8942 floating point or vector of floating point type. Not all targets support
8947 declare float @llvm.rint.f32(float %Val)
8948 declare double @llvm.rint.f64(double %Val)
8949 declare x86_fp80 @llvm.rint.f80(x86_fp80 %Val)
8950 declare fp128 @llvm.rint.f128(fp128 %Val)
8951 declare ppc_fp128 @llvm.rint.ppcf128(ppc_fp128 %Val)
8956 The '``llvm.rint.*``' intrinsics returns the operand rounded to the
8957 nearest integer. It may raise an inexact floating-point exception if the
8958 operand isn't an integer.
8963 The argument and return value are floating point numbers of the same
8969 This function returns the same values as the libm ``rint`` functions
8970 would, and handles error conditions in the same way.
8972 '``llvm.nearbyint.*``' Intrinsic
8973 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8978 This is an overloaded intrinsic. You can use ``llvm.nearbyint`` on any
8979 floating point or vector of floating point type. Not all targets support
8984 declare float @llvm.nearbyint.f32(float %Val)
8985 declare double @llvm.nearbyint.f64(double %Val)
8986 declare x86_fp80 @llvm.nearbyint.f80(x86_fp80 %Val)
8987 declare fp128 @llvm.nearbyint.f128(fp128 %Val)
8988 declare ppc_fp128 @llvm.nearbyint.ppcf128(ppc_fp128 %Val)
8993 The '``llvm.nearbyint.*``' intrinsics returns the operand rounded to the
8999 The argument and return value are floating point numbers of the same
9005 This function returns the same values as the libm ``nearbyint``
9006 functions would, and handles error conditions in the same way.
9008 '``llvm.round.*``' Intrinsic
9009 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9014 This is an overloaded intrinsic. You can use ``llvm.round`` on any
9015 floating point or vector of floating point type. Not all targets support
9020 declare float @llvm.round.f32(float %Val)
9021 declare double @llvm.round.f64(double %Val)
9022 declare x86_fp80 @llvm.round.f80(x86_fp80 %Val)
9023 declare fp128 @llvm.round.f128(fp128 %Val)
9024 declare ppc_fp128 @llvm.round.ppcf128(ppc_fp128 %Val)
9029 The '``llvm.round.*``' intrinsics returns the operand rounded to the
9035 The argument and return value are floating point numbers of the same
9041 This function returns the same values as the libm ``round``
9042 functions would, and handles error conditions in the same way.
9044 Bit Manipulation Intrinsics
9045 ---------------------------
9047 LLVM provides intrinsics for a few important bit manipulation
9048 operations. These allow efficient code generation for some algorithms.
9050 '``llvm.bswap.*``' Intrinsics
9051 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9056 This is an overloaded intrinsic function. You can use bswap on any
9057 integer type that is an even number of bytes (i.e. BitWidth % 16 == 0).
9061 declare i16 @llvm.bswap.i16(i16 <id>)
9062 declare i32 @llvm.bswap.i32(i32 <id>)
9063 declare i64 @llvm.bswap.i64(i64 <id>)
9068 The '``llvm.bswap``' family of intrinsics is used to byte swap integer
9069 values with an even number of bytes (positive multiple of 16 bits).
9070 These are useful for performing operations on data that is not in the
9071 target's native byte order.
9076 The ``llvm.bswap.i16`` intrinsic returns an i16 value that has the high
9077 and low byte of the input i16 swapped. Similarly, the ``llvm.bswap.i32``
9078 intrinsic returns an i32 value that has the four bytes of the input i32
9079 swapped, so that if the input bytes are numbered 0, 1, 2, 3 then the
9080 returned i32 will have its bytes in 3, 2, 1, 0 order. The
9081 ``llvm.bswap.i48``, ``llvm.bswap.i64`` and other intrinsics extend this
9082 concept to additional even-byte lengths (6 bytes, 8 bytes and more,
9085 '``llvm.ctpop.*``' Intrinsic
9086 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9091 This is an overloaded intrinsic. You can use llvm.ctpop on any integer
9092 bit width, or on any vector with integer elements. Not all targets
9093 support all bit widths or vector types, however.
9097 declare i8 @llvm.ctpop.i8(i8 <src>)
9098 declare i16 @llvm.ctpop.i16(i16 <src>)
9099 declare i32 @llvm.ctpop.i32(i32 <src>)
9100 declare i64 @llvm.ctpop.i64(i64 <src>)
9101 declare i256 @llvm.ctpop.i256(i256 <src>)
9102 declare <2 x i32> @llvm.ctpop.v2i32(<2 x i32> <src>)
9107 The '``llvm.ctpop``' family of intrinsics counts the number of bits set
9113 The only argument is the value to be counted. The argument may be of any
9114 integer type, or a vector with integer elements. The return type must
9115 match the argument type.
9120 The '``llvm.ctpop``' intrinsic counts the 1's in a variable, or within
9121 each element of a vector.
9123 '``llvm.ctlz.*``' Intrinsic
9124 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9129 This is an overloaded intrinsic. You can use ``llvm.ctlz`` on any
9130 integer bit width, or any vector whose elements are integers. Not all
9131 targets support all bit widths or vector types, however.
9135 declare i8 @llvm.ctlz.i8 (i8 <src>, i1 <is_zero_undef>)
9136 declare i16 @llvm.ctlz.i16 (i16 <src>, i1 <is_zero_undef>)
9137 declare i32 @llvm.ctlz.i32 (i32 <src>, i1 <is_zero_undef>)
9138 declare i64 @llvm.ctlz.i64 (i64 <src>, i1 <is_zero_undef>)
9139 declare i256 @llvm.ctlz.i256(i256 <src>, i1 <is_zero_undef>)
9140 declase <2 x i32> @llvm.ctlz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
9145 The '``llvm.ctlz``' family of intrinsic functions counts the number of
9146 leading zeros in a variable.
