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]* @.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 dynamcially
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 little
357 intrusive as possible. This calling convention behaves identical to the `C`
358 calling convention on how arguments and return values are passed, but it
359 uses a different set of caller/callee-saved registers. This alleviates the
360 burden of saving and recovering a large register set before and after the
361 call in the caller. If the arguments are passed in callee-saved registers,
362 then they will be preserved by the callee across the call. This doesn't
363 apply for values returned in callee-saved registers.
365 - On X86-64 the callee preserves all general purpose registers, except for
366 R11. R11 can be used as a scratch register. Floating-point registers
367 (XMMs/YMMs) are not preserved and need to be saved by the caller.
369 The idea behind this convention is to support calls to runtime functions
370 that have a hot path and a cold path. The hot path is usually a small piece
371 of code that doesn't use many registers. The cold path might need to call out to
372 another function and therefore only needs to preserve the caller-saved
373 registers, which haven't already been saved by the caller. The
374 `PreserveMost` calling convention is very similar to the `cold` calling
375 convention in terms of caller/callee-saved registers, but they are used for
376 different types of function calls. `coldcc` is for function calls that are
377 rarely executed, whereas `preserve_mostcc` function calls are intended to be
378 on the hot path and definitely executed a lot. Furthermore `preserve_mostcc`
379 doesn't prevent the inliner from inlining the function call.
381 This calling convention will be used by a future version of the ObjectiveC
382 runtime and should therefore still be considered experimental at this time.
383 Although this convention was created to optimize certain runtime calls to
384 the ObjectiveC runtime, it is not limited to this runtime and might be used
385 by other runtimes in the future too. The current implementation only
386 supports X86-64, but the intention is to support more architectures in the
388 "``preserve_allcc``" - The `PreserveAll` calling convention
389 This calling convention attempts to make the code in the caller even less
390 intrusive than the `PreserveMost` calling convention. This calling
391 convention also behaves identical to the `C` calling convention on how
392 arguments and return values are passed, but it uses a different set of
393 caller/callee-saved registers. This removes the burden of saving and
394 recovering a large register set before and after the call in the caller. If
395 the arguments are passed in callee-saved registers, then they will be
396 preserved by the callee across the call. This doesn't apply for values
397 returned in callee-saved registers.
399 - On X86-64 the callee preserves all general purpose registers, except for
400 R11. R11 can be used as a scratch register. Furthermore it also preserves
401 all floating-point registers (XMMs/YMMs).
403 The idea behind this convention is to support calls to runtime functions
404 that don't need to call out to any other functions.
406 This calling convention, like the `PreserveMost` calling convention, will be
407 used by a future version of the ObjectiveC runtime and should be considered
408 experimental at this time.
409 "``cc <n>``" - Numbered convention
410 Any calling convention may be specified by number, allowing
411 target-specific calling conventions to be used. Target specific
412 calling conventions start at 64.
414 More calling conventions can be added/defined on an as-needed basis, to
415 support Pascal conventions or any other well-known target-independent
418 .. _visibilitystyles:
423 All Global Variables and Functions have one of the following visibility
426 "``default``" - Default style
427 On targets that use the ELF object file format, default visibility
428 means that the declaration is visible to other modules and, in
429 shared libraries, means that the declared entity may be overridden.
430 On Darwin, default visibility means that the declaration is visible
431 to other modules. Default visibility corresponds to "external
432 linkage" in the language.
433 "``hidden``" - Hidden style
434 Two declarations of an object with hidden visibility refer to the
435 same object if they are in the same shared object. Usually, hidden
436 visibility indicates that the symbol will not be placed into the
437 dynamic symbol table, so no other module (executable or shared
438 library) can reference it directly.
439 "``protected``" - Protected style
440 On ELF, protected visibility indicates that the symbol will be
441 placed in the dynamic symbol table, but that references within the
442 defining module will bind to the local symbol. That is, the symbol
443 cannot be overridden by another module.
445 A symbol with ``internal`` or ``private`` linkage must have ``default``
453 All Global Variables, Functions and Aliases can have one of the following
457 "``dllimport``" causes the compiler to reference a function or variable via
458 a global pointer to a pointer that is set up by the DLL exporting the
459 symbol. On Microsoft Windows targets, the pointer name is formed by
460 combining ``__imp_`` and the function or variable name.
