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>`,
639 an optional :ref:`personality <personalityfn>`,
640 an opening curly brace, a list of basic blocks, and a closing curly brace.
642 LLVM function declarations consist of the "``declare``" keyword, an
643 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
644 style <visibility>`, an optional :ref:`DLL storage class <dllstorageclass>`,
645 an optional :ref:`calling convention <callingconv>`,
646 an optional ``unnamed_addr`` attribute, a return type, an optional
647 :ref:`parameter attribute <paramattrs>` for the return type, a function
648 name, a possibly empty list of arguments, an optional alignment, an optional
649 :ref:`garbage collector name <gc>`, an optional :ref:`prefix <prefixdata>`,
650 and an optional :ref:`prologue <prologuedata>`.
652 A function definition contains a list of basic blocks, forming the CFG (Control
653 Flow Graph) for the function. Each basic block may optionally start with a label
654 (giving the basic block a symbol table entry), contains a list of instructions,
655 and ends with a :ref:`terminator <terminators>` instruction (such as a branch or
656 function return). If an explicit label is not provided, a block is assigned an
657 implicit numbered label, using the next value from the same counter as used for
658 unnamed temporaries (:ref:`see above<identifiers>`). For example, if a function
659 entry block does not have an explicit label, it will be assigned label "%0",
660 then the first unnamed temporary in that block will be "%1", etc.
662 The first basic block in a function is special in two ways: it is
663 immediately executed on entrance to the function, and it is not allowed
664 to have predecessor basic blocks (i.e. there can not be any branches to
665 the entry block of a function). Because the block can have no
666 predecessors, it also cannot have any :ref:`PHI nodes <i_phi>`.
668 LLVM allows an explicit section to be specified for functions. If the
669 target supports it, it will emit functions to the section specified.
670 Additionally, the function can be placed in a COMDAT.
672 An explicit alignment may be specified for a function. If not present,
673 or if the alignment is set to zero, the alignment of the function is set
674 by the target to whatever it feels convenient. If an explicit alignment
675 is specified, the function is forced to have at least that much
676 alignment. All alignments must be a power of 2.
678 If the ``unnamed_addr`` attribute is given, the address is known to not
679 be significant and two identical functions can be merged.
683 define [linkage] [visibility] [DLLStorageClass]
685 <ResultType> @<FunctionName> ([argument list])
686 [unnamed_addr] [fn Attrs] [section "name"] [comdat [($name)]]
687 [align N] [gc] [prefix Constant] [prologue Constant]
688 [personality Constant] { ... }
690 The argument list is a comma seperated sequence of arguments where each
691 argument is of the following form
695 <type> [parameter Attrs] [name]
703 Aliases, unlike function or variables, don't create any new data. They
704 are just a new symbol and metadata for an existing position.
706 Aliases have a name and an aliasee that is either a global value or a
709 Aliases may have an optional :ref:`linkage type <linkage>`, an optional
710 :ref:`visibility style <visibility>`, an optional :ref:`DLL storage class
711 <dllstorageclass>` and an optional :ref:`tls model <tls_model>`.
715 @<Name> = [Linkage] [Visibility] [DLLStorageClass] [ThreadLocal] [unnamed_addr] alias <AliaseeTy> @<Aliasee>
717 The linkage must be one of ``private``, ``internal``, ``linkonce``, ``weak``,
718 ``linkonce_odr``, ``weak_odr``, ``external``. Note that some system linkers
719 might not correctly handle dropping a weak symbol that is aliased.
721 Aliases that are not ``unnamed_addr`` are guaranteed to have the same address as
722 the aliasee expression. ``unnamed_addr`` ones are only guaranteed to point
725 Since aliases are only a second name, some restrictions apply, of which
726 some can only be checked when producing an object file:
728 * The expression defining the aliasee must be computable at assembly
729 time. Since it is just a name, no relocations can be used.
731 * No alias in the expression can be weak as the possibility of the
732 intermediate alias being overridden cannot be represented in an
735 * No global value in the expression can be a declaration, since that
736 would require a relocation, which is not possible.
743 Comdat IR provides access to COFF and ELF object file COMDAT functionality.
745 Comdats have a name which represents the COMDAT key. All global objects that
746 specify this key will only end up in the final object file if the linker chooses
747 that key over some other key. Aliases are placed in the same COMDAT that their
748 aliasee computes to, if any.
750 Comdats have a selection kind to provide input on how the linker should
751 choose between keys in two different object files.
755 $<Name> = comdat SelectionKind
757 The selection kind must be one of the following:
760 The linker may choose any COMDAT key, the choice is arbitrary.
762 The linker may choose any COMDAT key but the sections must contain the
765 The linker will choose the section containing the largest COMDAT key.
767 The linker requires that only section with this COMDAT key exist.
769 The linker may choose any COMDAT key but the sections must contain the
772 Note that the Mach-O platform doesn't support COMDATs and ELF only supports
773 ``any`` as a selection kind.
775 Here is an example of a COMDAT group where a function will only be selected if
776 the COMDAT key's section is the largest:
780 $foo = comdat largest
781 @foo = global i32 2, comdat($foo)
783 define void @bar() comdat($foo) {
787 As a syntactic sugar the ``$name`` can be omitted if the name is the same as
793 @foo = global i32 2, comdat
796 In a COFF object file, this will create a COMDAT section with selection kind
797 ``IMAGE_COMDAT_SELECT_LARGEST`` containing the contents of the ``@foo`` symbol
798 and another COMDAT section with selection kind
799 ``IMAGE_COMDAT_SELECT_ASSOCIATIVE`` which is associated with the first COMDAT
800 section and contains the contents of the ``@bar`` symbol.
802 There are some restrictions on the properties of the global object.
803 It, or an alias to it, must have the same name as the COMDAT group when
805 The contents and size of this object may be used during link-time to determine
806 which COMDAT groups get selected depending on the selection kind.
807 Because the name of the object must match the name of the COMDAT group, the
808 linkage of the global object must not be local; local symbols can get renamed
809 if a collision occurs in the symbol table.
811 The combined use of COMDATS and section attributes may yield surprising results.
818 @g1 = global i32 42, section "sec", comdat($foo)
819 @g2 = global i32 42, section "sec", comdat($bar)
821 From the object file perspective, this requires the creation of two sections
822 with the same name. This is necessary because both globals belong to different
823 COMDAT groups and COMDATs, at the object file level, are represented by
826 Note that certain IR constructs like global variables and functions may
827 create COMDATs in the object file in addition to any which are specified using
828 COMDAT IR. This arises when the code generator is configured to emit globals
829 in individual sections (e.g. when `-data-sections` or `-function-sections`
830 is supplied to `llc`).
832 .. _namedmetadatastructure:
837 Named metadata is a collection of metadata. :ref:`Metadata
838 nodes <metadata>` (but not metadata strings) are the only valid
839 operands for a named metadata.
841 #. Named metadata are represented as a string of characters with the
842 metadata prefix. The rules for metadata names are the same as for
843 identifiers, but quoted names are not allowed. ``"\xx"`` type escapes
844 are still valid, which allows any character to be part of a name.
848 ; Some unnamed metadata nodes, which are referenced by the named metadata.
853 !name = !{!0, !1, !2}
860 The return type and each parameter of a function type may have a set of
861 *parameter attributes* associated with them. Parameter attributes are
862 used to communicate additional information about the result or
863 parameters of a function. Parameter attributes are considered to be part
864 of the function, not of the function type, so functions with different
865 parameter attributes can have the same function type.
867 Parameter attributes are simple keywords that follow the type specified.
868 If multiple parameter attributes are needed, they are space separated.
873 declare i32 @printf(i8* noalias nocapture, ...)
874 declare i32 @atoi(i8 zeroext)
875 declare signext i8 @returns_signed_char()
877 Note that any attributes for the function result (``nounwind``,
878 ``readonly``) come immediately after the argument list.
880 Currently, only the following parameter attributes are defined:
883 This indicates to the code generator that the parameter or return
884 value should be zero-extended to the extent required by the target's
885 ABI (which is usually 32-bits, but is 8-bits for a i1 on x86-64) by
886 the caller (for a parameter) or the callee (for a return value).
888 This indicates to the code generator that the parameter or return
889 value should be sign-extended to the extent required by the target's
890 ABI (which is usually 32-bits) by the caller (for a parameter) or
891 the callee (for a return value).
893 This indicates that this parameter or return value should be treated
894 in a special target-dependent fashion during while emitting code for
895 a function call or return (usually, by putting it in a register as
896 opposed to memory, though some targets use it to distinguish between
897 two different kinds of registers). Use of this attribute is
900 This indicates that the pointer parameter should really be passed by
901 value to the function. The attribute implies that a hidden copy of
902 the pointee is made between the caller and the callee, so the callee
903 is unable to modify the value in the caller. This attribute is only
904 valid on LLVM pointer arguments. It is generally used to pass
905 structs and arrays by value, but is also valid on pointers to
906 scalars. The copy is considered to belong to the caller not the
907 callee (for example, ``readonly`` functions should not write to
908 ``byval`` parameters). This is not a valid attribute for return
911 The byval attribute also supports specifying an alignment with the
912 align attribute. It indicates the alignment of the stack slot to
913 form and the known alignment of the pointer specified to the call
914 site. If the alignment is not specified, then the code generator
915 makes a target-specific assumption.
921 The ``inalloca`` argument attribute allows the caller to take the
922 address of outgoing stack arguments. An ``inalloca`` argument must
923 be a pointer to stack memory produced by an ``alloca`` instruction.
924 The alloca, or argument allocation, must also be tagged with the
925 inalloca keyword. Only the last argument may have the ``inalloca``
926 attribute, and that argument is guaranteed to be passed in memory.
928 An argument allocation may be used by a call at most once because
929 the call may deallocate it. The ``inalloca`` attribute cannot be
930 used in conjunction with other attributes that affect argument
931 storage, like ``inreg``, ``nest``, ``sret``, or ``byval``. The
932 ``inalloca`` attribute also disables LLVM's implicit lowering of
933 large aggregate return values, which means that frontend authors
934 must lower them with ``sret`` pointers.
936 When the call site is reached, the argument allocation must have
937 been the most recent stack allocation that is still live, or the
938 results are undefined. It is possible to allocate additional stack
939 space after an argument allocation and before its call site, but it
940 must be cleared off with :ref:`llvm.stackrestore
943 See :doc:`InAlloca` for more information on how to use this
947 This indicates that the pointer parameter specifies the address of a
948 structure that is the return value of the function in the source
949 program. This pointer must be guaranteed by the caller to be valid:
950 loads and stores to the structure may be assumed by the callee
951 not to trap and to be properly aligned. This may only be applied to
952 the first parameter. This is not a valid attribute for return
956 This indicates that the pointer value may be assumed by the optimizer to
957 have the specified alignment.
959 Note that this attribute has additional semantics when combined with the
965 This indicates that objects accessed via pointer values
966 :ref:`based <pointeraliasing>` on the argument or return value are not also
967 accessed, during the execution of the function, via pointer values not
968 *based* on the argument or return value. The attribute on a return value
969 also has additional semantics described below. The caller shares the
970 responsibility with the callee for ensuring that these requirements are met.
971 For further details, please see the discussion of the NoAlias response in
972 :ref:`alias analysis <Must, May, or No>`.
974 Note that this definition of ``noalias`` is intentionally similar
975 to the definition of ``restrict`` in C99 for function arguments.
977 For function return values, C99's ``restrict`` is not meaningful,
978 while LLVM's ``noalias`` is. Furthermore, the semantics of the ``noalias``
979 attribute on return values are stronger than the semantics of the attribute
980 when used on function arguments. On function return values, the ``noalias``
981 attribute indicates that the function acts like a system memory allocation
982 function, returning a pointer to allocated storage disjoint from the
983 storage for any other object accessible to the caller.
986 This indicates that the callee does not make any copies of the
987 pointer that outlive the callee itself. This is not a valid
988 attribute for return values.
993 This indicates that the pointer parameter can be excised using the
994 :ref:`trampoline intrinsics <int_trampoline>`. This is not a valid
995 attribute for return values and can only be applied to one parameter.
998 This indicates that the function always returns the argument as its return
999 value. This is an optimization hint to the code generator when generating
1000 the caller, allowing tail call optimization and omission of register saves
1001 and restores in some cases; it is not checked or enforced when generating
1002 the callee. The parameter and the function return type must be valid
1003 operands for the :ref:`bitcast instruction <i_bitcast>`. This is not a
1004 valid attribute for return values and can only be applied to one parameter.
1007 This indicates that the parameter or return pointer is not null. This
1008 attribute may only be applied to pointer typed parameters. This is not
1009 checked or enforced by LLVM, the caller must ensure that the pointer
1010 passed in is non-null, or the callee must ensure that the returned pointer
1013 ``dereferenceable(<n>)``
1014 This indicates that the parameter or return pointer is dereferenceable. This
1015 attribute may only be applied to pointer typed parameters. A pointer that
1016 is dereferenceable can be loaded from speculatively without a risk of
1017 trapping. The number of bytes known to be dereferenceable must be provided
1018 in parentheses. It is legal for the number of bytes to be less than the
1019 size of the pointee type. The ``nonnull`` attribute does not imply
1020 dereferenceability (consider a pointer to one element past the end of an
1021 array), however ``dereferenceable(<n>)`` does imply ``nonnull`` in
1022 ``addrspace(0)`` (which is the default address space).
1024 ``dereferenceable_or_null(<n>)``
1025 This indicates that the parameter or return value isn't both
1026 non-null and non-dereferenceable (up to ``<n>`` bytes) at the same
1027 time. All non-null pointers tagged with
1028 ``dereferenceable_or_null(<n>)`` are ``dereferenceable(<n>)``.
1029 For address space 0 ``dereferenceable_or_null(<n>)`` implies that
1030 a pointer is exactly one of ``dereferenceable(<n>)`` or ``null``,
1031 and in other address spaces ``dereferenceable_or_null(<n>)``
1032 implies that a pointer is at least one of ``dereferenceable(<n>)``
1033 or ``null`` (i.e. it may be both ``null`` and
1034 ``dereferenceable(<n>)``). This attribute may only be applied to
1035 pointer typed parameters.
1039 Garbage Collector Strategy Names
1040 --------------------------------
1042 Each function may specify a garbage collector strategy name, which is simply a
1045 .. code-block:: llvm
1047 define void @f() gc "name" { ... }
1049 The supported values of *name* includes those :ref:`built in to LLVM
1050 <builtin-gc-strategies>` and any provided by loaded plugins. Specifying a GC
1051 strategy will cause the compiler to alter its output in order to support the
1052 named garbage collection algorithm. Note that LLVM itself does not contain a
1053 garbage collector, this functionality is restricted to generating machine code
1054 which can interoperate with a collector provided externally.
1061 Prefix data is data associated with a function which the code
1062 generator will emit immediately before the function's entrypoint.
1063 The purpose of this feature is to allow frontends to associate
1064 language-specific runtime metadata with specific functions and make it
1065 available through the function pointer while still allowing the
1066 function pointer to be called.
1068 To access the data for a given function, a program may bitcast the
1069 function pointer to a pointer to the constant's type and dereference
1070 index -1. This implies that the IR symbol points just past the end of
1071 the prefix data. For instance, take the example of a function annotated
1072 with a single ``i32``,
1074 .. code-block:: llvm
1076 define void @f() prefix i32 123 { ... }
1078 The prefix data can be referenced as,
1080 .. code-block:: llvm
1082 %0 = bitcast void* () @f to i32*
1083 %a = getelementptr inbounds i32, i32* %0, i32 -1
1084 %b = load i32, i32* %a
1086 Prefix data is laid out as if it were an initializer for a global variable
1087 of the prefix data's type. The function will be placed such that the
1088 beginning of the prefix data is aligned. This means that if the size
1089 of the prefix data is not a multiple of the alignment size, the
1090 function's entrypoint will not be aligned. If alignment of the
1091 function's entrypoint is desired, padding must be added to the prefix
1094 A function may have prefix data but no body. This has similar semantics
1095 to the ``available_externally`` linkage in that the data may be used by the
1096 optimizers but will not be emitted in the object file.
1103 The ``prologue`` attribute allows arbitrary code (encoded as bytes) to
1104 be inserted prior to the function body. This can be used for enabling
1105 function hot-patching and instrumentation.
1107 To maintain the semantics of ordinary function calls, the prologue data must
1108 have a particular format. Specifically, it must begin with a sequence of
1109 bytes which decode to a sequence of machine instructions, valid for the
1110 module's target, which transfer control to the point immediately succeeding
1111 the prologue data, without performing any other visible action. This allows
1112 the inliner and other passes to reason about the semantics of the function
1113 definition without needing to reason about the prologue data. Obviously this
1114 makes the format of the prologue data highly target dependent.
1116 A trivial example of valid prologue data for the x86 architecture is ``i8 144``,
1117 which encodes the ``nop`` instruction:
1119 .. code-block:: llvm
1121 define void @f() prologue i8 144 { ... }
1123 Generally prologue data can be formed by encoding a relative branch instruction
1124 which skips the metadata, as in this example of valid prologue data for the
1125 x86_64 architecture, where the first two bytes encode ``jmp .+10``:
1127 .. code-block:: llvm
1129 %0 = type <{ i8, i8, i8* }>
1131 define void @f() prologue %0 <{ i8 235, i8 8, i8* @md}> { ... }
1133 A function may have prologue data but no body. This has similar semantics
1134 to the ``available_externally`` linkage in that the data may be used by the
1135 optimizers but will not be emitted in the object file.
1139 Personality Function
1140 --------------------
1142 The ``personality`` attribute permits functions to specify what function
1143 to use for exception handling.
1150 Attribute groups are groups of attributes that are referenced by objects within
1151 the IR. They are important for keeping ``.ll`` files readable, because a lot of
1152 functions will use the same set of attributes. In the degenerative case of a
1153 ``.ll`` file that corresponds to a single ``.c`` file, the single attribute
1154 group will capture the important command line flags used to build that file.
1156 An attribute group is a module-level object. To use an attribute group, an
1157 object references the attribute group's ID (e.g. ``#37``). An object may refer
1158 to more than one attribute group. In that situation, the attributes from the
1159 different groups are merged.
1161 Here is an example of attribute groups for a function that should always be
1162 inlined, has a stack alignment of 4, and which shouldn't use SSE instructions:
1164 .. code-block:: llvm
1166 ; Target-independent attributes:
1167 attributes #0 = { alwaysinline alignstack=4 }
1169 ; Target-dependent attributes:
1170 attributes #1 = { "no-sse" }
1172 ; Function @f has attributes: alwaysinline, alignstack=4, and "no-sse".
1173 define void @f() #0 #1 { ... }
1180 Function attributes are set to communicate additional information about
1181 a function. Function attributes are considered to be part of the
1182 function, not of the function type, so functions with different function
1183 attributes can have the same function type.
1185 Function attributes are simple keywords that follow the type specified.
1186 If multiple attributes are needed, they are space separated. For
1189 .. code-block:: llvm
1191 define void @f() noinline { ... }
1192 define void @f() alwaysinline { ... }
1193 define void @f() alwaysinline optsize { ... }
1194 define void @f() optsize { ... }
1197 This attribute indicates that, when emitting the prologue and
1198 epilogue, the backend should forcibly align the stack pointer.
1199 Specify the desired alignment, which must be a power of two, in
1202 This attribute indicates that the inliner should attempt to inline
1203 this function into callers whenever possible, ignoring any active
1204 inlining size threshold for this caller.
1206 This indicates that the callee function at a call site should be
1207 recognized as a built-in function, even though the function's declaration
1208 uses the ``nobuiltin`` attribute. This is only valid at call sites for
1209 direct calls to functions that are declared with the ``nobuiltin``
1212 This attribute indicates that this function is rarely called. When
1213 computing edge weights, basic blocks post-dominated by a cold
1214 function call are also considered to be cold; and, thus, given low
1217 This attribute indicates that the callee is dependent on a convergent
1218 thread execution pattern under certain parallel execution models.
1219 Transformations that are execution model agnostic may only move or
1220 tranform this call if the final location is control equivalent to its
1221 original position in the program, where control equivalence is defined as
1222 A dominates B and B post-dominates A, or vice versa.
1224 This attribute indicates that the source code contained a hint that
1225 inlining this function is desirable (such as the "inline" keyword in
1226 C/C++). It is just a hint; it imposes no requirements on the
1229 This attribute indicates that the function should be added to a
1230 jump-instruction table at code-generation time, and that all address-taken
1231 references to this function should be replaced with a reference to the
1232 appropriate jump-instruction-table function pointer. Note that this creates
1233 a new pointer for the original function, which means that code that depends
1234 on function-pointer identity can break. So, any function annotated with
1235 ``jumptable`` must also be ``unnamed_addr``.
1237 This attribute suggests that optimization passes and code generator
1238 passes make choices that keep the code size of this function as small
1239 as possible and perform optimizations that may sacrifice runtime
1240 performance in order to minimize the size of the generated code.
1242 This attribute disables prologue / epilogue emission for the
1243 function. This can have very system-specific consequences.
1245 This indicates that the callee function at a call site is not recognized as
1246 a built-in function. LLVM will retain the original call and not replace it
1247 with equivalent code based on the semantics of the built-in function, unless
1248 the call site uses the ``builtin`` attribute. This is valid at call sites
1249 and on function declarations and definitions.
1251 This attribute indicates that calls to the function cannot be
1252 duplicated. A call to a ``noduplicate`` function may be moved
1253 within its parent function, but may not be duplicated within
1254 its parent function.
1256 A function containing a ``noduplicate`` call may still
1257 be an inlining candidate, provided that the call is not
1258 duplicated by inlining. That implies that the function has
1259 internal linkage and only has one call site, so the original
1260 call is dead after inlining.
1262 This attributes disables implicit floating point instructions.
1264 This attribute indicates that the inliner should never inline this
1265 function in any situation. This attribute may not be used together
1266 with the ``alwaysinline`` attribute.
1268 This attribute suppresses lazy symbol binding for the function. This
1269 may make calls to the function faster, at the cost of extra program
1270 startup time if the function is not called during program startup.
1272 This attribute indicates that the code generator should not use a
1273 red zone, even if the target-specific ABI normally permits it.
1275 This function attribute indicates that the function never returns
1276 normally. This produces undefined behavior at runtime if the
1277 function ever does dynamically return.
1279 This function attribute indicates that the function never raises an
1280 exception. If the function does raise an exception, its runtime
1281 behavior is undefined. However, functions marked nounwind may still
1282 trap or generate asynchronous exceptions. Exception handling schemes
1283 that are recognized by LLVM to handle asynchronous exceptions, such
1284 as SEH, will still provide their implementation defined semantics.
1286 This function attribute indicates that the function is not optimized
1287 by any optimization or code generator passes with the
1288 exception of interprocedural optimization passes.
1289 This attribute cannot be used together with the ``alwaysinline``
1290 attribute; this attribute is also incompatible
1291 with the ``minsize`` attribute and the ``optsize`` attribute.
1293 This attribute requires the ``noinline`` attribute to be specified on
1294 the function as well, so the function is never inlined into any caller.
1295 Only functions with the ``alwaysinline`` attribute are valid
1296 candidates for inlining into the body of this function.
1298 This attribute suggests that optimization passes and code generator
1299 passes make choices that keep the code size of this function low,
1300 and otherwise do optimizations specifically to reduce code size as
1301 long as they do not significantly impact runtime performance.
1303 On a function, this attribute indicates that the function computes its
1304 result (or decides to unwind an exception) based strictly on its arguments,
1305 without dereferencing any pointer arguments or otherwise accessing
1306 any mutable state (e.g. memory, control registers, etc) visible to
1307 caller functions. It does not write through any pointer arguments
1308 (including ``byval`` arguments) and never changes any state visible
1309 to callers. This means that it cannot unwind exceptions by calling
1310 the ``C++`` exception throwing methods.
1312 On an argument, this attribute indicates that the function does not
1313 dereference that pointer argument, even though it may read or write the
1314 memory that the pointer points to if accessed through other pointers.
1316 On a function, this attribute indicates that the function does not write
1317 through any pointer arguments (including ``byval`` arguments) or otherwise
1318 modify any state (e.g. memory, control registers, etc) visible to
1319 caller functions. It may dereference pointer arguments and read
1320 state that may be set in the caller. A readonly function always
1321 returns the same value (or unwinds an exception identically) when
1322 called with the same set of arguments and global state. It cannot
1323 unwind an exception by calling the ``C++`` exception throwing
1326 On an argument, this attribute indicates that the function does not write
1327 through this pointer argument, even though it may write to the memory that
1328 the pointer points to.
1330 This attribute indicates that this function can return twice. The C
1331 ``setjmp`` is an example of such a function. The compiler disables
1332 some optimizations (like tail calls) in the caller of these
1335 This attribute indicates that
1336 `SafeStack <http://clang.llvm.org/docs/SafeStack.html>`_
1337 protection is enabled for this function.
1339 If a function that has a ``safestack`` attribute is inlined into a
1340 function that doesn't have a ``safestack`` attribute or which has an
1341 ``ssp``, ``sspstrong`` or ``sspreq`` attribute, then the resulting
1342 function will have a ``safestack`` attribute.
1343 ``sanitize_address``
1344 This attribute indicates that AddressSanitizer checks
1345 (dynamic address safety analysis) are enabled for this function.
1347 This attribute indicates that MemorySanitizer checks (dynamic detection
1348 of accesses to uninitialized memory) are enabled for this function.
1350 This attribute indicates that ThreadSanitizer checks
1351 (dynamic thread safety analysis) are enabled for this function.
1353 This attribute indicates that the function should emit a stack
1354 smashing protector. It is in the form of a "canary" --- a random value
1355 placed on the stack before the local variables that's checked upon
1356 return from the function to see if it has been overwritten. A
1357 heuristic is used to determine if a function needs stack protectors
1358 or not. The heuristic used will enable protectors for functions with:
1360 - Character arrays larger than ``ssp-buffer-size`` (default 8).
1361 - Aggregates containing character arrays larger than ``ssp-buffer-size``.
1362 - Calls to alloca() with variable sizes or constant sizes greater than
1363 ``ssp-buffer-size``.
1365 Variables that are identified as requiring a protector will be arranged
1366 on the stack such that they are adjacent to the stack protector guard.
1368 If a function that has an ``ssp`` attribute is inlined into a
1369 function that doesn't have an ``ssp`` attribute, then the resulting
1370 function will have an ``ssp`` attribute.
1372 This attribute indicates that the function should *always* emit a
1373 stack smashing protector. This overrides the ``ssp`` function
1376 Variables that are identified as requiring a protector will be arranged
1377 on the stack such that they are adjacent to the stack protector guard.
1378 The specific layout rules are:
1380 #. Large arrays and structures containing large arrays
1381 (``>= ssp-buffer-size``) are closest to the stack protector.
1382 #. Small arrays and structures containing small arrays
1383 (``< ssp-buffer-size``) are 2nd closest to the protector.
1384 #. Variables that have had their address taken are 3rd closest to the
1387 If a function that has an ``sspreq`` attribute is inlined into a
1388 function that doesn't have an ``sspreq`` attribute or which has an
1389 ``ssp`` or ``sspstrong`` attribute, then the resulting function will have
1390 an ``sspreq`` attribute.
1392 This attribute indicates that the function should emit a stack smashing
1393 protector. This attribute causes a strong heuristic to be used when
1394 determining if a function needs stack protectors. The strong heuristic
1395 will enable protectors for functions with:
1397 - Arrays of any size and type
1398 - Aggregates containing an array of any size and type.
1399 - Calls to alloca().
1400 - Local variables that have had their address taken.
1402 Variables that are identified as requiring a protector will be arranged
1403 on the stack such that they are adjacent to the stack protector guard.
1404 The specific layout rules are:
1406 #. Large arrays and structures containing large arrays
1407 (``>= ssp-buffer-size``) are closest to the stack protector.
1408 #. Small arrays and structures containing small arrays
1409 (``< ssp-buffer-size``) are 2nd closest to the protector.
1410 #. Variables that have had their address taken are 3rd closest to the
1413 This overrides the ``ssp`` function attribute.
1415 If a function that has an ``sspstrong`` attribute is inlined into a
1416 function that doesn't have an ``sspstrong`` attribute, then the
1417 resulting function will have an ``sspstrong`` attribute.
1419 This attribute indicates that the function will delegate to some other
1420 function with a tail call. The prototype of a thunk should not be used for
1421 optimization purposes. The caller is expected to cast the thunk prototype to
1422 match the thunk target prototype.
1424 This attribute indicates that the ABI being targeted requires that
1425 an unwind table entry be produce for this function even if we can
1426 show that no exceptions passes by it. This is normally the case for
1427 the ELF x86-64 abi, but it can be disabled for some compilation
1432 Module-Level Inline Assembly
1433 ----------------------------
1435 Modules may contain "module-level inline asm" blocks, which corresponds
1436 to the GCC "file scope inline asm" blocks. These blocks are internally
1437 concatenated by LLVM and treated as a single unit, but may be separated
1438 in the ``.ll`` file if desired. The syntax is very simple:
1440 .. code-block:: llvm
1442 module asm "inline asm code goes here"
1443 module asm "more can go here"
1445 The strings can contain any character by escaping non-printable
1446 characters. The escape sequence used is simply "\\xx" where "xx" is the
1447 two digit hex code for the number.
1449 The inline asm code is simply printed to the machine code .s file when
1450 assembly code is generated.
1452 .. _langref_datalayout:
1457 A module may specify a target specific data layout string that specifies
1458 how data is to be laid out in memory. The syntax for the data layout is
1461 .. code-block:: llvm
1463 target datalayout = "layout specification"
1465 The *layout specification* consists of a list of specifications
1466 separated by the minus sign character ('-'). Each specification starts
1467 with a letter and may include other information after the letter to
1468 define some aspect of the data layout. The specifications accepted are
1472 Specifies that the target lays out data in big-endian form. That is,
1473 the bits with the most significance have the lowest address
1476 Specifies that the target lays out data in little-endian form. That
1477 is, the bits with the least significance have the lowest address
1480 Specifies the natural alignment of the stack in bits. Alignment
1481 promotion of stack variables is limited to the natural stack
1482 alignment to avoid dynamic stack realignment. The stack alignment
1483 must be a multiple of 8-bits. If omitted, the natural stack
1484 alignment defaults to "unspecified", which does not prevent any
1485 alignment promotions.
1486 ``p[n]:<size>:<abi>:<pref>``
1487 This specifies the *size* of a pointer and its ``<abi>`` and
1488 ``<pref>``\erred alignments for address space ``n``. All sizes are in
1489 bits. The address space, ``n`` is optional, and if not specified,
1490 denotes the default address space 0. The value of ``n`` must be
1491 in the range [1,2^23).
1492 ``i<size>:<abi>:<pref>``
1493 This specifies the alignment for an integer type of a given bit
1494 ``<size>``. The value of ``<size>`` must be in the range [1,2^23).
1495 ``v<size>:<abi>:<pref>``
1496 This specifies the alignment for a vector type of a given bit
1498 ``f<size>:<abi>:<pref>``
1499 This specifies the alignment for a floating point type of a given bit
1500 ``<size>``. Only values of ``<size>`` that are supported by the target
1501 will work. 32 (float) and 64 (double) are supported on all targets; 80
1502 or 128 (different flavors of long double) are also supported on some
1505 This specifies the alignment for an object of aggregate type.
1507 If present, specifies that llvm names are mangled in the output. The
1510 * ``e``: ELF mangling: Private symbols get a ``.L`` prefix.
1511 * ``m``: Mips mangling: Private symbols get a ``$`` prefix.
1512 * ``o``: Mach-O mangling: Private symbols get ``L`` prefix. Other
1513 symbols get a ``_`` prefix.
1514 * ``w``: Windows COFF prefix: Similar to Mach-O, but stdcall and fastcall
1515 functions also get a suffix based on the frame size.
1516 ``n<size1>:<size2>:<size3>...``
1517 This specifies a set of native integer widths for the target CPU in
1518 bits. For example, it might contain ``n32`` for 32-bit PowerPC,
1519 ``n32:64`` for PowerPC 64, or ``n8:16:32:64`` for X86-64. Elements of
1520 this set are considered to support most general arithmetic operations
1523 On every specification that takes a ``<abi>:<pref>``, specifying the
1524 ``<pref>`` alignment is optional. If omitted, the preceding ``:``
1525 should be omitted too and ``<pref>`` will be equal to ``<abi>``.
1527 When constructing the data layout for a given target, LLVM starts with a
1528 default set of specifications which are then (possibly) overridden by
1529 the specifications in the ``datalayout`` keyword. The default
1530 specifications are given in this list:
1532 - ``E`` - big endian
1533 - ``p:64:64:64`` - 64-bit pointers with 64-bit alignment.
1534 - ``p[n]:64:64:64`` - Other address spaces are assumed to be the
1535 same as the default address space.
1536 - ``S0`` - natural stack alignment is unspecified
1537 - ``i1:8:8`` - i1 is 8-bit (byte) aligned
1538 - ``i8:8:8`` - i8 is 8-bit (byte) aligned
1539 - ``i16:16:16`` - i16 is 16-bit aligned
1540 - ``i32:32:32`` - i32 is 32-bit aligned
1541 - ``i64:32:64`` - i64 has ABI alignment of 32-bits but preferred
1542 alignment of 64-bits
1543 - ``f16:16:16`` - half is 16-bit aligned
1544 - ``f32:32:32`` - float is 32-bit aligned
1545 - ``f64:64:64`` - double is 64-bit aligned
1546 - ``f128:128:128`` - quad is 128-bit aligned
1547 - ``v64:64:64`` - 64-bit vector is 64-bit aligned
1548 - ``v128:128:128`` - 128-bit vector is 128-bit aligned
1549 - ``a:0:64`` - aggregates are 64-bit aligned
1551 When LLVM is determining the alignment for a given type, it uses the
1554 #. If the type sought is an exact match for one of the specifications,
1555 that specification is used.
1556 #. If no match is found, and the type sought is an integer type, then
1557 the smallest integer type that is larger than the bitwidth of the
1558 sought type is used. If none of the specifications are larger than
1559 the bitwidth then the largest integer type is used. For example,
1560 given the default specifications above, the i7 type will use the
1561 alignment of i8 (next largest) while both i65 and i256 will use the
1562 alignment of i64 (largest specified).
1563 #. If no match is found, and the type sought is a vector type, then the
1564 largest vector type that is smaller than the sought vector type will
1565 be used as a fall back. This happens because <128 x double> can be
1566 implemented in terms of 64 <2 x double>, for example.
1568 The function of the data layout string may not be what you expect.
1569 Notably, this is not a specification from the frontend of what alignment
1570 the code generator should use.
1572 Instead, if specified, the target data layout is required to match what
1573 the ultimate *code generator* expects. This string is used by the
1574 mid-level optimizers to improve code, and this only works if it matches
1575 what the ultimate code generator uses. There is no way to generate IR
1576 that does not embed this target-specific detail into the IR. If you
1577 don't specify the string, the default specifications will be used to
1578 generate a Data Layout and the optimization phases will operate
1579 accordingly and introduce target specificity into the IR with respect to
1580 these default specifications.
1587 A module may specify a target triple string that describes the target
1588 host. The syntax for the target triple is simply:
1590 .. code-block:: llvm
1592 target triple = "x86_64-apple-macosx10.7.0"
1594 The *target triple* string consists of a series of identifiers delimited
1595 by the minus sign character ('-'). The canonical forms are:
1599 ARCHITECTURE-VENDOR-OPERATING_SYSTEM
1600 ARCHITECTURE-VENDOR-OPERATING_SYSTEM-ENVIRONMENT
1602 This information is passed along to the backend so that it generates
1603 code for the proper architecture. It's possible to override this on the
1604 command line with the ``-mtriple`` command line option.
1606 .. _pointeraliasing:
1608 Pointer Aliasing Rules
1609 ----------------------
1611 Any memory access must be done through a pointer value associated with
1612 an address range of the memory access, otherwise the behavior is
1613 undefined. Pointer values are associated with address ranges according
1614 to the following rules:
1616 - A pointer value is associated with the addresses associated with any
1617 value it is *based* on.
1618 - An address of a global variable is associated with the address range
1619 of the variable's storage.
1620 - The result value of an allocation instruction is associated with the
1621 address range of the allocated storage.
1622 - A null pointer in the default address-space is associated with no
1624 - An integer constant other than zero or a pointer value returned from
1625 a function not defined within LLVM may be associated with address
1626 ranges allocated through mechanisms other than those provided by
1627 LLVM. Such ranges shall not overlap with any ranges of addresses
1628 allocated by mechanisms provided by LLVM.
1630 A pointer value is *based* on another pointer value according to the
1633 - A pointer value formed from a ``getelementptr`` operation is *based*
1634 on the first value operand of the ``getelementptr``.
1635 - The result value of a ``bitcast`` is *based* on the operand of the
1637 - A pointer value formed by an ``inttoptr`` is *based* on all pointer
1638 values that contribute (directly or indirectly) to the computation of
1639 the pointer's value.
1640 - The "*based* on" relationship is transitive.