9151 The first argument is the value to be counted. This argument may be of
9152 any integer type, or a vector with integer element type. The return
9153 type must match the first argument type.
9155 The second argument must be a constant and is a flag to indicate whether
9156 the intrinsic should ensure that a zero as the first argument produces a
9157 defined result. Historically some architectures did not provide a
9158 defined result for zero values as efficiently, and many algorithms are
9159 now predicated on avoiding zero-value inputs.
9164 The '``llvm.ctlz``' intrinsic counts the leading (most significant)
9165 zeros in a variable, or within each element of the vector. If
9166 ``src == 0`` then the result is the size in bits of the type of ``src``
9167 if ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
9168 ``llvm.ctlz(i32 2) = 30``.
9170 '``llvm.cttz.*``' Intrinsic
9171 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9176 This is an overloaded intrinsic. You can use ``llvm.cttz`` on any
9177 integer bit width, or any vector of integer elements. Not all targets
9178 support all bit widths or vector types, however.
9182 declare i8 @llvm.cttz.i8 (i8 <src>, i1 <is_zero_undef>)
9183 declare i16 @llvm.cttz.i16 (i16 <src>, i1 <is_zero_undef>)
9184 declare i32 @llvm.cttz.i32 (i32 <src>, i1 <is_zero_undef>)
9185 declare i64 @llvm.cttz.i64 (i64 <src>, i1 <is_zero_undef>)
9186 declare i256 @llvm.cttz.i256(i256 <src>, i1 <is_zero_undef>)
9187 declase <2 x i32> @llvm.cttz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
9192 The '``llvm.cttz``' family of intrinsic functions counts the number of
9198 The first argument is the value to be counted. This argument may be of
9199 any integer type, or a vector with integer element type. The return
9200 type must match the first argument type.
9202 The second argument must be a constant and is a flag to indicate whether
9203 the intrinsic should ensure that a zero as the first argument produces a
9204 defined result. Historically some architectures did not provide a
9205 defined result for zero values as efficiently, and many algorithms are
9206 now predicated on avoiding zero-value inputs.
9211 The '``llvm.cttz``' intrinsic counts the trailing (least significant)
9212 zeros in a variable, or within each element of a vector. If ``src == 0``
9213 then the result is the size in bits of the type of ``src`` if
9214 ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
9215 ``llvm.cttz(2) = 1``.
9219 Arithmetic with Overflow Intrinsics
9220 -----------------------------------
9222 LLVM provides intrinsics for some arithmetic with overflow operations.
9224 '``llvm.sadd.with.overflow.*``' Intrinsics
9225 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9230 This is an overloaded intrinsic. You can use ``llvm.sadd.with.overflow``
9231 on any integer bit width.
9235 declare {i16, i1} @llvm.sadd.with.overflow.i16(i16 %a, i16 %b)
9236 declare {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
9237 declare {i64, i1} @llvm.sadd.with.overflow.i64(i64 %a, i64 %b)
9242 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
9243 a signed addition of the two arguments, and indicate whether an overflow
9244 occurred during the signed summation.
9249 The arguments (%a and %b) and the first element of the result structure
9250 may be of integer types of any bit width, but they must have the same
9251 bit width. The second element of the result structure must be of type
9252 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
9258 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
9259 a signed addition of the two variables. They return a structure --- the
9260 first element of which is the signed summation, and the second element
9261 of which is a bit specifying if the signed summation resulted in an
9267 .. code-block:: llvm
9269 %res = call {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
9270 %sum = extractvalue {i32, i1} %res, 0
9271 %obit = extractvalue {i32, i1} %res, 1
9272 br i1 %obit, label %overflow, label %normal
9274 '``llvm.uadd.with.overflow.*``' Intrinsics
9275 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9280 This is an overloaded intrinsic. You can use ``llvm.uadd.with.overflow``
9281 on any integer bit width.
9285 declare {i16, i1} @llvm.uadd.with.overflow.i16(i16 %a, i16 %b)
9286 declare {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
9287 declare {i64, i1} @llvm.uadd.with.overflow.i64(i64 %a, i64 %b)
9292 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
9293 an unsigned addition of the two arguments, and indicate whether a carry
9294 occurred during the unsigned summation.
9299 The arguments (%a and %b) and the first element of the result structure
9300 may be of integer types of any bit width, but they must have the same
9301 bit width. The second element of the result structure must be of type
9302 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
9308 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
9309 an unsigned addition of the two arguments. They return a structure --- the
9310 first element of which is the sum, and the second element of which is a
9311 bit specifying if the unsigned summation resulted in a carry.
9316 .. code-block:: llvm
9318 %res = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
9319 %sum = extractvalue {i32, i1} %res, 0
9320 %obit = extractvalue {i32, i1} %res, 1
9321 br i1 %obit, label %carry, label %normal
9323 '``llvm.ssub.with.overflow.*``' Intrinsics
9324 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9329 This is an overloaded intrinsic. You can use ``llvm.ssub.with.overflow``
9330 on any integer bit width.
9334 declare {i16, i1} @llvm.ssub.with.overflow.i16(i16 %a, i16 %b)
9335 declare {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
9336 declare {i64, i1} @llvm.ssub.with.overflow.i64(i64 %a, i64 %b)
9341 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
9342 a signed subtraction of the two arguments, and indicate whether an
9343 overflow occurred during the signed subtraction.
9348 The arguments (%a and %b) and the first element of the result structure
9349 may be of integer types of any bit width, but they must have the same
9350 bit width. The second element of the result structure must be of type
9351 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
9357 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
9358 a signed subtraction of the two arguments. They return a structure --- the
9359 first element of which is the subtraction, and the second element of
9360 which is a bit specifying if the signed subtraction resulted in an
9366 .. code-block:: llvm
9368 %res = call {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
9369 %sum = extractvalue {i32, i1} %res, 0
9370 %obit = extractvalue {i32, i1} %res, 1
9371 br i1 %obit, label %overflow, label %normal
9373 '``llvm.usub.with.overflow.*``' Intrinsics
9374 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9379 This is an overloaded intrinsic. You can use ``llvm.usub.with.overflow``
9380 on any integer bit width.