462 "``dllexport``" causes the compiler to provide a global pointer to a pointer
463 in a DLL, so that it can be referenced with the ``dllimport`` attribute. On
464 Microsoft Windows targets, the pointer name is formed by combining
465 ``__imp_`` and the function or variable name. Since this storage class
466 exists for defining a dll interface, the compiler, assembler and linker know
467 it is externally referenced and must refrain from deleting the symbol.
471 Thread Local Storage Models
472 ---------------------------
474 A variable may be defined as ``thread_local``, which means that it will
475 not be shared by threads (each thread will have a separated copy of the
476 variable). Not all targets support thread-local variables. Optionally, a
477 TLS model may be specified:
480 For variables that are only used within the current shared library.
482 For variables in modules that will not be loaded dynamically.
484 For variables defined in the executable and only used within it.
486 If no explicit model is given, the "general dynamic" model is used.
488 The models correspond to the ELF TLS models; see `ELF Handling For
489 Thread-Local Storage <http://people.redhat.com/drepper/tls.pdf>`_ for
490 more information on under which circumstances the different models may
491 be used. The target may choose a different TLS model if the specified
492 model is not supported, or if a better choice of model can be made.
494 A model can also be specified in a alias, but then it only governs how
495 the alias is accessed. It will not have any effect in the aliasee.
502 LLVM IR allows you to specify both "identified" and "literal" :ref:`structure
503 types <t_struct>`. Literal types are uniqued structurally, but identified types
504 are never uniqued. An :ref:`opaque structural type <t_opaque>` can also be used
505 to forward declare a type that is not yet available.
507 An example of a identified structure specification is:
511 %mytype = type { %mytype*, i32 }
513 Prior to the LLVM 3.0 release, identified types were structurally uniqued. Only
514 literal types are uniqued in recent versions of LLVM.
521 Global variables define regions of memory allocated at compilation time
524 Global variable definitions must be initialized.
526 Global variables in other translation units can also be declared, in which
527 case they don't have an initializer.
529 Either global variable definitions or declarations may have an explicit section
530 to be placed in and may have an optional explicit alignment specified.
532 A variable may be defined as a global ``constant``, which indicates that
533 the contents of the variable will **never** be modified (enabling better
534 optimization, allowing the global data to be placed in the read-only
535 section of an executable, etc). Note that variables that need runtime
536 initialization cannot be marked ``constant`` as there is a store to the
539 LLVM explicitly allows *declarations* of global variables to be marked
540 constant, even if the final definition of the global is not. This
541 capability can be used to enable slightly better optimization of the
542 program, but requires the language definition to guarantee that
543 optimizations based on the 'constantness' are valid for the translation
544 units that do not include the definition.
546 As SSA values, global variables define pointer values that are in scope
547 (i.e. they dominate) all basic blocks in the program. Global variables
548 always define a pointer to their "content" type because they describe a
549 region of memory, and all memory objects in LLVM are accessed through
552 Global variables can be marked with ``unnamed_addr`` which indicates
553 that the address is not significant, only the content. Constants marked
554 like this can be merged with other constants if they have the same
555 initializer. Note that a constant with significant address *can* be
556 merged with a ``unnamed_addr`` constant, the result being a constant
557 whose address is significant.
559 A global variable may be declared to reside in a target-specific
560 numbered address space. For targets that support them, address spaces
561 may affect how optimizations are performed and/or what target
562 instructions are used to access the variable. The default address space
563 is zero. The address space qualifier must precede any other attributes.
565 LLVM allows an explicit section to be specified for globals. If the
566 target supports it, it will emit globals to the section specified.
567 Additionally, the global can placed in a comdat if the target has the necessary
570 By default, global initializers are optimized by assuming that global
571 variables defined within the module are not modified from their
572 initial values before the start of the global initializer. This is
573 true even for variables potentially accessible from outside the
574 module, including those with external linkage or appearing in
575 ``@llvm.used`` or dllexported variables. This assumption may be suppressed
576 by marking the variable with ``externally_initialized``.