1642 Note that this definition of *"based"* is intentionally similar to the
1643 definition of *"based"* in C99, though it is slightly weaker.
1645 LLVM IR does not associate types with memory. The result type of a
1646 ``load`` merely indicates the size and alignment of the memory from
1647 which to load, as well as the interpretation of the value. The first
1648 operand type of a ``store`` similarly only indicates the size and
1649 alignment of the store.
1651 Consequently, type-based alias analysis, aka TBAA, aka
1652 ``-fstrict-aliasing``, is not applicable to general unadorned LLVM IR.
1653 :ref:`Metadata <metadata>` may be used to encode additional information
1654 which specialized optimization passes may use to implement type-based
1659 Volatile Memory Accesses
1660 ------------------------
1662 Certain memory accesses, such as :ref:`load <i_load>`'s,
1663 :ref:`store <i_store>`'s, and :ref:`llvm.memcpy <int_memcpy>`'s may be
1664 marked ``volatile``. The optimizers must not change the number of
1665 volatile operations or change their order of execution relative to other
1666 volatile operations. The optimizers *may* change the order of volatile
1667 operations relative to non-volatile operations. This is not Java's
1668 "volatile" and has no cross-thread synchronization behavior.
1670 IR-level volatile loads and stores cannot safely be optimized into
1671 llvm.memcpy or llvm.memmove intrinsics even when those intrinsics are
1672 flagged volatile. Likewise, the backend should never split or merge
1673 target-legal volatile load/store instructions.
1675 .. admonition:: Rationale
1677 Platforms may rely on volatile loads and stores of natively supported
1678 data width to be executed as single instruction. For example, in C
1679 this holds for an l-value of volatile primitive type with native
1680 hardware support, but not necessarily for aggregate types. The
1681 frontend upholds these expectations, which are intentionally
1682 unspecified in the IR. The rules above ensure that IR transformation
1683 do not violate the frontend's contract with the language.
1687 Memory Model for Concurrent Operations
1688 --------------------------------------
1690 The LLVM IR does not define any way to start parallel threads of
1691 execution or to register signal handlers. Nonetheless, there are
1692 platform-specific ways to create them, and we define LLVM IR's behavior
1693 in their presence. This model is inspired by the C++0x memory model.
1695 For a more informal introduction to this model, see the :doc:`Atomics`.
1697 We define a *happens-before* partial order as the least partial order
1700 - Is a superset of single-thread program order, and
1701 - When a *synchronizes-with* ``b``, includes an edge from ``a`` to
1702 ``b``. *Synchronizes-with* pairs are introduced by platform-specific
1703 techniques, like pthread locks, thread creation, thread joining,
1704 etc., and by atomic instructions. (See also :ref:`Atomic Memory Ordering
1705 Constraints <ordering>`).
1707 Note that program order does not introduce *happens-before* edges
1708 between a thread and signals executing inside that thread.
1710 Every (defined) read operation (load instructions, memcpy, atomic
1711 loads/read-modify-writes, etc.) R reads a series of bytes written by
1712 (defined) write operations (store instructions, atomic
1713 stores/read-modify-writes, memcpy, etc.). For the purposes of this
1714 section, initialized globals are considered to have a write of the
1715 initializer which is atomic and happens before any other read or write
1716 of the memory in question. For each byte of a read R, R\ :sub:`byte`
1717 may see any write to the same byte, except:
1719 - If write\ :sub:`1` happens before write\ :sub:`2`, and
1720 write\ :sub:`2` happens before R\ :sub:`byte`, then
1721 R\ :sub:`byte` does not see write\ :sub:`1`.
1722 - If R\ :sub:`byte` happens before write\ :sub:`3`, then
1723 R\ :sub:`byte` does not see write\ :sub:`3`.
1725 Given that definition, R\ :sub:`byte` is defined as follows:
1727 - If R is volatile, the result is target-dependent. (Volatile is
1728 supposed to give guarantees which can support ``sig_atomic_t`` in
1729 C/C++, and may be used for accesses to addresses that do not behave
1730 like normal memory. It does not generally provide cross-thread
1732 - Otherwise, if there is no write to the same byte that happens before
1733 R\ :sub:`byte`, R\ :sub:`byte` returns ``undef`` for that byte.
1734 - Otherwise, if R\ :sub:`byte` may see exactly one write,
1735 R\ :sub:`byte` returns the value written by that write.
1736 - Otherwise, if R is atomic, and all the writes R\ :sub:`byte` may
1737 see are atomic, it chooses one of the values written. See the :ref:`Atomic
1738 Memory Ordering Constraints <ordering>` section for additional
1739 constraints on how the choice is made.
1740 - Otherwise R\ :sub:`byte` returns ``undef``.
1742 R returns the value composed of the series of bytes it read. This
1743 implies that some bytes within the value may be ``undef`` **without**
1744 the entire value being ``undef``. Note that this only defines the
1745 semantics of the operation; it doesn't mean that targets will emit more
1746 than one instruction to read the series of bytes.
1748 Note that in cases where none of the atomic intrinsics are used, this
1749 model places only one restriction on IR transformations on top of what
1750 is required for single-threaded execution: introducing a store to a byte
1751 which might not otherwise be stored is not allowed in general.
1752 (Specifically, in the case where another thread might write to and read
1753 from an address, introducing a store can change a load that may see
1754 exactly one write into a load that may see multiple writes.)
1758 Atomic Memory Ordering Constraints
1759 ----------------------------------
1761 Atomic instructions (:ref:`cmpxchg <i_cmpxchg>`,
1762 :ref:`atomicrmw <i_atomicrmw>`, :ref:`fence <i_fence>`,
1763 :ref:`atomic load <i_load>`, and :ref:`atomic store <i_store>`) take
1764 ordering parameters that determine which other atomic instructions on
1765 the same address they *synchronize with*. These semantics are borrowed
1766 from Java and C++0x, but are somewhat more colloquial. If these
1767 descriptions aren't precise enough, check those specs (see spec
1768 references in the :doc:`atomics guide <Atomics>`).
1769 :ref:`fence <i_fence>` instructions treat these orderings somewhat
1770 differently since they don't take an address. See that instruction's
1771 documentation for details.
1773 For a simpler introduction to the ordering constraints, see the
1777 The set of values that can be read is governed by the happens-before
1778 partial order. A value cannot be read unless some operation wrote
1779 it. This is intended to provide a guarantee strong enough to model
1780 Java's non-volatile shared variables. This ordering cannot be
1781 specified for read-modify-write operations; it is not strong enough
1782 to make them atomic in any interesting way.
1784 In addition to the guarantees of ``unordered``, there is a single
1785 total order for modifications by ``monotonic`` operations on each
1786 address. All modification orders must be compatible with the
1787 happens-before order. There is no guarantee that the modification
1788 orders can be combined to a global total order for the whole program
1789 (and this often will not be possible). The read in an atomic
1790 read-modify-write operation (:ref:`cmpxchg <i_cmpxchg>` and
1791 :ref:`atomicrmw <i_atomicrmw>`) reads the value in the modification
1792 order immediately before the value it writes. If one atomic read
1793 happens before another atomic read of the same address, the later
1794 read must see the same value or a later value in the address's
1795 modification order. This disallows reordering of ``monotonic`` (or
1796 stronger) operations on the same address. If an address is written
1797 ``monotonic``-ally by one thread, and other threads ``monotonic``-ally
1798 read that address repeatedly, the other threads must eventually see
1799 the write. This corresponds to the C++0x/C1x
1800 ``memory_order_relaxed``.
1802 In addition to the guarantees of ``monotonic``, a
1803 *synchronizes-with* edge may be formed with a ``release`` operation.
1804 This is intended to model C++'s ``memory_order_acquire``.
1806 In addition to the guarantees of ``monotonic``, if this operation
1807 writes a value which is subsequently read by an ``acquire``
1808 operation, it *synchronizes-with* that operation. (This isn't a
1809 complete description; see the C++0x definition of a release
1810 sequence.) This corresponds to the C++0x/C1x
1811 ``memory_order_release``.
1812 ``acq_rel`` (acquire+release)
1813 Acts as both an ``acquire`` and ``release`` operation on its
1814 address. This corresponds to the C++0x/C1x ``memory_order_acq_rel``.
1815 ``seq_cst`` (sequentially consistent)
1816 In addition to the guarantees of ``acq_rel`` (``acquire`` for an
1817 operation that only reads, ``release`` for an operation that only
1818 writes), there is a global total order on all
1819 sequentially-consistent operations on all addresses, which is
1820 consistent with the *happens-before* partial order and with the
1821 modification orders of all the affected addresses. Each
1822 sequentially-consistent read sees the last preceding write to the
1823 same address in this global order. This corresponds to the C++0x/C1x
1824 ``memory_order_seq_cst`` and Java volatile.
1828 If an atomic operation is marked ``singlethread``, it only *synchronizes
1829 with* or participates in modification and seq\_cst total orderings with
1830 other operations running in the same thread (for example, in signal
1838 LLVM IR floating-point binary ops (:ref:`fadd <i_fadd>`,
1839 :ref:`fsub <i_fsub>`, :ref:`fmul <i_fmul>`, :ref:`fdiv <i_fdiv>`,
1840 :ref:`frem <i_frem>`) have the following flags that can be set to enable
1841 otherwise unsafe floating point operations
1844 No NaNs - Allow optimizations to assume the arguments and result are not
1845 NaN. Such optimizations are required to retain defined behavior over
1846 NaNs, but the value of the result is undefined.
1849 No Infs - Allow optimizations to assume the arguments and result are not
1850 +/-Inf. Such optimizations are required to retain defined behavior over
1851 +/-Inf, but the value of the result is undefined.
1854 No Signed Zeros - Allow optimizations to treat the sign of a zero
1855 argument or result as insignificant.
1858 Allow Reciprocal - Allow optimizations to use the reciprocal of an
1859 argument rather than perform division.
1862 Fast - Allow algebraically equivalent transformations that may
1863 dramatically change results in floating point (e.g. reassociate). This
1864 flag implies all the others.
1868 Use-list Order Directives
1869 -------------------------
1871 Use-list directives encode the in-memory order of each use-list, allowing the
1872 order to be recreated. ``<order-indexes>`` is a comma-separated list of
1873 indexes that are assigned to the referenced value's uses. The referenced
1874 value's use-list is immediately sorted by these indexes.
1876 Use-list directives may appear at function scope or global scope. They are not
1877 instructions, and have no effect on the semantics of the IR. When they're at
1878 function scope, they must appear after the terminator of the final basic block.
1880 If basic blocks have their address taken via ``blockaddress()`` expressions,
1881 ``uselistorder_bb`` can be used to reorder their use-lists from outside their
1888 uselistorder <ty> <value>, { <order-indexes> }
1889 uselistorder_bb @function, %block { <order-indexes> }
1895 define void @foo(i32 %arg1, i32 %arg2) {
1897 ; ... instructions ...
1899 ; ... instructions ...
1901 ; At function scope.
1902 uselistorder i32 %arg1, { 1, 0, 2 }
1903 uselistorder label %bb, { 1, 0 }
1907 uselistorder i32* @global, { 1, 2, 0 }
1908 uselistorder i32 7, { 1, 0 }
1909 uselistorder i32 (i32) @bar, { 1, 0 }
1910 uselistorder_bb @foo, %bb, { 5, 1, 3, 2, 0, 4 }
1917 The LLVM type system is one of the most important features of the
1918 intermediate representation. Being typed enables a number of
1919 optimizations to be performed on the intermediate representation
1920 directly, without having to do extra analyses on the side before the
1921 transformation. A strong type system makes it easier to read the
1922 generated code and enables novel analyses and transformations that are
1923 not feasible to perform on normal three address code representations.
1933 The void type does not represent any value and has no size.
1951 The function type can be thought of as a function signature. It consists of a
1952 return type and a list of formal parameter types. The return type of a function
1953 type is a void type or first class type --- except for :ref:`label <t_label>`
1954 and :ref:`metadata <t_metadata>` types.
1960 <returntype> (<parameter list>)
1962 ...where '``<parameter list>``' is a comma-separated list of type
1963 specifiers. Optionally, the parameter list may include a type ``...``, which
1964 indicates that the function takes a variable number of arguments. Variable
1965 argument functions can access their arguments with the :ref:`variable argument
1966 handling intrinsic <int_varargs>` functions. '``<returntype>``' is any type
1967 except :ref:`label <t_label>` and :ref:`metadata <t_metadata>`.
1971 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1972 | ``i32 (i32)`` | function taking an ``i32``, returning an ``i32`` |
1973 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1974 | ``float (i16, i32 *) *`` | :ref:`Pointer <t_pointer>` to a function that takes an ``i16`` and a :ref:`pointer <t_pointer>` to ``i32``, returning ``float``. |
1975 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1976 | ``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. |
1977 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1978 | ``{i32, i32} (i32)`` | A function taking an ``i32``, returning a :ref:`structure <t_struct>` containing two ``i32`` values |
1979 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1986 The :ref:`first class <t_firstclass>` types are perhaps the most important.
1987 Values of these types are the only ones which can be produced by
1995 These are the types that are valid in registers from CodeGen's perspective.
2004 The integer type is a very simple type that simply specifies an
2005 arbitrary bit width for the integer type desired. Any bit width from 1
2006 bit to 2\ :sup:`23`\ -1 (about 8 million) can be specified.
2014 The number of bits the integer will occupy is specified by the ``N``
2020 +----------------+------------------------------------------------+
2021 | ``i1`` | a single-bit integer. |
2022 +----------------+------------------------------------------------+
2023 | ``i32`` | a 32-bit integer. |
2024 +----------------+------------------------------------------------+
2025 | ``i1942652`` | a really big integer of over 1 million bits. |
2026 +----------------+------------------------------------------------+
2030 Floating Point Types
2031 """"""""""""""""""""
2040 - 16-bit floating point value
2043 - 32-bit floating point value
2046 - 64-bit floating point value
2049 - 128-bit floating point value (112-bit mantissa)
2052 - 80-bit floating point value (X87)
2055 - 128-bit floating point value (two 64-bits)
2062 The x86_mmx type represents a value held in an MMX register on an x86
2063 machine. The operations allowed on it are quite limited: parameters and
2064 return values, load and store, and bitcast. User-specified MMX
2065 instructions are represented as intrinsic or asm calls with arguments
2066 and/or results of this type. There are no arrays, vectors or constants
2083 The pointer type is used to specify memory locations. Pointers are
2084 commonly used to reference objects in memory.
2086 Pointer types may have an optional address space attribute defining the
2087 numbered address space where the pointed-to object resides. The default
2088 address space is number zero. The semantics of non-zero address spaces
2089 are target-specific.
2091 Note that LLVM does not permit pointers to void (``void*``) nor does it
2092 permit pointers to labels (``label*``). Use ``i8*`` instead.
2102 +-------------------------+--------------------------------------------------------------------------------------------------------------+
2103 | ``[4 x i32]*`` | A :ref:`pointer <t_pointer>` to :ref:`array <t_array>` of four ``i32`` values. |
2104 +-------------------------+--------------------------------------------------------------------------------------------------------------+
2105 | ``i32 (i32*) *`` | A :ref:`pointer <t_pointer>` to a :ref:`function <t_function>` that takes an ``i32*``, returning an ``i32``. |
2106 +-------------------------+--------------------------------------------------------------------------------------------------------------+
2107 | ``i32 addrspace(5)*`` | A :ref:`pointer <t_pointer>` to an ``i32`` value that resides in address space #5. |
2108 +-------------------------+--------------------------------------------------------------------------------------------------------------+
2117 A vector type is a simple derived type that represents a vector of
2118 elements. Vector types are used when multiple primitive data are
2119 operated in parallel using a single instruction (SIMD). A vector type
2120 requires a size (number of elements) and an underlying primitive data
2121 type. Vector types are considered :ref:`first class <t_firstclass>`.
2127 < <# elements> x <elementtype> >
2129 The number of elements is a constant integer value larger than 0;
2130 elementtype may be any integer, floating point or pointer type. Vectors
2131 of size zero are not allowed.
2135 +-------------------+--------------------------------------------------+
2136 | ``<4 x i32>`` | Vector of 4 32-bit integer values. |
2137 +-------------------+--------------------------------------------------+
2138 | ``<8 x float>`` | Vector of 8 32-bit floating-point values. |
2139 +-------------------+--------------------------------------------------+
2140 | ``<2 x i64>`` | Vector of 2 64-bit integer values. |
2141 +-------------------+--------------------------------------------------+
2142 | ``<4 x i64*>`` | Vector of 4 pointers to 64-bit integer values. |
2143 +-------------------+--------------------------------------------------+
2152 The label type represents code labels.
2167 The metadata type represents embedded metadata. No derived types may be
2168 created from metadata except for :ref:`function <t_function>` arguments.
2181 Aggregate Types are a subset of derived types that can contain multiple
2182 member types. :ref:`Arrays <t_array>` and :ref:`structs <t_struct>` are
2183 aggregate types. :ref:`Vectors <t_vector>` are not considered to be
2193 The array type is a very simple derived type that arranges elements
2194 sequentially in memory. The array type requires a size (number of
2195 elements) and an underlying data type.
2201 [<# elements> x <elementtype>]
2203 The number of elements is a constant integer value; ``elementtype`` may
2204 be any type with a size.
2208 +------------------+--------------------------------------+
2209 | ``[40 x i32]`` | Array of 40 32-bit integer values. |
2210 +------------------+--------------------------------------+
2211 | ``[41 x i32]`` | Array of 41 32-bit integer values. |
2212 +------------------+--------------------------------------+
2213 | ``[4 x i8]`` | Array of 4 8-bit integer values. |
2214 +------------------+--------------------------------------+
2216 Here are some examples of multidimensional arrays:
2218 +-----------------------------+----------------------------------------------------------+
2219 | ``[3 x [4 x i32]]`` | 3x4 array of 32-bit integer values. |
2220 +-----------------------------+----------------------------------------------------------+
2221 | ``[12 x [10 x float]]`` | 12x10 array of single precision floating point values. |
2222 +-----------------------------+----------------------------------------------------------+
2223 | ``[2 x [3 x [4 x i16]]]`` | 2x3x4 array of 16-bit integer values. |
2224 +-----------------------------+----------------------------------------------------------+
2226 There is no restriction on indexing beyond the end of the array implied
2227 by a static type (though there are restrictions on indexing beyond the
2228 bounds of an allocated object in some cases). This means that
2229 single-dimension 'variable sized array' addressing can be implemented in
2230 LLVM with a zero length array type. An implementation of 'pascal style
2231 arrays' in LLVM could use the type "``{ i32, [0 x float]}``", for
2241 The structure type is used to represent a collection of data members
2242 together in memory. The elements of a structure may be any type that has
2245 Structures in memory are accessed using '``load``' and '``store``' by
2246 getting a pointer to a field with the '``getelementptr``' instruction.
2247 Structures in registers are accessed using the '``extractvalue``' and
2248 '``insertvalue``' instructions.
2250 Structures may optionally be "packed" structures, which indicate that
2251 the alignment of the struct is one byte, and that there is no padding
2252 between the elements. In non-packed structs, padding between field types
2253 is inserted as defined by the DataLayout string in the module, which is
2254 required to match what the underlying code generator expects.
2256 Structures can either be "literal" or "identified". A literal structure
2257 is defined inline with other types (e.g. ``{i32, i32}*``) whereas
2258 identified types are always defined at the top level with a name.
2259 Literal types are uniqued by their contents and can never be recursive
2260 or opaque since there is no way to write one. Identified types can be
2261 recursive, can be opaqued, and are never uniqued.
2267 %T1 = type { <type list> } ; Identified normal struct type
2268 %T2 = type <{ <type list> }> ; Identified packed struct type
2272 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2273 | ``{ i32, i32, i32 }`` | A triple of three ``i32`` values |
2274 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2275 | ``{ 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``. |
2276 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2277 | ``<{ i8, i32 }>`` | A packed struct known to be 5 bytes in size. |
2278 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2282 Opaque Structure Types
2283 """"""""""""""""""""""
2287 Opaque structure types are used to represent named structure types that
2288 do not have a body specified. This corresponds (for example) to the C
2289 notion of a forward declared structure.
2300 +--------------+-------------------+
2301 | ``opaque`` | An opaque type. |
2302 +--------------+-------------------+
2309 LLVM has several different basic types of constants. This section
2310 describes them all and their syntax.
2315 **Boolean constants**
2316 The two strings '``true``' and '``false``' are both valid constants
2318 **Integer constants**
2319 Standard integers (such as '4') are constants of the
2320 :ref:`integer <t_integer>` type. Negative numbers may be used with
2322 **Floating point constants**
2323 Floating point constants use standard decimal notation (e.g.
2324 123.421), exponential notation (e.g. 1.23421e+2), or a more precise
2325 hexadecimal notation (see below). The assembler requires the exact
2326 decimal value of a floating-point constant. For example, the
2327 assembler accepts 1.25 but rejects 1.3 because 1.3 is a repeating
2328 decimal in binary. Floating point constants must have a :ref:`floating
2329 point <t_floating>` type.
2330 **Null pointer constants**
2331 The identifier '``null``' is recognized as a null pointer constant
2332 and must be of :ref:`pointer type <t_pointer>`.
2334 The one non-intuitive notation for constants is the hexadecimal form of
2335 floating point constants. For example, the form
2336 '``double 0x432ff973cafa8000``' is equivalent to (but harder to read
2337 than) '``double 4.5e+15``'. The only time hexadecimal floating point
2338 constants are required (and the only time that they are generated by the
2339 disassembler) is when a floating point constant must be emitted but it
2340 cannot be represented as a decimal floating point number in a reasonable
2341 number of digits. For example, NaN's, infinities, and other special
2342 values are represented in their IEEE hexadecimal format so that assembly
2343 and disassembly do not cause any bits to change in the constants.
2345 When using the hexadecimal form, constants of types half, float, and
2346 double are represented using the 16-digit form shown above (which
2347 matches the IEEE754 representation for double); half and float values
2348 must, however, be exactly representable as IEEE 754 half and single
2349 precision, respectively. Hexadecimal format is always used for long
2350 double, and there are three forms of long double. The 80-bit format used
2351 by x86 is represented as ``0xK`` followed by 20 hexadecimal digits. The
2352 128-bit format used by PowerPC (two adjacent doubles) is represented by
2353 ``0xM`` followed by 32 hexadecimal digits. The IEEE 128-bit format is
2354 represented by ``0xL`` followed by 32 hexadecimal digits. Long doubles
2355 will only work if they match the long double format on your target.
2356 The IEEE 16-bit format (half precision) is represented by ``0xH``
2357 followed by 4 hexadecimal digits. All hexadecimal formats are big-endian
2358 (sign bit at the left).
2360 There are no constants of type x86_mmx.
2362 .. _complexconstants:
2367 Complex constants are a (potentially recursive) combination of simple
2368 constants and smaller complex constants.
2370 **Structure constants**
2371 Structure constants are represented with notation similar to
2372 structure type definitions (a comma separated list of elements,
2373 surrounded by braces (``{}``)). For example:
2374 "``{ i32 4, float 17.0, i32* @G }``", where "``@G``" is declared as
2375 "``@G = external global i32``". Structure constants must have
2376 :ref:`structure type <t_struct>`, and the number and types of elements
2377 must match those specified by the type.
2379 Array constants are represented with notation similar to array type
2380 definitions (a comma separated list of elements, surrounded by
2381 square brackets (``[]``)). For example:
2382 "``[ i32 42, i32 11, i32 74 ]``". Array constants must have
2383 :ref:`array type <t_array>`, and the number and types of elements must
2384 match those specified by the type. As a special case, character array
2385 constants may also be represented as a double-quoted string using the ``c``
2386 prefix. For example: "``c"Hello World\0A\00"``".
2387 **Vector constants**
2388 Vector constants are represented with notation similar to vector
2389 type definitions (a comma separated list of elements, surrounded by
2390 less-than/greater-than's (``<>``)). For example:
2391 "``< i32 42, i32 11, i32 74, i32 100 >``". Vector constants
2392 must have :ref:`vector type <t_vector>`, and the number and types of
2393 elements must match those specified by the type.
2394 **Zero initialization**
2395 The string '``zeroinitializer``' can be used to zero initialize a
2396 value to zero of *any* type, including scalar and
2397 :ref:`aggregate <t_aggregate>` types. This is often used to avoid
2398 having to print large zero initializers (e.g. for large arrays) and
2399 is always exactly equivalent to using explicit zero initializers.
2401 A metadata node is a constant tuple without types. For example:
2402 "``!{!0, !{!2, !0}, !"test"}``". Metadata can reference constant values,
2403 for example: "``!{!0, i32 0, i8* @global, i64 (i64)* @function, !"str"}``".
2404 Unlike other typed constants that are meant to be interpreted as part of
2405 the instruction stream, metadata is a place to attach additional
2406 information such as debug info.
2408 Global Variable and Function Addresses
2409 --------------------------------------
2411 The addresses of :ref:`global variables <globalvars>` and
2412 :ref:`functions <functionstructure>` are always implicitly valid
2413 (link-time) constants. These constants are explicitly referenced when
2414 the :ref:`identifier for the global <identifiers>` is used and always have
2415 :ref:`pointer <t_pointer>` type. For example, the following is a legal LLVM
2418 .. code-block:: llvm
2422 @Z = global [2 x i32*] [ i32* @X, i32* @Y ]
2429 The string '``undef``' can be used anywhere a constant is expected, and
2430 indicates that the user of the value may receive an unspecified
2431 bit-pattern. Undefined values may be of any type (other than '``label``'
2432 or '``void``') and be used anywhere a constant is permitted.
2434 Undefined values are useful because they indicate to the compiler that
2435 the program is well defined no matter what value is used. This gives the
2436 compiler more freedom to optimize. Here are some examples of
2437 (potentially surprising) transformations that are valid (in pseudo IR):
2439 .. code-block:: llvm
2449 This is safe because all of the output bits are affected by the undef
2450 bits. Any output bit can have a zero or one depending on the input bits.
2452 .. code-block:: llvm
2463 These logical operations have bits that are not always affected by the
2464 input. For example, if ``%X`` has a zero bit, then the output of the
2465 '``and``' operation will always be a zero for that bit, no matter what
2466 the corresponding bit from the '``undef``' is. As such, it is unsafe to
2467 optimize or assume that the result of the '``and``' is '``undef``'.
2468 However, it is safe to assume that all bits of the '``undef``' could be
2469 0, and optimize the '``and``' to 0. Likewise, it is safe to assume that
2470 all the bits of the '``undef``' operand to the '``or``' could be set,
2471 allowing the '``or``' to be folded to -1.
2473 .. code-block:: llvm
2475 %A = select undef, %X, %Y
2476 %B = select undef, 42, %Y
2477 %C = select %X, %Y, undef
2487 This set of examples shows that undefined '``select``' (and conditional
2488 branch) conditions can go *either way*, but they have to come from one
2489 of the two operands. In the ``%A`` example, if ``%X`` and ``%Y`` were
2490 both known to have a clear low bit, then ``%A`` would have to have a
2491 cleared low bit. However, in the ``%C`` example, the optimizer is
2492 allowed to assume that the '``undef``' operand could be the same as
2493 ``%Y``, allowing the whole '``select``' to be eliminated.
2495 .. code-block:: llvm
2497 %A = xor undef, undef
2514 This example points out that two '``undef``' operands are not
2515 necessarily the same. This can be surprising to people (and also matches
2516 C semantics) where they assume that "``X^X``" is always zero, even if
2517 ``X`` is undefined. This isn't true for a number of reasons, but the
2518 short answer is that an '``undef``' "variable" can arbitrarily change
2519 its value over its "live range". This is true because the variable
2520 doesn't actually *have a live range*. Instead, the value is logically
2521 read from arbitrary registers that happen to be around when needed, so
2522 the value is not necessarily consistent over time. In fact, ``%A`` and
2523 ``%C`` need to have the same semantics or the core LLVM "replace all
2524 uses with" concept would not hold.
2526 .. code-block:: llvm
2534 These examples show the crucial difference between an *undefined value*
2535 and *undefined behavior*. An undefined value (like '``undef``') is
2536 allowed to have an arbitrary bit-pattern. This means that the ``%A``
2537 operation can be constant folded to '``undef``', because the '``undef``'
2538 could be an SNaN, and ``fdiv`` is not (currently) defined on SNaN's.
2539 However, in the second example, we can make a more aggressive
2540 assumption: because the ``undef`` is allowed to be an arbitrary value,
2541 we are allowed to assume that it could be zero. Since a divide by zero
2542 has *undefined behavior*, we are allowed to assume that the operation
2543 does not execute at all. This allows us to delete the divide and all
2544 code after it. Because the undefined operation "can't happen", the
2545 optimizer can assume that it occurs in dead code.
2547 .. code-block:: llvm
2549 a: store undef -> %X
2550 b: store %X -> undef
2555 These examples reiterate the ``fdiv`` example: a store *of* an undefined
2556 value can be assumed to not have any effect; we can assume that the
2557 value is overwritten with bits that happen to match what was already
2558 there. However, a store *to* an undefined location could clobber
2559 arbitrary memory, therefore, it has undefined behavior.
2566 Poison values are similar to :ref:`undef values <undefvalues>`, however
2567 they also represent the fact that an instruction or constant expression
2568 that cannot evoke side effects has nevertheless detected a condition
2569 that results in undefined behavior.
2571 There is currently no way of representing a poison value in the IR; they
2572 only exist when produced by operations such as :ref:`add <i_add>` with
2575 Poison value behavior is defined in terms of value *dependence*:
2577 - Values other than :ref:`phi <i_phi>` nodes depend on their operands.
2578 - :ref:`Phi <i_phi>` nodes depend on the operand corresponding to
2579 their dynamic predecessor basic block.
2580 - Function arguments depend on the corresponding actual argument values
2581 in the dynamic callers of their functions.
2582 - :ref:`Call <i_call>` instructions depend on the :ref:`ret <i_ret>`
2583 instructions that dynamically transfer control back to them.
2584 - :ref:`Invoke <i_invoke>` instructions depend on the
2585 :ref:`ret <i_ret>`, :ref:`resume <i_resume>`, or exception-throwing
2586 call instructions that dynamically transfer control back to them.
2587 - Non-volatile loads and stores depend on the most recent stores to all
2588 of the referenced memory addresses, following the order in the IR
2589 (including loads and stores implied by intrinsics such as
2590 :ref:`@llvm.memcpy <int_memcpy>`.)
2591 - An instruction with externally visible side effects depends on the
2592 most recent preceding instruction with externally visible side
2593 effects, following the order in the IR. (This includes :ref:`volatile
2594 operations <volatile>`.)
2595 - An instruction *control-depends* on a :ref:`terminator
2596 instruction <terminators>` if the terminator instruction has
2597 multiple successors and the instruction is always executed when
2598 control transfers to one of the successors, and may not be executed
2599 when control is transferred to another.
2600 - Additionally, an instruction also *control-depends* on a terminator
2601 instruction if the set of instructions it otherwise depends on would
2602 be different if the terminator had transferred control to a different
2604 - Dependence is transitive.
2606 Poison values have the same behavior as :ref:`undef values <undefvalues>`,
2607 with the additional effect that any instruction that has a *dependence*
2608 on a poison value has undefined behavior.
2610 Here are some examples:
2612 .. code-block:: llvm
2615 %poison = sub nuw i32 0, 1 ; Results in a poison value.
2616 %still_poison = and i32 %poison, 0 ; 0, but also poison.
2617 %poison_yet_again = getelementptr i32, i32* @h, i32 %still_poison
2618 store i32 0, i32* %poison_yet_again ; memory at @h[0] is poisoned
2620 store i32 %poison, i32* @g ; Poison value stored to memory.
2621 %poison2 = load i32, i32* @g ; Poison value loaded back from memory.
2623 store volatile i32 %poison, i32* @g ; External observation; undefined behavior.
2625 %narrowaddr = bitcast i32* @g to i16*
2626 %wideaddr = bitcast i32* @g to i64*
2627 %poison3 = load i16, i16* %narrowaddr ; Returns a poison value.
2628 %poison4 = load i64, i64* %wideaddr ; Returns a poison value.
2630 %cmp = icmp slt i32 %poison, 0 ; Returns a poison value.
2631 br i1 %cmp, label %true, label %end ; Branch to either destination.
2634 store volatile i32 0, i32* @g ; This is control-dependent on %cmp, so
2635 ; it has undefined behavior.
2639 %p = phi i32 [ 0, %entry ], [ 1, %true ]
2640 ; Both edges into this PHI are
2641 ; control-dependent on %cmp, so this
2642 ; always results in a poison value.
2644 store volatile i32 0, i32* @g ; This would depend on the store in %true
2645 ; if %cmp is true, or the store in %entry
2646 ; otherwise, so this is undefined behavior.
2648 br i1 %cmp, label %second_true, label %second_end
2649 ; The same branch again, but this time the
2650 ; true block doesn't have side effects.
2657 store volatile i32 0, i32* @g ; This time, the instruction always depends
2658 ; on the store in %end. Also, it is
2659 ; control-equivalent to %end, so this is
2660 ; well-defined (ignoring earlier undefined
2661 ; behavior in this example).
2665 Addresses of Basic Blocks
2666 -------------------------
2668 ``blockaddress(@function, %block)``
2670 The '``blockaddress``' constant computes the address of the specified
2671 basic block in the specified function, and always has an ``i8*`` type.
2672 Taking the address of the entry block is illegal.
2674 This value only has defined behavior when used as an operand to the
2675 ':ref:`indirectbr <i_indirectbr>`' instruction, or for comparisons
2676 against null. Pointer equality tests between labels addresses results in
2677 undefined behavior --- though, again, comparison against null is ok, and
2678 no label is equal to the null pointer. This may be passed around as an
2679 opaque pointer sized value as long as the bits are not inspected. This
2680 allows ``ptrtoint`` and arithmetic to be performed on these values so
2681 long as the original value is reconstituted before the ``indirectbr``
2684 Finally, some targets may provide defined semantics when using the value
2685 as the operand to an inline assembly, but that is target specific.
2689 Constant Expressions
2690 --------------------
2692 Constant expressions are used to allow expressions involving other
2693 constants to be used as constants. Constant expressions may be of any
2694 :ref:`first class <t_firstclass>` type and may involve any LLVM operation
2695 that does not have side effects (e.g. load and call are not supported).
2696 The following is the syntax for constant expressions:
2698 ``trunc (CST to TYPE)``
2699 Truncate a constant to another type. The bit size of CST must be
2700 larger than the bit size of TYPE. Both types must be integers.
2701 ``zext (CST to TYPE)``
2702 Zero extend a constant to another type. The bit size of CST must be
2703 smaller than the bit size of TYPE. Both types must be integers.
2704 ``sext (CST to TYPE)``
2705 Sign extend a constant to another type. The bit size of CST must be
2706 smaller than the bit size of TYPE. Both types must be integers.
2707 ``fptrunc (CST to TYPE)``
2708 Truncate a floating point constant to another floating point type.
2709 The size of CST must be larger than the size of TYPE. Both types
2710 must be floating point.
2711 ``fpext (CST to TYPE)``
2712 Floating point extend a constant to another type. The size of CST
2713 must be smaller or equal to the size of TYPE. Both types must be
2715 ``fptoui (CST to TYPE)``
2716 Convert a floating point constant to the corresponding unsigned
2717 integer constant. TYPE must be a scalar or vector integer type. CST
2718 must be of scalar or vector floating point type. Both CST and TYPE
2719 must be scalars, or vectors of the same number of elements. If the
2720 value won't fit in the integer type, the results are undefined.
2721 ``fptosi (CST to TYPE)``
2722 Convert a floating point constant to the corresponding signed
2723 integer constant. TYPE must be a scalar or vector integer type. CST
2724 must be of scalar or vector floating point type. Both CST and TYPE
2725 must be scalars, or vectors of the same number of elements. If the
2726 value won't fit in the integer type, the results are undefined.
2727 ``uitofp (CST to TYPE)``
2728 Convert an unsigned integer constant to the corresponding floating
2729 point constant. TYPE must be a scalar or vector floating point type.
2730 CST must be of scalar or vector integer type. Both CST and TYPE must
2731 be scalars, or vectors of the same number of elements. If the value
2732 won't fit in the floating point type, the results are undefined.
2733 ``sitofp (CST to TYPE)``
2734 Convert a signed integer constant to the corresponding floating
2735 point constant. TYPE must be a scalar or vector floating point type.
2736 CST must be of scalar or vector integer type. Both CST and TYPE must
2737 be scalars, or vectors of the same number of elements. If the value
2738 won't fit in the floating point type, the results are undefined.
2739 ``ptrtoint (CST to TYPE)``
2740 Convert a pointer typed constant to the corresponding integer
2741 constant. ``TYPE`` must be an integer type. ``CST`` must be of
2742 pointer type. The ``CST`` value is zero extended, truncated, or
2743 unchanged to make it fit in ``TYPE``.
2744 ``inttoptr (CST to TYPE)``
2745 Convert an integer constant to a pointer constant. TYPE must be a
2746 pointer type. CST must be of integer type. The CST value is zero
2747 extended, truncated, or unchanged to make it fit in a pointer size.