9384 declare {i16, i1} @llvm.usub.with.overflow.i16(i16 %a, i16 %b)
9385 declare {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
9386 declare {i64, i1} @llvm.usub.with.overflow.i64(i64 %a, i64 %b)
9391 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
9392 an unsigned subtraction of the two arguments, and indicate whether an
9393 overflow occurred during the unsigned subtraction.
9398 The arguments (%a and %b) and the first element of the result structure
9399 may be of integer types of any bit width, but they must have the same
9400 bit width. The second element of the result structure must be of type
9401 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
9407 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
9408 an unsigned subtraction of the two arguments. They return a structure ---
9409 the first element of which is the subtraction, and the second element of
9410 which is a bit specifying if the unsigned subtraction resulted in an
9416 .. code-block:: llvm
9418 %res = call {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
9419 %sum = extractvalue {i32, i1} %res, 0
9420 %obit = extractvalue {i32, i1} %res, 1
9421 br i1 %obit, label %overflow, label %normal
9423 '``llvm.smul.with.overflow.*``' Intrinsics
9424 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9429 This is an overloaded intrinsic. You can use ``llvm.smul.with.overflow``
9430 on any integer bit width.
9434 declare {i16, i1} @llvm.smul.with.overflow.i16(i16 %a, i16 %b)
9435 declare {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
9436 declare {i64, i1} @llvm.smul.with.overflow.i64(i64 %a, i64 %b)
9441 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
9442 a signed multiplication of the two arguments, and indicate whether an
9443 overflow occurred during the signed multiplication.
9448 The arguments (%a and %b) and the first element of the result structure
9449 may be of integer types of any bit width, but they must have the same
9450 bit width. The second element of the result structure must be of type
9451 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
9457 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
9458 a signed multiplication of the two arguments. They return a structure ---
9459 the first element of which is the multiplication, and the second element
9460 of which is a bit specifying if the signed multiplication resulted in an
9466 .. code-block:: llvm
9468 %res = call {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
9469 %sum = extractvalue {i32, i1} %res, 0
9470 %obit = extractvalue {i32, i1} %res, 1
9471 br i1 %obit, label %overflow, label %normal
9473 '``llvm.umul.with.overflow.*``' Intrinsics
9474 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9479 This is an overloaded intrinsic. You can use ``llvm.umul.with.overflow``
9480 on any integer bit width.
9484 declare {i16, i1} @llvm.umul.with.overflow.i16(i16 %a, i16 %b)
9485 declare {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
9486 declare {i64, i1} @llvm.umul.with.overflow.i64(i64 %a, i64 %b)
9491 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
9492 a unsigned multiplication of the two arguments, and indicate whether an
9493 overflow occurred during the unsigned multiplication.
9498 The arguments (%a and %b) and the first element of the result structure
9499 may be of integer types of any bit width, but they must have the same
9500 bit width. The second element of the result structure must be of type
9501 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
9507 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
9508 an unsigned multiplication of the two arguments. They return a structure ---
9509 the first element of which is the multiplication, and the second
9510 element of which is a bit specifying if the unsigned multiplication
9511 resulted in an overflow.
9516 .. code-block:: llvm
9518 %res = call {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
9519 %sum = extractvalue {i32, i1} %res, 0
9520 %obit = extractvalue {i32, i1} %res, 1
9521 br i1 %obit, label %overflow, label %normal
9523 Specialised Arithmetic Intrinsics
9524 ---------------------------------
9526 '``llvm.fmuladd.*``' Intrinsic
9527 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9534 declare float @llvm.fmuladd.f32(float %a, float %b, float %c)
9535 declare double @llvm.fmuladd.f64(double %a, double %b, double %c)
9540 The '``llvm.fmuladd.*``' intrinsic functions represent multiply-add
9541 expressions that can be fused if the code generator determines that (a) the
9542 target instruction set has support for a fused operation, and (b) that the
9543 fused operation is more efficient than the equivalent, separate pair of mul
9544 and add instructions.
9549 The '``llvm.fmuladd.*``' intrinsics each take three arguments: two
9550 multiplicands, a and b, and an addend c.
9559 %0 = call float @llvm.fmuladd.f32(%a, %b, %c)
9561 is equivalent to the expression a \* b + c, except that rounding will
9562 not be performed between the multiplication and addition steps if the
9563 code generator fuses the operations. Fusion is not guaranteed, even if
9564 the target platform supports it. If a fused multiply-add is required the
9565 corresponding llvm.fma.\* intrinsic function should be used
9566 instead. This never sets errno, just as '``llvm.fma.*``'.
9571 .. code-block:: llvm
9573 %r2 = call float @llvm.fmuladd.f32(float %a, float %b, float %c) ; yields float:r2 = (a * b) + c
9575 Half Precision Floating Point Intrinsics
9576 ----------------------------------------
9578 For most target platforms, half precision floating point is a
9579 storage-only format. This means that it is a dense encoding (in memory)
9580 but does not support computation in the format.
9582 This means that code must first load the half-precision floating point
9583 value as an i16, then convert it to float with
9584 :ref:`llvm.convert.from.fp16 <int_convert_from_fp16>`. Computation can
9585 then be performed on the float value (including extending to double
9586 etc). To store the value back to memory, it is first converted to float
9587 if needed, then converted to i16 with
9588 :ref:`llvm.convert.to.fp16 <int_convert_to_fp16>`, then storing as an
9591 .. _int_convert_to_fp16:
9593 '``llvm.convert.to.fp16``' Intrinsic
9594 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9601 declare i16 @llvm.convert.to.fp16.f32(float %a)
9602 declare i16 @llvm.convert.to.fp16.f64(double %a)
9607 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion from a
9608 conventional floating point type to half precision floating point format.