578 An explicit alignment may be specified for a global, which must be a
579 power of 2. If not present, or if the alignment is set to zero, the
580 alignment of the global is set by the target to whatever it feels
581 convenient. If an explicit alignment is specified, the global is forced
582 to have exactly that alignment. Targets and optimizers are not allowed
583 to over-align the global if the global has an assigned section. In this
584 case, the extra alignment could be observable: for example, code could
585 assume that the globals are densely packed in their section and try to
586 iterate over them as an array, alignment padding would break this
587 iteration. The maximum alignment is ``1 << 29``.
589 Globals can also have a :ref:`DLL storage class <dllstorageclass>`.
591 Variables and aliases can have a
592 :ref:`Thread Local Storage Model <tls_model>`.
596 [@<GlobalVarName> =] [Linkage] [Visibility] [DLLStorageClass] [ThreadLocal]
597 [unnamed_addr] [AddrSpace] [ExternallyInitialized]
598 <global | constant> <Type> [<InitializerConstant>]
599 [, section "name"] [, comdat [($name)]]
600 [, align <Alignment>]
602 For example, the following defines a global in a numbered address space
603 with an initializer, section, and alignment:
607 @G = addrspace(5) constant float 1.0, section "foo", align 4
609 The following example just declares a global variable
613 @G = external global i32
615 The following example defines a thread-local global with the
616 ``initialexec`` TLS model:
620 @G = thread_local(initialexec) global i32 0, align 4
622 .. _functionstructure:
627 LLVM function definitions consist of the "``define``" keyword, an
628 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
629 style <visibility>`, an optional :ref:`DLL storage class <dllstorageclass>`,
630 an optional :ref:`calling convention <callingconv>`,
631 an optional ``unnamed_addr`` attribute, a return type, an optional
632 :ref:`parameter attribute <paramattrs>` for the return type, a function
633 name, a (possibly empty) argument list (each with optional :ref:`parameter
634 attributes <paramattrs>`), optional :ref:`function attributes <fnattrs>`,
635 an optional section, an optional alignment,
636 an optional :ref:`comdat <langref_comdats>`,
637 an optional :ref:`garbage collector name <gc>`, an optional :ref:`prefix <prefixdata>`,
638 an optional :ref:`prologue <prologuedata>`, an opening
639 curly brace, a list of basic blocks, and a closing curly brace.
641 LLVM function declarations consist of the "``declare``" keyword, an
642 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
643 style <visibility>`, an optional :ref:`DLL storage class <dllstorageclass>`,
644 an optional :ref:`calling convention <callingconv>`,
645 an optional ``unnamed_addr`` attribute, a return type, an optional
646 :ref:`parameter attribute <paramattrs>` for the return type, a function
647 name, a possibly empty list of arguments, an optional alignment, an optional
648 :ref:`garbage collector name <gc>`, an optional :ref:`prefix <prefixdata>`,
649 and an optional :ref:`prologue <prologuedata>`.
651 A function definition contains a list of basic blocks, forming the CFG (Control
652 Flow Graph) for the function. Each basic block may optionally start with a label
653 (giving the basic block a symbol table entry), contains a list of instructions,
654 and ends with a :ref:`terminator <terminators>` instruction (such as a branch or
655 function return). If an explicit label is not provided, a block is assigned an
656 implicit numbered label, using the next value from the same counter as used for
657 unnamed temporaries (:ref:`see above<identifiers>`). For example, if a function
658 entry block does not have an explicit label, it will be assigned label "%0",
659 then the first unnamed temporary in that block will be "%1", etc.
661 The first basic block in a function is special in two ways: it is
662 immediately executed on entrance to the function, and it is not allowed
663 to have predecessor basic blocks (i.e. there can not be any branches to
664 the entry block of a function). Because the block can have no
665 predecessors, it also cannot have any :ref:`PHI nodes <i_phi>`.
667 LLVM allows an explicit section to be specified for functions. If the
668 target supports it, it will emit functions to the section specified.