2748 This one is *really* dangerous!
2749 ``bitcast (CST to TYPE)``
2750 Convert a constant, CST, to another TYPE. The constraints of the
2751 operands are the same as those for the :ref:`bitcast
2752 instruction <i_bitcast>`.
2753 ``addrspacecast (CST to TYPE)``
2754 Convert a constant pointer or constant vector of pointer, CST, to another
2755 TYPE in a different address space. The constraints of the operands are the
2756 same as those for the :ref:`addrspacecast instruction <i_addrspacecast>`.
2757 ``getelementptr (TY, CSTPTR, IDX0, IDX1, ...)``, ``getelementptr inbounds (TY, CSTPTR, IDX0, IDX1, ...)``
2758 Perform the :ref:`getelementptr operation <i_getelementptr>` on
2759 constants. As with the :ref:`getelementptr <i_getelementptr>`
2760 instruction, the index list may have zero or more indexes, which are
2761 required to make sense for the type of "pointer to TY".
2762 ``select (COND, VAL1, VAL2)``
2763 Perform the :ref:`select operation <i_select>` on constants.
2764 ``icmp COND (VAL1, VAL2)``
2765 Performs the :ref:`icmp operation <i_icmp>` on constants.
2766 ``fcmp COND (VAL1, VAL2)``
2767 Performs the :ref:`fcmp operation <i_fcmp>` on constants.
2768 ``extractelement (VAL, IDX)``
2769 Perform the :ref:`extractelement operation <i_extractelement>` on
2771 ``insertelement (VAL, ELT, IDX)``
2772 Perform the :ref:`insertelement operation <i_insertelement>` on
2774 ``shufflevector (VEC1, VEC2, IDXMASK)``
2775 Perform the :ref:`shufflevector operation <i_shufflevector>` on
2777 ``extractvalue (VAL, IDX0, IDX1, ...)``
2778 Perform the :ref:`extractvalue operation <i_extractvalue>` on
2779 constants. The index list is interpreted in a similar manner as
2780 indices in a ':ref:`getelementptr <i_getelementptr>`' operation. At
2781 least one index value must be specified.
2782 ``insertvalue (VAL, ELT, IDX0, IDX1, ...)``
2783 Perform the :ref:`insertvalue operation <i_insertvalue>` on constants.
2784 The index list is interpreted in a similar manner as indices in a
2785 ':ref:`getelementptr <i_getelementptr>`' operation. At least one index
2786 value must be specified.
2787 ``OPCODE (LHS, RHS)``
2788 Perform the specified operation of the LHS and RHS constants. OPCODE
2789 may be any of the :ref:`binary <binaryops>` or :ref:`bitwise
2790 binary <bitwiseops>` operations. The constraints on operands are
2791 the same as those for the corresponding instruction (e.g. no bitwise
2792 operations on floating point values are allowed).
2799 Inline Assembler Expressions
2800 ----------------------------
2802 LLVM supports inline assembler expressions (as opposed to :ref:`Module-Level
2803 Inline Assembly <moduleasm>`) through the use of a special value. This
2804 value represents the inline assembler as a string (containing the
2805 instructions to emit), a list of operand constraints (stored as a
2806 string), a flag that indicates whether or not the inline asm expression
2807 has side effects, and a flag indicating whether the function containing
2808 the asm needs to align its stack conservatively. An example inline
2809 assembler expression is:
2811 .. code-block:: llvm
2813 i32 (i32) asm "bswap $0", "=r,r"
2815 Inline assembler expressions may **only** be used as the callee operand
2816 of a :ref:`call <i_call>` or an :ref:`invoke <i_invoke>` instruction.
2817 Thus, typically we have:
2819 .. code-block:: llvm
2821 %X = call i32 asm "bswap $0", "=r,r"(i32 %Y)
2823 Inline asms with side effects not visible in the constraint list must be
2824 marked as having side effects. This is done through the use of the
2825 '``sideeffect``' keyword, like so:
2827 .. code-block:: llvm
2829 call void asm sideeffect "eieio", ""()
2831 In some cases inline asms will contain code that will not work unless
2832 the stack is aligned in some way, such as calls or SSE instructions on
2833 x86, yet will not contain code that does that alignment within the asm.
2834 The compiler should make conservative assumptions about what the asm
2835 might contain and should generate its usual stack alignment code in the
2836 prologue if the '``alignstack``' keyword is present:
2838 .. code-block:: llvm
2840 call void asm alignstack "eieio", ""()
2842 Inline asms also support using non-standard assembly dialects. The
2843 assumed dialect is ATT. When the '``inteldialect``' keyword is present,
2844 the inline asm is using the Intel dialect. Currently, ATT and Intel are
2845 the only supported dialects. An example is:
2847 .. code-block:: llvm
2849 call void asm inteldialect "eieio", ""()
2851 If multiple keywords appear the '``sideeffect``' keyword must come
2852 first, the '``alignstack``' keyword second and the '``inteldialect``'
2858 The call instructions that wrap inline asm nodes may have a
2859 "``!srcloc``" MDNode attached to it that contains a list of constant
2860 integers. If present, the code generator will use the integer as the
2861 location cookie value when report errors through the ``LLVMContext``
2862 error reporting mechanisms. This allows a front-end to correlate backend
2863 errors that occur with inline asm back to the source code that produced
2866 .. code-block:: llvm
2868 call void asm sideeffect "something bad", ""(), !srcloc !42
2870 !42 = !{ i32 1234567 }
2872 It is up to the front-end to make sense of the magic numbers it places
2873 in the IR. If the MDNode contains multiple constants, the code generator
2874 will use the one that corresponds to the line of the asm that the error
2882 LLVM IR allows metadata to be attached to instructions in the program
2883 that can convey extra information about the code to the optimizers and
2884 code generator. One example application of metadata is source-level
2885 debug information. There are two metadata primitives: strings and nodes.
2887 Metadata does not have a type, and is not a value. If referenced from a
2888 ``call`` instruction, it uses the ``metadata`` type.
2890 All metadata are identified in syntax by a exclamation point ('``!``').
2892 .. _metadata-string:
2894 Metadata Nodes and Metadata Strings
2895 -----------------------------------
2897 A metadata string is a string surrounded by double quotes. It can
2898 contain any character by escaping non-printable characters with
2899 "``\xx``" where "``xx``" is the two digit hex code. For example:
2902 Metadata nodes are represented with notation similar to structure
2903 constants (a comma separated list of elements, surrounded by braces and
2904 preceded by an exclamation point). Metadata nodes can have any values as
2905 their operand. For example:
2907 .. code-block:: llvm
2909 !{ !"test\00", i32 10}
2911 Metadata nodes that aren't uniqued use the ``distinct`` keyword. For example:
2913 .. code-block:: llvm
2915 !0 = distinct !{!"test\00", i32 10}
2917 ``distinct`` nodes are useful when nodes shouldn't be merged based on their
2918 content. They can also occur when transformations cause uniquing collisions
2919 when metadata operands change.
2921 A :ref:`named metadata <namedmetadatastructure>` is a collection of
2922 metadata nodes, which can be looked up in the module symbol table. For
2925 .. code-block:: llvm
2929 Metadata can be used as function arguments. Here ``llvm.dbg.value``
2930 function is using two metadata arguments:
2932 .. code-block:: llvm
2934 call void @llvm.dbg.value(metadata !24, i64 0, metadata !25)
2936 Metadata can be attached with an instruction. Here metadata ``!21`` is
2937 attached to the ``add`` instruction using the ``!dbg`` identifier:
2939 .. code-block:: llvm
2941 %indvar.next = add i64 %indvar, 1, !dbg !21
2943 More information about specific metadata nodes recognized by the
2944 optimizers and code generator is found below.
2946 .. _specialized-metadata:
2948 Specialized Metadata Nodes
2949 ^^^^^^^^^^^^^^^^^^^^^^^^^^
2951 Specialized metadata nodes are custom data structures in metadata (as opposed
2952 to generic tuples). Their fields are labelled, and can be specified in any
2955 These aren't inherently debug info centric, but currently all the specialized
2956 metadata nodes are related to debug info.
2963 ``DICompileUnit`` nodes represent a compile unit. The ``enums:``,
2964 ``retainedTypes:``, ``subprograms:``, ``globals:`` and ``imports:`` fields are
2965 tuples containing the debug info to be emitted along with the compile unit,
2966 regardless of code optimizations (some nodes are only emitted if there are
2967 references to them from instructions).
2969 .. code-block:: llvm
2971 !0 = !DICompileUnit(language: DW_LANG_C99, file: !1, producer: "clang",
2972 isOptimized: true, flags: "-O2", runtimeVersion: 2,
2973 splitDebugFilename: "abc.debug", emissionKind: 1,
2974 enums: !2, retainedTypes: !3, subprograms: !4,
2975 globals: !5, imports: !6)
2977 Compile unit descriptors provide the root scope for objects declared in a
2978 specific compilation unit. File descriptors are defined using this scope.
2979 These descriptors are collected by a named metadata ``!llvm.dbg.cu``. They
2980 keep track of subprograms, global variables, type information, and imported
2981 entities (declarations and namespaces).
2988 ``DIFile`` nodes represent files. The ``filename:`` can include slashes.
2990 .. code-block:: llvm
2992 !0 = !DIFile(filename: "path/to/file", directory: "/path/to/dir")
2994 Files are sometimes used in ``scope:`` fields, and are the only valid target
2995 for ``file:`` fields.
3002 ``DIBasicType`` nodes represent primitive types, such as ``int``, ``bool`` and
3003 ``float``. ``tag:`` defaults to ``DW_TAG_base_type``.
3005 .. code-block:: llvm
3007 !0 = !DIBasicType(name: "unsigned char", size: 8, align: 8,
3008 encoding: DW_ATE_unsigned_char)
3009 !1 = !DIBasicType(tag: DW_TAG_unspecified_type, name: "decltype(nullptr)")
3011 The ``encoding:`` describes the details of the type. Usually it's one of the
3014 .. code-block:: llvm
3020 DW_ATE_signed_char = 6
3022 DW_ATE_unsigned_char = 8
3024 .. _DISubroutineType:
3029 ``DISubroutineType`` nodes represent subroutine types. Their ``types:`` field
3030 refers to a tuple; the first operand is the return type, while the rest are the
3031 types of the formal arguments in order. If the first operand is ``null``, that
3032 represents a function with no return value (such as ``void foo() {}`` in C++).
3034 .. code-block:: llvm
3036 !0 = !BasicType(name: "int", size: 32, align: 32, DW_ATE_signed)
3037 !1 = !BasicType(name: "char", size: 8, align: 8, DW_ATE_signed_char)
3038 !2 = !DISubroutineType(types: !{null, !0, !1}) ; void (int, char)
3045 ``DIDerivedType`` nodes represent types derived from other types, such as
3048 .. code-block:: llvm
3050 !0 = !DIBasicType(name: "unsigned char", size: 8, align: 8,
3051 encoding: DW_ATE_unsigned_char)
3052 !1 = !DIDerivedType(tag: DW_TAG_pointer_type, baseType: !0, size: 32,
3055 The following ``tag:`` values are valid:
3057 .. code-block:: llvm
3059 DW_TAG_formal_parameter = 5
3061 DW_TAG_pointer_type = 15
3062 DW_TAG_reference_type = 16
3064 DW_TAG_ptr_to_member_type = 31
3065 DW_TAG_const_type = 38
3066 DW_TAG_volatile_type = 53
3067 DW_TAG_restrict_type = 55
3069 ``DW_TAG_member`` is used to define a member of a :ref:`composite type
3070 <DICompositeType>` or :ref:`subprogram <DISubprogram>`. The type of the member
3071 is the ``baseType:``. The ``offset:`` is the member's bit offset.
3072 ``DW_TAG_formal_parameter`` is used to define a member which is a formal
3073 argument of a subprogram.
3075 ``DW_TAG_typedef`` is used to provide a name for the ``baseType:``.
3077 ``DW_TAG_pointer_type``, ``DW_TAG_reference_type``, ``DW_TAG_const_type``,
3078 ``DW_TAG_volatile_type`` and ``DW_TAG_restrict_type`` are used to qualify the
3081 Note that the ``void *`` type is expressed as a type derived from NULL.
3083 .. _DICompositeType:
3088 ``DICompositeType`` nodes represent types composed of other types, like
3089 structures and unions. ``elements:`` points to a tuple of the composed types.
3091 If the source language supports ODR, the ``identifier:`` field gives the unique
3092 identifier used for type merging between modules. When specified, other types
3093 can refer to composite types indirectly via a :ref:`metadata string
3094 <metadata-string>` that matches their identifier.
3096 .. code-block:: llvm
3098 !0 = !DIEnumerator(name: "SixKind", value: 7)
3099 !1 = !DIEnumerator(name: "SevenKind", value: 7)
3100 !2 = !DIEnumerator(name: "NegEightKind", value: -8)
3101 !3 = !DICompositeType(tag: DW_TAG_enumeration_type, name: "Enum", file: !12,
3102 line: 2, size: 32, align: 32, identifier: "_M4Enum",
3103 elements: !{!0, !1, !2})
3105 The following ``tag:`` values are valid:
3107 .. code-block:: llvm
3109 DW_TAG_array_type = 1
3110 DW_TAG_class_type = 2
3111 DW_TAG_enumeration_type = 4
3112 DW_TAG_structure_type = 19
3113 DW_TAG_union_type = 23
3114 DW_TAG_subroutine_type = 21
3115 DW_TAG_inheritance = 28
3118 For ``DW_TAG_array_type``, the ``elements:`` should be :ref:`subrange
3119 descriptors <DISubrange>`, each representing the range of subscripts at that
3120 level of indexing. The ``DIFlagVector`` flag to ``flags:`` indicates that an
3121 array type is a native packed vector.
3123 For ``DW_TAG_enumeration_type``, the ``elements:`` should be :ref:`enumerator
3124 descriptors <DIEnumerator>`, each representing the definition of an enumeration
3125 value for the set. All enumeration type descriptors are collected in the
3126 ``enums:`` field of the :ref:`compile unit <DICompileUnit>`.
3128 For ``DW_TAG_structure_type``, ``DW_TAG_class_type``, and
3129 ``DW_TAG_union_type``, the ``elements:`` should be :ref:`derived types
3130 <DIDerivedType>` with ``tag: DW_TAG_member`` or ``tag: DW_TAG_inheritance``.
3137 ``DISubrange`` nodes are the elements for ``DW_TAG_array_type`` variants of
3138 :ref:`DICompositeType`. ``count: -1`` indicates an empty array.
3140 .. code-block:: llvm
3142 !0 = !DISubrange(count: 5, lowerBound: 0) ; array counting from 0
3143 !1 = !DISubrange(count: 5, lowerBound: 1) ; array counting from 1
3144 !2 = !DISubrange(count: -1) ; empty array.
3151 ``DIEnumerator`` nodes are the elements for ``DW_TAG_enumeration_type``
3152 variants of :ref:`DICompositeType`.
3154 .. code-block:: llvm
3156 !0 = !DIEnumerator(name: "SixKind", value: 7)
3157 !1 = !DIEnumerator(name: "SevenKind", value: 7)
3158 !2 = !DIEnumerator(name: "NegEightKind", value: -8)
3160 DITemplateTypeParameter
3161 """""""""""""""""""""""
3163 ``DITemplateTypeParameter`` nodes represent type parameters to generic source
3164 language constructs. They are used (optionally) in :ref:`DICompositeType` and
3165 :ref:`DISubprogram` ``templateParams:`` fields.
3167 .. code-block:: llvm
3169 !0 = !DITemplateTypeParameter(name: "Ty", type: !1)
3171 DITemplateValueParameter
3172 """"""""""""""""""""""""
3174 ``DITemplateValueParameter`` nodes represent value parameters to generic source
3175 language constructs. ``tag:`` defaults to ``DW_TAG_template_value_parameter``,
3176 but if specified can also be set to ``DW_TAG_GNU_template_template_param`` or
3177 ``DW_TAG_GNU_template_param_pack``. They are used (optionally) in
3178 :ref:`DICompositeType` and :ref:`DISubprogram` ``templateParams:`` fields.
3180 .. code-block:: llvm
3182 !0 = !DITemplateValueParameter(name: "Ty", type: !1, value: i32 7)
3187 ``DINamespace`` nodes represent namespaces in the source language.
3189 .. code-block:: llvm
3191 !0 = !DINamespace(name: "myawesomeproject", scope: !1, file: !2, line: 7)
3196 ``DIGlobalVariable`` nodes represent global variables in the source language.
3198 .. code-block:: llvm
3200 !0 = !DIGlobalVariable(name: "foo", linkageName: "foo", scope: !1,
3201 file: !2, line: 7, type: !3, isLocal: true,
3202 isDefinition: false, variable: i32* @foo,
3205 All global variables should be referenced by the `globals:` field of a
3206 :ref:`compile unit <DICompileUnit>`.
3213 ``DISubprogram`` nodes represent functions from the source language. The
3214 ``variables:`` field points at :ref:`variables <DILocalVariable>` that must be
3215 retained, even if their IR counterparts are optimized out of the IR. The
3216 ``type:`` field must point at an :ref:`DISubroutineType`.
3218 .. code-block:: llvm
3220 !0 = !DISubprogram(name: "foo", linkageName: "_Zfoov", scope: !1,
3221 file: !2, line: 7, type: !3, isLocal: true,
3222 isDefinition: false, scopeLine: 8, containingType: !4,
3223 virtuality: DW_VIRTUALITY_pure_virtual, virtualIndex: 10,
3224 flags: DIFlagPrototyped, isOptimized: true,
3225 function: void ()* @_Z3foov,
3226 templateParams: !5, declaration: !6, variables: !7)
3233 ``DILexicalBlock`` nodes describe nested blocks within a :ref:`subprogram
3234 <DISubprogram>`. The line number and column numbers are used to dinstinguish
3235 two lexical blocks at same depth. They are valid targets for ``scope:``
3238 .. code-block:: llvm
3240 !0 = distinct !DILexicalBlock(scope: !1, file: !2, line: 7, column: 35)
3242 Usually lexical blocks are ``distinct`` to prevent node merging based on
3245 .. _DILexicalBlockFile:
3250 ``DILexicalBlockFile`` nodes are used to discriminate between sections of a
3251 :ref:`lexical block <DILexicalBlock>`. The ``file:`` field can be changed to
3252 indicate textual inclusion, or the ``discriminator:`` field can be used to
3253 discriminate between control flow within a single block in the source language.
3255 .. code-block:: llvm
3257 !0 = !DILexicalBlock(scope: !3, file: !4, line: 7, column: 35)
3258 !1 = !DILexicalBlockFile(scope: !0, file: !4, discriminator: 0)
3259 !2 = !DILexicalBlockFile(scope: !0, file: !4, discriminator: 1)
3266 ``DILocation`` nodes represent source debug locations. The ``scope:`` field is
3267 mandatory, and points at an :ref:`DILexicalBlockFile`, an
3268 :ref:`DILexicalBlock`, or an :ref:`DISubprogram`.
3270 .. code-block:: llvm
3272 !0 = !DILocation(line: 2900, column: 42, scope: !1, inlinedAt: !2)
3274 .. _DILocalVariable:
3279 ``DILocalVariable`` nodes represent local variables in the source language.
3280 Instead of ``DW_TAG_variable``, they use LLVM-specific fake tags to
3281 discriminate between local variables (``DW_TAG_auto_variable``) and subprogram
3282 arguments (``DW_TAG_arg_variable``). In the latter case, the ``arg:`` field
3283 specifies the argument position, and this variable will be included in the
3284 ``variables:`` field of its :ref:`DISubprogram`.
3286 .. code-block:: llvm
3288 !0 = !DILocalVariable(tag: DW_TAG_arg_variable, name: "this", arg: 0,
3289 scope: !3, file: !2, line: 7, type: !3,
3290 flags: DIFlagArtificial)
3291 !1 = !DILocalVariable(tag: DW_TAG_arg_variable, name: "x", arg: 1,
3292 scope: !4, file: !2, line: 7, type: !3)
3293 !1 = !DILocalVariable(tag: DW_TAG_auto_variable, name: "y",
3294 scope: !5, file: !2, line: 7, type: !3)
3299 ``DIExpression`` nodes represent DWARF expression sequences. They are used in
3300 :ref:`debug intrinsics<dbg_intrinsics>` (such as ``llvm.dbg.declare``) to
3301 describe how the referenced LLVM variable relates to the source language
3304 The current supported vocabulary is limited:
3306 - ``DW_OP_deref`` dereferences the working expression.
3307 - ``DW_OP_plus, 93`` adds ``93`` to the working expression.
3308 - ``DW_OP_bit_piece, 16, 8`` specifies the offset and size (``16`` and ``8``
3309 here, respectively) of the variable piece from the working expression.
3311 .. code-block:: llvm
3313 !0 = !DIExpression(DW_OP_deref)
3314 !1 = !DIExpression(DW_OP_plus, 3)
3315 !2 = !DIExpression(DW_OP_bit_piece, 3, 7)
3316 !3 = !DIExpression(DW_OP_deref, DW_OP_plus, 3, DW_OP_bit_piece, 3, 7)
3321 ``DIObjCProperty`` nodes represent Objective-C property nodes.
3323 .. code-block:: llvm
3325 !3 = !DIObjCProperty(name: "foo", file: !1, line: 7, setter: "setFoo",
3326 getter: "getFoo", attributes: 7, type: !2)
3331 ``DIImportedEntity`` nodes represent entities (such as modules) imported into a
3334 .. code-block:: llvm
3336 !2 = !DIImportedEntity(tag: DW_TAG_imported_module, name: "foo", scope: !0,
3337 entity: !1, line: 7)
3342 In LLVM IR, memory does not have types, so LLVM's own type system is not
3343 suitable for doing TBAA. Instead, metadata is added to the IR to
3344 describe a type system of a higher level language. This can be used to
3345 implement typical C/C++ TBAA, but it can also be used to implement
3346 custom alias analysis behavior for other languages.
3348 The current metadata format is very simple. TBAA metadata nodes have up
3349 to three fields, e.g.:
3351 .. code-block:: llvm
3353 !0 = !{ !"an example type tree" }
3354 !1 = !{ !"int", !0 }
3355 !2 = !{ !"float", !0 }
3356 !3 = !{ !"const float", !2, i64 1 }
3358 The first field is an identity field. It can be any value, usually a
3359 metadata string, which uniquely identifies the type. The most important
3360 name in the tree is the name of the root node. Two trees with different
3361 root node names are entirely disjoint, even if they have leaves with
3364 The second field identifies the type's parent node in the tree, or is
3365 null or omitted for a root node. A type is considered to alias all of
3366 its descendants and all of its ancestors in the tree. Also, a type is
3367 considered to alias all types in other trees, so that bitcode produced
3368 from multiple front-ends is handled conservatively.
3370 If the third field is present, it's an integer which if equal to 1
3371 indicates that the type is "constant" (meaning
3372 ``pointsToConstantMemory`` should return true; see `other useful
3373 AliasAnalysis methods <AliasAnalysis.html#OtherItfs>`_).
3375 '``tbaa.struct``' Metadata
3376 ^^^^^^^^^^^^^^^^^^^^^^^^^^
3378 The :ref:`llvm.memcpy <int_memcpy>` is often used to implement
3379 aggregate assignment operations in C and similar languages, however it
3380 is defined to copy a contiguous region of memory, which is more than
3381 strictly necessary for aggregate types which contain holes due to
3382 padding. Also, it doesn't contain any TBAA information about the fields
3385 ``!tbaa.struct`` metadata can describe which memory subregions in a
3386 memcpy are padding and what the TBAA tags of the struct are.
3388 The current metadata format is very simple. ``!tbaa.struct`` metadata
3389 nodes are a list of operands which are in conceptual groups of three.
3390 For each group of three, the first operand gives the byte offset of a
3391 field in bytes, the second gives its size in bytes, and the third gives
3394 .. code-block:: llvm
3396 !4 = !{ i64 0, i64 4, !1, i64 8, i64 4, !2 }
3398 This describes a struct with two fields. The first is at offset 0 bytes
3399 with size 4 bytes, and has tbaa tag !1. The second is at offset 8 bytes
3400 and has size 4 bytes and has tbaa tag !2.
3402 Note that the fields need not be contiguous. In this example, there is a
3403 4 byte gap between the two fields. This gap represents padding which
3404 does not carry useful data and need not be preserved.
3406 '``noalias``' and '``alias.scope``' Metadata
3407 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3409 ``noalias`` and ``alias.scope`` metadata provide the ability to specify generic
3410 noalias memory-access sets. This means that some collection of memory access
3411 instructions (loads, stores, memory-accessing calls, etc.) that carry
3412 ``noalias`` metadata can specifically be specified not to alias with some other
3413 collection of memory access instructions that carry ``alias.scope`` metadata.
3414 Each type of metadata specifies a list of scopes where each scope has an id and
3415 a domain. When evaluating an aliasing query, if for some domain, the set
3416 of scopes with that domain in one instruction's ``alias.scope`` list is a
3417 subset of (or equal to) the set of scopes for that domain in another
3418 instruction's ``noalias`` list, then the two memory accesses are assumed not to
3421 The metadata identifying each domain is itself a list containing one or two
3422 entries. The first entry is the name of the domain. Note that if the name is a
3423 string then it can be combined accross functions and translation units. A
3424 self-reference can be used to create globally unique domain names. A
3425 descriptive string may optionally be provided as a second list entry.
3427 The metadata identifying each scope is also itself a list containing two or
3428 three entries. The first entry is the name of the scope. Note that if the name
3429 is a string then it can be combined accross functions and translation units. A
3430 self-reference can be used to create globally unique scope names. A metadata
3431 reference to the scope's domain is the second entry. A descriptive string may
3432 optionally be provided as a third list entry.
3436 .. code-block:: llvm
3438 ; Two scope domains:
3442 ; Some scopes in these domains:
3448 !5 = !{!4} ; A list containing only scope !4
3452 ; These two instructions don't alias:
3453 %0 = load float, float* %c, align 4, !alias.scope !5
3454 store float %0, float* %arrayidx.i, align 4, !noalias !5
3456 ; These two instructions also don't alias (for domain !1, the set of scopes
3457 ; in the !alias.scope equals that in the !noalias list):
3458 %2 = load float, float* %c, align 4, !alias.scope !5
3459 store float %2, float* %arrayidx.i2, align 4, !noalias !6
3461 ; These two instructions may alias (for domain !0, the set of scopes in
3462 ; the !noalias list is not a superset of, or equal to, the scopes in the
3463 ; !alias.scope list):
3464 %2 = load float, float* %c, align 4, !alias.scope !6
3465 store float %0, float* %arrayidx.i, align 4, !noalias !7
3467 '``fpmath``' Metadata
3468 ^^^^^^^^^^^^^^^^^^^^^
3470 ``fpmath`` metadata may be attached to any instruction of floating point
3471 type. It can be used to express the maximum acceptable error in the
3472 result of that instruction, in ULPs, thus potentially allowing the
3473 compiler to use a more efficient but less accurate method of computing
3474 it. ULP is defined as follows:
3476 If ``x`` is a real number that lies between two finite consecutive
3477 floating-point numbers ``a`` and ``b``, without being equal to one
3478 of them, then ``ulp(x) = |b - a|``, otherwise ``ulp(x)`` is the
3479 distance between the two non-equal finite floating-point numbers
3480 nearest ``x``. Moreover, ``ulp(NaN)`` is ``NaN``.
3482 The metadata node shall consist of a single positive floating point
3483 number representing the maximum relative error, for example:
3485 .. code-block:: llvm
3487 !0 = !{ float 2.5 } ; maximum acceptable inaccuracy is 2.5 ULPs
3491 '``range``' Metadata
3492 ^^^^^^^^^^^^^^^^^^^^
3494 ``range`` metadata may be attached only to ``load``, ``call`` and ``invoke`` of
3495 integer types. It expresses the possible ranges the loaded value or the value
3496 returned by the called function at this call site is in. The ranges are
3497 represented with a flattened list of integers. The loaded value or the value
3498 returned is known to be in the union of the ranges defined by each consecutive
3499 pair. Each pair has the following properties:
3501 - The type must match the type loaded by the instruction.
3502 - The pair ``a,b`` represents the range ``[a,b)``.
3503 - Both ``a`` and ``b`` are constants.
3504 - The range is allowed to wrap.
3505 - The range should not represent the full or empty set. That is,
3508 In addition, the pairs must be in signed order of the lower bound and
3509 they must be non-contiguous.
3513 .. code-block:: llvm
3515 %a = load i8, i8* %x, align 1, !range !0 ; Can only be 0 or 1
3516 %b = load i8, i8* %y, align 1, !range !1 ; Can only be 255 (-1), 0 or 1
3517 %c = call i8 @foo(), !range !2 ; Can only be 0, 1, 3, 4 or 5
3518 %d = invoke i8 @bar() to label %cont
3519 unwind label %lpad, !range !3 ; Can only be -2, -1, 3, 4 or 5
3521 !0 = !{ i8 0, i8 2 }
3522 !1 = !{ i8 255, i8 2 }
3523 !2 = !{ i8 0, i8 2, i8 3, i8 6 }
3524 !3 = !{ i8 -2, i8 0, i8 3, i8 6 }
3529 It is sometimes useful to attach information to loop constructs. Currently,
3530 loop metadata is implemented as metadata attached to the branch instruction
3531 in the loop latch block. This type of metadata refer to a metadata node that is
3532 guaranteed to be separate for each loop. The loop identifier metadata is
3533 specified with the name ``llvm.loop``.
3535 The loop identifier metadata is implemented using a metadata that refers to
3536 itself to avoid merging it with any other identifier metadata, e.g.,
3537 during module linkage or function inlining. That is, each loop should refer
3538 to their own identification metadata even if they reside in separate functions.
3539 The following example contains loop identifier metadata for two separate loop
3542 .. code-block:: llvm
3547 The loop identifier metadata can be used to specify additional
3548 per-loop metadata. Any operands after the first operand can be treated
3549 as user-defined metadata. For example the ``llvm.loop.unroll.count``
3550 suggests an unroll factor to the loop unroller:
3552 .. code-block:: llvm
3554 br i1 %exitcond, label %._crit_edge, label %.lr.ph, !llvm.loop !0
3557 !1 = !{!"llvm.loop.unroll.count", i32 4}
3559 '``llvm.loop.vectorize``' and '``llvm.loop.interleave``'
3560 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3562 Metadata prefixed with ``llvm.loop.vectorize`` or ``llvm.loop.interleave`` are
3563 used to control per-loop vectorization and interleaving parameters such as
3564 vectorization width and interleave count. These metadata should be used in
3565 conjunction with ``llvm.loop`` loop identification metadata. The
3566 ``llvm.loop.vectorize`` and ``llvm.loop.interleave`` metadata are only
3567 optimization hints and the optimizer will only interleave and vectorize loops if
3568 it believes it is safe to do so. The ``llvm.mem.parallel_loop_access`` metadata
3569 which contains information about loop-carried memory dependencies can be helpful
3570 in determining the safety of these transformations.
3572 '``llvm.loop.interleave.count``' Metadata
3573 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3575 This metadata suggests an interleave count to the loop interleaver.
3576 The first operand is the string ``llvm.loop.interleave.count`` and the
3577 second operand is an integer specifying the interleave count. For
3580 .. code-block:: llvm
3582 !0 = !{!"llvm.loop.interleave.count", i32 4}
3584 Note that setting ``llvm.loop.interleave.count`` to 1 disables interleaving
3585 multiple iterations of the loop. If ``llvm.loop.interleave.count`` is set to 0
3586 then the interleave count will be determined automatically.
3588 '``llvm.loop.vectorize.enable``' Metadata
3589 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3591 This metadata selectively enables or disables vectorization for the loop. The
3592 first operand is the string ``llvm.loop.vectorize.enable`` and the second operand
3593 is a bit. If the bit operand value is 1 vectorization is enabled. A value of
3594 0 disables vectorization:
3596 .. code-block:: llvm
3598 !0 = !{!"llvm.loop.vectorize.enable", i1 0}
3599 !1 = !{!"llvm.loop.vectorize.enable", i1 1}
3601 '``llvm.loop.vectorize.width``' Metadata
3602 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3604 This metadata sets the target width of the vectorizer. The first
3605 operand is the string ``llvm.loop.vectorize.width`` and the second
3606 operand is an integer specifying the width. For example:
3608 .. code-block:: llvm
3610 !0 = !{!"llvm.loop.vectorize.width", i32 4}
3612 Note that setting ``llvm.loop.vectorize.width`` to 1 disables
3613 vectorization of the loop. If ``llvm.loop.vectorize.width`` is set to
3614 0 or if the loop does not have this metadata the width will be
3615 determined automatically.
3617 '``llvm.loop.unroll``'
3618 ^^^^^^^^^^^^^^^^^^^^^^
3620 Metadata prefixed with ``llvm.loop.unroll`` are loop unrolling
3621 optimization hints such as the unroll factor. ``llvm.loop.unroll``
3622 metadata should be used in conjunction with ``llvm.loop`` loop
3623 identification metadata. The ``llvm.loop.unroll`` metadata are only
3624 optimization hints and the unrolling will only be performed if the
3625 optimizer believes it is safe to do so.
3627 '``llvm.loop.unroll.count``' Metadata
3628 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3630 This metadata suggests an unroll factor to the loop unroller. The
3631 first operand is the string ``llvm.loop.unroll.count`` and the second
3632 operand is a positive integer specifying the unroll factor. For
3635 .. code-block:: llvm
3637 !0 = !{!"llvm.loop.unroll.count", i32 4}
3639 If the trip count of the loop is less than the unroll count the loop
3640 will be partially unrolled.
3642 '``llvm.loop.unroll.disable``' Metadata
3643 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3645 This metadata disables loop unrolling. The metadata has a single operand
3646 which is the string ``llvm.loop.unroll.disable``. For example:
3648 .. code-block:: llvm
3650 !0 = !{!"llvm.loop.unroll.disable"}
3652 '``llvm.loop.unroll.runtime.disable``' Metadata
3653 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3655 This metadata disables runtime loop unrolling. The metadata has a single
3656 operand which is the string ``llvm.loop.unroll.runtime.disable``. For example:
3658 .. code-block:: llvm
3660 !0 = !{!"llvm.loop.unroll.runtime.disable"}
3662 '``llvm.loop.unroll.full``' Metadata
3663 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3665 This metadata suggests that the loop should be unrolled fully. The
3666 metadata has a single operand which is the string ``llvm.loop.unroll.full``.
3669 .. code-block:: llvm
3671 !0 = !{!"llvm.loop.unroll.full"}
3676 Metadata types used to annotate memory accesses with information helpful
3677 for optimizations are prefixed with ``llvm.mem``.
3679 '``llvm.mem.parallel_loop_access``' Metadata
3680 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3682 The ``llvm.mem.parallel_loop_access`` metadata refers to a loop identifier,
3683 or metadata containing a list of loop identifiers for nested loops.
3684 The metadata is attached to memory accessing instructions and denotes that
3685 no loop carried memory dependence exist between it and other instructions denoted
3686 with the same loop identifier.
3688 Precisely, given two instructions ``m1`` and ``m2`` that both have the
3689 ``llvm.mem.parallel_loop_access`` metadata, with ``L1`` and ``L2`` being the
3690 set of loops associated with that metadata, respectively, then there is no loop
3691 carried dependence between ``m1`` and ``m2`` for loops in both ``L1`` and
3694 As a special case, if all memory accessing instructions in a loop have
3695 ``llvm.mem.parallel_loop_access`` metadata that refers to that loop, then the
3696 loop has no loop carried memory dependences and is considered to be a parallel
3699 Note that if not all memory access instructions have such metadata referring to
3700 the loop, then the loop is considered not being trivially parallel. Additional
3701 memory dependence analysis is required to make that determination. As a fail
3702 safe mechanism, this causes loops that were originally parallel to be considered
3703 sequential (if optimization passes that are unaware of the parallel semantics
3704 insert new memory instructions into the loop body).
3706 Example of a loop that is considered parallel due to its correct use of
3707 both ``llvm.loop`` and ``llvm.mem.parallel_loop_access``
3708 metadata types that refer to the same loop identifier metadata.
3710 .. code-block:: llvm
3714 %val0 = load i32, i32* %arrayidx, !llvm.mem.parallel_loop_access !0
3716 store i32 %val0, i32* %arrayidx1, !llvm.mem.parallel_loop_access !0
3718 br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !0
3724 It is also possible to have nested parallel loops. In that case the
3725 memory accesses refer to a list of loop identifier metadata nodes instead of
3726 the loop identifier metadata node directly:
3728 .. code-block:: llvm
3732 %val1 = load i32, i32* %arrayidx3, !llvm.mem.parallel_loop_access !2
3734 br label %inner.for.body
3738 %val0 = load i32, i32* %arrayidx1, !llvm.mem.parallel_loop_access !0
3740 store i32 %val0, i32* %arrayidx2, !llvm.mem.parallel_loop_access !0
3742 br i1 %exitcond, label %inner.for.end, label %inner.for.body, !llvm.loop !1
3746 store i32 %val1, i32* %arrayidx4, !llvm.mem.parallel_loop_access !2
3748 br i1 %exitcond, label %outer.for.end, label %outer.for.body, !llvm.loop !2
3750 outer.for.end: ; preds = %for.body
3752 !0 = !{!1, !2} ; a list of loop identifiers
3753 !1 = !{!1} ; an identifier for the inner loop
3754 !2 = !{!2} ; an identifier for the outer loop
3759 The ``llvm.bitsets`` global metadata is used to implement
3760 :doc:`bitsets <BitSets>`.