9613 The intrinsic function contains single argument - the value to be
9619 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion from a
9620 conventional floating point format to half precision floating point format. The
9621 return value is an ``i16`` which contains the converted number.
9626 .. code-block:: llvm
9628 %res = call i16 @llvm.convert.to.fp16.f32(float %a)
9629 store i16 %res, i16* @x, align 2
9631 .. _int_convert_from_fp16:
9633 '``llvm.convert.from.fp16``' Intrinsic
9634 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9641 declare float @llvm.convert.from.fp16.f32(i16 %a)
9642 declare double @llvm.convert.from.fp16.f64(i16 %a)
9647 The '``llvm.convert.from.fp16``' intrinsic function performs a
9648 conversion from half precision floating point format to single precision
9649 floating point format.
9654 The intrinsic function contains single argument - the value to be
9660 The '``llvm.convert.from.fp16``' intrinsic function performs a
9661 conversion from half single precision floating point format to single
9662 precision floating point format. The input half-float value is
9663 represented by an ``i16`` value.
9668 .. code-block:: llvm
9670 %a = load i16, i16* @x, align 2
9671 %res = call float @llvm.convert.from.fp16(i16 %a)
9678 The LLVM debugger intrinsics (which all start with ``llvm.dbg.``
9679 prefix), are described in the `LLVM Source Level
9680 Debugging <SourceLevelDebugging.html#format_common_intrinsics>`_
9683 Exception Handling Intrinsics
9684 -----------------------------
9686 The LLVM exception handling intrinsics (which all start with
9687 ``llvm.eh.`` prefix), are described in the `LLVM Exception
9688 Handling <ExceptionHandling.html#format_common_intrinsics>`_ document.
9692 Trampoline Intrinsics
9693 ---------------------
9695 These intrinsics make it possible to excise one parameter, marked with
9696 the :ref:`nest <nest>` attribute, from a function. The result is a
9697 callable function pointer lacking the nest parameter - the caller does
9698 not need to provide a value for it. Instead, the value to use is stored
9699 in advance in a "trampoline", a block of memory usually allocated on the
9700 stack, which also contains code to splice the nest value into the
9701 argument list. This is used to implement the GCC nested function address
9704 For example, if the function is ``i32 f(i8* nest %c, i32 %x, i32 %y)``
9705 then the resulting function pointer has signature ``i32 (i32, i32)*``.
9706 It can be created as follows:
9708 .. code-block:: llvm
9710 %tramp = alloca [10 x i8], align 4 ; size and alignment only correct for X86
9711 %tramp1 = getelementptr [10 x i8], [10 x i8]* %tramp, i32 0, i32 0
9712 call i8* @llvm.init.trampoline(i8* %tramp1, i8* bitcast (i32 (i8*, i32, i32)* @f to i8*), i8* %nval)
9713 %p = call i8* @llvm.adjust.trampoline(i8* %tramp1)
9714 %fp = bitcast i8* %p to i32 (i32, i32)*
9716 The call ``%val = call i32 %fp(i32 %x, i32 %y)`` is then equivalent to
9717 ``%val = call i32 %f(i8* %nval, i32 %x, i32 %y)``.
9721 '``llvm.init.trampoline``' Intrinsic
9722 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9729 declare void @llvm.init.trampoline(i8* <tramp>, i8* <func>, i8* <nval>)
9734 This fills the memory pointed to by ``tramp`` with executable code,
9735 turning it into a trampoline.
9740 The ``llvm.init.trampoline`` intrinsic takes three arguments, all
9741 pointers. The ``tramp`` argument must point to a sufficiently large and
9742 sufficiently aligned block of memory; this memory is written to by the
9743 intrinsic. Note that the size and the alignment are target-specific -
9744 LLVM currently provides no portable way of determining them, so a
9745 front-end that generates this intrinsic needs to have some
9746 target-specific knowledge. The ``func`` argument must hold a function
9747 bitcast to an ``i8*``.
9752 The block of memory pointed to by ``tramp`` is filled with target
9753 dependent code, turning it into a function. Then ``tramp`` needs to be
9754 passed to :ref:`llvm.adjust.trampoline <int_at>` to get a pointer which can
9755 be :ref:`bitcast (to a new function) and called <int_trampoline>`. The new
9756 function's signature is the same as that of ``func`` with any arguments
9757 marked with the ``nest`` attribute removed. At most one such ``nest``
9758 argument is allowed, and it must be of pointer type. Calling the new
9759 function is equivalent to calling ``func`` with the same argument list,
9760 but with ``nval`` used for the missing ``nest`` argument. If, after
9761 calling ``llvm.init.trampoline``, the memory pointed to by ``tramp`` is
9762 modified, then the effect of any later call to the returned function
9763 pointer is undefined.
9767 '``llvm.adjust.trampoline``' Intrinsic
9768 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9775 declare i8* @llvm.adjust.trampoline(i8* <tramp>)
9780 This performs any required machine-specific adjustment to the address of
9781 a trampoline (passed as ``tramp``).
9786 ``tramp`` must point to a block of memory which already has trampoline
9787 code filled in by a previous call to
9788 :ref:`llvm.init.trampoline <int_it>`.
9793 On some architectures the address of the code to be executed needs to be
9794 different than the address where the trampoline is actually stored. This
9795 intrinsic returns the executable address corresponding to ``tramp``
9796 after performing the required machine specific adjustments. The pointer
9797 returned can then be :ref:`bitcast and executed <int_trampoline>`.