669 Additionally, the function can be placed in a COMDAT.
671 An explicit alignment may be specified for a function. If not present,
672 or if the alignment is set to zero, the alignment of the function is set
673 by the target to whatever it feels convenient. If an explicit alignment
674 is specified, the function is forced to have at least that much
675 alignment. All alignments must be a power of 2.
677 If the ``unnamed_addr`` attribute is given, the address is known to not
678 be significant and two identical functions can be merged.
682 define [linkage] [visibility] [DLLStorageClass]
684 <ResultType> @<FunctionName> ([argument list])
685 [unnamed_addr] [fn Attrs] [section "name"] [comdat [($name)]]
686 [align N] [gc] [prefix Constant] [prologue Constant] { ... }
688 The argument list is a comma seperated sequence of arguments where each
689 argument is of the following form
693 <type> [parameter Attrs] [name]
701 Aliases, unlike function or variables, don't create any new data. They
702 are just a new symbol and metadata for an existing position.
704 Aliases have a name and an aliasee that is either a global value or a
707 Aliases may have an optional :ref:`linkage type <linkage>`, an optional
708 :ref:`visibility style <visibility>`, an optional :ref:`DLL storage class
709 <dllstorageclass>` and an optional :ref:`tls model <tls_model>`.
713 @<Name> = [Linkage] [Visibility] [DLLStorageClass] [ThreadLocal] [unnamed_addr] alias <AliaseeTy> @<Aliasee>
715 The linkage must be one of ``private``, ``internal``, ``linkonce``, ``weak``,
716 ``linkonce_odr``, ``weak_odr``, ``external``. Note that some system linkers
717 might not correctly handle dropping a weak symbol that is aliased.
719 Aliases that are not ``unnamed_addr`` are guaranteed to have the same address as
720 the aliasee expression. ``unnamed_addr`` ones are only guaranteed to point
723 Since aliases are only a second name, some restrictions apply, of which
724 some can only be checked when producing an object file:
726 * The expression defining the aliasee must be computable at assembly
727 time. Since it is just a name, no relocations can be used.
729 * No alias in the expression can be weak as the possibility of the
730 intermediate alias being overridden cannot be represented in an
733 * No global value in the expression can be a declaration, since that
734 would require a relocation, which is not possible.
741 Comdat IR provides access to COFF and ELF object file COMDAT functionality.
743 Comdats have a name which represents the COMDAT key. All global objects that
744 specify this key will only end up in the final object file if the linker chooses
745 that key over some other key. Aliases are placed in the same COMDAT that their
746 aliasee computes to, if any.
748 Comdats have a selection kind to provide input on how the linker should
749 choose between keys in two different object files.
753 $<Name> = comdat SelectionKind
755 The selection kind must be one of the following:
758 The linker may choose any COMDAT key, the choice is arbitrary.
760 The linker may choose any COMDAT key but the sections must contain the
763 The linker will choose the section containing the largest COMDAT key.
765 The linker requires that only section with this COMDAT key exist.
767 The linker may choose any COMDAT key but the sections must contain the
770 Note that the Mach-O platform doesn't support COMDATs and ELF only supports
771 ``any`` as a selection kind.
773 Here is an example of a COMDAT group where a function will only be selected if
774 the COMDAT key's section is the largest:
778 $foo = comdat largest
779 @foo = global i32 2, comdat($foo)
781 define void @bar() comdat($foo) {
785 As a syntactic sugar the ``$name`` can be omitted if the name is the same as
791 @foo = global i32 2, comdat
794 In a COFF object file, this will create a COMDAT section with selection kind
795 ``IMAGE_COMDAT_SELECT_LARGEST`` containing the contents of the ``@foo`` symbol
796 and another COMDAT section with selection kind
797 ``IMAGE_COMDAT_SELECT_ASSOCIATIVE`` which is associated with the first COMDAT
798 section and contains the contents of the ``@bar`` symbol.
800 There are some restrictions on the properties of the global object.
801 It, or an alias to it, must have the same name as the COMDAT group when
803 The contents and size of this object may be used during link-time to determine
804 which COMDAT groups get selected depending on the selection kind.