3762 Module Flags Metadata
3763 =====================
3765 Information about the module as a whole is difficult to convey to LLVM's
3766 subsystems. The LLVM IR isn't sufficient to transmit this information.
3767 The ``llvm.module.flags`` named metadata exists in order to facilitate
3768 this. These flags are in the form of key / value pairs --- much like a
3769 dictionary --- making it easy for any subsystem who cares about a flag to
3772 The ``llvm.module.flags`` metadata contains a list of metadata triplets.
3773 Each triplet has the following form:
3775 - The first element is a *behavior* flag, which specifies the behavior
3776 when two (or more) modules are merged together, and it encounters two
3777 (or more) metadata with the same ID. The supported behaviors are
3779 - The second element is a metadata string that is a unique ID for the
3780 metadata. Each module may only have one flag entry for each unique ID (not
3781 including entries with the **Require** behavior).
3782 - The third element is the value of the flag.
3784 When two (or more) modules are merged together, the resulting
3785 ``llvm.module.flags`` metadata is the union of the modules' flags. That is, for
3786 each unique metadata ID string, there will be exactly one entry in the merged
3787 modules ``llvm.module.flags`` metadata table, and the value for that entry will
3788 be determined by the merge behavior flag, as described below. The only exception
3789 is that entries with the *Require* behavior are always preserved.
3791 The following behaviors are supported:
3802 Emits an error if two values disagree, otherwise the resulting value
3803 is that of the operands.
3807 Emits a warning if two values disagree. The result value will be the
3808 operand for the flag from the first module being linked.
3812 Adds a requirement that another module flag be present and have a
3813 specified value after linking is performed. The value must be a
3814 metadata pair, where the first element of the pair is the ID of the
3815 module flag to be restricted, and the second element of the pair is
3816 the value the module flag should be restricted to. This behavior can
3817 be used to restrict the allowable results (via triggering of an
3818 error) of linking IDs with the **Override** behavior.
3822 Uses the specified value, regardless of the behavior or value of the
3823 other module. If both modules specify **Override**, but the values
3824 differ, an error will be emitted.
3828 Appends the two values, which are required to be metadata nodes.
3832 Appends the two values, which are required to be metadata
3833 nodes. However, duplicate entries in the second list are dropped
3834 during the append operation.
3836 It is an error for a particular unique flag ID to have multiple behaviors,
3837 except in the case of **Require** (which adds restrictions on another metadata
3838 value) or **Override**.
3840 An example of module flags:
3842 .. code-block:: llvm
3844 !0 = !{ i32 1, !"foo", i32 1 }
3845 !1 = !{ i32 4, !"bar", i32 37 }
3846 !2 = !{ i32 2, !"qux", i32 42 }
3847 !3 = !{ i32 3, !"qux",
3852 !llvm.module.flags = !{ !0, !1, !2, !3 }
3854 - Metadata ``!0`` has the ID ``!"foo"`` and the value '1'. The behavior
3855 if two or more ``!"foo"`` flags are seen is to emit an error if their
3856 values are not equal.
3858 - Metadata ``!1`` has the ID ``!"bar"`` and the value '37'. The
3859 behavior if two or more ``!"bar"`` flags are seen is to use the value
3862 - Metadata ``!2`` has the ID ``!"qux"`` and the value '42'. The
3863 behavior if two or more ``!"qux"`` flags are seen is to emit a
3864 warning if their values are not equal.
3866 - Metadata ``!3`` has the ID ``!"qux"`` and the value:
3872 The behavior is to emit an error if the ``llvm.module.flags`` does not
3873 contain a flag with the ID ``!"foo"`` that has the value '1' after linking is
3876 Objective-C Garbage Collection Module Flags Metadata
3877 ----------------------------------------------------
3879 On the Mach-O platform, Objective-C stores metadata about garbage
3880 collection in a special section called "image info". The metadata
3881 consists of a version number and a bitmask specifying what types of
3882 garbage collection are supported (if any) by the file. If two or more
3883 modules are linked together their garbage collection metadata needs to
3884 be merged rather than appended together.
3886 The Objective-C garbage collection module flags metadata consists of the
3887 following key-value pairs:
3896 * - ``Objective-C Version``
3897 - **[Required]** --- The Objective-C ABI version. Valid values are 1 and 2.
3899 * - ``Objective-C Image Info Version``
3900 - **[Required]** --- The version of the image info section. Currently
3903 * - ``Objective-C Image Info Section``
3904 - **[Required]** --- The section to place the metadata. Valid values are
3905 ``"__OBJC, __image_info, regular"`` for Objective-C ABI version 1, and
3906 ``"__DATA,__objc_imageinfo, regular, no_dead_strip"`` for
3907 Objective-C ABI version 2.
3909 * - ``Objective-C Garbage Collection``
3910 - **[Required]** --- Specifies whether garbage collection is supported or
3911 not. Valid values are 0, for no garbage collection, and 2, for garbage
3912 collection supported.
3914 * - ``Objective-C GC Only``
3915 - **[Optional]** --- Specifies that only garbage collection is supported.
3916 If present, its value must be 6. This flag requires that the
3917 ``Objective-C Garbage Collection`` flag have the value 2.
3919 Some important flag interactions:
3921 - If a module with ``Objective-C Garbage Collection`` set to 0 is
3922 merged with a module with ``Objective-C Garbage Collection`` set to
3923 2, then the resulting module has the
3924 ``Objective-C Garbage Collection`` flag set to 0.
3925 - A module with ``Objective-C Garbage Collection`` set to 0 cannot be
3926 merged with a module with ``Objective-C GC Only`` set to 6.
3928 Automatic Linker Flags Module Flags Metadata
3929 --------------------------------------------
3931 Some targets support embedding flags to the linker inside individual object
3932 files. Typically this is used in conjunction with language extensions which
3933 allow source files to explicitly declare the libraries they depend on, and have
3934 these automatically be transmitted to the linker via object files.
3936 These flags are encoded in the IR using metadata in the module flags section,
3937 using the ``Linker Options`` key. The merge behavior for this flag is required
3938 to be ``AppendUnique``, and the value for the key is expected to be a metadata
3939 node which should be a list of other metadata nodes, each of which should be a
3940 list of metadata strings defining linker options.
3942 For example, the following metadata section specifies two separate sets of
3943 linker options, presumably to link against ``libz`` and the ``Cocoa``
3946 !0 = !{ i32 6, !"Linker Options",
3949 !{ !"-framework", !"Cocoa" } } }
3950 !llvm.module.flags = !{ !0 }
3952 The metadata encoding as lists of lists of options, as opposed to a collapsed
3953 list of options, is chosen so that the IR encoding can use multiple option
3954 strings to specify e.g., a single library, while still having that specifier be
3955 preserved as an atomic element that can be recognized by a target specific
3956 assembly writer or object file emitter.
3958 Each individual option is required to be either a valid option for the target's
3959 linker, or an option that is reserved by the target specific assembly writer or
3960 object file emitter. No other aspect of these options is defined by the IR.
3962 C type width Module Flags Metadata
3963 ----------------------------------
3965 The ARM backend emits a section into each generated object file describing the
3966 options that it was compiled with (in a compiler-independent way) to prevent
3967 linking incompatible objects, and to allow automatic library selection. Some
3968 of these options are not visible at the IR level, namely wchar_t width and enum
3971 To pass this information to the backend, these options are encoded in module
3972 flags metadata, using the following key-value pairs:
3982 - * 0 --- sizeof(wchar_t) == 4
3983 * 1 --- sizeof(wchar_t) == 2
3986 - * 0 --- Enums are at least as large as an ``int``.
3987 * 1 --- Enums are stored in the smallest integer type which can
3988 represent all of its values.
3990 For example, the following metadata section specifies that the module was
3991 compiled with a ``wchar_t`` width of 4 bytes, and the underlying type of an
3992 enum is the smallest type which can represent all of its values::
3994 !llvm.module.flags = !{!0, !1}
3995 !0 = !{i32 1, !"short_wchar", i32 1}
3996 !1 = !{i32 1, !"short_enum", i32 0}
3998 .. _intrinsicglobalvariables:
4000 Intrinsic Global Variables
4001 ==========================
4003 LLVM has a number of "magic" global variables that contain data that
4004 affect code generation or other IR semantics. These are documented here.
4005 All globals of this sort should have a section specified as
4006 "``llvm.metadata``". This section and all globals that start with
4007 "``llvm.``" are reserved for use by LLVM.
4011 The '``llvm.used``' Global Variable
4012 -----------------------------------
4014 The ``@llvm.used`` global is an array which has
4015 :ref:`appending linkage <linkage_appending>`. This array contains a list of
4016 pointers to named global variables, functions and aliases which may optionally
4017 have a pointer cast formed of bitcast or getelementptr. For example, a legal
4020 .. code-block:: llvm
4025 @llvm.used = appending global [2 x i8*] [
4027 i8* bitcast (i32* @Y to i8*)
4028 ], section "llvm.metadata"
4030 If a symbol appears in the ``@llvm.used`` list, then the compiler, assembler,
4031 and linker are required to treat the symbol as if there is a reference to the
4032 symbol that it cannot see (which is why they have to be named). For example, if
4033 a variable has internal linkage and no references other than that from the
4034 ``@llvm.used`` list, it cannot be deleted. This is commonly used to represent
4035 references from inline asms and other things the compiler cannot "see", and
4036 corresponds to "``attribute((used))``" in GNU C.
4038 On some targets, the code generator must emit a directive to the
4039 assembler or object file to prevent the assembler and linker from
4040 molesting the symbol.
4042 .. _gv_llvmcompilerused:
4044 The '``llvm.compiler.used``' Global Variable
4045 --------------------------------------------
4047 The ``@llvm.compiler.used`` directive is the same as the ``@llvm.used``
4048 directive, except that it only prevents the compiler from touching the
4049 symbol. On targets that support it, this allows an intelligent linker to
4050 optimize references to the symbol without being impeded as it would be
4053 This is a rare construct that should only be used in rare circumstances,
4054 and should not be exposed to source languages.
4056 .. _gv_llvmglobalctors:
4058 The '``llvm.global_ctors``' Global Variable
4059 -------------------------------------------
4061 .. code-block:: llvm
4063 %0 = type { i32, void ()*, i8* }
4064 @llvm.global_ctors = appending global [1 x %0] [%0 { i32 65535, void ()* @ctor, i8* @data }]
4066 The ``@llvm.global_ctors`` array contains a list of constructor
4067 functions, priorities, and an optional associated global or function.
4068 The functions referenced by this array will be called in ascending order
4069 of priority (i.e. lowest first) when the module is loaded. The order of
4070 functions with the same priority is not defined.
4072 If the third field is present, non-null, and points to a global variable
4073 or function, the initializer function will only run if the associated
4074 data from the current module is not discarded.
4076 .. _llvmglobaldtors:
4078 The '``llvm.global_dtors``' Global Variable
4079 -------------------------------------------
4081 .. code-block:: llvm
4083 %0 = type { i32, void ()*, i8* }
4084 @llvm.global_dtors = appending global [1 x %0] [%0 { i32 65535, void ()* @dtor, i8* @data }]
4086 The ``@llvm.global_dtors`` array contains a list of destructor
4087 functions, priorities, and an optional associated global or function.
4088 The functions referenced by this array will be called in descending
4089 order of priority (i.e. highest first) when the module is unloaded. The
4090 order of functions with the same priority is not defined.
4092 If the third field is present, non-null, and points to a global variable
4093 or function, the destructor function will only run if the associated
4094 data from the current module is not discarded.
4096 Instruction Reference
4097 =====================
4099 The LLVM instruction set consists of several different classifications
4100 of instructions: :ref:`terminator instructions <terminators>`, :ref:`binary
4101 instructions <binaryops>`, :ref:`bitwise binary
4102 instructions <bitwiseops>`, :ref:`memory instructions <memoryops>`, and
4103 :ref:`other instructions <otherops>`.
4107 Terminator Instructions
4108 -----------------------
4110 As mentioned :ref:`previously <functionstructure>`, every basic block in a
4111 program ends with a "Terminator" instruction, which indicates which
4112 block should be executed after the current block is finished. These
4113 terminator instructions typically yield a '``void``' value: they produce
4114 control flow, not values (the one exception being the
4115 ':ref:`invoke <i_invoke>`' instruction).
4117 The terminator instructions are: ':ref:`ret <i_ret>`',
4118 ':ref:`br <i_br>`', ':ref:`switch <i_switch>`',
4119 ':ref:`indirectbr <i_indirectbr>`', ':ref:`invoke <i_invoke>`',
4120 ':ref:`resume <i_resume>`', and ':ref:`unreachable <i_unreachable>`'.
4124 '``ret``' Instruction
4125 ^^^^^^^^^^^^^^^^^^^^^
4132 ret <type> <value> ; Return a value from a non-void function
4133 ret void ; Return from void function
4138 The '``ret``' instruction is used to return control flow (and optionally
4139 a value) from a function back to the caller.
4141 There are two forms of the '``ret``' instruction: one that returns a
4142 value and then causes control flow, and one that just causes control
4148 The '``ret``' instruction optionally accepts a single argument, the
4149 return value. The type of the return value must be a ':ref:`first
4150 class <t_firstclass>`' type.
4152 A function is not :ref:`well formed <wellformed>` if it it has a non-void
4153 return type and contains a '``ret``' instruction with no return value or
4154 a return value with a type that does not match its type, or if it has a
4155 void return type and contains a '``ret``' instruction with a return
4161 When the '``ret``' instruction is executed, control flow returns back to
4162 the calling function's context. If the caller is a
4163 ":ref:`call <i_call>`" instruction, execution continues at the
4164 instruction after the call. If the caller was an
4165 ":ref:`invoke <i_invoke>`" instruction, execution continues at the
4166 beginning of the "normal" destination block. If the instruction returns
4167 a value, that value shall set the call or invoke instruction's return
4173 .. code-block:: llvm
4175 ret i32 5 ; Return an integer value of 5
4176 ret void ; Return from a void function
4177 ret { i32, i8 } { i32 4, i8 2 } ; Return a struct of values 4 and 2
4181 '``br``' Instruction
4182 ^^^^^^^^^^^^^^^^^^^^
4189 br i1 <cond>, label <iftrue>, label <iffalse>
4190 br label <dest> ; Unconditional branch
4195 The '``br``' instruction is used to cause control flow to transfer to a
4196 different basic block in the current function. There are two forms of
4197 this instruction, corresponding to a conditional branch and an
4198 unconditional branch.
4203 The conditional branch form of the '``br``' instruction takes a single
4204 '``i1``' value and two '``label``' values. The unconditional form of the
4205 '``br``' instruction takes a single '``label``' value as a target.
4210 Upon execution of a conditional '``br``' instruction, the '``i1``'
4211 argument is evaluated. If the value is ``true``, control flows to the
4212 '``iftrue``' ``label`` argument. If "cond" is ``false``, control flows
4213 to the '``iffalse``' ``label`` argument.
4218 .. code-block:: llvm
4221 %cond = icmp eq i32 %a, %b
4222 br i1 %cond, label %IfEqual, label %IfUnequal
4230 '``switch``' Instruction
4231 ^^^^^^^^^^^^^^^^^^^^^^^^
4238 switch <intty> <value>, label <defaultdest> [ <intty> <val>, label <dest> ... ]
4243 The '``switch``' instruction is used to transfer control flow to one of
4244 several different places. It is a generalization of the '``br``'
4245 instruction, allowing a branch to occur to one of many possible
4251 The '``switch``' instruction uses three parameters: an integer
4252 comparison value '``value``', a default '``label``' destination, and an
4253 array of pairs of comparison value constants and '``label``'s. The table
4254 is not allowed to contain duplicate constant entries.
4259 The ``switch`` instruction specifies a table of values and destinations.
4260 When the '``switch``' instruction is executed, this table is searched
4261 for the given value. If the value is found, control flow is transferred
4262 to the corresponding destination; otherwise, control flow is transferred
4263 to the default destination.
4268 Depending on properties of the target machine and the particular
4269 ``switch`` instruction, this instruction may be code generated in
4270 different ways. For example, it could be generated as a series of
4271 chained conditional branches or with a lookup table.
4276 .. code-block:: llvm
4278 ; Emulate a conditional br instruction
4279 %Val = zext i1 %value to i32
4280 switch i32 %Val, label %truedest [ i32 0, label %falsedest ]
4282 ; Emulate an unconditional br instruction
4283 switch i32 0, label %dest [ ]
4285 ; Implement a jump table:
4286 switch i32 %val, label %otherwise [ i32 0, label %onzero
4288 i32 2, label %ontwo ]
4292 '``indirectbr``' Instruction
4293 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4300 indirectbr <somety>* <address>, [ label <dest1>, label <dest2>, ... ]
4305 The '``indirectbr``' instruction implements an indirect branch to a
4306 label within the current function, whose address is specified by
4307 "``address``". Address must be derived from a
4308 :ref:`blockaddress <blockaddress>` constant.
4313 The '``address``' argument is the address of the label to jump to. The
4314 rest of the arguments indicate the full set of possible destinations
4315 that the address may point to. Blocks are allowed to occur multiple
4316 times in the destination list, though this isn't particularly useful.
4318 This destination list is required so that dataflow analysis has an
4319 accurate understanding of the CFG.
4324 Control transfers to the block specified in the address argument. All
4325 possible destination blocks must be listed in the label list, otherwise
4326 this instruction has undefined behavior. This implies that jumps to
4327 labels defined in other functions have undefined behavior as well.
4332 This is typically implemented with a jump through a register.
4337 .. code-block:: llvm
4339 indirectbr i8* %Addr, [ label %bb1, label %bb2, label %bb3 ]
4343 '``invoke``' Instruction
4344 ^^^^^^^^^^^^^^^^^^^^^^^^
4351 <result> = invoke [cconv] [ret attrs] <ptr to function ty> <function ptr val>(<function args>) [fn attrs]
4352 to label <normal label> unwind label <exception label>
4357 The '``invoke``' instruction causes control to transfer to a specified
4358 function, with the possibility of control flow transfer to either the
4359 '``normal``' label or the '``exception``' label. If the callee function
4360 returns with the "``ret``" instruction, control flow will return to the
4361 "normal" label. If the callee (or any indirect callees) returns via the
4362 ":ref:`resume <i_resume>`" instruction or other exception handling
4363 mechanism, control is interrupted and continued at the dynamically
4364 nearest "exception" label.
4366 The '``exception``' label is a `landing
4367 pad <ExceptionHandling.html#overview>`_ for the exception. As such,
4368 '``exception``' label is required to have the
4369 ":ref:`landingpad <i_landingpad>`" instruction, which contains the
4370 information about the behavior of the program after unwinding happens,
4371 as its first non-PHI instruction. The restrictions on the
4372 "``landingpad``" instruction's tightly couples it to the "``invoke``"
4373 instruction, so that the important information contained within the
4374 "``landingpad``" instruction can't be lost through normal code motion.
4379 This instruction requires several arguments:
4381 #. The optional "cconv" marker indicates which :ref:`calling
4382 convention <callingconv>` the call should use. If none is
4383 specified, the call defaults to using C calling conventions.
4384 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
4385 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
4387 #. '``ptr to function ty``': shall be the signature of the pointer to
4388 function value being invoked. In most cases, this is a direct
4389 function invocation, but indirect ``invoke``'s are just as possible,
4390 branching off an arbitrary pointer to function value.
4391 #. '``function ptr val``': An LLVM value containing a pointer to a
4392 function to be invoked.
4393 #. '``function args``': argument list whose types match the function
4394 signature argument types and parameter attributes. All arguments must
4395 be of :ref:`first class <t_firstclass>` type. If the function signature
4396 indicates the function accepts a variable number of arguments, the
4397 extra arguments can be specified.
4398 #. '``normal label``': the label reached when the called function
4399 executes a '``ret``' instruction.
4400 #. '``exception label``': the label reached when a callee returns via
4401 the :ref:`resume <i_resume>` instruction or other exception handling
4403 #. The optional :ref:`function attributes <fnattrs>` list. Only
4404 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
4405 attributes are valid here.
4410 This instruction is designed to operate as a standard '``call``'
4411 instruction in most regards. The primary difference is that it
4412 establishes an association with a label, which is used by the runtime
4413 library to unwind the stack.
4415 This instruction is used in languages with destructors to ensure that
4416 proper cleanup is performed in the case of either a ``longjmp`` or a
4417 thrown exception. Additionally, this is important for implementation of
4418 '``catch``' clauses in high-level languages that support them.
4420 For the purposes of the SSA form, the definition of the value returned
4421 by the '``invoke``' instruction is deemed to occur on the edge from the
4422 current block to the "normal" label. If the callee unwinds then no
4423 return value is available.
4428 .. code-block:: llvm
4430 %retval = invoke i32 @Test(i32 15) to label %Continue
4431 unwind label %TestCleanup ; i32:retval set
4432 %retval = invoke coldcc i32 %Testfnptr(i32 15) to label %Continue
4433 unwind label %TestCleanup ; i32:retval set
4437 '``resume``' Instruction
4438 ^^^^^^^^^^^^^^^^^^^^^^^^
4445 resume <type> <value>
4450 The '``resume``' instruction is a terminator instruction that has no
4456 The '``resume``' instruction requires one argument, which must have the
4457 same type as the result of any '``landingpad``' instruction in the same
4463 The '``resume``' instruction resumes propagation of an existing
4464 (in-flight) exception whose unwinding was interrupted with a
4465 :ref:`landingpad <i_landingpad>` instruction.
4470 .. code-block:: llvm
4472 resume { i8*, i32 } %exn
4476 '``unreachable``' Instruction
4477 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4489 The '``unreachable``' instruction has no defined semantics. This
4490 instruction is used to inform the optimizer that a particular portion of
4491 the code is not reachable. This can be used to indicate that the code
4492 after a no-return function cannot be reached, and other facts.
4497 The '``unreachable``' instruction has no defined semantics.
4504 Binary operators are used to do most of the computation in a program.
4505 They require two operands of the same type, execute an operation on
4506 them, and produce a single value. The operands might represent multiple
4507 data, as is the case with the :ref:`vector <t_vector>` data type. The
4508 result value has the same type as its operands.
4510 There are several different binary operators:
4514 '``add``' Instruction
4515 ^^^^^^^^^^^^^^^^^^^^^
4522 <result> = add <ty> <op1>, <op2> ; yields ty:result
4523 <result> = add nuw <ty> <op1>, <op2> ; yields ty:result
4524 <result> = add nsw <ty> <op1>, <op2> ; yields ty:result
4525 <result> = add nuw nsw <ty> <op1>, <op2> ; yields ty:result
4530 The '``add``' instruction returns the sum of its two operands.
4535 The two arguments to the '``add``' instruction must be
4536 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4537 arguments must have identical types.
4542 The value produced is the integer sum of the two operands.
4544 If the sum has unsigned overflow, the result returned is the
4545 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
4548 Because LLVM integers use a two's complement representation, this
4549 instruction is appropriate for both signed and unsigned integers.
4551 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
4552 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
4553 result value of the ``add`` is a :ref:`poison value <poisonvalues>` if
4554 unsigned and/or signed overflow, respectively, occurs.
4559 .. code-block:: llvm
4561 <result> = add i32 4, %var ; yields i32:result = 4 + %var
4565 '``fadd``' Instruction
4566 ^^^^^^^^^^^^^^^^^^^^^^
4573 <result> = fadd [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
4578 The '``fadd``' instruction returns the sum of its two operands.
4583 The two arguments to the '``fadd``' instruction must be :ref:`floating
4584 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
4585 Both arguments must have identical types.
4590 The value produced is the floating point sum of the two operands. This
4591 instruction can also take any number of :ref:`fast-math flags <fastmath>`,
4592 which are optimization hints to enable otherwise unsafe floating point
4598 .. code-block:: llvm
4600 <result> = fadd float 4.0, %var ; yields float:result = 4.0 + %var
4602 '``sub``' Instruction
4603 ^^^^^^^^^^^^^^^^^^^^^
4610 <result> = sub <ty> <op1>, <op2> ; yields ty:result
4611 <result> = sub nuw <ty> <op1>, <op2> ; yields ty:result
4612 <result> = sub nsw <ty> <op1>, <op2> ; yields ty:result
4613 <result> = sub nuw nsw <ty> <op1>, <op2> ; yields ty:result
4618 The '``sub``' instruction returns the difference of its two operands.
4620 Note that the '``sub``' instruction is used to represent the '``neg``'
4621 instruction present in most other intermediate representations.
4626 The two arguments to the '``sub``' instruction must be
4627 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4628 arguments must have identical types.
4633 The value produced is the integer difference of the two operands.
4635 If the difference has unsigned overflow, the result returned is the
4636 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
4639 Because LLVM integers use a two's complement representation, this
4640 instruction is appropriate for both signed and unsigned integers.
4642 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
4643 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
4644 result value of the ``sub`` is a :ref:`poison value <poisonvalues>` if
4645 unsigned and/or signed overflow, respectively, occurs.
4650 .. code-block:: llvm
4652 <result> = sub i32 4, %var ; yields i32:result = 4 - %var
4653 <result> = sub i32 0, %val ; yields i32:result = -%var
4657 '``fsub``' Instruction
4658 ^^^^^^^^^^^^^^^^^^^^^^
4665 <result> = fsub [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
4670 The '``fsub``' instruction returns the difference of its two operands.
4672 Note that the '``fsub``' instruction is used to represent the '``fneg``'
4673 instruction present in most other intermediate representations.
4678 The two arguments to the '``fsub``' instruction must be :ref:`floating
4679 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
4680 Both arguments must have identical types.
4685 The value produced is the floating point difference of the two operands.
4686 This instruction can also take any number of :ref:`fast-math
4687 flags <fastmath>`, which are optimization hints to enable otherwise
4688 unsafe floating point optimizations:
4693 .. code-block:: llvm
4695 <result> = fsub float 4.0, %var ; yields float:result = 4.0 - %var
4696 <result> = fsub float -0.0, %val ; yields float:result = -%var
4698 '``mul``' Instruction
4699 ^^^^^^^^^^^^^^^^^^^^^
4706 <result> = mul <ty> <op1>, <op2> ; yields ty:result
4707 <result> = mul nuw <ty> <op1>, <op2> ; yields ty:result
4708 <result> = mul nsw <ty> <op1>, <op2> ; yields ty:result
4709 <result> = mul nuw nsw <ty> <op1>, <op2> ; yields ty:result
4714 The '``mul``' instruction returns the product of its two operands.
4719 The two arguments to the '``mul``' instruction must be
4720 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4721 arguments must have identical types.
4726 The value produced is the integer product of the two operands.
4728 If the result of the multiplication has unsigned overflow, the result
4729 returned is the mathematical result modulo 2\ :sup:`n`\ , where n is the
4730 bit width of the result.
4732 Because LLVM integers use a two's complement representation, and the
4733 result is the same width as the operands, this instruction returns the
4734 correct result for both signed and unsigned integers. If a full product
4735 (e.g. ``i32`` * ``i32`` -> ``i64``) is needed, the operands should be
4736 sign-extended or zero-extended as appropriate to the width of the full
4739 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
4740 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
4741 result value of the ``mul`` is a :ref:`poison value <poisonvalues>` if
4742 unsigned and/or signed overflow, respectively, occurs.
4747 .. code-block:: llvm
4749 <result> = mul i32 4, %var ; yields i32:result = 4 * %var
4753 '``fmul``' Instruction
4754 ^^^^^^^^^^^^^^^^^^^^^^
4761 <result> = fmul [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
4766 The '``fmul``' instruction returns the product of its two operands.
4771 The two arguments to the '``fmul``' instruction must be :ref:`floating
4772 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
4773 Both arguments must have identical types.
4778 The value produced is the floating point product of the two operands.
4779 This instruction can also take any number of :ref:`fast-math
4780 flags <fastmath>`, which are optimization hints to enable otherwise
4781 unsafe floating point optimizations:
4786 .. code-block:: llvm
4788 <result> = fmul float 4.0, %var ; yields float:result = 4.0 * %var
4790 '``udiv``' Instruction
4791 ^^^^^^^^^^^^^^^^^^^^^^
4798 <result> = udiv <ty> <op1>, <op2> ; yields ty:result
4799 <result> = udiv exact <ty> <op1>, <op2> ; yields ty:result
4804 The '``udiv``' instruction returns the quotient of its two operands.
4809 The two arguments to the '``udiv``' instruction must be
4810 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4811 arguments must have identical types.
4816 The value produced is the unsigned integer quotient of the two operands.
4818 Note that unsigned integer division and signed integer division are
4819 distinct operations; for signed integer division, use '``sdiv``'.
4821 Division by zero leads to undefined behavior.
4823 If the ``exact`` keyword is present, the result value of the ``udiv`` is
4824 a :ref:`poison value <poisonvalues>` if %op1 is not a multiple of %op2 (as
4825 such, "((a udiv exact b) mul b) == a").
4830 .. code-block:: llvm
4832 <result> = udiv i32 4, %var ; yields i32:result = 4 / %var
4834 '``sdiv``' Instruction
4835 ^^^^^^^^^^^^^^^^^^^^^^
4842 <result> = sdiv <ty> <op1>, <op2> ; yields ty:result
4843 <result> = sdiv exact <ty> <op1>, <op2> ; yields ty:result
4848 The '``sdiv``' instruction returns the quotient of its two operands.
4853 The two arguments to the '``sdiv``' instruction must be
4854 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4855 arguments must have identical types.
4860 The value produced is the signed integer quotient of the two operands
4861 rounded towards zero.
4863 Note that signed integer division and unsigned integer division are
4864 distinct operations; for unsigned integer division, use '``udiv``'.
4866 Division by zero leads to undefined behavior. Overflow also leads to
4867 undefined behavior; this is a rare case, but can occur, for example, by
4868 doing a 32-bit division of -2147483648 by -1.
4870 If the ``exact`` keyword is present, the result value of the ``sdiv`` is
4871 a :ref:`poison value <poisonvalues>` if the result would be rounded.
4876 .. code-block:: llvm
4878 <result> = sdiv i32 4, %var ; yields i32:result = 4 / %var
4882 '``fdiv``' Instruction
4883 ^^^^^^^^^^^^^^^^^^^^^^
4890 <result> = fdiv [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
4895 The '``fdiv``' instruction returns the quotient of its two operands.
4900 The two arguments to the '``fdiv``' instruction must be :ref:`floating
4901 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
4902 Both arguments must have identical types.
4907 The value produced is the floating point quotient of the two operands.
4908 This instruction can also take any number of :ref:`fast-math
4909 flags <fastmath>`, which are optimization hints to enable otherwise
4910 unsafe floating point optimizations:
4915 .. code-block:: llvm
4917 <result> = fdiv float 4.0, %var ; yields float:result = 4.0 / %var
4919 '``urem``' Instruction
4920 ^^^^^^^^^^^^^^^^^^^^^^
4927 <result> = urem <ty> <op1>, <op2> ; yields ty:result
4932 The '``urem``' instruction returns the remainder from the unsigned
4933 division of its two arguments.
4938 The two arguments to the '``urem``' instruction must be
4939 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4940 arguments must have identical types.
4945 This instruction returns the unsigned integer *remainder* of a division.
4946 This instruction always performs an unsigned division to get the
4949 Note that unsigned integer remainder and signed integer remainder are
4950 distinct operations; for signed integer remainder, use '``srem``'.
4952 Taking the remainder of a division by zero leads to undefined behavior.
4957 .. code-block:: llvm
4959 <result> = urem i32 4, %var ; yields i32:result = 4 % %var
4961 '``srem``' Instruction
4962 ^^^^^^^^^^^^^^^^^^^^^^
4969 <result> = srem <ty> <op1>, <op2> ; yields ty:result
4974 The '``srem``' instruction returns the remainder from the signed
4975 division of its two operands. This instruction can also take
4976 :ref:`vector <t_vector>` versions of the values in which case the elements
4982 The two arguments to the '``srem``' instruction must be
4983 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4984 arguments must have identical types.
4989 This instruction returns the *remainder* of a division (where the result
4990 is either zero or has the same sign as the dividend, ``op1``), not the
4991 *modulo* operator (where the result is either zero or has the same sign
4992 as the divisor, ``op2``) of a value. For more information about the
4993 difference, see `The Math
4994 Forum <http://mathforum.org/dr.math/problems/anne.4.28.99.html>`_. For a
4995 table of how this is implemented in various languages, please see
4997 operation <http://en.wikipedia.org/wiki/Modulo_operation>`_.
4999 Note that signed integer remainder and unsigned integer remainder are
5000 distinct operations; for unsigned integer remainder, use '``urem``'.
5002 Taking the remainder of a division by zero leads to undefined behavior.
5003 Overflow also leads to undefined behavior; this is a rare case, but can
5004 occur, for example, by taking the remainder of a 32-bit division of
5005 -2147483648 by -1. (The remainder doesn't actually overflow, but this
5006 rule lets srem be implemented using instructions that return both the
5007 result of the division and the remainder.)
5012 .. code-block:: llvm
5014 <result> = srem i32 4, %var ; yields i32:result = 4 % %var
5018 '``frem``' Instruction
5019 ^^^^^^^^^^^^^^^^^^^^^^
5026 <result> = frem [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
5031 The '``frem``' instruction returns the remainder from the division of
5037 The two arguments to the '``frem``' instruction must be :ref:`floating
5038 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
5039 Both arguments must have identical types.
5044 This instruction returns the *remainder* of a division. The remainder
5045 has the same sign as the dividend. This instruction can also take any
5046 number of :ref:`fast-math flags <fastmath>`, which are optimization hints
5047 to enable otherwise unsafe floating point optimizations:
5052 .. code-block:: llvm
5054 <result> = frem float 4.0, %var ; yields float:result = 4.0 % %var
5058 Bitwise Binary Operations
5059 -------------------------
5061 Bitwise binary operators are used to do various forms of bit-twiddling
5062 in a program. They are generally very efficient instructions and can
5063 commonly be strength reduced from other instructions. They require two
5064 operands of the same type, execute an operation on them, and produce a
5065 single value. The resulting value is the same type as its operands.
5067 '``shl``' Instruction
5068 ^^^^^^^^^^^^^^^^^^^^^
5075 <result> = shl <ty> <op1>, <op2> ; yields ty:result
5076 <result> = shl nuw <ty> <op1>, <op2> ; yields ty:result
5077 <result> = shl nsw <ty> <op1>, <op2> ; yields ty:result
5078 <result> = shl nuw nsw <ty> <op1>, <op2> ; yields ty:result
5083 The '``shl``' instruction returns the first operand shifted to the left
5084 a specified number of bits.
5089 Both arguments to the '``shl``' instruction must be the same
5090 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
5091 '``op2``' is treated as an unsigned value.
5096 The value produced is ``op1`` \* 2\ :sup:`op2` mod 2\ :sup:`n`,
5097 where ``n`` is the width of the result. If ``op2`` is (statically or
5098 dynamically) equal to or larger than the number of bits in
5099 ``op1``, the result is undefined. If the arguments are vectors, each
5100 vector element of ``op1`` is shifted by the corresponding shift amount
5103 If the ``nuw`` keyword is present, then the shift produces a :ref:`poison
5104 value <poisonvalues>` if it shifts out any non-zero bits. If the
5105 ``nsw`` keyword is present, then the shift produces a :ref:`poison
5106 value <poisonvalues>` if it shifts out any bits that disagree with the
5107 resultant sign bit. As such, NUW/NSW have the same semantics as they
5108 would if the shift were expressed as a mul instruction with the same
5109 nsw/nuw bits in (mul %op1, (shl 1, %op2)).
5114 .. code-block:: llvm
5116 <result> = shl i32 4, %var ; yields i32: 4 << %var
5117 <result> = shl i32 4, 2 ; yields i32: 16
5118 <result> = shl i32 1, 10 ; yields i32: 1024
5119 <result> = shl i32 1, 32 ; undefined
5120 <result> = shl <2 x i32> < i32 1, i32 1>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 2, i32 4>
5122 '``lshr``' Instruction
5123 ^^^^^^^^^^^^^^^^^^^^^^
5130 <result> = lshr <ty> <op1>, <op2> ; yields ty:result
5131 <result> = lshr exact <ty> <op1>, <op2> ; yields ty:result
5136 The '``lshr``' instruction (logical shift right) returns the first
5137 operand shifted to the right a specified number of bits with zero fill.
5142 Both arguments to the '``lshr``' instruction must be the same
5143 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
5144 '``op2``' is treated as an unsigned value.