9799 .. _int_mload_mstore:
9801 Masked Vector Load and Store Intrinsics
9802 ---------------------------------------
9804 LLVM provides intrinsics for predicated vector load and store operations. The predicate is specified by a mask operand, which holds one bit per vector element, switching the associated vector lane on or off. The memory addresses corresponding to the "off" lanes are not accessed. When all bits of the mask are on, the intrinsic is identical to a regular vector load or store. When all bits are off, no memory is accessed.
9808 '``llvm.masked.load.*``' Intrinsics
9809 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9813 This is an overloaded intrinsic. The loaded data is a vector of any integer or floating point data type.
9817 declare <16 x float> @llvm.masked.load.v16f32 (<16 x float>* <ptr>, i32 <alignment>, <16 x i1> <mask>, <16 x float> <passthru>)
9818 declare <2 x double> @llvm.masked.load.v2f64 (<2 x double>* <ptr>, i32 <alignment>, <2 x i1> <mask>, <2 x double> <passthru>)
9823 Reads a vector from memory according to the provided mask. The mask holds a bit for each vector lane, and is used to prevent memory accesses to the masked-off lanes. The masked-off lanes in the result vector are taken from the corresponding lanes of the '``passthru``' operand.
9829 The first operand is the base pointer for the load. The second operand is the alignment of the source location. It must be a constant integer value. The third operand, mask, is a vector of boolean values with the same number of elements as the return type. The fourth is a pass-through value that is used to fill the masked-off lanes of the result. The return type, underlying type of the base pointer and the type of the '``passthru``' operand are the same vector types.
9835 The '``llvm.masked.load``' intrinsic is designed for conditional reading of selected vector elements in a single IR operation. It is useful for targets that support vector masked loads and allows vectorizing predicated basic blocks on these targets. Other targets may support this intrinsic differently, for example by lowering it into a sequence of branches that guard scalar load operations.
9836 The result of this operation is equivalent to a regular vector load instruction followed by a 'select' between the loaded and the passthru values, predicated on the same mask. However, using this intrinsic prevents exceptions on memory access to masked-off lanes.
9841 %res = call <16 x float> @llvm.masked.load.v16f32 (<16 x float>* %ptr, i32 4, <16 x i1>%mask, <16 x float> %passthru)
9843 ;; The result of the two following instructions is identical aside from potential memory access exception
9844 %loadlal = load <16 x float>, <16 x float>* %ptr, align 4
9845 %res = select <16 x i1> %mask, <16 x float> %loadlal, <16 x float> %passthru
9849 '``llvm.masked.store.*``' Intrinsics
9850 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9854 This is an overloaded intrinsic. The data stored in memory is a vector of any integer or floating point data type.
9858 declare void @llvm.masked.store.v8i32 (<8 x i32> <value>, <8 x i32> * <ptr>, i32 <alignment>, <8 x i1> <mask>)
9859 declare void @llvm.masked.store.v16f32(<16 x i32> <value>, <16 x i32>* <ptr>, i32 <alignment>, <16 x i1> <mask>)
9864 Writes a vector to memory according to the provided mask. The mask holds a bit for each vector lane, and is used to prevent memory accesses to the masked-off lanes.
9869 The first operand is the vector value to be written to memory. The second operand is the base pointer for the store, it has the same underlying type as the value operand. The third operand is the alignment of the destination location. The fourth operand, mask, is a vector of boolean values. The types of the mask and the value operand must have the same number of vector elements.
9875 The '``llvm.masked.store``' intrinsics is designed for conditional writing of selected vector elements in a single IR operation. It is useful for targets that support vector masked store and allows vectorizing predicated basic blocks on these targets. Other targets may support this intrinsic differently, for example by lowering it into a sequence of branches that guard scalar store operations.
9876 The result of this operation is equivalent to a load-modify-store sequence. However, using this intrinsic prevents exceptions and data races on memory access to masked-off lanes.
9880 call void @llvm.masked.store.v16f32(<16 x float> %value, <16 x float>* %ptr, i32 4, <16 x i1> %mask)
9882 ;; The result of the following instructions is identical aside from potential data races and memory access exceptions
9883 %oldval = load <16 x float>, <16 x float>* %ptr, align 4
9884 %res = select <16 x i1> %mask, <16 x float> %value, <16 x float> %oldval
9885 store <16 x float> %res, <16 x float>* %ptr, align 4
9888 Masked Vector Gather and Scatter Intrinsics
9889 -------------------------------------------
9891 LLVM provides intrinsics for vector gather and scatter operations. They are similar to :ref:`Masked Vector Load and Store <int_mload_mstore>`, except they are designed for arbitrary memory accesses, rather than sequential memory accesses. Gather and scatter also employ a mask operand, which holds one bit per vector element, switching the associated vector lane on or off. The memory addresses corresponding to the "off" lanes are not accessed. When all bits are off, no memory is accessed.
9895 '``llvm.masked.gather.*``' Intrinsics
9896 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9900 This is an overloaded intrinsic. The loaded data are multiple scalar values of any integer or floating point data type gathered together into one vector.
9904 declare <16 x float> @llvm.masked.gather.v16f32 (<16 x float*> <ptrs>, i32 <alignment>, <16 x i1> <mask>, <16 x float> <passthru>)
9905 declare <2 x double> @llvm.masked.gather.v2f64 (<2 x double*> <ptrs>, i32 <alignment>, <2 x i1> <mask>, <2 x double> <passthru>)
9910 Reads scalar values from arbitrary memory locations and gathers them into one vector. The memory locations are provided in the vector of pointers '``ptrs``'. The memory is accessed according to the provided mask. The mask holds a bit for each vector lane, and is used to prevent memory accesses to the masked-off lanes. The masked-off lanes in the result vector are taken from the corresponding lanes of the '``passthru``' operand.