805 Because the name of the object must match the name of the COMDAT group, the
806 linkage of the global object must not be local; local symbols can get renamed
807 if a collision occurs in the symbol table.
809 The combined use of COMDATS and section attributes may yield surprising results.
816 @g1 = global i32 42, section "sec", comdat($foo)
817 @g2 = global i32 42, section "sec", comdat($bar)
819 From the object file perspective, this requires the creation of two sections
820 with the same name. This is necessary because both globals belong to different
821 COMDAT groups and COMDATs, at the object file level, are represented by
824 Note that certain IR constructs like global variables and functions may create
825 COMDATs in the object file in addition to any which are specified using COMDAT
826 IR. This arises, for example, when a global variable has linkonce_odr linkage.
828 .. _namedmetadatastructure:
833 Named metadata is a collection of metadata. :ref:`Metadata
834 nodes <metadata>` (but not metadata strings) are the only valid
835 operands for a named metadata.
839 ; Some unnamed metadata nodes, which are referenced by the named metadata.
844 !name = !{!0, !1, !2}
851 The return type and each parameter of a function type may have a set of
852 *parameter attributes* associated with them. Parameter attributes are
853 used to communicate additional information about the result or
854 parameters of a function. Parameter attributes are considered to be part
855 of the function, not of the function type, so functions with different
856 parameter attributes can have the same function type.
858 Parameter attributes are simple keywords that follow the type specified.
859 If multiple parameter attributes are needed, they are space separated.
864 declare i32 @printf(i8* noalias nocapture, ...)
865 declare i32 @atoi(i8 zeroext)
866 declare signext i8 @returns_signed_char()
868 Note that any attributes for the function result (``nounwind``,
869 ``readonly``) come immediately after the argument list.
871 Currently, only the following parameter attributes are defined:
874 This indicates to the code generator that the parameter or return
875 value should be zero-extended to the extent required by the target's
876 ABI (which is usually 32-bits, but is 8-bits for a i1 on x86-64) by
877 the caller (for a parameter) or the callee (for a return value).
879 This indicates to the code generator that the parameter or return
880 value should be sign-extended to the extent required by the target's
881 ABI (which is usually 32-bits) by the caller (for a parameter) or
882 the callee (for a return value).
884 This indicates that this parameter or return value should be treated
885 in a special target-dependent fashion during while emitting code for
886 a function call or return (usually, by putting it in a register as
887 opposed to memory, though some targets use it to distinguish between
888 two different kinds of registers). Use of this attribute is
891 This indicates that the pointer parameter should really be passed by
892 value to the function. The attribute implies that a hidden copy of
893 the pointee is made between the caller and the callee, so the callee
894 is unable to modify the value in the caller. This attribute is only
895 valid on LLVM pointer arguments. It is generally used to pass
896 structs and arrays by value, but is also valid on pointers to
897 scalars. The copy is considered to belong to the caller not the
898 callee (for example, ``readonly`` functions should not write to
899 ``byval`` parameters). This is not a valid attribute for return
902 The byval attribute also supports specifying an alignment with the
903 align attribute. It indicates the alignment of the stack slot to
904 form and the known alignment of the pointer specified to the call
905 site. If the alignment is not specified, then the code generator
906 makes a target-specific assumption.
912 The ``inalloca`` argument attribute allows the caller to take the
913 address of outgoing stack arguments. An ``inalloca`` argument must
914 be a pointer to stack memory produced by an ``alloca`` instruction.
915 The alloca, or argument allocation, must also be tagged with the
916 inalloca keyword. Only the last argument may have the ``inalloca``
917 attribute, and that argument is guaranteed to be passed in memory.
919 An argument allocation may be used by a call at most once because
920 the call may deallocate it. The ``inalloca`` attribute cannot be
921 used in conjunction with other attributes that affect argument
922 storage, like ``inreg``, ``nest``, ``sret``, or ``byval``. The
923 ``inalloca`` attribute also disables LLVM's implicit lowering of
924 large aggregate return values, which means that frontend authors
925 must lower them with ``sret`` pointers.