5149 This instruction always performs a logical shift right operation. The
5150 most significant bits of the result will be filled with zero bits after
5151 the shift. If ``op2`` is (statically or dynamically) equal to or larger
5152 than the number of bits in ``op1``, the result is undefined. If the
5153 arguments are vectors, each vector element of ``op1`` is shifted by the
5154 corresponding shift amount in ``op2``.
5156 If the ``exact`` keyword is present, the result value of the ``lshr`` is
5157 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
5163 .. code-block:: llvm
5165 <result> = lshr i32 4, 1 ; yields i32:result = 2
5166 <result> = lshr i32 4, 2 ; yields i32:result = 1
5167 <result> = lshr i8 4, 3 ; yields i8:result = 0
5168 <result> = lshr i8 -2, 1 ; yields i8:result = 0x7F
5169 <result> = lshr i32 1, 32 ; undefined
5170 <result> = lshr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 0x7FFFFFFF, i32 1>
5172 '``ashr``' Instruction
5173 ^^^^^^^^^^^^^^^^^^^^^^
5180 <result> = ashr <ty> <op1>, <op2> ; yields ty:result
5181 <result> = ashr exact <ty> <op1>, <op2> ; yields ty:result
5186 The '``ashr``' instruction (arithmetic shift right) returns the first
5187 operand shifted to the right a specified number of bits with sign
5193 Both arguments to the '``ashr``' instruction must be the same
5194 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
5195 '``op2``' is treated as an unsigned value.
5200 This instruction always performs an arithmetic shift right operation,
5201 The most significant bits of the result will be filled with the sign bit
5202 of ``op1``. If ``op2`` is (statically or dynamically) equal to or larger
5203 than the number of bits in ``op1``, the result is undefined. If the
5204 arguments are vectors, each vector element of ``op1`` is shifted by the
5205 corresponding shift amount in ``op2``.
5207 If the ``exact`` keyword is present, the result value of the ``ashr`` is
5208 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
5214 .. code-block:: llvm
5216 <result> = ashr i32 4, 1 ; yields i32:result = 2
5217 <result> = ashr i32 4, 2 ; yields i32:result = 1
5218 <result> = ashr i8 4, 3 ; yields i8:result = 0
5219 <result> = ashr i8 -2, 1 ; yields i8:result = -1
5220 <result> = ashr i32 1, 32 ; undefined
5221 <result> = ashr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 3> ; yields: result=<2 x i32> < i32 -1, i32 0>
5223 '``and``' Instruction
5224 ^^^^^^^^^^^^^^^^^^^^^
5231 <result> = and <ty> <op1>, <op2> ; yields ty:result
5236 The '``and``' instruction returns the bitwise logical and of its two
5242 The two arguments to the '``and``' instruction must be
5243 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
5244 arguments must have identical types.
5249 The truth table used for the '``and``' instruction is:
5266 .. code-block:: llvm
5268 <result> = and i32 4, %var ; yields i32:result = 4 & %var
5269 <result> = and i32 15, 40 ; yields i32:result = 8
5270 <result> = and i32 4, 8 ; yields i32:result = 0
5272 '``or``' Instruction
5273 ^^^^^^^^^^^^^^^^^^^^
5280 <result> = or <ty> <op1>, <op2> ; yields ty:result
5285 The '``or``' instruction returns the bitwise logical inclusive or of its
5291 The two arguments to the '``or``' instruction must be
5292 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
5293 arguments must have identical types.
5298 The truth table used for the '``or``' instruction is:
5317 <result> = or i32 4, %var ; yields i32:result = 4 | %var
5318 <result> = or i32 15, 40 ; yields i32:result = 47
5319 <result> = or i32 4, 8 ; yields i32:result = 12
5321 '``xor``' Instruction
5322 ^^^^^^^^^^^^^^^^^^^^^
5329 <result> = xor <ty> <op1>, <op2> ; yields ty:result
5334 The '``xor``' instruction returns the bitwise logical exclusive or of
5335 its two operands. The ``xor`` is used to implement the "one's
5336 complement" operation, which is the "~" operator in C.
5341 The two arguments to the '``xor``' instruction must be
5342 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
5343 arguments must have identical types.
5348 The truth table used for the '``xor``' instruction is:
5365 .. code-block:: llvm
5367 <result> = xor i32 4, %var ; yields i32:result = 4 ^ %var
5368 <result> = xor i32 15, 40 ; yields i32:result = 39
5369 <result> = xor i32 4, 8 ; yields i32:result = 12
5370 <result> = xor i32 %V, -1 ; yields i32:result = ~%V
5375 LLVM supports several instructions to represent vector operations in a
5376 target-independent manner. These instructions cover the element-access
5377 and vector-specific operations needed to process vectors effectively.
5378 While LLVM does directly support these vector operations, many
5379 sophisticated algorithms will want to use target-specific intrinsics to
5380 take full advantage of a specific target.
5382 .. _i_extractelement:
5384 '``extractelement``' Instruction
5385 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5392 <result> = extractelement <n x <ty>> <val>, <ty2> <idx> ; yields <ty>
5397 The '``extractelement``' instruction extracts a single scalar element
5398 from a vector at a specified index.
5403 The first operand of an '``extractelement``' instruction is a value of
5404 :ref:`vector <t_vector>` type. The second operand is an index indicating
5405 the position from which to extract the element. The index may be a
5406 variable of any integer type.
5411 The result is a scalar of the same type as the element type of ``val``.
5412 Its value is the value at position ``idx`` of ``val``. If ``idx``
5413 exceeds the length of ``val``, the results are undefined.
5418 .. code-block:: llvm
5420 <result> = extractelement <4 x i32> %vec, i32 0 ; yields i32
5422 .. _i_insertelement:
5424 '``insertelement``' Instruction
5425 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5432 <result> = insertelement <n x <ty>> <val>, <ty> <elt>, <ty2> <idx> ; yields <n x <ty>>
5437 The '``insertelement``' instruction inserts a scalar element into a
5438 vector at a specified index.
5443 The first operand of an '``insertelement``' instruction is a value of
5444 :ref:`vector <t_vector>` type. The second operand is a scalar value whose
5445 type must equal the element type of the first operand. The third operand
5446 is an index indicating the position at which to insert the value. The
5447 index may be a variable of any integer type.
5452 The result is a vector of the same type as ``val``. Its element values
5453 are those of ``val`` except at position ``idx``, where it gets the value
5454 ``elt``. If ``idx`` exceeds the length of ``val``, the results are
5460 .. code-block:: llvm
5462 <result> = insertelement <4 x i32> %vec, i32 1, i32 0 ; yields <4 x i32>
5464 .. _i_shufflevector:
5466 '``shufflevector``' Instruction
5467 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5474 <result> = shufflevector <n x <ty>> <v1>, <n x <ty>> <v2>, <m x i32> <mask> ; yields <m x <ty>>
5479 The '``shufflevector``' instruction constructs a permutation of elements
5480 from two input vectors, returning a vector with the same element type as
5481 the input and length that is the same as the shuffle mask.
5486 The first two operands of a '``shufflevector``' instruction are vectors
5487 with the same type. The third argument is a shuffle mask whose element
5488 type is always 'i32'. The result of the instruction is a vector whose
5489 length is the same as the shuffle mask and whose element type is the
5490 same as the element type of the first two operands.
5492 The shuffle mask operand is required to be a constant vector with either
5493 constant integer or undef values.
5498 The elements of the two input vectors are numbered from left to right
5499 across both of the vectors. The shuffle mask operand specifies, for each
5500 element of the result vector, which element of the two input vectors the
5501 result element gets. The element selector may be undef (meaning "don't
5502 care") and the second operand may be undef if performing a shuffle from
5508 .. code-block:: llvm
5510 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
5511 <4 x i32> <i32 0, i32 4, i32 1, i32 5> ; yields <4 x i32>
5512 <result> = shufflevector <4 x i32> %v1, <4 x i32> undef,
5513 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32> - Identity shuffle.
5514 <result> = shufflevector <8 x i32> %v1, <8 x i32> undef,
5515 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32>
5516 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
5517 <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7 > ; yields <8 x i32>
5519 Aggregate Operations
5520 --------------------
5522 LLVM supports several instructions for working with
5523 :ref:`aggregate <t_aggregate>` values.
5527 '``extractvalue``' Instruction
5528 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5535 <result> = extractvalue <aggregate type> <val>, <idx>{, <idx>}*
5540 The '``extractvalue``' instruction extracts the value of a member field
5541 from an :ref:`aggregate <t_aggregate>` value.
5546 The first operand of an '``extractvalue``' instruction is a value of
5547 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The operands are
5548 constant indices to specify which value to extract in a similar manner
5549 as indices in a '``getelementptr``' instruction.
5551 The major differences to ``getelementptr`` indexing are:
5553 - Since the value being indexed is not a pointer, the first index is
5554 omitted and assumed to be zero.
5555 - At least one index must be specified.
5556 - Not only struct indices but also array indices must be in bounds.
5561 The result is the value at the position in the aggregate specified by
5567 .. code-block:: llvm
5569 <result> = extractvalue {i32, float} %agg, 0 ; yields i32
5573 '``insertvalue``' Instruction
5574 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5581 <result> = insertvalue <aggregate type> <val>, <ty> <elt>, <idx>{, <idx>}* ; yields <aggregate type>
5586 The '``insertvalue``' instruction inserts a value into a member field in
5587 an :ref:`aggregate <t_aggregate>` value.
5592 The first operand of an '``insertvalue``' instruction is a value of
5593 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The second operand is
5594 a first-class value to insert. The following operands are constant
5595 indices indicating the position at which to insert the value in a
5596 similar manner as indices in a '``extractvalue``' instruction. The value
5597 to insert must have the same type as the value identified by the
5603 The result is an aggregate of the same type as ``val``. Its value is
5604 that of ``val`` except that the value at the position specified by the
5605 indices is that of ``elt``.
5610 .. code-block:: llvm
5612 %agg1 = insertvalue {i32, float} undef, i32 1, 0 ; yields {i32 1, float undef}
5613 %agg2 = insertvalue {i32, float} %agg1, float %val, 1 ; yields {i32 1, float %val}
5614 %agg3 = insertvalue {i32, {float}} undef, float %val, 1, 0 ; yields {i32 undef, {float %val}}
5618 Memory Access and Addressing Operations
5619 ---------------------------------------
5621 A key design point of an SSA-based representation is how it represents
5622 memory. In LLVM, no memory locations are in SSA form, which makes things
5623 very simple. This section describes how to read, write, and allocate
5628 '``alloca``' Instruction
5629 ^^^^^^^^^^^^^^^^^^^^^^^^
5636 <result> = alloca [inalloca] <type> [, <ty> <NumElements>] [, align <alignment>] ; yields type*:result
5641 The '``alloca``' instruction allocates memory on the stack frame of the
5642 currently executing function, to be automatically released when this
5643 function returns to its caller. The object is always allocated in the
5644 generic address space (address space zero).
5649 The '``alloca``' instruction allocates ``sizeof(<type>)*NumElements``
5650 bytes of memory on the runtime stack, returning a pointer of the
5651 appropriate type to the program. If "NumElements" is specified, it is
5652 the number of elements allocated, otherwise "NumElements" is defaulted
5653 to be one. If a constant alignment is specified, the value result of the
5654 allocation is guaranteed to be aligned to at least that boundary. The
5655 alignment may not be greater than ``1 << 29``. If not specified, or if
5656 zero, the target can choose to align the allocation on any convenient
5657 boundary compatible with the type.
5659 '``type``' may be any sized type.
5664 Memory is allocated; a pointer is returned. The operation is undefined
5665 if there is insufficient stack space for the allocation. '``alloca``'d
5666 memory is automatically released when the function returns. The
5667 '``alloca``' instruction is commonly used to represent automatic
5668 variables that must have an address available. When the function returns
5669 (either with the ``ret`` or ``resume`` instructions), the memory is
5670 reclaimed. Allocating zero bytes is legal, but the result is undefined.
5671 The order in which memory is allocated (ie., which way the stack grows)
5677 .. code-block:: llvm
5679 %ptr = alloca i32 ; yields i32*:ptr
5680 %ptr = alloca i32, i32 4 ; yields i32*:ptr
5681 %ptr = alloca i32, i32 4, align 1024 ; yields i32*:ptr
5682 %ptr = alloca i32, align 1024 ; yields i32*:ptr
5686 '``load``' Instruction
5687 ^^^^^^^^^^^^^^^^^^^^^^
5694 <result> = load [volatile] <ty>, <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>][, !invariant.load !<index>][, !nonnull !<index>][, !dereferenceable !<index>][, !dereferenceable_or_null !<index>]
5695 <result> = load atomic [volatile] <ty>* <pointer> [singlethread] <ordering>, align <alignment>
5696 !<index> = !{ i32 1 }
5701 The '``load``' instruction is used to read from memory.
5706 The argument to the ``load`` instruction specifies the memory address
5707 from which to load. The type specified must be a :ref:`first
5708 class <t_firstclass>` type. If the ``load`` is marked as ``volatile``,
5709 then the optimizer is not allowed to modify the number or order of
5710 execution of this ``load`` with other :ref:`volatile
5711 operations <volatile>`.
5713 If the ``load`` is marked as ``atomic``, it takes an extra
5714 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
5715 ``release`` and ``acq_rel`` orderings are not valid on ``load``
5716 instructions. Atomic loads produce :ref:`defined <memmodel>` results
5717 when they may see multiple atomic stores. The type of the pointee must
5718 be an integer type whose bit width is a power of two greater than or
5719 equal to eight and less than or equal to a target-specific size limit.
5720 ``align`` must be explicitly specified on atomic loads, and the load has
5721 undefined behavior if the alignment is not set to a value which is at
5722 least the size in bytes of the pointee. ``!nontemporal`` does not have
5723 any defined semantics for atomic loads.
5725 The optional constant ``align`` argument specifies the alignment of the
5726 operation (that is, the alignment of the memory address). A value of 0
5727 or an omitted ``align`` argument means that the operation has the ABI
5728 alignment for the target. It is the responsibility of the code emitter
5729 to ensure that the alignment information is correct. Overestimating the
5730 alignment results in undefined behavior. Underestimating the alignment
5731 may produce less efficient code. An alignment of 1 is always safe. The
5732 maximum possible alignment is ``1 << 29``.
5734 The optional ``!nontemporal`` metadata must reference a single
5735 metadata name ``<index>`` corresponding to a metadata node with one
5736 ``i32`` entry of value 1. The existence of the ``!nontemporal``
5737 metadata on the instruction tells the optimizer and code generator
5738 that this load is not expected to be reused in the cache. The code
5739 generator may select special instructions to save cache bandwidth, such
5740 as the ``MOVNT`` instruction on x86.
5742 The optional ``!invariant.load`` metadata must reference a single
5743 metadata name ``<index>`` corresponding to a metadata node with no
5744 entries. The existence of the ``!invariant.load`` metadata on the
5745 instruction tells the optimizer and code generator that the address
5746 operand to this load points to memory which can be assumed unchanged.
5747 Being invariant does not imply that a location is dereferenceable,
5748 but it does imply that once the location is known dereferenceable
5749 its value is henceforth unchanging.
5751 The optional ``!nonnull`` metadata must reference a single
5752 metadata name ``<index>`` corresponding to a metadata node with no
5753 entries. The existence of the ``!nonnull`` metadata on the
5754 instruction tells the optimizer that the value loaded is known to
5755 never be null. This is analogous to the ''nonnull'' attribute
5756 on parameters and return values. This metadata can only be applied
5757 to loads of a pointer type.
5759 The optional ``!dereferenceable`` metadata must reference a single
5760 metadata name ``<index>`` corresponding to a metadata node with one ``i64``
5761 entry. The existence of the ``!dereferenceable`` metadata on the instruction
5762 tells the optimizer that the value loaded is known to be dereferenceable.
5763 The number of bytes known to be dereferenceable is specified by the integer
5764 value in the metadata node. This is analogous to the ''dereferenceable''
5765 attribute on parameters and return values. This metadata can only be applied
5766 to loads of a pointer type.
5768 The optional ``!dereferenceable_or_null`` metadata must reference a single
5769 metadata name ``<index>`` corresponding to a metadata node with one ``i64``
5770 entry. The existence of the ``!dereferenceable_or_null`` metadata on the
5771 instruction tells the optimizer that the value loaded is known to be either
5772 dereferenceable or null.
5773 The number of bytes known to be dereferenceable is specified by the integer
5774 value in the metadata node. This is analogous to the ''dereferenceable_or_null''
5775 attribute on parameters and return values. This metadata can only be applied
5776 to loads of a pointer type.
5781 The location of memory pointed to is loaded. If the value being loaded
5782 is of scalar type then the number of bytes read does not exceed the
5783 minimum number of bytes needed to hold all bits of the type. For
5784 example, loading an ``i24`` reads at most three bytes. When loading a
5785 value of a type like ``i20`` with a size that is not an integral number
5786 of bytes, the result is undefined if the value was not originally
5787 written using a store of the same type.
5792 .. code-block:: llvm
5794 %ptr = alloca i32 ; yields i32*:ptr
5795 store i32 3, i32* %ptr ; yields void
5796 %val = load i32, i32* %ptr ; yields i32:val = i32 3
5800 '``store``' Instruction
5801 ^^^^^^^^^^^^^^^^^^^^^^^
5808 store [volatile] <ty> <value>, <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>] ; yields void
5809 store atomic [volatile] <ty> <value>, <ty>* <pointer> [singlethread] <ordering>, align <alignment> ; yields void
5814 The '``store``' instruction is used to write to memory.
5819 There are two arguments to the ``store`` instruction: a value to store
5820 and an address at which to store it. The type of the ``<pointer>``
5821 operand must be a pointer to the :ref:`first class <t_firstclass>` type of
5822 the ``<value>`` operand. If the ``store`` is marked as ``volatile``,
5823 then the optimizer is not allowed to modify the number or order of
5824 execution of this ``store`` with other :ref:`volatile
5825 operations <volatile>`.
5827 If the ``store`` is marked as ``atomic``, it takes an extra
5828 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
5829 ``acquire`` and ``acq_rel`` orderings aren't valid on ``store``
5830 instructions. Atomic loads produce :ref:`defined <memmodel>` results
5831 when they may see multiple atomic stores. The type of the pointee must
5832 be an integer type whose bit width is a power of two greater than or
5833 equal to eight and less than or equal to a target-specific size limit.
5834 ``align`` must be explicitly specified on atomic stores, and the store
5835 has undefined behavior if the alignment is not set to a value which is
5836 at least the size in bytes of the pointee. ``!nontemporal`` does not
5837 have any defined semantics for atomic stores.
5839 The optional constant ``align`` argument specifies the alignment of the
5840 operation (that is, the alignment of the memory address). A value of 0
5841 or an omitted ``align`` argument means that the operation has the ABI
5842 alignment for the target. It is the responsibility of the code emitter
5843 to ensure that the alignment information is correct. Overestimating the
5844 alignment results in undefined behavior. Underestimating the
5845 alignment may produce less efficient code. An alignment of 1 is always
5846 safe. The maximum possible alignment is ``1 << 29``.
5848 The optional ``!nontemporal`` metadata must reference a single metadata
5849 name ``<index>`` corresponding to a metadata node with one ``i32`` entry of
5850 value 1. The existence of the ``!nontemporal`` metadata on the instruction
5851 tells the optimizer and code generator that this load is not expected to
5852 be reused in the cache. The code generator may select special
5853 instructions to save cache bandwidth, such as the MOVNT instruction on
5859 The contents of memory are updated to contain ``<value>`` at the
5860 location specified by the ``<pointer>`` operand. If ``<value>`` is
5861 of scalar type then the number of bytes written does not exceed the
5862 minimum number of bytes needed to hold all bits of the type. For
5863 example, storing an ``i24`` writes at most three bytes. When writing a
5864 value of a type like ``i20`` with a size that is not an integral number
5865 of bytes, it is unspecified what happens to the extra bits that do not
5866 belong to the type, but they will typically be overwritten.
5871 .. code-block:: llvm
5873 %ptr = alloca i32 ; yields i32*:ptr
5874 store i32 3, i32* %ptr ; yields void
5875 %val = load i32* %ptr ; yields i32:val = i32 3
5879 '``fence``' Instruction
5880 ^^^^^^^^^^^^^^^^^^^^^^^
5887 fence [singlethread] <ordering> ; yields void
5892 The '``fence``' instruction is used to introduce happens-before edges
5898 '``fence``' instructions take an :ref:`ordering <ordering>` argument which
5899 defines what *synchronizes-with* edges they add. They can only be given
5900 ``acquire``, ``release``, ``acq_rel``, and ``seq_cst`` orderings.
5905 A fence A which has (at least) ``release`` ordering semantics
5906 *synchronizes with* a fence B with (at least) ``acquire`` ordering
5907 semantics if and only if there exist atomic operations X and Y, both
5908 operating on some atomic object M, such that A is sequenced before X, X
5909 modifies M (either directly or through some side effect of a sequence
5910 headed by X), Y is sequenced before B, and Y observes M. This provides a
5911 *happens-before* dependency between A and B. Rather than an explicit
5912 ``fence``, one (but not both) of the atomic operations X or Y might
5913 provide a ``release`` or ``acquire`` (resp.) ordering constraint and
5914 still *synchronize-with* the explicit ``fence`` and establish the
5915 *happens-before* edge.
5917 A ``fence`` which has ``seq_cst`` ordering, in addition to having both
5918 ``acquire`` and ``release`` semantics specified above, participates in
5919 the global program order of other ``seq_cst`` operations and/or fences.
5921 The optional ":ref:`singlethread <singlethread>`" argument specifies
5922 that the fence only synchronizes with other fences in the same thread.
5923 (This is useful for interacting with signal handlers.)
5928 .. code-block:: llvm
5930 fence acquire ; yields void
5931 fence singlethread seq_cst ; yields void
5935 '``cmpxchg``' Instruction
5936 ^^^^^^^^^^^^^^^^^^^^^^^^^
5943 cmpxchg [weak] [volatile] <ty>* <pointer>, <ty> <cmp>, <ty> <new> [singlethread] <success ordering> <failure ordering> ; yields { ty, i1 }
5948 The '``cmpxchg``' instruction is used to atomically modify memory. It
5949 loads a value in memory and compares it to a given value. If they are
5950 equal, it tries to store a new value into the memory.
5955 There are three arguments to the '``cmpxchg``' instruction: an address
5956 to operate on, a value to compare to the value currently be at that
5957 address, and a new value to place at that address if the compared values
5958 are equal. The type of '<cmp>' must be an integer type whose bit width
5959 is a power of two greater than or equal to eight and less than or equal
5960 to a target-specific size limit. '<cmp>' and '<new>' must have the same
5961 type, and the type of '<pointer>' must be a pointer to that type. If the
5962 ``cmpxchg`` is marked as ``volatile``, then the optimizer is not allowed
5963 to modify the number or order of execution of this ``cmpxchg`` with
5964 other :ref:`volatile operations <volatile>`.
5966 The success and failure :ref:`ordering <ordering>` arguments specify how this
5967 ``cmpxchg`` synchronizes with other atomic operations. Both ordering parameters
5968 must be at least ``monotonic``, the ordering constraint on failure must be no
5969 stronger than that on success, and the failure ordering cannot be either
5970 ``release`` or ``acq_rel``.
5972 The optional "``singlethread``" argument declares that the ``cmpxchg``
5973 is only atomic with respect to code (usually signal handlers) running in
5974 the same thread as the ``cmpxchg``. Otherwise the cmpxchg is atomic with
5975 respect to all other code in the system.
5977 The pointer passed into cmpxchg must have alignment greater than or
5978 equal to the size in memory of the operand.
5983 The contents of memory at the location specified by the '``<pointer>``' operand
5984 is read and compared to '``<cmp>``'; if the read value is the equal, the
5985 '``<new>``' is written. The original value at the location is returned, together
5986 with a flag indicating success (true) or failure (false).
5988 If the cmpxchg operation is marked as ``weak`` then a spurious failure is
5989 permitted: the operation may not write ``<new>`` even if the comparison
5992 If the cmpxchg operation is strong (the default), the i1 value is 1 if and only
5993 if the value loaded equals ``cmp``.
5995 A successful ``cmpxchg`` is a read-modify-write instruction for the purpose of
5996 identifying release sequences. A failed ``cmpxchg`` is equivalent to an atomic
5997 load with an ordering parameter determined the second ordering parameter.
6002 .. code-block:: llvm
6005 %orig = atomic load i32, i32* %ptr unordered ; yields i32
6009 %cmp = phi i32 [ %orig, %entry ], [%old, %loop]
6010 %squared = mul i32 %cmp, %cmp
6011 %val_success = cmpxchg i32* %ptr, i32 %cmp, i32 %squared acq_rel monotonic ; yields { i32, i1 }
6012 %value_loaded = extractvalue { i32, i1 } %val_success, 0
6013 %success = extractvalue { i32, i1 } %val_success, 1
6014 br i1 %success, label %done, label %loop
6021 '``atomicrmw``' Instruction
6022 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6029 atomicrmw [volatile] <operation> <ty>* <pointer>, <ty> <value> [singlethread] <ordering> ; yields ty
6034 The '``atomicrmw``' instruction is used to atomically modify memory.
6039 There are three arguments to the '``atomicrmw``' instruction: an
6040 operation to apply, an address whose value to modify, an argument to the
6041 operation. The operation must be one of the following keywords:
6055 The type of '<value>' must be an integer type whose bit width is a power
6056 of two greater than or equal to eight and less than or equal to a
6057 target-specific size limit. The type of the '``<pointer>``' operand must
6058 be a pointer to that type. If the ``atomicrmw`` is marked as
6059 ``volatile``, then the optimizer is not allowed to modify the number or
6060 order of execution of this ``atomicrmw`` with other :ref:`volatile
6061 operations <volatile>`.
6066 The contents of memory at the location specified by the '``<pointer>``'
6067 operand are atomically read, modified, and written back. The original
6068 value at the location is returned. The modification is specified by the
6071 - xchg: ``*ptr = val``
6072 - add: ``*ptr = *ptr + val``
6073 - sub: ``*ptr = *ptr - val``
6074 - and: ``*ptr = *ptr & val``
6075 - nand: ``*ptr = ~(*ptr & val)``
6076 - or: ``*ptr = *ptr | val``
6077 - xor: ``*ptr = *ptr ^ val``
6078 - max: ``*ptr = *ptr > val ? *ptr : val`` (using a signed comparison)
6079 - min: ``*ptr = *ptr < val ? *ptr : val`` (using a signed comparison)
6080 - umax: ``*ptr = *ptr > val ? *ptr : val`` (using an unsigned
6082 - umin: ``*ptr = *ptr < val ? *ptr : val`` (using an unsigned
6088 .. code-block:: llvm
6090 %old = atomicrmw add i32* %ptr, i32 1 acquire ; yields i32
6092 .. _i_getelementptr:
6094 '``getelementptr``' Instruction
6095 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6102 <result> = getelementptr <ty>, <ty>* <ptrval>{, <ty> <idx>}*
6103 <result> = getelementptr inbounds <ty>, <ty>* <ptrval>{, <ty> <idx>}*
6104 <result> = getelementptr <ty>, <ptr vector> <ptrval>, <vector index type> <idx>
6109 The '``getelementptr``' instruction is used to get the address of a
6110 subelement of an :ref:`aggregate <t_aggregate>` data structure. It performs
6111 address calculation only and does not access memory.
6116 The first argument is always a type used as the basis for the calculations.
6117 The second argument is always a pointer or a vector of pointers, and is the
6118 base address to start from. The remaining arguments are indices
6119 that indicate which of the elements of the aggregate object are indexed.
6120 The interpretation of each index is dependent on the type being indexed
6121 into. The first index always indexes the pointer value given as the
6122 first argument, the second index indexes a value of the type pointed to
6123 (not necessarily the value directly pointed to, since the first index
6124 can be non-zero), etc. The first type indexed into must be a pointer
6125 value, subsequent types can be arrays, vectors, and structs. Note that
6126 subsequent types being indexed into can never be pointers, since that
6127 would require loading the pointer before continuing calculation.
6129 The type of each index argument depends on the type it is indexing into.
6130 When indexing into a (optionally packed) structure, only ``i32`` integer
6131 **constants** are allowed (when using a vector of indices they must all
6132 be the **same** ``i32`` integer constant). When indexing into an array,
6133 pointer or vector, integers of any width are allowed, and they are not
6134 required to be constant. These integers are treated as signed values
6137 For example, let's consider a C code fragment and how it gets compiled
6153 int *foo(struct ST *s) {
6154 return &s[1].Z.B[5][13];
6157 The LLVM code generated by Clang is:
6159 .. code-block:: llvm
6161 %struct.RT = type { i8, [10 x [20 x i32]], i8 }
6162 %struct.ST = type { i32, double, %struct.RT }
6164 define i32* @foo(%struct.ST* %s) nounwind uwtable readnone optsize ssp {
6166 %arrayidx = getelementptr inbounds %struct.ST, %struct.ST* %s, i64 1, i32 2, i32 1, i64 5, i64 13
6173 In the example above, the first index is indexing into the
6174 '``%struct.ST*``' type, which is a pointer, yielding a '``%struct.ST``'
6175 = '``{ i32, double, %struct.RT }``' type, a structure. The second index
6176 indexes into the third element of the structure, yielding a
6177 '``%struct.RT``' = '``{ i8 , [10 x [20 x i32]], i8 }``' type, another
6178 structure. The third index indexes into the second element of the
6179 structure, yielding a '``[10 x [20 x i32]]``' type, an array. The two
6180 dimensions of the array are subscripted into, yielding an '``i32``'
6181 type. The '``getelementptr``' instruction returns a pointer to this
6182 element, thus computing a value of '``i32*``' type.
6184 Note that it is perfectly legal to index partially through a structure,
6185 returning a pointer to an inner element. Because of this, the LLVM code
6186 for the given testcase is equivalent to:
6188 .. code-block:: llvm
6190 define i32* @foo(%struct.ST* %s) {
6191 %t1 = getelementptr %struct.ST, %struct.ST* %s, i32 1 ; yields %struct.ST*:%t1
6192 %t2 = getelementptr %struct.ST, %struct.ST* %t1, i32 0, i32 2 ; yields %struct.RT*:%t2
6193 %t3 = getelementptr %struct.RT, %struct.RT* %t2, i32 0, i32 1 ; yields [10 x [20 x i32]]*:%t3
6194 %t4 = getelementptr [10 x [20 x i32]], [10 x [20 x i32]]* %t3, i32 0, i32 5 ; yields [20 x i32]*:%t4
6195 %t5 = getelementptr [20 x i32], [20 x i32]* %t4, i32 0, i32 13 ; yields i32*:%t5
6199 If the ``inbounds`` keyword is present, the result value of the
6200 ``getelementptr`` is a :ref:`poison value <poisonvalues>` if the base
6201 pointer is not an *in bounds* address of an allocated object, or if any
6202 of the addresses that would be formed by successive addition of the
6203 offsets implied by the indices to the base address with infinitely
6204 precise signed arithmetic are not an *in bounds* address of that
6205 allocated object. The *in bounds* addresses for an allocated object are
6206 all the addresses that point into the object, plus the address one byte
6207 past the end. In cases where the base is a vector of pointers the
6208 ``inbounds`` keyword applies to each of the computations element-wise.
6210 If the ``inbounds`` keyword is not present, the offsets are added to the
6211 base address with silently-wrapping two's complement arithmetic. If the
6212 offsets have a different width from the pointer, they are sign-extended
6213 or truncated to the width of the pointer. The result value of the
6214 ``getelementptr`` may be outside the object pointed to by the base
6215 pointer. The result value may not necessarily be used to access memory
6216 though, even if it happens to point into allocated storage. See the
6217 :ref:`Pointer Aliasing Rules <pointeraliasing>` section for more
6220 The getelementptr instruction is often confusing. For some more insight
6221 into how it works, see :doc:`the getelementptr FAQ <GetElementPtr>`.
6226 .. code-block:: llvm
6228 ; yields [12 x i8]*:aptr
6229 %aptr = getelementptr {i32, [12 x i8]}, {i32, [12 x i8]}* %saptr, i64 0, i32 1
6231 %vptr = getelementptr {i32, <2 x i8>}, {i32, <2 x i8>}* %svptr, i64 0, i32 1, i32 1
6233 %eptr = getelementptr [12 x i8], [12 x i8]* %aptr, i64 0, i32 1
6235 %iptr = getelementptr [10 x i32], [10 x i32]* @arr, i16 0, i16 0
6237 In cases where the pointer argument is a vector of pointers, each index
6238 must be a vector with the same number of elements. For example:
6240 .. code-block:: llvm
6242 %A = getelementptr i8, <4 x i8*> %ptrs, <4 x i64> %offsets,
6244 Conversion Operations
6245 ---------------------
6247 The instructions in this category are the conversion instructions
6248 (casting) which all take a single operand and a type. They perform
6249 various bit conversions on the operand.
6251 '``trunc .. to``' Instruction
6252 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6259 <result> = trunc <ty> <value> to <ty2> ; yields ty2
6264 The '``trunc``' instruction truncates its operand to the type ``ty2``.
6269 The '``trunc``' instruction takes a value to trunc, and a type to trunc
6270 it to. Both types must be of :ref:`integer <t_integer>` types, or vectors
6271 of the same number of integers. The bit size of the ``value`` must be
6272 larger than the bit size of the destination type, ``ty2``. Equal sized
6273 types are not allowed.
6278 The '``trunc``' instruction truncates the high order bits in ``value``
6279 and converts the remaining bits to ``ty2``. Since the source size must
6280 be larger than the destination size, ``trunc`` cannot be a *no-op cast*.
6281 It will always truncate bits.
6286 .. code-block:: llvm
6288 %X = trunc i32 257 to i8 ; yields i8:1
6289 %Y = trunc i32 123 to i1 ; yields i1:true
6290 %Z = trunc i32 122 to i1 ; yields i1:false
6291 %W = trunc <2 x i16> <i16 8, i16 7> to <2 x i8> ; yields <i8 8, i8 7>
6293 '``zext .. to``' Instruction
6294 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6301 <result> = zext <ty> <value> to <ty2> ; yields ty2
6306 The '``zext``' instruction zero extends its operand to type ``ty2``.
6311 The '``zext``' instruction takes a value to cast, and a type to cast it
6312 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
6313 the same number of integers. The bit size of the ``value`` must be
6314 smaller than the bit size of the destination type, ``ty2``.
6319 The ``zext`` fills the high order bits of the ``value`` with zero bits
6320 until it reaches the size of the destination type, ``ty2``.
6322 When zero extending from i1, the result will always be either 0 or 1.
6327 .. code-block:: llvm
6329 %X = zext i32 257 to i64 ; yields i64:257
6330 %Y = zext i1 true to i32 ; yields i32:1
6331 %Z = zext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
6333 '``sext .. to``' Instruction
6334 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6341 <result> = sext <ty> <value> to <ty2> ; yields ty2
6346 The '``sext``' sign extends ``value`` to the type ``ty2``.
6351 The '``sext``' instruction takes a value to cast, and a type to cast it
6352 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
6353 the same number of integers. The bit size of the ``value`` must be
6354 smaller than the bit size of the destination type, ``ty2``.
6359 The '``sext``' instruction performs a sign extension by copying the sign
6360 bit (highest order bit) of the ``value`` until it reaches the bit size
6361 of the type ``ty2``.
6363 When sign extending from i1, the extension always results in -1 or 0.
6368 .. code-block:: llvm
6370 %X = sext i8 -1 to i16 ; yields i16 :65535
6371 %Y = sext i1 true to i32 ; yields i32:-1
6372 %Z = sext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
6374 '``fptrunc .. to``' Instruction
6375 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6382 <result> = fptrunc <ty> <value> to <ty2> ; yields ty2
6387 The '``fptrunc``' instruction truncates ``value`` to type ``ty2``.
6392 The '``fptrunc``' instruction takes a :ref:`floating point <t_floating>`
6393 value to cast and a :ref:`floating point <t_floating>` type to cast it to.
6394 The size of ``value`` must be larger than the size of ``ty2``. This
6395 implies that ``fptrunc`` cannot be used to make a *no-op cast*.
6400 The '``fptrunc``' instruction truncates a ``value`` from a larger
6401 :ref:`floating point <t_floating>` type to a smaller :ref:`floating
6402 point <t_floating>` type. If the value cannot fit within the
6403 destination type, ``ty2``, then the results are undefined.
6408 .. code-block:: llvm
6410 %X = fptrunc double 123.0 to float ; yields float:123.0
6411 %Y = fptrunc double 1.0E+300 to float ; yields undefined
6413 '``fpext .. to``' Instruction
6414 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6421 <result> = fpext <ty> <value> to <ty2> ; yields ty2
6426 The '``fpext``' extends a floating point ``value`` to a larger floating
6432 The '``fpext``' instruction takes a :ref:`floating point <t_floating>`
6433 ``value`` to cast, and a :ref:`floating point <t_floating>` type to cast it
6434 to. The source type must be smaller than the destination type.