9916 The first operand is a vector of pointers which holds all memory addresses to read. The second operand is an alignment of the source addresses. It must be a constant integer value. The third operand, mask, is a vector of boolean values with the same number of elements as the return type. The fourth is a pass-through value that is used to fill the masked-off lanes of the result. The return type, underlying type of the vector of pointers and the type of the '``passthru``' operand are the same vector types.
9922 The '``llvm.masked.gather``' intrinsic is designed for conditional reading of multiple scalar values from arbitrary memory locations in a single IR operation. It is useful for targets that support vector masked gathers and allows vectorizing basic blocks with data and control divergence. Other targets may support this intrinsic differently, for example by lowering it into a sequence of scalar load operations.
9923 The semantics of this operation are equivalent to a sequence of conditional scalar loads with subsequent gathering all loaded values into a single vector. The mask restricts memory access to certain lanes and facilitates vectorization of predicated basic blocks.
9928 %res = call <4 x double> @llvm.masked.gather.v4f64 (<4 x double*> %ptrs, i32 8, <4 x i1>%mask, <4 x double> <true, true, true, true>)
9930 ;; The gather with all-true mask is equivalent to the following instruction sequence
9931 %ptr0 = extractelement <4 x double*> %ptrs, i32 0
9932 %ptr1 = extractelement <4 x double*> %ptrs, i32 1
9933 %ptr2 = extractelement <4 x double*> %ptrs, i32 2
9934 %ptr3 = extractelement <4 x double*> %ptrs, i32 3
9936 %val0 = load double, double* %ptr0, align 8
9937 %val1 = load double, double* %ptr1, align 8
9938 %val2 = load double, double* %ptr2, align 8
9939 %val3 = load double, double* %ptr3, align 8
9941 %vec0 = insertelement <4 x double>undef, %val0, 0
9942 %vec01 = insertelement <4 x double>%vec0, %val1, 1
9943 %vec012 = insertelement <4 x double>%vec01, %val2, 2
9944 %vec0123 = insertelement <4 x double>%vec012, %val3, 3
9948 '``llvm.masked.scatter.*``' Intrinsics
9949 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9953 This is an overloaded intrinsic. The data stored in memory is a vector of any integer or floating point data type. Each vector element is stored in an arbitrary memory addresses. Scatter with overlapping addresses is guaranteed to be ordered from least-significant to most-significant element.
9957 declare void @llvm.masked.scatter.v8i32 (<8 x i32> <value>, <8 x i32*> <ptrs>, i32 <alignment>, <8 x i1> <mask>)
9958 declare void @llvm.masked.scatter.v16f32(<16 x i32> <value>, <16 x i32*> <ptrs>, i32 <alignment>, <16 x i1> <mask>)
9963 Writes each element from the value vector to the corresponding memory address. The memory addresses are represented as a vector of pointers. Writing is done according to the provided mask. The mask holds a bit for each vector lane, and is used to prevent memory accesses to the masked-off lanes.
9968 The first operand is a vector value to be written to memory. The second operand is a vector of pointers, pointing to where the value elements should be stored. It has the same underlying type as the value operand. The third operand is an alignment of the destination addresses. The fourth operand, mask, is a vector of boolean values. The types of the mask and the value operand must have the same number of vector elements.
9974 The '``llvm.masked.scatter``' intrinsics is designed for writing selected vector elements to arbitrary memory addresses in a single IR operation. The operation may be conditional, when not all bits in the mask are switched on. It is useful for targets that support vector masked scatter and allows vectorizing basic blocks with data and control divergency. Other targets may support this intrinsic differently, for example by lowering it into a sequence of branches that guard scalar store operations.
9978 ;; This instruction unconditionaly stores data vector in multiple addresses
9979 call @llvm.masked.scatter.v8i32 (<8 x i32> %value, <8 x i32*> %ptrs, i32 4, <8 x i1> <true, true, .. true>)
9981 ;; It is equivalent to a list of scalar stores
9982 %val0 = extractelement <8 x i32> %value, i32 0
9983 %val1 = extractelement <8 x i32> %value, i32 1
9985 %val7 = extractelement <8 x i32> %value, i32 7
9986 %ptr0 = extractelement <8 x i32*> %ptrs, i32 0
9987 %ptr1 = extractelement <8 x i32*> %ptrs, i32 1
9989 %ptr7 = extractelement <8 x i32*> %ptrs, i32 7
9990 ;; Note: the order of the following stores is important when they overlap:
9991 store i32 %val0, i32* %ptr0, align 4
9992 store i32 %val1, i32* %ptr1, align 4
9994 store i32 %val7, i32* %ptr7, align 4
10000 This class of intrinsics provides information about the lifetime of
10001 memory objects and ranges where variables are immutable.
10005 '``llvm.lifetime.start``' Intrinsic
10006 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10013 declare void @llvm.lifetime.start(i64 <size>, i8* nocapture <ptr>)
10018 The '``llvm.lifetime.start``' intrinsic specifies the start of a memory
10024 The first argument is a constant integer representing the size of the
10025 object, or -1 if it is variable sized. The second argument is a pointer
10031 This intrinsic indicates that before this point in the code, the value
10032 of the memory pointed to by ``ptr`` is dead. This means that it is known
10033 to never be used and has an undefined value. A load from the pointer
10034 that precedes this intrinsic can be replaced with ``'undef'``.
10038 '``llvm.lifetime.end``' Intrinsic
10039 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10046 declare void @llvm.lifetime.end(i64 <size>, i8* nocapture <ptr>)
10051 The '``llvm.lifetime.end``' intrinsic specifies the end of a memory
10057 The first argument is a constant integer representing the size of the
10058 object, or -1 if it is variable sized. The second argument is a pointer
10064 This intrinsic indicates that after this point in the code, the value of
10065 the memory pointed to by ``ptr`` is dead. This means that it is known to
10066 never be used and has an undefined value. Any stores into the memory
10067 object following this intrinsic may be removed as dead.