6439 The '``fpext``' instruction extends the ``value`` from a smaller
6440 :ref:`floating point <t_floating>` type to a larger :ref:`floating
6441 point <t_floating>` type. The ``fpext`` cannot be used to make a
6442 *no-op cast* because it always changes bits. Use ``bitcast`` to make a
6443 *no-op cast* for a floating point cast.
6448 .. code-block:: llvm
6450 %X = fpext float 3.125 to double ; yields double:3.125000e+00
6451 %Y = fpext double %X to fp128 ; yields fp128:0xL00000000000000004000900000000000
6453 '``fptoui .. to``' Instruction
6454 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6461 <result> = fptoui <ty> <value> to <ty2> ; yields ty2
6466 The '``fptoui``' converts a floating point ``value`` to its unsigned
6467 integer equivalent of type ``ty2``.
6472 The '``fptoui``' instruction takes a value to cast, which must be a
6473 scalar or vector :ref:`floating point <t_floating>` value, and a type to
6474 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
6475 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
6476 type with the same number of elements as ``ty``
6481 The '``fptoui``' instruction converts its :ref:`floating
6482 point <t_floating>` operand into the nearest (rounding towards zero)
6483 unsigned integer value. If the value cannot fit in ``ty2``, the results
6489 .. code-block:: llvm
6491 %X = fptoui double 123.0 to i32 ; yields i32:123
6492 %Y = fptoui float 1.0E+300 to i1 ; yields undefined:1
6493 %Z = fptoui float 1.04E+17 to i8 ; yields undefined:1
6495 '``fptosi .. to``' Instruction
6496 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6503 <result> = fptosi <ty> <value> to <ty2> ; yields ty2
6508 The '``fptosi``' instruction converts :ref:`floating point <t_floating>`
6509 ``value`` to type ``ty2``.
6514 The '``fptosi``' instruction takes a value to cast, which must be a
6515 scalar or vector :ref:`floating point <t_floating>` value, and a type to
6516 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
6517 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
6518 type with the same number of elements as ``ty``
6523 The '``fptosi``' instruction converts its :ref:`floating
6524 point <t_floating>` operand into the nearest (rounding towards zero)
6525 signed integer value. If the value cannot fit in ``ty2``, the results
6531 .. code-block:: llvm
6533 %X = fptosi double -123.0 to i32 ; yields i32:-123
6534 %Y = fptosi float 1.0E-247 to i1 ; yields undefined:1
6535 %Z = fptosi float 1.04E+17 to i8 ; yields undefined:1
6537 '``uitofp .. to``' Instruction
6538 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6545 <result> = uitofp <ty> <value> to <ty2> ; yields ty2
6550 The '``uitofp``' instruction regards ``value`` as an unsigned integer
6551 and converts that value to the ``ty2`` type.
6556 The '``uitofp``' instruction takes a value to cast, which must be a
6557 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
6558 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
6559 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
6560 type with the same number of elements as ``ty``
6565 The '``uitofp``' instruction interprets its operand as an unsigned
6566 integer quantity and converts it to the corresponding floating point
6567 value. If the value cannot fit in the floating point value, the results
6573 .. code-block:: llvm
6575 %X = uitofp i32 257 to float ; yields float:257.0
6576 %Y = uitofp i8 -1 to double ; yields double:255.0
6578 '``sitofp .. to``' Instruction
6579 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6586 <result> = sitofp <ty> <value> to <ty2> ; yields ty2
6591 The '``sitofp``' instruction regards ``value`` as a signed integer and
6592 converts that value to the ``ty2`` type.
6597 The '``sitofp``' instruction takes a value to cast, which must be a
6598 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
6599 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
6600 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
6601 type with the same number of elements as ``ty``
6606 The '``sitofp``' instruction interprets its operand as a signed integer
6607 quantity and converts it to the corresponding floating point value. If
6608 the value cannot fit in the floating point value, the results are
6614 .. code-block:: llvm
6616 %X = sitofp i32 257 to float ; yields float:257.0
6617 %Y = sitofp i8 -1 to double ; yields double:-1.0
6621 '``ptrtoint .. to``' Instruction
6622 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6629 <result> = ptrtoint <ty> <value> to <ty2> ; yields ty2
6634 The '``ptrtoint``' instruction converts the pointer or a vector of
6635 pointers ``value`` to the integer (or vector of integers) type ``ty2``.
6640 The '``ptrtoint``' instruction takes a ``value`` to cast, which must be
6641 a value of type :ref:`pointer <t_pointer>` or a vector of pointers, and a
6642 type to cast it to ``ty2``, which must be an :ref:`integer <t_integer>` or
6643 a vector of integers type.
6648 The '``ptrtoint``' instruction converts ``value`` to integer type
6649 ``ty2`` by interpreting the pointer value as an integer and either
6650 truncating or zero extending that value to the size of the integer type.
6651 If ``value`` is smaller than ``ty2`` then a zero extension is done. If
6652 ``value`` is larger than ``ty2`` then a truncation is done. If they are
6653 the same size, then nothing is done (*no-op cast*) other than a type
6659 .. code-block:: llvm
6661 %X = ptrtoint i32* %P to i8 ; yields truncation on 32-bit architecture
6662 %Y = ptrtoint i32* %P to i64 ; yields zero extension on 32-bit architecture
6663 %Z = ptrtoint <4 x i32*> %P to <4 x i64>; yields vector zero extension for a vector of addresses on 32-bit architecture
6667 '``inttoptr .. to``' Instruction
6668 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6675 <result> = inttoptr <ty> <value> to <ty2> ; yields ty2
6680 The '``inttoptr``' instruction converts an integer ``value`` to a
6681 pointer type, ``ty2``.
6686 The '``inttoptr``' instruction takes an :ref:`integer <t_integer>` value to
6687 cast, and a type to cast it to, which must be a :ref:`pointer <t_pointer>`
6693 The '``inttoptr``' instruction converts ``value`` to type ``ty2`` by
6694 applying either a zero extension or a truncation depending on the size
6695 of the integer ``value``. If ``value`` is larger than the size of a
6696 pointer then a truncation is done. If ``value`` is smaller than the size
6697 of a pointer then a zero extension is done. If they are the same size,
6698 nothing is done (*no-op cast*).
6703 .. code-block:: llvm
6705 %X = inttoptr i32 255 to i32* ; yields zero extension on 64-bit architecture
6706 %Y = inttoptr i32 255 to i32* ; yields no-op on 32-bit architecture
6707 %Z = inttoptr i64 0 to i32* ; yields truncation on 32-bit architecture
6708 %Z = inttoptr <4 x i32> %G to <4 x i8*>; yields truncation of vector G to four pointers
6712 '``bitcast .. to``' Instruction
6713 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6720 <result> = bitcast <ty> <value> to <ty2> ; yields ty2
6725 The '``bitcast``' instruction converts ``value`` to type ``ty2`` without
6731 The '``bitcast``' instruction takes a value to cast, which must be a
6732 non-aggregate first class value, and a type to cast it to, which must
6733 also be a non-aggregate :ref:`first class <t_firstclass>` type. The
6734 bit sizes of ``value`` and the destination type, ``ty2``, must be
6735 identical. If the source type is a pointer, the destination type must
6736 also be a pointer of the same size. This instruction supports bitwise
6737 conversion of vectors to integers and to vectors of other types (as
6738 long as they have the same size).
6743 The '``bitcast``' instruction converts ``value`` to type ``ty2``. It
6744 is always a *no-op cast* because no bits change with this
6745 conversion. The conversion is done as if the ``value`` had been stored
6746 to memory and read back as type ``ty2``. Pointer (or vector of
6747 pointers) types may only be converted to other pointer (or vector of
6748 pointers) types with the same address space through this instruction.
6749 To convert pointers to other types, use the :ref:`inttoptr <i_inttoptr>`
6750 or :ref:`ptrtoint <i_ptrtoint>` instructions first.
6755 .. code-block:: llvm
6757 %X = bitcast i8 255 to i8 ; yields i8 :-1
6758 %Y = bitcast i32* %x to sint* ; yields sint*:%x
6759 %Z = bitcast <2 x int> %V to i64; ; yields i64: %V
6760 %Z = bitcast <2 x i32*> %V to <2 x i64*> ; yields <2 x i64*>
6762 .. _i_addrspacecast:
6764 '``addrspacecast .. to``' Instruction
6765 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6772 <result> = addrspacecast <pty> <ptrval> to <pty2> ; yields pty2
6777 The '``addrspacecast``' instruction converts ``ptrval`` from ``pty`` in
6778 address space ``n`` to type ``pty2`` in address space ``m``.
6783 The '``addrspacecast``' instruction takes a pointer or vector of pointer value
6784 to cast and a pointer type to cast it to, which must have a different
6790 The '``addrspacecast``' instruction converts the pointer value
6791 ``ptrval`` to type ``pty2``. It can be a *no-op cast* or a complex
6792 value modification, depending on the target and the address space
6793 pair. Pointer conversions within the same address space must be
6794 performed with the ``bitcast`` instruction. Note that if the address space
6795 conversion is legal then both result and operand refer to the same memory
6801 .. code-block:: llvm
6803 %X = addrspacecast i32* %x to i32 addrspace(1)* ; yields i32 addrspace(1)*:%x
6804 %Y = addrspacecast i32 addrspace(1)* %y to i64 addrspace(2)* ; yields i64 addrspace(2)*:%y
6805 %Z = addrspacecast <4 x i32*> %z to <4 x float addrspace(3)*> ; yields <4 x float addrspace(3)*>:%z
6812 The instructions in this category are the "miscellaneous" instructions,
6813 which defy better classification.
6817 '``icmp``' Instruction
6818 ^^^^^^^^^^^^^^^^^^^^^^
6825 <result> = icmp <cond> <ty> <op1>, <op2> ; yields i1 or <N x i1>:result
6830 The '``icmp``' instruction returns a boolean value or a vector of
6831 boolean values based on comparison of its two integer, integer vector,
6832 pointer, or pointer vector operands.
6837 The '``icmp``' instruction takes three operands. The first operand is
6838 the condition code indicating the kind of comparison to perform. It is
6839 not a value, just a keyword. The possible condition code are:
6842 #. ``ne``: not equal
6843 #. ``ugt``: unsigned greater than
6844 #. ``uge``: unsigned greater or equal
6845 #. ``ult``: unsigned less than
6846 #. ``ule``: unsigned less or equal
6847 #. ``sgt``: signed greater than
6848 #. ``sge``: signed greater or equal
6849 #. ``slt``: signed less than
6850 #. ``sle``: signed less or equal
6852 The remaining two arguments must be :ref:`integer <t_integer>` or
6853 :ref:`pointer <t_pointer>` or integer :ref:`vector <t_vector>` typed. They
6854 must also be identical types.
6859 The '``icmp``' compares ``op1`` and ``op2`` according to the condition
6860 code given as ``cond``. The comparison performed always yields either an
6861 :ref:`i1 <t_integer>` or vector of ``i1`` result, as follows:
6863 #. ``eq``: yields ``true`` if the operands are equal, ``false``
6864 otherwise. No sign interpretation is necessary or performed.
6865 #. ``ne``: yields ``true`` if the operands are unequal, ``false``
6866 otherwise. No sign interpretation is necessary or performed.
6867 #. ``ugt``: interprets the operands as unsigned values and yields
6868 ``true`` if ``op1`` is greater than ``op2``.
6869 #. ``uge``: interprets the operands as unsigned values and yields
6870 ``true`` if ``op1`` is greater than or equal to ``op2``.
6871 #. ``ult``: interprets the operands as unsigned values and yields
6872 ``true`` if ``op1`` is less than ``op2``.
6873 #. ``ule``: interprets the operands as unsigned values and yields
6874 ``true`` if ``op1`` is less than or equal to ``op2``.
6875 #. ``sgt``: interprets the operands as signed values and yields ``true``
6876 if ``op1`` is greater than ``op2``.
6877 #. ``sge``: interprets the operands as signed values and yields ``true``
6878 if ``op1`` is greater than or equal to ``op2``.
6879 #. ``slt``: interprets the operands as signed values and yields ``true``
6880 if ``op1`` is less than ``op2``.
6881 #. ``sle``: interprets the operands as signed values and yields ``true``
6882 if ``op1`` is less than or equal to ``op2``.
6884 If the operands are :ref:`pointer <t_pointer>` typed, the pointer values
6885 are compared as if they were integers.
6887 If the operands are integer vectors, then they are compared element by
6888 element. The result is an ``i1`` vector with the same number of elements
6889 as the values being compared. Otherwise, the result is an ``i1``.
6894 .. code-block:: llvm
6896 <result> = icmp eq i32 4, 5 ; yields: result=false
6897 <result> = icmp ne float* %X, %X ; yields: result=false
6898 <result> = icmp ult i16 4, 5 ; yields: result=true
6899 <result> = icmp sgt i16 4, 5 ; yields: result=false
6900 <result> = icmp ule i16 -4, 5 ; yields: result=false
6901 <result> = icmp sge i16 4, 5 ; yields: result=false
6903 Note that the code generator does not yet support vector types with the
6904 ``icmp`` instruction.
6908 '``fcmp``' Instruction
6909 ^^^^^^^^^^^^^^^^^^^^^^
6916 <result> = fcmp <cond> <ty> <op1>, <op2> ; yields i1 or <N x i1>:result
6921 The '``fcmp``' instruction returns a boolean value or vector of boolean
6922 values based on comparison of its operands.
6924 If the operands are floating point scalars, then the result type is a
6925 boolean (:ref:`i1 <t_integer>`).
6927 If the operands are floating point vectors, then the result type is a
6928 vector of boolean with the same number of elements as the operands being
6934 The '``fcmp``' instruction takes three operands. The first operand is
6935 the condition code indicating the kind of comparison to perform. It is
6936 not a value, just a keyword. The possible condition code are:
6938 #. ``false``: no comparison, always returns false
6939 #. ``oeq``: ordered and equal
6940 #. ``ogt``: ordered and greater than
6941 #. ``oge``: ordered and greater than or equal
6942 #. ``olt``: ordered and less than
6943 #. ``ole``: ordered and less than or equal
6944 #. ``one``: ordered and not equal
6945 #. ``ord``: ordered (no nans)
6946 #. ``ueq``: unordered or equal
6947 #. ``ugt``: unordered or greater than
6948 #. ``uge``: unordered or greater than or equal
6949 #. ``ult``: unordered or less than
6950 #. ``ule``: unordered or less than or equal
6951 #. ``une``: unordered or not equal
6952 #. ``uno``: unordered (either nans)
6953 #. ``true``: no comparison, always returns true
6955 *Ordered* means that neither operand is a QNAN while *unordered* means
6956 that either operand may be a QNAN.
6958 Each of ``val1`` and ``val2`` arguments must be either a :ref:`floating
6959 point <t_floating>` type or a :ref:`vector <t_vector>` of floating point
6960 type. They must have identical types.
6965 The '``fcmp``' instruction compares ``op1`` and ``op2`` according to the
6966 condition code given as ``cond``. If the operands are vectors, then the
6967 vectors are compared element by element. Each comparison performed
6968 always yields an :ref:`i1 <t_integer>` result, as follows:
6970 #. ``false``: always yields ``false``, regardless of operands.
6971 #. ``oeq``: yields ``true`` if both operands are not a QNAN and ``op1``
6972 is equal to ``op2``.
6973 #. ``ogt``: yields ``true`` if both operands are not a QNAN and ``op1``
6974 is greater than ``op2``.
6975 #. ``oge``: yields ``true`` if both operands are not a QNAN and ``op1``
6976 is greater than or equal to ``op2``.
6977 #. ``olt``: yields ``true`` if both operands are not a QNAN and ``op1``
6978 is less than ``op2``.
6979 #. ``ole``: yields ``true`` if both operands are not a QNAN and ``op1``
6980 is less than or equal to ``op2``.
6981 #. ``one``: yields ``true`` if both operands are not a QNAN and ``op1``
6982 is not equal to ``op2``.
6983 #. ``ord``: yields ``true`` if both operands are not a QNAN.
6984 #. ``ueq``: yields ``true`` if either operand is a QNAN or ``op1`` is
6986 #. ``ugt``: yields ``true`` if either operand is a QNAN or ``op1`` is
6987 greater than ``op2``.
6988 #. ``uge``: yields ``true`` if either operand is a QNAN or ``op1`` is
6989 greater than or equal to ``op2``.
6990 #. ``ult``: yields ``true`` if either operand is a QNAN or ``op1`` is
6992 #. ``ule``: yields ``true`` if either operand is a QNAN or ``op1`` is
6993 less than or equal to ``op2``.
6994 #. ``une``: yields ``true`` if either operand is a QNAN or ``op1`` is
6995 not equal to ``op2``.
6996 #. ``uno``: yields ``true`` if either operand is a QNAN.
6997 #. ``true``: always yields ``true``, regardless of operands.
7002 .. code-block:: llvm
7004 <result> = fcmp oeq float 4.0, 5.0 ; yields: result=false
7005 <result> = fcmp one float 4.0, 5.0 ; yields: result=true
7006 <result> = fcmp olt float 4.0, 5.0 ; yields: result=true
7007 <result> = fcmp ueq double 1.0, 2.0 ; yields: result=false
7009 Note that the code generator does not yet support vector types with the
7010 ``fcmp`` instruction.
7014 '``phi``' Instruction
7015 ^^^^^^^^^^^^^^^^^^^^^
7022 <result> = phi <ty> [ <val0>, <label0>], ...
7027 The '``phi``' instruction is used to implement the φ node in the SSA
7028 graph representing the function.
7033 The type of the incoming values is specified with the first type field.
7034 After this, the '``phi``' instruction takes a list of pairs as
7035 arguments, with one pair for each predecessor basic block of the current
7036 block. Only values of :ref:`first class <t_firstclass>` type may be used as
7037 the value arguments to the PHI node. Only labels may be used as the
7040 There must be no non-phi instructions between the start of a basic block
7041 and the PHI instructions: i.e. PHI instructions must be first in a basic
7044 For the purposes of the SSA form, the use of each incoming value is
7045 deemed to occur on the edge from the corresponding predecessor block to
7046 the current block (but after any definition of an '``invoke``'
7047 instruction's return value on the same edge).
7052 At runtime, the '``phi``' instruction logically takes on the value
7053 specified by the pair corresponding to the predecessor basic block that
7054 executed just prior to the current block.
7059 .. code-block:: llvm
7061 Loop: ; Infinite loop that counts from 0 on up...
7062 %indvar = phi i32 [ 0, %LoopHeader ], [ %nextindvar, %Loop ]
7063 %nextindvar = add i32 %indvar, 1
7068 '``select``' Instruction
7069 ^^^^^^^^^^^^^^^^^^^^^^^^
7076 <result> = select selty <cond>, <ty> <val1>, <ty> <val2> ; yields ty
7078 selty is either i1 or {<N x i1>}
7083 The '``select``' instruction is used to choose one value based on a
7084 condition, without IR-level branching.
7089 The '``select``' instruction requires an 'i1' value or a vector of 'i1'
7090 values indicating the condition, and two values of the same :ref:`first
7091 class <t_firstclass>` type.
7096 If the condition is an i1 and it evaluates to 1, the instruction returns
7097 the first value argument; otherwise, it returns the second value
7100 If the condition is a vector of i1, then the value arguments must be
7101 vectors of the same size, and the selection is done element by element.
7103 If the condition is an i1 and the value arguments are vectors of the
7104 same size, then an entire vector is selected.
7109 .. code-block:: llvm
7111 %X = select i1 true, i8 17, i8 42 ; yields i8:17
7115 '``call``' Instruction
7116 ^^^^^^^^^^^^^^^^^^^^^^
7123 <result> = [tail | musttail] call [cconv] [ret attrs] <ty> [<fnty>*] <fnptrval>(<function args>) [fn attrs]
7128 The '``call``' instruction represents a simple function call.
7133 This instruction requires several arguments:
7135 #. The optional ``tail`` and ``musttail`` markers indicate that the optimizers
7136 should perform tail call optimization. The ``tail`` marker is a hint that
7137 `can be ignored <CodeGenerator.html#sibcallopt>`_. The ``musttail`` marker
7138 means that the call must be tail call optimized in order for the program to
7139 be correct. The ``musttail`` marker provides these guarantees:
7141 #. The call will not cause unbounded stack growth if it is part of a
7142 recursive cycle in the call graph.
7143 #. Arguments with the :ref:`inalloca <attr_inalloca>` attribute are
7146 Both markers imply that the callee does not access allocas or varargs from
7147 the caller. Calls marked ``musttail`` must obey the following additional
7150 - The call must immediately precede a :ref:`ret <i_ret>` instruction,
7151 or a pointer bitcast followed by a ret instruction.
7152 - The ret instruction must return the (possibly bitcasted) value
7153 produced by the call or void.
7154 - The caller and callee prototypes must match. Pointer types of
7155 parameters or return types may differ in pointee type, but not
7157 - The calling conventions of the caller and callee must match.
7158 - All ABI-impacting function attributes, such as sret, byval, inreg,
7159 returned, and inalloca, must match.
7160 - The callee must be varargs iff the caller is varargs. Bitcasting a
7161 non-varargs function to the appropriate varargs type is legal so
7162 long as the non-varargs prefixes obey the other rules.
7164 Tail call optimization for calls marked ``tail`` is guaranteed to occur if
7165 the following conditions are met:
7167 - Caller and callee both have the calling convention ``fastcc``.
7168 - The call is in tail position (ret immediately follows call and ret
7169 uses value of call or is void).
7170 - Option ``-tailcallopt`` is enabled, or
7171 ``llvm::GuaranteedTailCallOpt`` is ``true``.
7172 - `Platform-specific constraints are
7173 met. <CodeGenerator.html#tailcallopt>`_
7175 #. The optional "cconv" marker indicates which :ref:`calling
7176 convention <callingconv>` the call should use. If none is
7177 specified, the call defaults to using C calling conventions. The
7178 calling convention of the call must match the calling convention of
7179 the target function, or else the behavior is undefined.
7180 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
7181 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
7183 #. '``ty``': the type of the call instruction itself which is also the
7184 type of the return value. Functions that return no value are marked
7186 #. '``fnty``': shall be the signature of the pointer to function value
7187 being invoked. The argument types must match the types implied by
7188 this signature. This type can be omitted if the function is not
7189 varargs and if the function type does not return a pointer to a
7191 #. '``fnptrval``': An LLVM value containing a pointer to a function to
7192 be invoked. In most cases, this is a direct function invocation, but
7193 indirect ``call``'s are just as possible, calling an arbitrary pointer
7195 #. '``function args``': argument list whose types match the function
7196 signature argument types and parameter attributes. All arguments must
7197 be of :ref:`first class <t_firstclass>` type. If the function signature
7198 indicates the function accepts a variable number of arguments, the
7199 extra arguments can be specified.
7200 #. The optional :ref:`function attributes <fnattrs>` list. Only
7201 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
7202 attributes are valid here.
7207 The '``call``' instruction is used to cause control flow to transfer to
7208 a specified function, with its incoming arguments bound to the specified
7209 values. Upon a '``ret``' instruction in the called function, control
7210 flow continues with the instruction after the function call, and the
7211 return value of the function is bound to the result argument.
7216 .. code-block:: llvm
7218 %retval = call i32 @test(i32 %argc)
7219 call i32 (i8*, ...)* @printf(i8* %msg, i32 12, i8 42) ; yields i32
7220 %X = tail call i32 @foo() ; yields i32
7221 %Y = tail call fastcc i32 @foo() ; yields i32
7222 call void %foo(i8 97 signext)
7224 %struct.A = type { i32, i8 }
7225 %r = call %struct.A @foo() ; yields { i32, i8 }
7226 %gr = extractvalue %struct.A %r, 0 ; yields i32
7227 %gr1 = extractvalue %struct.A %r, 1 ; yields i8
7228 %Z = call void @foo() noreturn ; indicates that %foo never returns normally
7229 %ZZ = call zeroext i32 @bar() ; Return value is %zero extended
7231 llvm treats calls to some functions with names and arguments that match
7232 the standard C99 library as being the C99 library functions, and may
7233 perform optimizations or generate code for them under that assumption.
7234 This is something we'd like to change in the future to provide better
7235 support for freestanding environments and non-C-based languages.
7239 '``va_arg``' Instruction
7240 ^^^^^^^^^^^^^^^^^^^^^^^^
7247 <resultval> = va_arg <va_list*> <arglist>, <argty>
7252 The '``va_arg``' instruction is used to access arguments passed through
7253 the "variable argument" area of a function call. It is used to implement
7254 the ``va_arg`` macro in C.
7259 This instruction takes a ``va_list*`` value and the type of the
7260 argument. It returns a value of the specified argument type and
7261 increments the ``va_list`` to point to the next argument. The actual
7262 type of ``va_list`` is target specific.
7267 The '``va_arg``' instruction loads an argument of the specified type
7268 from the specified ``va_list`` and causes the ``va_list`` to point to
7269 the next argument. For more information, see the variable argument
7270 handling :ref:`Intrinsic Functions <int_varargs>`.
7272 It is legal for this instruction to be called in a function which does
7273 not take a variable number of arguments, for example, the ``vfprintf``
7276 ``va_arg`` is an LLVM instruction instead of an :ref:`intrinsic
7277 function <intrinsics>` because it takes a type as an argument.
7282 See the :ref:`variable argument processing <int_varargs>` section.
7284 Note that the code generator does not yet fully support va\_arg on many
7285 targets. Also, it does not currently support va\_arg with aggregate
7286 types on any target.
7290 '``landingpad``' Instruction
7291 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7298 <resultval> = landingpad <resultty> <clause>+
7299 <resultval> = landingpad <resultty> cleanup <clause>*
7301 <clause> := catch <type> <value>
7302 <clause> := filter <array constant type> <array constant>
7307 The '``landingpad``' instruction is used by `LLVM's exception handling
7308 system <ExceptionHandling.html#overview>`_ to specify that a basic block
7309 is a landing pad --- one where the exception lands, and corresponds to the
7310 code found in the ``catch`` portion of a ``try``/``catch`` sequence. It
7311 defines values supplied by the :ref:`personality function <personalityfn>` upon
7312 re-entry to the function. The ``resultval`` has the type ``resultty``.
7318 ``cleanup`` flag indicates that the landing pad block is a cleanup.
7320 A ``clause`` begins with the clause type --- ``catch`` or ``filter`` --- and
7321 contains the global variable representing the "type" that may be caught
7322 or filtered respectively. Unlike the ``catch`` clause, the ``filter``
7323 clause takes an array constant as its argument. Use
7324 "``[0 x i8**] undef``" for a filter which cannot throw. The
7325 '``landingpad``' instruction must contain *at least* one ``clause`` or
7326 the ``cleanup`` flag.
7331 The '``landingpad``' instruction defines the values which are set by the
7332 :ref:`personality function <personalityfn>` upon re-entry to the function, and
7333 therefore the "result type" of the ``landingpad`` instruction. As with
7334 calling conventions, how the personality function results are
7335 represented in LLVM IR is target specific.
7337 The clauses are applied in order from top to bottom. If two
7338 ``landingpad`` instructions are merged together through inlining, the
7339 clauses from the calling function are appended to the list of clauses.
7340 When the call stack is being unwound due to an exception being thrown,
7341 the exception is compared against each ``clause`` in turn. If it doesn't
7342 match any of the clauses, and the ``cleanup`` flag is not set, then
7343 unwinding continues further up the call stack.
7345 The ``landingpad`` instruction has several restrictions:
7347 - A landing pad block is a basic block which is the unwind destination
7348 of an '``invoke``' instruction.
7349 - A landing pad block must have a '``landingpad``' instruction as its
7350 first non-PHI instruction.
7351 - There can be only one '``landingpad``' instruction within the landing
7353 - A basic block that is not a landing pad block may not include a
7354 '``landingpad``' instruction.
7359 .. code-block:: llvm
7361 ;; A landing pad which can catch an integer.
7362 %res = landingpad { i8*, i32 }
7364 ;; A landing pad that is a cleanup.
7365 %res = landingpad { i8*, i32 }
7367 ;; A landing pad which can catch an integer and can only throw a double.
7368 %res = landingpad { i8*, i32 }
7370 filter [1 x i8**] [@_ZTId]
7377 LLVM supports the notion of an "intrinsic function". These functions
7378 have well known names and semantics and are required to follow certain
7379 restrictions. Overall, these intrinsics represent an extension mechanism
7380 for the LLVM language that does not require changing all of the
7381 transformations in LLVM when adding to the language (or the bitcode
7382 reader/writer, the parser, etc...).
7384 Intrinsic function names must all start with an "``llvm.``" prefix. This
7385 prefix is reserved in LLVM for intrinsic names; thus, function names may
7386 not begin with this prefix. Intrinsic functions must always be external
7387 functions: you cannot define the body of intrinsic functions. Intrinsic
7388 functions may only be used in call or invoke instructions: it is illegal
7389 to take the address of an intrinsic function. Additionally, because
7390 intrinsic functions are part of the LLVM language, it is required if any
7391 are added that they be documented here.
7393 Some intrinsic functions can be overloaded, i.e., the intrinsic
7394 represents a family of functions that perform the same operation but on
7395 different data types. Because LLVM can represent over 8 million
7396 different integer types, overloading is used commonly to allow an
7397 intrinsic function to operate on any integer type. One or more of the
7398 argument types or the result type can be overloaded to accept any
7399 integer type. Argument types may also be defined as exactly matching a
7400 previous argument's type or the result type. This allows an intrinsic
7401 function which accepts multiple arguments, but needs all of them to be
7402 of the same type, to only be overloaded with respect to a single
7403 argument or the result.
7405 Overloaded intrinsics will have the names of its overloaded argument
7406 types encoded into its function name, each preceded by a period. Only
7407 those types which are overloaded result in a name suffix. Arguments
7408 whose type is matched against another type do not. For example, the
7409 ``llvm.ctpop`` function can take an integer of any width and returns an
7410 integer of exactly the same integer width. This leads to a family of
7411 functions such as ``i8 @llvm.ctpop.i8(i8 %val)`` and
7412 ``i29 @llvm.ctpop.i29(i29 %val)``. Only one type, the return type, is
7413 overloaded, and only one type suffix is required. Because the argument's
7414 type is matched against the return type, it does not require its own
7417 To learn how to add an intrinsic function, please see the `Extending
7418 LLVM Guide <ExtendingLLVM.html>`_.
7422 Variable Argument Handling Intrinsics
7423 -------------------------------------
7425 Variable argument support is defined in LLVM with the
7426 :ref:`va_arg <i_va_arg>` instruction and these three intrinsic
7427 functions. These functions are related to the similarly named macros
7428 defined in the ``<stdarg.h>`` header file.
7430 All of these functions operate on arguments that use a target-specific
7431 value type "``va_list``". The LLVM assembly language reference manual
7432 does not define what this type is, so all transformations should be
7433 prepared to handle these functions regardless of the type used.
7435 This example shows how the :ref:`va_arg <i_va_arg>` instruction and the
7436 variable argument handling intrinsic functions are used.
7438 .. code-block:: llvm
7440 ; This struct is different for every platform. For most platforms,
7441 ; it is merely an i8*.
7442 %struct.va_list = type { i8* }
7444 ; For Unix x86_64 platforms, va_list is the following struct:
7445 ; %struct.va_list = type { i32, i32, i8*, i8* }
7447 define i32 @test(i32 %X, ...) {
7448 ; Initialize variable argument processing
7449 %ap = alloca %struct.va_list
7450 %ap2 = bitcast %struct.va_list* %ap to i8*
7451 call void @llvm.va_start(i8* %ap2)
7453 ; Read a single integer argument
7454 %tmp = va_arg i8* %ap2, i32
7456 ; Demonstrate usage of llvm.va_copy and llvm.va_end
7458 %aq2 = bitcast i8** %aq to i8*
7459 call void @llvm.va_copy(i8* %aq2, i8* %ap2)
7460 call void @llvm.va_end(i8* %aq2)
7462 ; Stop processing of arguments.
7463 call void @llvm.va_end(i8* %ap2)
7467 declare void @llvm.va_start(i8*)
7468 declare void @llvm.va_copy(i8*, i8*)
7469 declare void @llvm.va_end(i8*)
7473 '``llvm.va_start``' Intrinsic
7474 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7481 declare void @llvm.va_start(i8* <arglist>)
7486 The '``llvm.va_start``' intrinsic initializes ``*<arglist>`` for
7487 subsequent use by ``va_arg``.
7492 The argument is a pointer to a ``va_list`` element to initialize.
7497 The '``llvm.va_start``' intrinsic works just like the ``va_start`` macro
7498 available in C. In a target-dependent way, it initializes the
7499 ``va_list`` element to which the argument points, so that the next call
7500 to ``va_arg`` will produce the first variable argument passed to the
7501 function. Unlike the C ``va_start`` macro, this intrinsic does not need
7502 to know the last argument of the function as the compiler can figure
7505 '``llvm.va_end``' Intrinsic
7506 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7513 declare void @llvm.va_end(i8* <arglist>)
7518 The '``llvm.va_end``' intrinsic destroys ``*<arglist>``, which has been
7519 initialized previously with ``llvm.va_start`` or ``llvm.va_copy``.
7524 The argument is a pointer to a ``va_list`` to destroy.
7529 The '``llvm.va_end``' intrinsic works just like the ``va_end`` macro
7530 available in C. In a target-dependent way, it destroys the ``va_list``
7531 element to which the argument points. Calls to
7532 :ref:`llvm.va_start <int_va_start>` and
7533 :ref:`llvm.va_copy <int_va_copy>` must be matched exactly with calls to
7538 '``llvm.va_copy``' Intrinsic
7539 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7546 declare void @llvm.va_copy(i8* <destarglist>, i8* <srcarglist>)
7551 The '``llvm.va_copy``' intrinsic copies the current argument position
7552 from the source argument list to the destination argument list.
7557 The first argument is a pointer to a ``va_list`` element to initialize.
7558 The second argument is a pointer to a ``va_list`` element to copy from.
7563 The '``llvm.va_copy``' intrinsic works just like the ``va_copy`` macro
7564 available in C. In a target-dependent way, it copies the source
7565 ``va_list`` element into the destination ``va_list`` element. This
7566 intrinsic is necessary because the `` llvm.va_start`` intrinsic may be
7567 arbitrarily complex and require, for example, memory allocation.
7569 Accurate Garbage Collection Intrinsics
7570 --------------------------------------
7572 LLVM's support for `Accurate Garbage Collection <GarbageCollection.html>`_
7573 (GC) requires the frontend to generate code containing appropriate intrinsic
7574 calls and select an appropriate GC strategy which knows how to lower these
7575 intrinsics in a manner which is appropriate for the target collector.
7577 These intrinsics allow identification of :ref:`GC roots on the
7578 stack <int_gcroot>`, as well as garbage collector implementations that
7579 require :ref:`read <int_gcread>` and :ref:`write <int_gcwrite>` barriers.
7580 Frontends for type-safe garbage collected languages should generate
7581 these intrinsics to make use of the LLVM garbage collectors. For more
7582 details, see `Garbage Collection with LLVM <GarbageCollection.html>`_.
7584 Experimental Statepoint Intrinsics
7585 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7587 LLVM provides an second experimental set of intrinsics for describing garbage
7588 collection safepoints in compiled code. These intrinsics are an alternative
7589 to the ``llvm.gcroot`` intrinsics, but are compatible with the ones for
7590 :ref:`read <int_gcread>` and :ref:`write <int_gcwrite>` barriers. The
7591 differences in approach are covered in the `Garbage Collection with LLVM
7592 <GarbageCollection.html>`_ documentation. The intrinsics themselves are
7593 described in :doc:`Statepoints`.
7597 '``llvm.gcroot``' Intrinsic
7598 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7605 declare void @llvm.gcroot(i8** %ptrloc, i8* %metadata)
7610 The '``llvm.gcroot``' intrinsic declares the existence of a GC root to
7611 the code generator, and allows some metadata to be associated with it.
7616 The first argument specifies the address of a stack object that contains
7617 the root pointer. The second pointer (which must be either a constant or
7618 a global value address) contains the meta-data to be associated with the
7624 At runtime, a call to this intrinsic stores a null pointer into the
7625 "ptrloc" location. At compile-time, the code generator generates
7626 information to allow the runtime to find the pointer at GC safe points.
7627 The '``llvm.gcroot``' intrinsic may only be used in a function which
7628 :ref:`specifies a GC algorithm <gc>`.
7632 '``llvm.gcread``' Intrinsic
7633 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7640 declare i8* @llvm.gcread(i8* %ObjPtr, i8** %Ptr)
7645 The '``llvm.gcread``' intrinsic identifies reads of references from heap
7646 locations, allowing garbage collector implementations that require read
7652 The second argument is the address to read from, which should be an
7653 address allocated from the garbage collector. The first object is a
7654 pointer to the start of the referenced object, if needed by the language
7655 runtime (otherwise null).
7660 The '``llvm.gcread``' intrinsic has the same semantics as a load
7661 instruction, but may be replaced with substantially more complex code by
7662 the garbage collector runtime, as needed. The '``llvm.gcread``'
7663 intrinsic may only be used in a function which :ref:`specifies a GC
7668 '``llvm.gcwrite``' Intrinsic
7669 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7676 declare void @llvm.gcwrite(i8* %P1, i8* %Obj, i8** %P2)
7681 The '``llvm.gcwrite``' intrinsic identifies writes of references to heap
7682 locations, allowing garbage collector implementations that require write
7683 barriers (such as generational or reference counting collectors).
7688 The first argument is the reference to store, the second is the start of
7689 the object to store it to, and the third is the address of the field of
7690 Obj to store to. If the runtime does not require a pointer to the
7691 object, Obj may be null.
7696 The '``llvm.gcwrite``' intrinsic has the same semantics as a store
7697 instruction, but may be replaced with substantially more complex code by
7698 the garbage collector runtime, as needed. The '``llvm.gcwrite``'
7699 intrinsic may only be used in a function which :ref:`specifies a GC
7702 Code Generator Intrinsics
7703 -------------------------
7705 These intrinsics are provided by LLVM to expose special features that
7706 may only be implemented with code generator support.
7708 '``llvm.returnaddress``' Intrinsic
7709 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7716 declare i8 *@llvm.returnaddress(i32 <level>)
7721 The '``llvm.returnaddress``' intrinsic attempts to compute a
7722 target-specific value indicating the return address of the current
7723 function or one of its callers.
7728 The argument to this intrinsic indicates which function to return the
7729 address for. Zero indicates the calling function, one indicates its
7730 caller, etc. The argument is **required** to be a constant integer
7736 The '``llvm.returnaddress``' intrinsic either returns a pointer
7737 indicating the return address of the specified call frame, or zero if it
7738 cannot be identified. The value returned by this intrinsic is likely to
7739 be incorrect or 0 for arguments other than zero, so it should only be
7740 used for debugging purposes.
7742 Note that calling this intrinsic does not prevent function inlining or
7743 other aggressive transformations, so the value returned may not be that
7744 of the obvious source-language caller.
7746 '``llvm.frameaddress``' Intrinsic
7747 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7754 declare i8* @llvm.frameaddress(i32 <level>)
7759 The '``llvm.frameaddress``' intrinsic attempts to return the
7760 target-specific frame pointer value for the specified stack frame.
7765 The argument to this intrinsic indicates which function to return the
7766 frame pointer for. Zero indicates the calling function, one indicates
7767 its caller, etc. The argument is **required** to be a constant integer
7773 The '``llvm.frameaddress``' intrinsic either returns a pointer
7774 indicating the frame address of the specified call frame, or zero if it
7775 cannot be identified. The value returned by this intrinsic is likely to
7776 be incorrect or 0 for arguments other than zero, so it should only be
7777 used for debugging purposes.
7779 Note that calling this intrinsic does not prevent function inlining or
7780 other aggressive transformations, so the value returned may not be that
7781 of the obvious source-language caller.
7783 '``llvm.frameescape``' and '``llvm.framerecover``' Intrinsics
7784 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7791 declare void @llvm.frameescape(...)
7792 declare i8* @llvm.framerecover(i8* %func, i8* %fp, i32 %idx)
7797 The '``llvm.frameescape``' intrinsic escapes offsets of a collection of static
7798 allocas, and the '``llvm.framerecover``' intrinsic applies those offsets to a
7799 live frame pointer to recover the address of the allocation. The offset is
7800 computed during frame layout of the caller of ``llvm.frameescape``.
7805 All arguments to '``llvm.frameescape``' must be pointers to static allocas or
7806 casts of static allocas. Each function can only call '``llvm.frameescape``'
7807 once, and it can only do so from the entry block.
7809 The ``func`` argument to '``llvm.framerecover``' must be a constant
7810 bitcasted pointer to a function defined in the current module. The code
7811 generator cannot determine the frame allocation offset of functions defined in
7814 The ``fp`` argument to '``llvm.framerecover``' must be a frame
7815 pointer of a call frame that is currently live. The return value of
7816 '``llvm.frameaddress``' is one way to produce such a value, but most platforms
7817 also expose the frame pointer through stack unwinding mechanisms.
7819 The ``idx`` argument to '``llvm.framerecover``' indicates which alloca passed to
7820 '``llvm.frameescape``' to recover. It is zero-indexed.
7825 These intrinsics allow a group of functions to access one stack memory
7826 allocation in an ancestor stack frame. The memory returned from
7827 '``llvm.frameallocate``' may be allocated prior to stack realignment, so the
7828 memory is only aligned to the ABI-required stack alignment. Each function may
7829 only call '``llvm.frameallocate``' one or zero times from the function entry
7830 block. The frame allocation intrinsic inhibits inlining, as any frame
7831 allocations in the inlined function frame are likely to be at a different
7832 offset from the one used by '``llvm.framerecover``' called with the
7835 .. _int_read_register:
7836 .. _int_write_register:
7838 '``llvm.read_register``' and '``llvm.write_register``' Intrinsics
7839 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7846 declare i32 @llvm.read_register.i32(metadata)
7847 declare i64 @llvm.read_register.i64(metadata)
7848 declare void @llvm.write_register.i32(metadata, i32 @value)
7849 declare void @llvm.write_register.i64(metadata, i64 @value)
7855 The '``llvm.read_register``' and '``llvm.write_register``' intrinsics
7856 provides access to the named register. The register must be valid on
7857 the architecture being compiled to. The type needs to be compatible
7858 with the register being read.
7863 The '``llvm.read_register``' intrinsic returns the current value of the
7864 register, where possible. The '``llvm.write_register``' intrinsic sets
7865 the current value of the register, where possible.
7867 This is useful to implement named register global variables that need
7868 to always be mapped to a specific register, as is common practice on
7869 bare-metal programs including OS kernels.
7871 The compiler doesn't check for register availability or use of the used
7872 register in surrounding code, including inline assembly. Because of that,
7873 allocatable registers are not supported.
7875 Warning: So far it only works with the stack pointer on selected
7876 architectures (ARM, AArch64, PowerPC and x86_64). Significant amount of
7877 work is needed to support other registers and even more so, allocatable
7882 '``llvm.stacksave``' Intrinsic
7883 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7890 declare i8* @llvm.stacksave()
7895 The '``llvm.stacksave``' intrinsic is used to remember the current state
7896 of the function stack, for use with
7897 :ref:`llvm.stackrestore <int_stackrestore>`. This is useful for
7898 implementing language features like scoped automatic variable sized
7904 This intrinsic returns a opaque pointer value that can be passed to
7905 :ref:`llvm.stackrestore <int_stackrestore>`. When an
7906 ``llvm.stackrestore`` intrinsic is executed with a value saved from
7907 ``llvm.stacksave``, it effectively restores the state of the stack to
7908 the state it was in when the ``llvm.stacksave`` intrinsic executed. In
7909 practice, this pops any :ref:`alloca <i_alloca>` blocks from the stack that
7910 were allocated after the ``llvm.stacksave`` was executed.
7912 .. _int_stackrestore:
7914 '``llvm.stackrestore``' Intrinsic
7915 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7922 declare void @llvm.stackrestore(i8* %ptr)
7927 The '``llvm.stackrestore``' intrinsic is used to restore the state of
7928 the function stack to the state it was in when the corresponding
7929 :ref:`llvm.stacksave <int_stacksave>` intrinsic executed. This is
7930 useful for implementing language features like scoped automatic variable
7931 sized arrays in C99.
7936 See the description for :ref:`llvm.stacksave <int_stacksave>`.
7938 '``llvm.prefetch``' Intrinsic
7939 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7946 declare void @llvm.prefetch(i8* <address>, i32 <rw>, i32 <locality>, i32 <cache type>)
7951 The '``llvm.prefetch``' intrinsic is a hint to the code generator to
7952 insert a prefetch instruction if supported; otherwise, it is a noop.
7953 Prefetches have no effect on the behavior of the program but can change
7954 its performance characteristics.
7959 ``address`` is the address to be prefetched, ``rw`` is the specifier
7960 determining if the fetch should be for a read (0) or write (1), and
7961 ``locality`` is a temporal locality specifier ranging from (0) - no
7962 locality, to (3) - extremely local keep in cache. The ``cache type``
7963 specifies whether the prefetch is performed on the data (1) or
7964 instruction (0) cache. The ``rw``, ``locality`` and ``cache type``
7965 arguments must be constant integers.
7970 This intrinsic does not modify the behavior of the program. In
7971 particular, prefetches cannot trap and do not produce a value. On
7972 targets that support this intrinsic, the prefetch can provide hints to
7973 the processor cache for better performance.
7975 '``llvm.pcmarker``' Intrinsic
7976 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7983 declare void @llvm.pcmarker(i32 <id>)
7988 The '``llvm.pcmarker``' intrinsic is a method to export a Program
7989 Counter (PC) in a region of code to simulators and other tools. The
7990 method is target specific, but it is expected that the marker will use
7991 exported symbols to transmit the PC of the marker. The marker makes no
7992 guarantees that it will remain with any specific instruction after
7993 optimizations. It is possible that the presence of a marker will inhibit
7994 optimizations. The intended use is to be inserted after optimizations to
7995 allow correlations of simulation runs.
8000 ``id`` is a numerical id identifying the marker.
8005 This intrinsic does not modify the behavior of the program. Backends
8006 that do not support this intrinsic may ignore it.
8008 '``llvm.readcyclecounter``' Intrinsic
8009 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8016 declare i64 @llvm.readcyclecounter()
8021 The '``llvm.readcyclecounter``' intrinsic provides access to the cycle
8022 counter register (or similar low latency, high accuracy clocks) on those
8023 targets that support it. On X86, it should map to RDTSC. On Alpha, it
8024 should map to RPCC. As the backing counters overflow quickly (on the
8025 order of 9 seconds on alpha), this should only be used for small
8031 When directly supported, reading the cycle counter should not modify any
8032 memory. Implementations are allowed to either return a application
8033 specific value or a system wide value. On backends without support, this
8034 is lowered to a constant 0.
8036 Note that runtime support may be conditional on the privilege-level code is
8037 running at and the host platform.
8039 '``llvm.clear_cache``' Intrinsic
8040 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8047 declare void @llvm.clear_cache(i8*, i8*)
8052 The '``llvm.clear_cache``' intrinsic ensures visibility of modifications
8053 in the specified range to the execution unit of the processor. On
8054 targets with non-unified instruction and data cache, the implementation
8055 flushes the instruction cache.
8060 On platforms with coherent instruction and data caches (e.g. x86), this
8061 intrinsic is a nop. On platforms with non-coherent instruction and data
8062 cache (e.g. ARM, MIPS), the intrinsic is lowered either to appropriate
8063 instructions or a system call, if cache flushing requires special
8066 The default behavior is to emit a call to ``__clear_cache`` from the run
8069 This instrinsic does *not* empty the instruction pipeline. Modifications
8070 of the current function are outside the scope of the intrinsic.
8072 '``llvm.instrprof_increment``' Intrinsic
8073 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8080 declare void @llvm.instrprof_increment(i8* <name>, i64 <hash>,
8081 i32 <num-counters>, i32 <index>)
8086 The '``llvm.instrprof_increment``' intrinsic can be emitted by a
8087 frontend for use with instrumentation based profiling. These will be
8088 lowered by the ``-instrprof`` pass to generate execution counts of a
8094 The first argument is a pointer to a global variable containing the
8095 name of the entity being instrumented. This should generally be the
8096 (mangled) function name for a set of counters.
8098 The second argument is a hash value that can be used by the consumer
8099 of the profile data to detect changes to the instrumented source, and
8100 the third is the number of counters associated with ``name``. It is an
8101 error if ``hash`` or ``num-counters`` differ between two instances of
8102 ``instrprof_increment`` that refer to the same name.
8104 The last argument refers to which of the counters for ``name`` should
8105 be incremented. It should be a value between 0 and ``num-counters``.
8110 This intrinsic represents an increment of a profiling counter. It will
8111 cause the ``-instrprof`` pass to generate the appropriate data
8112 structures and the code to increment the appropriate value, in a
8113 format that can be written out by a compiler runtime and consumed via
8114 the ``llvm-profdata`` tool.
8116 Standard C Library Intrinsics
8117 -----------------------------
8119 LLVM provides intrinsics for a few important standard C library
8120 functions. These intrinsics allow source-language front-ends to pass
8121 information about the alignment of the pointer arguments to the code
8122 generator, providing opportunity for more efficient code generation.
8126 '``llvm.memcpy``' Intrinsic
8127 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8132 This is an overloaded intrinsic. You can use ``llvm.memcpy`` on any
8133 integer bit width and for different address spaces. Not all targets
8134 support all bit widths however.
8138 declare void @llvm.memcpy.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
8139 i32 <len>, i32 <align>, i1 <isvolatile>)
8140 declare void @llvm.memcpy.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
8141 i64 <len>, i32 <align>, i1 <isvolatile>)
8146 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
8147 source location to the destination location.
8149 Note that, unlike the standard libc function, the ``llvm.memcpy.*``
8150 intrinsics do not return a value, takes extra alignment/isvolatile
8151 arguments and the pointers can be in specified address spaces.
8156 The first argument is a pointer to the destination, the second is a
8157 pointer to the source. The third argument is an integer argument
8158 specifying the number of bytes to copy, the fourth argument is the
8159 alignment of the source and destination locations, and the fifth is a
8160 boolean indicating a volatile access.
8162 If the call to this intrinsic has an alignment value that is not 0 or 1,
8163 then the caller guarantees that both the source and destination pointers
8164 are aligned to that boundary.
8166 If the ``isvolatile`` parameter is ``true``, the ``llvm.memcpy`` call is
8167 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
8168 very cleanly specified and it is unwise to depend on it.
8173 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
8174 source location to the destination location, which are not allowed to
8175 overlap. It copies "len" bytes of memory over. If the argument is known
8176 to be aligned to some boundary, this can be specified as the fourth
8177 argument, otherwise it should be set to 0 or 1 (both meaning no alignment).
8179 '``llvm.memmove``' Intrinsic
8180 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8185 This is an overloaded intrinsic. You can use llvm.memmove on any integer
8186 bit width and for different address space. Not all targets support all
8191 declare void @llvm.memmove.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
8192 i32 <len>, i32 <align>, i1 <isvolatile>)
8193 declare void @llvm.memmove.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
8194 i64 <len>, i32 <align>, i1 <isvolatile>)
8199 The '``llvm.memmove.*``' intrinsics move a block of memory from the
8200 source location to the destination location. It is similar to the
8201 '``llvm.memcpy``' intrinsic but allows the two memory locations to
8204 Note that, unlike the standard libc function, the ``llvm.memmove.*``
8205 intrinsics do not return a value, takes extra alignment/isvolatile
8206 arguments and the pointers can be in specified address spaces.
8211 The first argument is a pointer to the destination, the second is a
8212 pointer to the source. The third argument is an integer argument
8213 specifying the number of bytes to copy, the fourth argument is the
8214 alignment of the source and destination locations, and the fifth is a
8215 boolean indicating a volatile access.
8217 If the call to this intrinsic has an alignment value that is not 0 or 1,
8218 then the caller guarantees that the source and destination pointers are
8219 aligned to that boundary.
8221 If the ``isvolatile`` parameter is ``true``, the ``llvm.memmove`` call
8222 is a :ref:`volatile operation <volatile>`. The detailed access behavior is
8223 not very cleanly specified and it is unwise to depend on it.
8228 The '``llvm.memmove.*``' intrinsics copy a block of memory from the
8229 source location to the destination location, which may overlap. It
8230 copies "len" bytes of memory over. If the argument is known to be
8231 aligned to some boundary, this can be specified as the fourth argument,
8232 otherwise it should be set to 0 or 1 (both meaning no alignment).
8234 '``llvm.memset.*``' Intrinsics
8235 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8240 This is an overloaded intrinsic. You can use llvm.memset on any integer
8241 bit width and for different address spaces. However, not all targets
8242 support all bit widths.
8246 declare void @llvm.memset.p0i8.i32(i8* <dest>, i8 <val>,
8247 i32 <len>, i32 <align>, i1 <isvolatile>)
8248 declare void @llvm.memset.p0i8.i64(i8* <dest>, i8 <val>,
8249 i64 <len>, i32 <align>, i1 <isvolatile>)
8254 The '``llvm.memset.*``' intrinsics fill a block of memory with a
8255 particular byte value.
8257 Note that, unlike the standard libc function, the ``llvm.memset``
8258 intrinsic does not return a value and takes extra alignment/volatile
8259 arguments. Also, the destination can be in an arbitrary address space.
8264 The first argument is a pointer to the destination to fill, the second
8265 is the byte value with which to fill it, the third argument is an
8266 integer argument specifying the number of bytes to fill, and the fourth
8267 argument is the known alignment of the destination location.
8269 If the call to this intrinsic has an alignment value that is not 0 or 1,
8270 then the caller guarantees that the destination pointer is aligned to
8273 If the ``isvolatile`` parameter is ``true``, the ``llvm.memset`` call is
8274 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
8275 very cleanly specified and it is unwise to depend on it.
8280 The '``llvm.memset.*``' intrinsics fill "len" bytes of memory starting
8281 at the destination location. If the argument is known to be aligned to
8282 some boundary, this can be specified as the fourth argument, otherwise
8283 it should be set to 0 or 1 (both meaning no alignment).
8285 '``llvm.sqrt.*``' Intrinsic
8286 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8291 This is an overloaded intrinsic. You can use ``llvm.sqrt`` on any
8292 floating point or vector of floating point type. Not all targets support
8297 declare float @llvm.sqrt.f32(float %Val)
8298 declare double @llvm.sqrt.f64(double %Val)
8299 declare x86_fp80 @llvm.sqrt.f80(x86_fp80 %Val)
8300 declare fp128 @llvm.sqrt.f128(fp128 %Val)
8301 declare ppc_fp128 @llvm.sqrt.ppcf128(ppc_fp128 %Val)
8306 The '``llvm.sqrt``' intrinsics return the sqrt of the specified operand,
8307 returning the same value as the libm '``sqrt``' functions would. Unlike
8308 ``sqrt`` in libm, however, ``llvm.sqrt`` has undefined behavior for
8309 negative numbers other than -0.0 (which allows for better optimization,
8310 because there is no need to worry about errno being set).
8311 ``llvm.sqrt(-0.0)`` is defined to return -0.0 like IEEE sqrt.
8316 The argument and return value are floating point numbers of the same
8322 This function returns the sqrt of the specified operand if it is a
8323 nonnegative floating point number.
8325 '``llvm.powi.*``' Intrinsic
8326 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8331 This is an overloaded intrinsic. You can use ``llvm.powi`` on any
8332 floating point or vector of floating point type. Not all targets support
8337 declare float @llvm.powi.f32(float %Val, i32 %power)
8338 declare double @llvm.powi.f64(double %Val, i32 %power)
8339 declare x86_fp80 @llvm.powi.f80(x86_fp80 %Val, i32 %power)
8340 declare fp128 @llvm.powi.f128(fp128 %Val, i32 %power)
8341 declare ppc_fp128 @llvm.powi.ppcf128(ppc_fp128 %Val, i32 %power)
8346 The '``llvm.powi.*``' intrinsics return the first operand raised to the
8347 specified (positive or negative) power. The order of evaluation of
8348 multiplications is not defined. When a vector of floating point type is
8349 used, the second argument remains a scalar integer value.
8354 The second argument is an integer power, and the first is a value to
8355 raise to that power.
8360 This function returns the first value raised to the second power with an
8361 unspecified sequence of rounding operations.
8363 '``llvm.sin.*``' Intrinsic
8364 ^^^^^^^^^^^^^^^^^^^^^^^^^^
8369 This is an overloaded intrinsic. You can use ``llvm.sin`` on any
8370 floating point or vector of floating point type. Not all targets support
8375 declare float @llvm.sin.f32(float %Val)
8376 declare double @llvm.sin.f64(double %Val)
8377 declare x86_fp80 @llvm.sin.f80(x86_fp80 %Val)
8378 declare fp128 @llvm.sin.f128(fp128 %Val)
8379 declare ppc_fp128 @llvm.sin.ppcf128(ppc_fp128 %Val)
8384 The '``llvm.sin.*``' intrinsics return the sine of the operand.
8389 The argument and return value are floating point numbers of the same
8395 This function returns the sine of the specified operand, returning the
8396 same values as the libm ``sin`` functions would, and handles error
8397 conditions in the same way.
8399 '``llvm.cos.*``' Intrinsic
8400 ^^^^^^^^^^^^^^^^^^^^^^^^^^
8405 This is an overloaded intrinsic. You can use ``llvm.cos`` on any
8406 floating point or vector of floating point type. Not all targets support
8411 declare float @llvm.cos.f32(float %Val)
8412 declare double @llvm.cos.f64(double %Val)
8413 declare x86_fp80 @llvm.cos.f80(x86_fp80 %Val)
8414 declare fp128 @llvm.cos.f128(fp128 %Val)
8415 declare ppc_fp128 @llvm.cos.ppcf128(ppc_fp128 %Val)
8420 The '``llvm.cos.*``' intrinsics return the cosine of the operand.
8425 The argument and return value are floating point numbers of the same
8431 This function returns the cosine of the specified operand, returning the
8432 same values as the libm ``cos`` functions would, and handles error
8433 conditions in the same way.
8435 '``llvm.pow.*``' Intrinsic
8436 ^^^^^^^^^^^^^^^^^^^^^^^^^^
8441 This is an overloaded intrinsic. You can use ``llvm.pow`` on any
8442 floating point or vector of floating point type. Not all targets support
8447 declare float @llvm.pow.f32(float %Val, float %Power)
8448 declare double @llvm.pow.f64(double %Val, double %Power)
8449 declare x86_fp80 @llvm.pow.f80(x86_fp80 %Val, x86_fp80 %Power)
8450 declare fp128 @llvm.pow.f128(fp128 %Val, fp128 %Power)
8451 declare ppc_fp128 @llvm.pow.ppcf128(ppc_fp128 %Val, ppc_fp128 Power)
8456 The '``llvm.pow.*``' intrinsics return the first operand raised to the
8457 specified (positive or negative) power.
8462 The second argument is a floating point power, and the first is a value
8463 to raise to that power.
8468 This function returns the first value raised to the second power,
8469 returning the same values as the libm ``pow`` functions would, and
8470 handles error conditions in the same way.
8472 '``llvm.exp.*``' Intrinsic
8473 ^^^^^^^^^^^^^^^^^^^^^^^^^^
8478 This is an overloaded intrinsic. You can use ``llvm.exp`` on any
8479 floating point or vector of floating point type. Not all targets support
8484 declare float @llvm.exp.f32(float %Val)
8485 declare double @llvm.exp.f64(double %Val)
8486 declare x86_fp80 @llvm.exp.f80(x86_fp80 %Val)
8487 declare fp128 @llvm.exp.f128(fp128 %Val)
8488 declare ppc_fp128 @llvm.exp.ppcf128(ppc_fp128 %Val)
8493 The '``llvm.exp.*``' intrinsics perform the exp function.
8498 The argument and return value are floating point numbers of the same
8504 This function returns the same values as the libm ``exp`` functions
8505 would, and handles error conditions in the same way.
8507 '``llvm.exp2.*``' Intrinsic
8508 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8513 This is an overloaded intrinsic. You can use ``llvm.exp2`` on any
8514 floating point or vector of floating point type. Not all targets support
8519 declare float @llvm.exp2.f32(float %Val)
8520 declare double @llvm.exp2.f64(double %Val)
8521 declare x86_fp80 @llvm.exp2.f80(x86_fp80 %Val)
8522 declare fp128 @llvm.exp2.f128(fp128 %Val)
8523 declare ppc_fp128 @llvm.exp2.ppcf128(ppc_fp128 %Val)
8528 The '``llvm.exp2.*``' intrinsics perform the exp2 function.
8533 The argument and return value are floating point numbers of the same
8539 This function returns the same values as the libm ``exp2`` functions
8540 would, and handles error conditions in the same way.
8542 '``llvm.log.*``' Intrinsic
8543 ^^^^^^^^^^^^^^^^^^^^^^^^^^
8548 This is an overloaded intrinsic. You can use ``llvm.log`` on any
8549 floating point or vector of floating point type. Not all targets support
8554 declare float @llvm.log.f32(float %Val)
8555 declare double @llvm.log.f64(double %Val)
8556 declare x86_fp80 @llvm.log.f80(x86_fp80 %Val)
8557 declare fp128 @llvm.log.f128(fp128 %Val)
8558 declare ppc_fp128 @llvm.log.ppcf128(ppc_fp128 %Val)
8563 The '``llvm.log.*``' intrinsics perform the log function.
8568 The argument and return value are floating point numbers of the same
8574 This function returns the same values as the libm ``log`` functions
8575 would, and handles error conditions in the same way.
8577 '``llvm.log10.*``' Intrinsic
8578 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8583 This is an overloaded intrinsic. You can use ``llvm.log10`` on any
8584 floating point or vector of floating point type. Not all targets support
8589 declare float @llvm.log10.f32(float %Val)
8590 declare double @llvm.log10.f64(double %Val)
8591 declare x86_fp80 @llvm.log10.f80(x86_fp80 %Val)
8592 declare fp128 @llvm.log10.f128(fp128 %Val)
8593 declare ppc_fp128 @llvm.log10.ppcf128(ppc_fp128 %Val)
8598 The '``llvm.log10.*``' intrinsics perform the log10 function.
8603 The argument and return value are floating point numbers of the same
8609 This function returns the same values as the libm ``log10`` functions
8610 would, and handles error conditions in the same way.
8612 '``llvm.log2.*``' Intrinsic
8613 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8618 This is an overloaded intrinsic. You can use ``llvm.log2`` on any
8619 floating point or vector of floating point type. Not all targets support
8624 declare float @llvm.log2.f32(float %Val)
8625 declare double @llvm.log2.f64(double %Val)
8626 declare x86_fp80 @llvm.log2.f80(x86_fp80 %Val)
8627 declare fp128 @llvm.log2.f128(fp128 %Val)
8628 declare ppc_fp128 @llvm.log2.ppcf128(ppc_fp128 %Val)
8633 The '``llvm.log2.*``' intrinsics perform the log2 function.
8638 The argument and return value are floating point numbers of the same
8644 This function returns the same values as the libm ``log2`` functions
8645 would, and handles error conditions in the same way.
8647 '``llvm.fma.*``' Intrinsic
8648 ^^^^^^^^^^^^^^^^^^^^^^^^^^
8653 This is an overloaded intrinsic. You can use ``llvm.fma`` on any
8654 floating point or vector of floating point type. Not all targets support
8659 declare float @llvm.fma.f32(float %a, float %b, float %c)
8660 declare double @llvm.fma.f64(double %a, double %b, double %c)
8661 declare x86_fp80 @llvm.fma.f80(x86_fp80 %a, x86_fp80 %b, x86_fp80 %c)
8662 declare fp128 @llvm.fma.f128(fp128 %a, fp128 %b, fp128 %c)
8663 declare ppc_fp128 @llvm.fma.ppcf128(ppc_fp128 %a, ppc_fp128 %b, ppc_fp128 %c)
8668 The '``llvm.fma.*``' intrinsics perform the fused multiply-add
8674 The argument and return value are floating point numbers of the same
8680 This function returns the same values as the libm ``fma`` functions
8681 would, and does not set errno.
8683 '``llvm.fabs.*``' Intrinsic
8684 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8689 This is an overloaded intrinsic. You can use ``llvm.fabs`` on any
8690 floating point or vector of floating point type. Not all targets support
8695 declare float @llvm.fabs.f32(float %Val)
8696 declare double @llvm.fabs.f64(double %Val)
8697 declare x86_fp80 @llvm.fabs.f80(x86_fp80 %Val)
8698 declare fp128 @llvm.fabs.f128(fp128 %Val)
8699 declare ppc_fp128 @llvm.fabs.ppcf128(ppc_fp128 %Val)
8704 The '``llvm.fabs.*``' intrinsics return the absolute value of the
8710 The argument and return value are floating point numbers of the same
8716 This function returns the same values as the libm ``fabs`` functions
8717 would, and handles error conditions in the same way.
8719 '``llvm.minnum.*``' Intrinsic
8720 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8725 This is an overloaded intrinsic. You can use ``llvm.minnum`` on any
8726 floating point or vector of floating point type. Not all targets support
8731 declare float @llvm.minnum.f32(float %Val0, float %Val1)
8732 declare double @llvm.minnum.f64(double %Val0, double %Val1)
8733 declare x86_fp80 @llvm.minnum.f80(x86_fp80 %Val0, x86_fp80 %Val1)
8734 declare fp128 @llvm.minnum.f128(fp128 %Val0, fp128 %Val1)
8735 declare ppc_fp128 @llvm.minnum.ppcf128(ppc_fp128 %Val0, ppc_fp128 %Val1)
8740 The '``llvm.minnum.*``' intrinsics return the minimum of the two
8747 The arguments and return value are floating point numbers of the same
8753 Follows the IEEE-754 semantics for minNum, which also match for libm's
8756 If either operand is a NaN, returns the other non-NaN operand. Returns
8757 NaN only if both operands are NaN. If the operands compare equal,
8758 returns a value that compares equal to both operands. This means that
8759 fmin(+/-0.0, +/-0.0) could return either -0.0 or 0.0.
8761 '``llvm.maxnum.*``' Intrinsic
8762 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8767 This is an overloaded intrinsic. You can use ``llvm.maxnum`` on any
8768 floating point or vector of floating point type. Not all targets support
8773 declare float @llvm.maxnum.f32(float %Val0, float %Val1l)
8774 declare double @llvm.maxnum.f64(double %Val0, double %Val1)
8775 declare x86_fp80 @llvm.maxnum.f80(x86_fp80 %Val0, x86_fp80 %Val1)
8776 declare fp128 @llvm.maxnum.f128(fp128 %Val0, fp128 %Val1)
8777 declare ppc_fp128 @llvm.maxnum.ppcf128(ppc_fp128 %Val0, ppc_fp128 %Val1)
8782 The '``llvm.maxnum.*``' intrinsics return the maximum of the two
8789 The arguments and return value are floating point numbers of the same
8794 Follows the IEEE-754 semantics for maxNum, which also match for libm's
8797 If either operand is a NaN, returns the other non-NaN operand. Returns
8798 NaN only if both operands are NaN. If the operands compare equal,
8799 returns a value that compares equal to both operands. This means that
8800 fmax(+/-0.0, +/-0.0) could return either -0.0 or 0.0.
8802 '``llvm.copysign.*``' Intrinsic
8803 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8808 This is an overloaded intrinsic. You can use ``llvm.copysign`` on any
8809 floating point or vector of floating point type. Not all targets support
8814 declare float @llvm.copysign.f32(float %Mag, float %Sgn)
8815 declare double @llvm.copysign.f64(double %Mag, double %Sgn)
8816 declare x86_fp80 @llvm.copysign.f80(x86_fp80 %Mag, x86_fp80 %Sgn)
8817 declare fp128 @llvm.copysign.f128(fp128 %Mag, fp128 %Sgn)
8818 declare ppc_fp128 @llvm.copysign.ppcf128(ppc_fp128 %Mag, ppc_fp128 %Sgn)
8823 The '``llvm.copysign.*``' intrinsics return a value with the magnitude of the
8824 first operand and the sign of the second operand.
8829 The arguments and return value are floating point numbers of the same
8835 This function returns the same values as the libm ``copysign``
8836 functions would, and handles error conditions in the same way.
8838 '``llvm.floor.*``' Intrinsic
8839 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8844 This is an overloaded intrinsic. You can use ``llvm.floor`` on any
8845 floating point or vector of floating point type. Not all targets support
8850 declare float @llvm.floor.f32(float %Val)
8851 declare double @llvm.floor.f64(double %Val)
8852 declare x86_fp80 @llvm.floor.f80(x86_fp80 %Val)
8853 declare fp128 @llvm.floor.f128(fp128 %Val)
8854 declare ppc_fp128 @llvm.floor.ppcf128(ppc_fp128 %Val)
8859 The '``llvm.floor.*``' intrinsics return the floor of the operand.
8864 The argument and return value are floating point numbers of the same
8870 This function returns the same values as the libm ``floor`` functions
8871 would, and handles error conditions in the same way.
8873 '``llvm.ceil.*``' Intrinsic
8874 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8879 This is an overloaded intrinsic. You can use ``llvm.ceil`` on any
8880 floating point or vector of floating point type. Not all targets support
8885 declare float @llvm.ceil.f32(float %Val)
8886 declare double @llvm.ceil.f64(double %Val)
8887 declare x86_fp80 @llvm.ceil.f80(x86_fp80 %Val)
8888 declare fp128 @llvm.ceil.f128(fp128 %Val)
8889 declare ppc_fp128 @llvm.ceil.ppcf128(ppc_fp128 %Val)
8894 The '``llvm.ceil.*``' intrinsics return the ceiling of the operand.
8899 The argument and return value are floating point numbers of the same
8905 This function returns the same values as the libm ``ceil`` functions
8906 would, and handles error conditions in the same way.
8908 '``llvm.trunc.*``' Intrinsic
8909 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8914 This is an overloaded intrinsic. You can use ``llvm.trunc`` on any
8915 floating point or vector of floating point type. Not all targets support
8920 declare float @llvm.trunc.f32(float %Val)
8921 declare double @llvm.trunc.f64(double %Val)
8922 declare x86_fp80 @llvm.trunc.f80(x86_fp80 %Val)
8923 declare fp128 @llvm.trunc.f128(fp128 %Val)
8924 declare ppc_fp128 @llvm.trunc.ppcf128(ppc_fp128 %Val)
8929 The '``llvm.trunc.*``' intrinsics returns the operand rounded to the
8930 nearest integer not larger in magnitude than the operand.
8935 The argument and return value are floating point numbers of the same
8941 This function returns the same values as the libm ``trunc`` functions
8942 would, and handles error conditions in the same way.
8944 '``llvm.rint.*``' Intrinsic
8945 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8950 This is an overloaded intrinsic. You can use ``llvm.rint`` on any
8951 floating point or vector of floating point type. Not all targets support
8956 declare float @llvm.rint.f32(float %Val)
8957 declare double @llvm.rint.f64(double %Val)
8958 declare x86_fp80 @llvm.rint.f80(x86_fp80 %Val)
8959 declare fp128 @llvm.rint.f128(fp128 %Val)
8960 declare ppc_fp128 @llvm.rint.ppcf128(ppc_fp128 %Val)
8965 The '``llvm.rint.*``' intrinsics returns the operand rounded to the
8966 nearest integer. It may raise an inexact floating-point exception if the
8967 operand isn't an integer.
8972 The argument and return value are floating point numbers of the same
8978 This function returns the same values as the libm ``rint`` functions
8979 would, and handles error conditions in the same way.
8981 '``llvm.nearbyint.*``' Intrinsic
8982 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8987 This is an overloaded intrinsic. You can use ``llvm.nearbyint`` on any
8988 floating point or vector of floating point type. Not all targets support
8993 declare float @llvm.nearbyint.f32(float %Val)
8994 declare double @llvm.nearbyint.f64(double %Val)
8995 declare x86_fp80 @llvm.nearbyint.f80(x86_fp80 %Val)
8996 declare fp128 @llvm.nearbyint.f128(fp128 %Val)
8997 declare ppc_fp128 @llvm.nearbyint.ppcf128(ppc_fp128 %Val)
9002 The '``llvm.nearbyint.*``' intrinsics returns the operand rounded to the
9008 The argument and return value are floating point numbers of the same
9014 This function returns the same values as the libm ``nearbyint``
9015 functions would, and handles error conditions in the same way.
9017 '``llvm.round.*``' Intrinsic
9018 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9023 This is an overloaded intrinsic. You can use ``llvm.round`` on any
9024 floating point or vector of floating point type. Not all targets support
9029 declare float @llvm.round.f32(float %Val)
9030 declare double @llvm.round.f64(double %Val)
9031 declare x86_fp80 @llvm.round.f80(x86_fp80 %Val)
9032 declare fp128 @llvm.round.f128(fp128 %Val)
9033 declare ppc_fp128 @llvm.round.ppcf128(ppc_fp128 %Val)
9038 The '``llvm.round.*``' intrinsics returns the operand rounded to the
9044 The argument and return value are floating point numbers of the same
9050 This function returns the same values as the libm ``round``
9051 functions would, and handles error conditions in the same way.
9053 Bit Manipulation Intrinsics
9054 ---------------------------
9056 LLVM provides intrinsics for a few important bit manipulation
9057 operations. These allow efficient code generation for some algorithms.
9059 '``llvm.bswap.*``' Intrinsics
9060 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9065 This is an overloaded intrinsic function. You can use bswap on any
9066 integer type that is an even number of bytes (i.e. BitWidth % 16 == 0).
9070 declare i16 @llvm.bswap.i16(i16 <id>)
9071 declare i32 @llvm.bswap.i32(i32 <id>)
9072 declare i64 @llvm.bswap.i64(i64 <id>)
9077 The '``llvm.bswap``' family of intrinsics is used to byte swap integer
9078 values with an even number of bytes (positive multiple of 16 bits).