10069 '``llvm.invariant.start``' Intrinsic
10070 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10077 declare {}* @llvm.invariant.start(i64 <size>, i8* nocapture <ptr>)
10082 The '``llvm.invariant.start``' intrinsic specifies that the contents of
10083 a memory object will not change.
10088 The first argument is a constant integer representing the size of the
10089 object, or -1 if it is variable sized. The second argument is a pointer
10095 This intrinsic indicates that until an ``llvm.invariant.end`` that uses
10096 the return value, the referenced memory location is constant and
10099 '``llvm.invariant.end``' Intrinsic
10100 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10107 declare void @llvm.invariant.end({}* <start>, i64 <size>, i8* nocapture <ptr>)
10112 The '``llvm.invariant.end``' intrinsic specifies that the contents of a
10113 memory object are mutable.
10118 The first argument is the matching ``llvm.invariant.start`` intrinsic.
10119 The second argument is a constant integer representing the size of the
10120 object, or -1 if it is variable sized and the third argument is a
10121 pointer to the object.
10126 This intrinsic indicates that the memory is mutable again.
10131 This class of intrinsics is designed to be generic and has no specific
10134 '``llvm.var.annotation``' Intrinsic
10135 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10142 declare void @llvm.var.annotation(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
10147 The '``llvm.var.annotation``' intrinsic.
10152 The first argument is a pointer to a value, the second is a pointer to a
10153 global string, the third is a pointer to a global string which is the
10154 source file name, and the last argument is the line number.
10159 This intrinsic allows annotation of local variables with arbitrary
10160 strings. This can be useful for special purpose optimizations that want
10161 to look for these annotations. These have no other defined use; they are
10162 ignored by code generation and optimization.
10164 '``llvm.ptr.annotation.*``' Intrinsic
10165 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10170 This is an overloaded intrinsic. You can use '``llvm.ptr.annotation``' on a
10171 pointer to an integer of any width. *NOTE* you must specify an address space for
10172 the pointer. The identifier for the default address space is the integer
10177 declare i8* @llvm.ptr.annotation.p<address space>i8(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
10178 declare i16* @llvm.ptr.annotation.p<address space>i16(i16* <val>, i8* <str>, i8* <str>, i32 <int>)
10179 declare i32* @llvm.ptr.annotation.p<address space>i32(i32* <val>, i8* <str>, i8* <str>, i32 <int>)
10180 declare i64* @llvm.ptr.annotation.p<address space>i64(i64* <val>, i8* <str>, i8* <str>, i32 <int>)
10181 declare i256* @llvm.ptr.annotation.p<address space>i256(i256* <val>, i8* <str>, i8* <str>, i32 <int>)
10186 The '``llvm.ptr.annotation``' intrinsic.
10191 The first argument is a pointer to an integer value of arbitrary bitwidth
10192 (result of some expression), the second is a pointer to a global string, the
10193 third is a pointer to a global string which is the source file name, and the
10194 last argument is the line number. It returns the value of the first argument.
10199 This intrinsic allows annotation of a pointer to an integer with arbitrary
10200 strings. This can be useful for special purpose optimizations that want to look
10201 for these annotations. These have no other defined use; they are ignored by code
10202 generation and optimization.
10204 '``llvm.annotation.*``' Intrinsic
10205 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10210 This is an overloaded intrinsic. You can use '``llvm.annotation``' on
10211 any integer bit width.
10215 declare i8 @llvm.annotation.i8(i8 <val>, i8* <str>, i8* <str>, i32 <int>)
10216 declare i16 @llvm.annotation.i16(i16 <val>, i8* <str>, i8* <str>, i32 <int>)
10217 declare i32 @llvm.annotation.i32(i32 <val>, i8* <str>, i8* <str>, i32 <int>)
10218 declare i64 @llvm.annotation.i64(i64 <val>, i8* <str>, i8* <str>, i32 <int>)
10219 declare i256 @llvm.annotation.i256(i256 <val>, i8* <str>, i8* <str>, i32 <int>)
10224 The '``llvm.annotation``' intrinsic.
10229 The first argument is an integer value (result of some expression), the
10230 second is a pointer to a global string, the third is a pointer to a
10231 global string which is the source file name, and the last argument is
10232 the line number. It returns the value of the first argument.
10237 This intrinsic allows annotations to be put on arbitrary expressions
10238 with arbitrary strings. This can be useful for special purpose
10239 optimizations that want to look for these annotations. These have no
10240 other defined use; they are ignored by code generation and optimization.
10242 '``llvm.trap``' Intrinsic
10243 ^^^^^^^^^^^^^^^^^^^^^^^^^
10250 declare void @llvm.trap() noreturn nounwind
10255 The '``llvm.trap``' intrinsic.
10265 This intrinsic is lowered to the target dependent trap instruction. If
10266 the target does not have a trap instruction, this intrinsic will be
10267 lowered to a call of the ``abort()`` function.
10269 '``llvm.debugtrap``' Intrinsic
10270 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10277 declare void @llvm.debugtrap() nounwind
10282 The '``llvm.debugtrap``' intrinsic.
10292 This intrinsic is lowered to code which is intended to cause an
10293 execution trap with the intention of requesting the attention of a
10296 '``llvm.stackprotector``' Intrinsic
10297 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10304 declare void @llvm.stackprotector(i8* <guard>, i8** <slot>)
10309 The ``llvm.stackprotector`` intrinsic takes the ``guard`` and stores it
10310 onto the stack at ``slot``. The stack slot is adjusted to ensure that it
10311 is placed on the stack before local variables.
10316 The ``llvm.stackprotector`` intrinsic requires two pointer arguments.
10317 The first argument is the value loaded from the stack guard
10318 ``@__stack_chk_guard``. The second variable is an ``alloca`` that has
10319 enough space to hold the value of the guard.