9079 These are useful for performing operations on data that is not in the
9080 target's native byte order.
9085 The ``llvm.bswap.i16`` intrinsic returns an i16 value that has the high
9086 and low byte of the input i16 swapped. Similarly, the ``llvm.bswap.i32``
9087 intrinsic returns an i32 value that has the four bytes of the input i32
9088 swapped, so that if the input bytes are numbered 0, 1, 2, 3 then the
9089 returned i32 will have its bytes in 3, 2, 1, 0 order. The
9090 ``llvm.bswap.i48``, ``llvm.bswap.i64`` and other intrinsics extend this
9091 concept to additional even-byte lengths (6 bytes, 8 bytes and more,
9094 '``llvm.ctpop.*``' Intrinsic
9095 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9100 This is an overloaded intrinsic. You can use llvm.ctpop on any integer
9101 bit width, or on any vector with integer elements. Not all targets
9102 support all bit widths or vector types, however.
9106 declare i8 @llvm.ctpop.i8(i8 <src>)
9107 declare i16 @llvm.ctpop.i16(i16 <src>)
9108 declare i32 @llvm.ctpop.i32(i32 <src>)
9109 declare i64 @llvm.ctpop.i64(i64 <src>)
9110 declare i256 @llvm.ctpop.i256(i256 <src>)
9111 declare <2 x i32> @llvm.ctpop.v2i32(<2 x i32> <src>)
9116 The '``llvm.ctpop``' family of intrinsics counts the number of bits set
9122 The only argument is the value to be counted. The argument may be of any
9123 integer type, or a vector with integer elements. The return type must
9124 match the argument type.
9129 The '``llvm.ctpop``' intrinsic counts the 1's in a variable, or within
9130 each element of a vector.
9132 '``llvm.ctlz.*``' Intrinsic
9133 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9138 This is an overloaded intrinsic. You can use ``llvm.ctlz`` on any
9139 integer bit width, or any vector whose elements are integers. Not all
9140 targets support all bit widths or vector types, however.
9144 declare i8 @llvm.ctlz.i8 (i8 <src>, i1 <is_zero_undef>)
9145 declare i16 @llvm.ctlz.i16 (i16 <src>, i1 <is_zero_undef>)
9146 declare i32 @llvm.ctlz.i32 (i32 <src>, i1 <is_zero_undef>)
9147 declare i64 @llvm.ctlz.i64 (i64 <src>, i1 <is_zero_undef>)
9148 declare i256 @llvm.ctlz.i256(i256 <src>, i1 <is_zero_undef>)
9149 declase <2 x i32> @llvm.ctlz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
9154 The '``llvm.ctlz``' family of intrinsic functions counts the number of
9155 leading zeros in a variable.
9160 The first argument is the value to be counted. This argument may be of
9161 any integer type, or a vector with integer element type. The return
9162 type must match the first argument type.
9164 The second argument must be a constant and is a flag to indicate whether
9165 the intrinsic should ensure that a zero as the first argument produces a
9166 defined result. Historically some architectures did not provide a
9167 defined result for zero values as efficiently, and many algorithms are
9168 now predicated on avoiding zero-value inputs.
9173 The '``llvm.ctlz``' intrinsic counts the leading (most significant)
9174 zeros in a variable, or within each element of the vector. If
9175 ``src == 0`` then the result is the size in bits of the type of ``src``
9176 if ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
9177 ``llvm.ctlz(i32 2) = 30``.
9179 '``llvm.cttz.*``' Intrinsic
9180 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9185 This is an overloaded intrinsic. You can use ``llvm.cttz`` on any
9186 integer bit width, or any vector of integer elements. Not all targets
9187 support all bit widths or vector types, however.
9191 declare i8 @llvm.cttz.i8 (i8 <src>, i1 <is_zero_undef>)
9192 declare i16 @llvm.cttz.i16 (i16 <src>, i1 <is_zero_undef>)
9193 declare i32 @llvm.cttz.i32 (i32 <src>, i1 <is_zero_undef>)
9194 declare i64 @llvm.cttz.i64 (i64 <src>, i1 <is_zero_undef>)
9195 declare i256 @llvm.cttz.i256(i256 <src>, i1 <is_zero_undef>)
9196 declase <2 x i32> @llvm.cttz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
9201 The '``llvm.cttz``' family of intrinsic functions counts the number of
9207 The first argument is the value to be counted. This argument may be of
9208 any integer type, or a vector with integer element type. The return
9209 type must match the first argument type.
9211 The second argument must be a constant and is a flag to indicate whether
9212 the intrinsic should ensure that a zero as the first argument produces a
9213 defined result. Historically some architectures did not provide a
9214 defined result for zero values as efficiently, and many algorithms are
9215 now predicated on avoiding zero-value inputs.
9220 The '``llvm.cttz``' intrinsic counts the trailing (least significant)
9221 zeros in a variable, or within each element of a vector. If ``src == 0``
9222 then the result is the size in bits of the type of ``src`` if
9223 ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
9224 ``llvm.cttz(2) = 1``.
9228 Arithmetic with Overflow Intrinsics
9229 -----------------------------------
9231 LLVM provides intrinsics for some arithmetic with overflow operations.
9233 '``llvm.sadd.with.overflow.*``' Intrinsics
9234 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9239 This is an overloaded intrinsic. You can use ``llvm.sadd.with.overflow``
9240 on any integer bit width.
9244 declare {i16, i1} @llvm.sadd.with.overflow.i16(i16 %a, i16 %b)
9245 declare {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
9246 declare {i64, i1} @llvm.sadd.with.overflow.i64(i64 %a, i64 %b)
9251 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
9252 a signed addition of the two arguments, and indicate whether an overflow
9253 occurred during the signed summation.
9258 The arguments (%a and %b) and the first element of the result structure
9259 may be of integer types of any bit width, but they must have the same
9260 bit width. The second element of the result structure must be of type
9261 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
9267 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
9268 a signed addition of the two variables. They return a structure --- the
9269 first element of which is the signed summation, and the second element
9270 of which is a bit specifying if the signed summation resulted in an
9276 .. code-block:: llvm
9278 %res = call {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
9279 %sum = extractvalue {i32, i1} %res, 0
9280 %obit = extractvalue {i32, i1} %res, 1
9281 br i1 %obit, label %overflow, label %normal
9283 '``llvm.uadd.with.overflow.*``' Intrinsics
9284 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9289 This is an overloaded intrinsic. You can use ``llvm.uadd.with.overflow``
9290 on any integer bit width.
9294 declare {i16, i1} @llvm.uadd.with.overflow.i16(i16 %a, i16 %b)
9295 declare {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
9296 declare {i64, i1} @llvm.uadd.with.overflow.i64(i64 %a, i64 %b)
9301 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
9302 an unsigned addition of the two arguments, and indicate whether a carry
9303 occurred during the unsigned summation.
9308 The arguments (%a and %b) and the first element of the result structure
9309 may be of integer types of any bit width, but they must have the same
9310 bit width. The second element of the result structure must be of type
9311 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
9317 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
9318 an unsigned addition of the two arguments. They return a structure --- the
9319 first element of which is the sum, and the second element of which is a
9320 bit specifying if the unsigned summation resulted in a carry.
9325 .. code-block:: llvm
9327 %res = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
9328 %sum = extractvalue {i32, i1} %res, 0
9329 %obit = extractvalue {i32, i1} %res, 1
9330 br i1 %obit, label %carry, label %normal
9332 '``llvm.ssub.with.overflow.*``' Intrinsics
9333 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9338 This is an overloaded intrinsic. You can use ``llvm.ssub.with.overflow``
9339 on any integer bit width.
9343 declare {i16, i1} @llvm.ssub.with.overflow.i16(i16 %a, i16 %b)
9344 declare {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
9345 declare {i64, i1} @llvm.ssub.with.overflow.i64(i64 %a, i64 %b)
9350 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
9351 a signed subtraction of the two arguments, and indicate whether an
9352 overflow occurred during the signed subtraction.
9357 The arguments (%a and %b) and the first element of the result structure
9358 may be of integer types of any bit width, but they must have the same
9359 bit width. The second element of the result structure must be of type
9360 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
9366 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
9367 a signed subtraction of the two arguments. They return a structure --- the
9368 first element of which is the subtraction, and the second element of
9369 which is a bit specifying if the signed subtraction resulted in an
9375 .. code-block:: llvm
9377 %res = call {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
9378 %sum = extractvalue {i32, i1} %res, 0
9379 %obit = extractvalue {i32, i1} %res, 1
9380 br i1 %obit, label %overflow, label %normal
9382 '``llvm.usub.with.overflow.*``' Intrinsics
9383 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9388 This is an overloaded intrinsic. You can use ``llvm.usub.with.overflow``
9389 on any integer bit width.
9393 declare {i16, i1} @llvm.usub.with.overflow.i16(i16 %a, i16 %b)
9394 declare {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
9395 declare {i64, i1} @llvm.usub.with.overflow.i64(i64 %a, i64 %b)
9400 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
9401 an unsigned subtraction of the two arguments, and indicate whether an
9402 overflow occurred during the unsigned subtraction.
9407 The arguments (%a and %b) and the first element of the result structure
9408 may be of integer types of any bit width, but they must have the same
9409 bit width. The second element of the result structure must be of type
9410 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
9416 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
9417 an unsigned subtraction of the two arguments. They return a structure ---
9418 the first element of which is the subtraction, and the second element of
9419 which is a bit specifying if the unsigned subtraction resulted in an
9425 .. code-block:: llvm
9427 %res = call {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
9428 %sum = extractvalue {i32, i1} %res, 0
9429 %obit = extractvalue {i32, i1} %res, 1
9430 br i1 %obit, label %overflow, label %normal
9432 '``llvm.smul.with.overflow.*``' Intrinsics
9433 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9438 This is an overloaded intrinsic. You can use ``llvm.smul.with.overflow``
9439 on any integer bit width.
9443 declare {i16, i1} @llvm.smul.with.overflow.i16(i16 %a, i16 %b)
9444 declare {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
9445 declare {i64, i1} @llvm.smul.with.overflow.i64(i64 %a, i64 %b)
9450 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
9451 a signed multiplication of the two arguments, and indicate whether an
9452 overflow occurred during the signed multiplication.
9457 The arguments (%a and %b) and the first element of the result structure
9458 may be of integer types of any bit width, but they must have the same
9459 bit width. The second element of the result structure must be of type
9460 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
9466 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
9467 a signed multiplication of the two arguments. They return a structure ---
9468 the first element of which is the multiplication, and the second element
9469 of which is a bit specifying if the signed multiplication resulted in an
9475 .. code-block:: llvm
9477 %res = call {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
9478 %sum = extractvalue {i32, i1} %res, 0
9479 %obit = extractvalue {i32, i1} %res, 1
9480 br i1 %obit, label %overflow, label %normal
9482 '``llvm.umul.with.overflow.*``' Intrinsics
9483 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9488 This is an overloaded intrinsic. You can use ``llvm.umul.with.overflow``
9489 on any integer bit width.
9493 declare {i16, i1} @llvm.umul.with.overflow.i16(i16 %a, i16 %b)
9494 declare {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
9495 declare {i64, i1} @llvm.umul.with.overflow.i64(i64 %a, i64 %b)
9500 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
9501 a unsigned multiplication of the two arguments, and indicate whether an
9502 overflow occurred during the unsigned multiplication.
9507 The arguments (%a and %b) and the first element of the result structure
9508 may be of integer types of any bit width, but they must have the same
9509 bit width. The second element of the result structure must be of type
9510 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
9516 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
9517 an unsigned multiplication of the two arguments. They return a structure ---
9518 the first element of which is the multiplication, and the second
9519 element of which is a bit specifying if the unsigned multiplication
9520 resulted in an overflow.
9525 .. code-block:: llvm
9527 %res = call {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
9528 %sum = extractvalue {i32, i1} %res, 0
9529 %obit = extractvalue {i32, i1} %res, 1
9530 br i1 %obit, label %overflow, label %normal
9532 Specialised Arithmetic Intrinsics
9533 ---------------------------------
9535 '``llvm.fmuladd.*``' Intrinsic
9536 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9543 declare float @llvm.fmuladd.f32(float %a, float %b, float %c)
9544 declare double @llvm.fmuladd.f64(double %a, double %b, double %c)
9549 The '``llvm.fmuladd.*``' intrinsic functions represent multiply-add
9550 expressions that can be fused if the code generator determines that (a) the
9551 target instruction set has support for a fused operation, and (b) that the
9552 fused operation is more efficient than the equivalent, separate pair of mul
9553 and add instructions.
9558 The '``llvm.fmuladd.*``' intrinsics each take three arguments: two
9559 multiplicands, a and b, and an addend c.
9568 %0 = call float @llvm.fmuladd.f32(%a, %b, %c)
9570 is equivalent to the expression a \* b + c, except that rounding will
9571 not be performed between the multiplication and addition steps if the
9572 code generator fuses the operations. Fusion is not guaranteed, even if
9573 the target platform supports it. If a fused multiply-add is required the
9574 corresponding llvm.fma.\* intrinsic function should be used
9575 instead. This never sets errno, just as '``llvm.fma.*``'.
9580 .. code-block:: llvm
9582 %r2 = call float @llvm.fmuladd.f32(float %a, float %b, float %c) ; yields float:r2 = (a * b) + c
9584 Half Precision Floating Point Intrinsics
9585 ----------------------------------------
9587 For most target platforms, half precision floating point is a
9588 storage-only format. This means that it is a dense encoding (in memory)
9589 but does not support computation in the format.
9591 This means that code must first load the half-precision floating point
9592 value as an i16, then convert it to float with
9593 :ref:`llvm.convert.from.fp16 <int_convert_from_fp16>`. Computation can
9594 then be performed on the float value (including extending to double
9595 etc). To store the value back to memory, it is first converted to float
9596 if needed, then converted to i16 with
9597 :ref:`llvm.convert.to.fp16 <int_convert_to_fp16>`, then storing as an
9600 .. _int_convert_to_fp16:
9602 '``llvm.convert.to.fp16``' Intrinsic
9603 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9610 declare i16 @llvm.convert.to.fp16.f32(float %a)
9611 declare i16 @llvm.convert.to.fp16.f64(double %a)
9616 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion from a
9617 conventional floating point type to half precision floating point format.
9622 The intrinsic function contains single argument - the value to be
9628 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion from a
9629 conventional floating point format to half precision floating point format. The
9630 return value is an ``i16`` which contains the converted number.
9635 .. code-block:: llvm
9637 %res = call i16 @llvm.convert.to.fp16.f32(float %a)
9638 store i16 %res, i16* @x, align 2
9640 .. _int_convert_from_fp16:
9642 '``llvm.convert.from.fp16``' Intrinsic
9643 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9650 declare float @llvm.convert.from.fp16.f32(i16 %a)
9651 declare double @llvm.convert.from.fp16.f64(i16 %a)
9656 The '``llvm.convert.from.fp16``' intrinsic function performs a
9657 conversion from half precision floating point format to single precision
9658 floating point format.
9663 The intrinsic function contains single argument - the value to be
9669 The '``llvm.convert.from.fp16``' intrinsic function performs a
9670 conversion from half single precision floating point format to single
9671 precision floating point format. The input half-float value is
9672 represented by an ``i16`` value.
9677 .. code-block:: llvm
9679 %a = load i16, i16* @x, align 2
9680 %res = call float @llvm.convert.from.fp16(i16 %a)
9687 The LLVM debugger intrinsics (which all start with ``llvm.dbg.``
9688 prefix), are described in the `LLVM Source Level
9689 Debugging <SourceLevelDebugging.html#format_common_intrinsics>`_
9692 Exception Handling Intrinsics
9693 -----------------------------
9695 The LLVM exception handling intrinsics (which all start with
9696 ``llvm.eh.`` prefix), are described in the `LLVM Exception
9697 Handling <ExceptionHandling.html#format_common_intrinsics>`_ document.
9701 Trampoline Intrinsics
9702 ---------------------
9704 These intrinsics make it possible to excise one parameter, marked with
9705 the :ref:`nest <nest>` attribute, from a function. The result is a
9706 callable function pointer lacking the nest parameter - the caller does
9707 not need to provide a value for it. Instead, the value to use is stored
9708 in advance in a "trampoline", a block of memory usually allocated on the
9709 stack, which also contains code to splice the nest value into the
9710 argument list. This is used to implement the GCC nested function address
9713 For example, if the function is ``i32 f(i8* nest %c, i32 %x, i32 %y)``
9714 then the resulting function pointer has signature ``i32 (i32, i32)*``.
9715 It can be created as follows:
9717 .. code-block:: llvm
9719 %tramp = alloca [10 x i8], align 4 ; size and alignment only correct for X86
9720 %tramp1 = getelementptr [10 x i8], [10 x i8]* %tramp, i32 0, i32 0
9721 call i8* @llvm.init.trampoline(i8* %tramp1, i8* bitcast (i32 (i8*, i32, i32)* @f to i8*), i8* %nval)
9722 %p = call i8* @llvm.adjust.trampoline(i8* %tramp1)
9723 %fp = bitcast i8* %p to i32 (i32, i32)*
9725 The call ``%val = call i32 %fp(i32 %x, i32 %y)`` is then equivalent to
9726 ``%val = call i32 %f(i8* %nval, i32 %x, i32 %y)``.
9730 '``llvm.init.trampoline``' Intrinsic
9731 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9738 declare void @llvm.init.trampoline(i8* <tramp>, i8* <func>, i8* <nval>)
9743 This fills the memory pointed to by ``tramp`` with executable code,
9744 turning it into a trampoline.
9749 The ``llvm.init.trampoline`` intrinsic takes three arguments, all
9750 pointers. The ``tramp`` argument must point to a sufficiently large and
9751 sufficiently aligned block of memory; this memory is written to by the
9752 intrinsic. Note that the size and the alignment are target-specific -
9753 LLVM currently provides no portable way of determining them, so a
9754 front-end that generates this intrinsic needs to have some
9755 target-specific knowledge. The ``func`` argument must hold a function
9756 bitcast to an ``i8*``.
9761 The block of memory pointed to by ``tramp`` is filled with target
9762 dependent code, turning it into a function. Then ``tramp`` needs to be
9763 passed to :ref:`llvm.adjust.trampoline <int_at>` to get a pointer which can
9764 be :ref:`bitcast (to a new function) and called <int_trampoline>`. The new
9765 function's signature is the same as that of ``func`` with any arguments
9766 marked with the ``nest`` attribute removed. At most one such ``nest``
9767 argument is allowed, and it must be of pointer type. Calling the new
9768 function is equivalent to calling ``func`` with the same argument list,
9769 but with ``nval`` used for the missing ``nest`` argument. If, after
9770 calling ``llvm.init.trampoline``, the memory pointed to by ``tramp`` is
9771 modified, then the effect of any later call to the returned function
9772 pointer is undefined.
9776 '``llvm.adjust.trampoline``' Intrinsic
9777 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9784 declare i8* @llvm.adjust.trampoline(i8* <tramp>)
9789 This performs any required machine-specific adjustment to the address of
9790 a trampoline (passed as ``tramp``).
9795 ``tramp`` must point to a block of memory which already has trampoline
9796 code filled in by a previous call to
9797 :ref:`llvm.init.trampoline <int_it>`.
9802 On some architectures the address of the code to be executed needs to be
9803 different than the address where the trampoline is actually stored. This
9804 intrinsic returns the executable address corresponding to ``tramp``
9805 after performing the required machine specific adjustments. The pointer
9806 returned can then be :ref:`bitcast and executed <int_trampoline>`.
9808 .. _int_mload_mstore:
9810 Masked Vector Load and Store Intrinsics
9811 ---------------------------------------
9813 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.
9817 '``llvm.masked.load.*``' Intrinsics
9818 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9822 This is an overloaded intrinsic. The loaded data is a vector of any integer or floating point data type.
9826 declare <16 x float> @llvm.masked.load.v16f32 (<16 x float>* <ptr>, i32 <alignment>, <16 x i1> <mask>, <16 x float> <passthru>)
9827 declare <2 x double> @llvm.masked.load.v2f64 (<2 x double>* <ptr>, i32 <alignment>, <2 x i1> <mask>, <2 x double> <passthru>)
9832 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.
9838 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.
9844 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.
9845 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.
9850 %res = call <16 x float> @llvm.masked.load.v16f32 (<16 x float>* %ptr, i32 4, <16 x i1>%mask, <16 x float> %passthru)
9852 ;; The result of the two following instructions is identical aside from potential memory access exception
9853 %loadlal = load <16 x float>, <16 x float>* %ptr, align 4
9854 %res = select <16 x i1> %mask, <16 x float> %loadlal, <16 x float> %passthru
9858 '``llvm.masked.store.*``' Intrinsics
9859 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9863 This is an overloaded intrinsic. The data stored in memory is a vector of any integer or floating point data type.
9867 declare void @llvm.masked.store.v8i32 (<8 x i32> <value>, <8 x i32> * <ptr>, i32 <alignment>, <8 x i1> <mask>)
9868 declare void @llvm.masked.store.v16f32(<16 x i32> <value>, <16 x i32>* <ptr>, i32 <alignment>, <16 x i1> <mask>)
9873 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.
9878 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.
9884 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.
9885 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.
9889 call void @llvm.masked.store.v16f32(<16 x float> %value, <16 x float>* %ptr, i32 4, <16 x i1> %mask)
9891 ;; The result of the following instructions is identical aside from potential data races and memory access exceptions
9892 %oldval = load <16 x float>, <16 x float>* %ptr, align 4
9893 %res = select <16 x i1> %mask, <16 x float> %value, <16 x float> %oldval
9894 store <16 x float> %res, <16 x float>* %ptr, align 4
9897 Masked Vector Gather and Scatter Intrinsics
9898 -------------------------------------------
9900 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.
9904 '``llvm.masked.gather.*``' Intrinsics
9905 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9909 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.
9913 declare <16 x float> @llvm.masked.gather.v16f32 (<16 x float*> <ptrs>, i32 <alignment>, <16 x i1> <mask>, <16 x float> <passthru>)
9914 declare <2 x double> @llvm.masked.gather.v2f64 (<2 x double*> <ptrs>, i32 <alignment>, <2 x i1> <mask>, <2 x double> <passthru>)
9919 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.
9925 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.
9931 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.
9932 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.
9937 %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>)
9939 ;; The gather with all-true mask is equivalent to the following instruction sequence
9940 %ptr0 = extractelement <4 x double*> %ptrs, i32 0
9941 %ptr1 = extractelement <4 x double*> %ptrs, i32 1
9942 %ptr2 = extractelement <4 x double*> %ptrs, i32 2
9943 %ptr3 = extractelement <4 x double*> %ptrs, i32 3
9945 %val0 = load double, double* %ptr0, align 8
9946 %val1 = load double, double* %ptr1, align 8
9947 %val2 = load double, double* %ptr2, align 8
9948 %val3 = load double, double* %ptr3, align 8
9950 %vec0 = insertelement <4 x double>undef, %val0, 0
9951 %vec01 = insertelement <4 x double>%vec0, %val1, 1
9952 %vec012 = insertelement <4 x double>%vec01, %val2, 2
9953 %vec0123 = insertelement <4 x double>%vec012, %val3, 3
9957 '``llvm.masked.scatter.*``' Intrinsics
9958 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9962 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.
9966 declare void @llvm.masked.scatter.v8i32 (<8 x i32> <value>, <8 x i32*> <ptrs>, i32 <alignment>, <8 x i1> <mask>)
9967 declare void @llvm.masked.scatter.v16f32(<16 x i32> <value>, <16 x i32*> <ptrs>, i32 <alignment>, <16 x i1> <mask>)
9972 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.
9977 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.
9983 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.
9987 ;; This instruction unconditionaly stores data vector in multiple addresses
9988 call @llvm.masked.scatter.v8i32 (<8 x i32> %value, <8 x i32*> %ptrs, i32 4, <8 x i1> <true, true, .. true>)
9990 ;; It is equivalent to a list of scalar stores
9991 %val0 = extractelement <8 x i32> %value, i32 0
9992 %val1 = extractelement <8 x i32> %value, i32 1
9994 %val7 = extractelement <8 x i32> %value, i32 7
9995 %ptr0 = extractelement <8 x i32*> %ptrs, i32 0
9996 %ptr1 = extractelement <8 x i32*> %ptrs, i32 1
9998 %ptr7 = extractelement <8 x i32*> %ptrs, i32 7
9999 ;; Note: the order of the following stores is important when they overlap:
10000 store i32 %val0, i32* %ptr0, align 4
10001 store i32 %val1, i32* %ptr1, align 4
10003 store i32 %val7, i32* %ptr7, align 4
10009 This class of intrinsics provides information about the lifetime of
10010 memory objects and ranges where variables are immutable.
10014 '``llvm.lifetime.start``' Intrinsic
10015 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10022 declare void @llvm.lifetime.start(i64 <size>, i8* nocapture <ptr>)
10027 The '``llvm.lifetime.start``' intrinsic specifies the start of a memory
10033 The first argument is a constant integer representing the size of the
10034 object, or -1 if it is variable sized. The second argument is a pointer
10040 This intrinsic indicates that before this point in the code, the value
10041 of the memory pointed to by ``ptr`` is dead. This means that it is known
10042 to never be used and has an undefined value. A load from the pointer
10043 that precedes this intrinsic can be replaced with ``'undef'``.
10047 '``llvm.lifetime.end``' Intrinsic
10048 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10055 declare void @llvm.lifetime.end(i64 <size>, i8* nocapture <ptr>)
10060 The '``llvm.lifetime.end``' intrinsic specifies the end of a memory
10066 The first argument is a constant integer representing the size of the
10067 object, or -1 if it is variable sized. The second argument is a pointer
10073 This intrinsic indicates that after this point in the code, the value of
10074 the memory pointed to by ``ptr`` is dead. This means that it is known to
10075 never be used and has an undefined value. Any stores into the memory
10076 object following this intrinsic may be removed as dead.
10078 '``llvm.invariant.start``' Intrinsic
10079 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10086 declare {}* @llvm.invariant.start(i64 <size>, i8* nocapture <ptr>)
10091 The '``llvm.invariant.start``' intrinsic specifies that the contents of
10092 a memory object will not change.
10097 The first argument is a constant integer representing the size of the
10098 object, or -1 if it is variable sized. The second argument is a pointer
10104 This intrinsic indicates that until an ``llvm.invariant.end`` that uses
10105 the return value, the referenced memory location is constant and
10108 '``llvm.invariant.end``' Intrinsic
10109 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10116 declare void @llvm.invariant.end({}* <start>, i64 <size>, i8* nocapture <ptr>)
10121 The '``llvm.invariant.end``' intrinsic specifies that the contents of a
10122 memory object are mutable.
10127 The first argument is the matching ``llvm.invariant.start`` intrinsic.
10128 The second argument is a constant integer representing the size of the
10129 object, or -1 if it is variable sized and the third argument is a
10130 pointer to the object.
10135 This intrinsic indicates that the memory is mutable again.
10140 This class of intrinsics is designed to be generic and has no specific
10143 '``llvm.var.annotation``' Intrinsic
10144 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10151 declare void @llvm.var.annotation(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
10156 The '``llvm.var.annotation``' intrinsic.
10161 The first argument is a pointer to a value, the second is a pointer to a
10162 global string, the third is a pointer to a global string which is the
10163 source file name, and the last argument is the line number.
10168 This intrinsic allows annotation of local variables with arbitrary
10169 strings. This can be useful for special purpose optimizations that want
10170 to look for these annotations. These have no other defined use; they are
10171 ignored by code generation and optimization.
10173 '``llvm.ptr.annotation.*``' Intrinsic
10174 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10179 This is an overloaded intrinsic. You can use '``llvm.ptr.annotation``' on a
10180 pointer to an integer of any width. *NOTE* you must specify an address space for
10181 the pointer. The identifier for the default address space is the integer
10186 declare i8* @llvm.ptr.annotation.p<address space>i8(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
10187 declare i16* @llvm.ptr.annotation.p<address space>i16(i16* <val>, i8* <str>, i8* <str>, i32 <int>)
10188 declare i32* @llvm.ptr.annotation.p<address space>i32(i32* <val>, i8* <str>, i8* <str>, i32 <int>)
10189 declare i64* @llvm.ptr.annotation.p<address space>i64(i64* <val>, i8* <str>, i8* <str>, i32 <int>)
10190 declare i256* @llvm.ptr.annotation.p<address space>i256(i256* <val>, i8* <str>, i8* <str>, i32 <int>)
10195 The '``llvm.ptr.annotation``' intrinsic.
10200 The first argument is a pointer to an integer value of arbitrary bitwidth
10201 (result of some expression), the second is a pointer to a global string, the
10202 third is a pointer to a global string which is the source file name, and the
10203 last argument is the line number. It returns the value of the first argument.
10208 This intrinsic allows annotation of a pointer to an integer with arbitrary
10209 strings. This can be useful for special purpose optimizations that want to look
10210 for these annotations. These have no other defined use; they are ignored by code
10211 generation and optimization.
10213 '``llvm.annotation.*``' Intrinsic
10214 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10219 This is an overloaded intrinsic. You can use '``llvm.annotation``' on
10220 any integer bit width.
10224 declare i8 @llvm.annotation.i8(i8 <val>, i8* <str>, i8* <str>, i32 <int>)
10225 declare i16 @llvm.annotation.i16(i16 <val>, i8* <str>, i8* <str>, i32 <int>)
10226 declare i32 @llvm.annotation.i32(i32 <val>, i8* <str>, i8* <str>, i32 <int>)
10227 declare i64 @llvm.annotation.i64(i64 <val>, i8* <str>, i8* <str>, i32 <int>)
10228 declare i256 @llvm.annotation.i256(i256 <val>, i8* <str>, i8* <str>, i32 <int>)
10233 The '``llvm.annotation``' intrinsic.
10238 The first argument is an integer value (result of some expression), the
10239 second is a pointer to a global string, the third is a pointer to a
10240 global string which is the source file name, and the last argument is
10241 the line number. It returns the value of the first argument.
10246 This intrinsic allows annotations to be put on arbitrary expressions
10247 with arbitrary strings. This can be useful for special purpose
10248 optimizations that want to look for these annotations. These have no
10249 other defined use; they are ignored by code generation and optimization.
10251 '``llvm.trap``' Intrinsic
10252 ^^^^^^^^^^^^^^^^^^^^^^^^^
10259 declare void @llvm.trap() noreturn nounwind
10264 The '``llvm.trap``' intrinsic.
10274 This intrinsic is lowered to the target dependent trap instruction. If
10275 the target does not have a trap instruction, this intrinsic will be
10276 lowered to a call of the ``abort()`` function.
10278 '``llvm.debugtrap``' Intrinsic
10279 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10286 declare void @llvm.debugtrap() nounwind
10291 The '``llvm.debugtrap``' intrinsic.
10301 This intrinsic is lowered to code which is intended to cause an
10302 execution trap with the intention of requesting the attention of a
10305 '``llvm.stackprotector``' Intrinsic
10306 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10313 declare void @llvm.stackprotector(i8* <guard>, i8** <slot>)
10318 The ``llvm.stackprotector`` intrinsic takes the ``guard`` and stores it
10319 onto the stack at ``slot``. The stack slot is adjusted to ensure that it
10320 is placed on the stack before local variables.
10325 The ``llvm.stackprotector`` intrinsic requires two pointer arguments.
10326 The first argument is the value loaded from the stack guard
10327 ``@__stack_chk_guard``. The second variable is an ``alloca`` that has
10328 enough space to hold the value of the guard.
10333 This intrinsic causes the prologue/epilogue inserter to force the position of
10334 the ``AllocaInst`` stack slot to be before local variables on the stack. This is
10335 to ensure that if a local variable on the stack is overwritten, it will destroy
10336 the value of the guard. When the function exits, the guard on the stack is
10337 checked against the original guard by ``llvm.stackprotectorcheck``. If they are
10338 different, then ``llvm.stackprotectorcheck`` causes the program to abort by
10339 calling the ``__stack_chk_fail()`` function.
10341 '``llvm.stackprotectorcheck``' Intrinsic
10342 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10349 declare void @llvm.stackprotectorcheck(i8** <guard>)
10354 The ``llvm.stackprotectorcheck`` intrinsic compares ``guard`` against an already
10355 created stack protector and if they are not equal calls the
10356 ``__stack_chk_fail()`` function.
10361 The ``llvm.stackprotectorcheck`` intrinsic requires one pointer argument, the
10362 the variable ``@__stack_chk_guard``.
10367 This intrinsic is provided to perform the stack protector check by comparing
10368 ``guard`` with the stack slot created by ``llvm.stackprotector`` and if the
10369 values do not match call the ``__stack_chk_fail()`` function.
10371 The reason to provide this as an IR level intrinsic instead of implementing it
10372 via other IR operations is that in order to perform this operation at the IR
10373 level without an intrinsic, one would need to create additional basic blocks to
10374 handle the success/failure cases. This makes it difficult to stop the stack
10375 protector check from disrupting sibling tail calls in Codegen. With this
10376 intrinsic, we are able to generate the stack protector basic blocks late in
10377 codegen after the tail call decision has occurred.
10379 '``llvm.objectsize``' Intrinsic
10380 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10387 declare i32 @llvm.objectsize.i32(i8* <object>, i1 <min>)
10388 declare i64 @llvm.objectsize.i64(i8* <object>, i1 <min>)
10393 The ``llvm.objectsize`` intrinsic is designed to provide information to
10394 the optimizers to determine at compile time whether a) an operation
10395 (like memcpy) will overflow a buffer that corresponds to an object, or
10396 b) that a runtime check for overflow isn't necessary. An object in this
10397 context means an allocation of a specific class, structure, array, or
10403 The ``llvm.objectsize`` intrinsic takes two arguments. The first
10404 argument is a pointer to or into the ``object``. The second argument is
10405 a boolean and determines whether ``llvm.objectsize`` returns 0 (if true)
10406 or -1 (if false) when the object size is unknown. The second argument
10407 only accepts constants.
10412 The ``llvm.objectsize`` intrinsic is lowered to a constant representing
10413 the size of the object concerned. If the size cannot be determined at
10414 compile time, ``llvm.objectsize`` returns ``i32/i64 -1 or 0`` (depending
10415 on the ``min`` argument).
10417 '``llvm.expect``' Intrinsic
10418 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
10423 This is an overloaded intrinsic. You can use ``llvm.expect`` on any
10428 declare i1 @llvm.expect.i1(i1 <val>, i1 <expected_val>)
10429 declare i32 @llvm.expect.i32(i32 <val>, i32 <expected_val>)
10430 declare i64 @llvm.expect.i64(i64 <val>, i64 <expected_val>)
10435 The ``llvm.expect`` intrinsic provides information about expected (the
10436 most probable) value of ``val``, which can be used by optimizers.
10441 The ``llvm.expect`` intrinsic takes two arguments. The first argument is
10442 a value. The second argument is an expected value, this needs to be a
10443 constant value, variables are not allowed.
10448 This intrinsic is lowered to the ``val``.
10452 '``llvm.assume``' Intrinsic
10453 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10460 declare void @llvm.assume(i1 %cond)
10465 The ``llvm.assume`` allows the optimizer to assume that the provided
10466 condition is true. This information can then be used in simplifying other parts
10472 The condition which the optimizer may assume is always true.
10477 The intrinsic allows the optimizer to assume that the provided condition is
10478 always true whenever the control flow reaches the intrinsic call. No code is
10479 generated for this intrinsic, and instructions that contribute only to the
10480 provided condition are not used for code generation. If the condition is
10481 violated during execution, the behavior is undefined.
10483 Note that the optimizer might limit the transformations performed on values
10484 used by the ``llvm.assume`` intrinsic in order to preserve the instructions
10485 only used to form the intrinsic's input argument. This might prove undesirable
10486 if the extra information provided by the ``llvm.assume`` intrinsic does not cause
10487 sufficient overall improvement in code quality. For this reason,
10488 ``llvm.assume`` should not be used to document basic mathematical invariants
10489 that the optimizer can otherwise deduce or facts that are of little use to the
10494 '``llvm.bitset.test``' Intrinsic
10495 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10502 declare i1 @llvm.bitset.test(i8* %ptr, metadata %bitset) nounwind readnone
10508 The first argument is a pointer to be tested. The second argument is a
10509 metadata string containing the name of a :doc:`bitset <BitSets>`.
10514 The ``llvm.bitset.test`` intrinsic tests whether the given pointer is a
10515 member of the given bitset.
10517 '``llvm.donothing``' Intrinsic
10518 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10525 declare void @llvm.donothing() nounwind readnone
10530 The ``llvm.donothing`` intrinsic doesn't perform any operation. It's one of only
10531 two intrinsics (besides ``llvm.experimental.patchpoint``) that can be called
10532 with an invoke instruction.
10542 This intrinsic does nothing, and it's removed by optimizers and ignored
10545 Stack Map Intrinsics
10546 --------------------
10548 LLVM provides experimental intrinsics to support runtime patching
10549 mechanisms commonly desired in dynamic language JITs. These intrinsics
10550 are described in :doc:`StackMaps`.