10324 This intrinsic causes the prologue/epilogue inserter to force the position of
10325 the ``AllocaInst`` stack slot to be before local variables on the stack. This is
10326 to ensure that if a local variable on the stack is overwritten, it will destroy
10327 the value of the guard. When the function exits, the guard on the stack is
10328 checked against the original guard by ``llvm.stackprotectorcheck``. If they are
10329 different, then ``llvm.stackprotectorcheck`` causes the program to abort by
10330 calling the ``__stack_chk_fail()`` function.
10332 '``llvm.stackprotectorcheck``' Intrinsic
10333 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10340 declare void @llvm.stackprotectorcheck(i8** <guard>)
10345 The ``llvm.stackprotectorcheck`` intrinsic compares ``guard`` against an already
10346 created stack protector and if they are not equal calls the
10347 ``__stack_chk_fail()`` function.
10352 The ``llvm.stackprotectorcheck`` intrinsic requires one pointer argument, the
10353 the variable ``@__stack_chk_guard``.
10358 This intrinsic is provided to perform the stack protector check by comparing
10359 ``guard`` with the stack slot created by ``llvm.stackprotector`` and if the
10360 values do not match call the ``__stack_chk_fail()`` function.
10362 The reason to provide this as an IR level intrinsic instead of implementing it
10363 via other IR operations is that in order to perform this operation at the IR
10364 level without an intrinsic, one would need to create additional basic blocks to
10365 handle the success/failure cases. This makes it difficult to stop the stack
10366 protector check from disrupting sibling tail calls in Codegen. With this
10367 intrinsic, we are able to generate the stack protector basic blocks late in
10368 codegen after the tail call decision has occurred.
10370 '``llvm.objectsize``' Intrinsic
10371 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10378 declare i32 @llvm.objectsize.i32(i8* <object>, i1 <min>)
10379 declare i64 @llvm.objectsize.i64(i8* <object>, i1 <min>)
10384 The ``llvm.objectsize`` intrinsic is designed to provide information to
10385 the optimizers to determine at compile time whether a) an operation
10386 (like memcpy) will overflow a buffer that corresponds to an object, or
10387 b) that a runtime check for overflow isn't necessary. An object in this
10388 context means an allocation of a specific class, structure, array, or
10394 The ``llvm.objectsize`` intrinsic takes two arguments. The first
10395 argument is a pointer to or into the ``object``. The second argument is
10396 a boolean and determines whether ``llvm.objectsize`` returns 0 (if true)
10397 or -1 (if false) when the object size is unknown. The second argument
10398 only accepts constants.
10403 The ``llvm.objectsize`` intrinsic is lowered to a constant representing
10404 the size of the object concerned. If the size cannot be determined at
10405 compile time, ``llvm.objectsize`` returns ``i32/i64 -1 or 0`` (depending
10406 on the ``min`` argument).
10408 '``llvm.expect``' Intrinsic
10409 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
10414 This is an overloaded intrinsic. You can use ``llvm.expect`` on any
10419 declare i1 @llvm.expect.i1(i1 <val>, i1 <expected_val>)
10420 declare i32 @llvm.expect.i32(i32 <val>, i32 <expected_val>)
10421 declare i64 @llvm.expect.i64(i64 <val>, i64 <expected_val>)
10426 The ``llvm.expect`` intrinsic provides information about expected (the
10427 most probable) value of ``val``, which can be used by optimizers.
10432 The ``llvm.expect`` intrinsic takes two arguments. The first argument is
10433 a value. The second argument is an expected value, this needs to be a
10434 constant value, variables are not allowed.
10439 This intrinsic is lowered to the ``val``.
10443 '``llvm.assume``' Intrinsic
10444 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10451 declare void @llvm.assume(i1 %cond)
10456 The ``llvm.assume`` allows the optimizer to assume that the provided
10457 condition is true. This information can then be used in simplifying other parts
10463 The condition which the optimizer may assume is always true.
10468 The intrinsic allows the optimizer to assume that the provided condition is
10469 always true whenever the control flow reaches the intrinsic call. No code is
10470 generated for this intrinsic, and instructions that contribute only to the
10471 provided condition are not used for code generation. If the condition is
10472 violated during execution, the behavior is undefined.
10474 Note that the optimizer might limit the transformations performed on values
10475 used by the ``llvm.assume`` intrinsic in order to preserve the instructions
10476 only used to form the intrinsic's input argument. This might prove undesirable
10477 if the extra information provided by the ``llvm.assume`` intrinsic does not cause
10478 sufficient overall improvement in code quality. For this reason,
10479 ``llvm.assume`` should not be used to document basic mathematical invariants
10480 that the optimizer can otherwise deduce or facts that are of little use to the
10485 '``llvm.bitset.test``' Intrinsic
10486 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10493 declare i1 @llvm.bitset.test(i8* %ptr, metadata %bitset) nounwind readnone
10499 The first argument is a pointer to be tested. The second argument is a
10500 metadata string containing the name of a :doc:`bitset <BitSets>`.
10505 The ``llvm.bitset.test`` intrinsic tests whether the given pointer is a
10506 member of the given bitset.
10508 '``llvm.donothing``' Intrinsic
10509 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10516 declare void @llvm.donothing() nounwind readnone
10521 The ``llvm.donothing`` intrinsic doesn't perform any operation. It's one of only
10522 two intrinsics (besides ``llvm.experimental.patchpoint``) that can be called
10523 with an invoke instruction.
10533 This intrinsic does nothing, and it's removed by optimizers and ignored
10536 Stack Map Intrinsics
10537 --------------------
10539 LLVM provides experimental intrinsics to support runtime patching
10540 mechanisms commonly desired in dynamic language JITs. These intrinsics
10541 are described in :doc:`StackMaps`.