927 When the call site is reached, the argument allocation must have
928 been the most recent stack allocation that is still live, or the
929 results are undefined. It is possible to allocate additional stack
930 space after an argument allocation and before its call site, but it
931 must be cleared off with :ref:`llvm.stackrestore
934 See :doc:`InAlloca` for more information on how to use this
938 This indicates that the pointer parameter specifies the address of a
939 structure that is the return value of the function in the source
940 program. This pointer must be guaranteed by the caller to be valid:
941 loads and stores to the structure may be assumed by the callee
942 not to trap and to be properly aligned. This may only be applied to
943 the first parameter. This is not a valid attribute for return
947 This indicates that the pointer value may be assumed by the optimizer to
948 have the specified alignment.
950 Note that this attribute has additional semantics when combined with the
956 This indicates that objects accessed via pointer values
957 :ref:`based <pointeraliasing>` on the argument or return value are not also
958 accessed, during the execution of the function, via pointer values not
959 *based* on the argument or return value. The attribute on a return value
960 also has additional semantics described below. The caller shares the
961 responsibility with the callee for ensuring that these requirements are met.
962 For further details, please see the discussion of the NoAlias response in
963 :ref:`alias analysis <Must, May, or No>`.
965 Note that this definition of ``noalias`` is intentionally similar
966 to the definition of ``restrict`` in C99 for function arguments.
968 For function return values, C99's ``restrict`` is not meaningful,
969 while LLVM's ``noalias`` is. Furthermore, the semantics of the ``noalias``
970 attribute on return values are stronger than the semantics of the attribute
971 when used on function arguments. On function return values, the ``noalias``
972 attribute indicates that the function acts like a system memory allocation
973 function, returning a pointer to allocated storage disjoint from the
974 storage for any other object accessible to the caller.
977 This indicates that the callee does not make any copies of the
978 pointer that outlive the callee itself. This is not a valid
979 attribute for return values.
984 This indicates that the pointer parameter can be excised using the
985 :ref:`trampoline intrinsics <int_trampoline>`. This is not a valid
986 attribute for return values and can only be applied to one parameter.
989 This indicates that the function always returns the argument as its return
990 value. This is an optimization hint to the code generator when generating
991 the caller, allowing tail call optimization and omission of register saves
992 and restores in some cases; it is not checked or enforced when generating
993 the callee. The parameter and the function return type must be valid
994 operands for the :ref:`bitcast instruction <i_bitcast>`. This is not a
995 valid attribute for return values and can only be applied to one parameter.
998 This indicates that the parameter or return pointer is not null. This
999 attribute may only be applied to pointer typed parameters. This is not
1000 checked or enforced by LLVM, the caller must ensure that the pointer
1001 passed in is non-null, or the callee must ensure that the returned pointer
1004 ``dereferenceable(<n>)``
1005 This indicates that the parameter or return pointer is dereferenceable. This
1006 attribute may only be applied to pointer typed parameters. A pointer that
1007 is dereferenceable can be loaded from speculatively without a risk of
1008 trapping. The number of bytes known to be dereferenceable must be provided
1009 in parentheses. It is legal for the number of bytes to be less than the
1010 size of the pointee type. The ``nonnull`` attribute does not imply
1011 dereferenceability (consider a pointer to one element past the end of an
1012 array), however ``dereferenceable(<n>)`` does imply ``nonnull`` in
1013 ``addrspace(0)`` (which is the default address space).
1017 Garbage Collector Strategy Names
1018 --------------------------------
1020 Each function may specify a garbage collector strategy name, which is simply a
1023 .. code-block:: llvm
1025 define void @f() gc "name" { ... }
1027 The supported values of *name* includes those :ref:`built in to LLVM
1028 <builtin-gc-strategies>` and any provided by loaded plugins. Specifying a GC
1029 strategy will cause the compiler to alter its output in order to support the
1030 named garbage collection algorithm. Note that LLVM itself does not contain a
1031 garbage collector, this functionality is restricted to generating machine code
1032 which can interoperate with a collector provided externally.
1039 Prefix data is data associated with a function which the code
1040 generator will emit immediately before the function's entrypoint.
1041 The purpose of this feature is to allow frontends to associate
1042 language-specific runtime metadata with specific functions and make it
1043 available through the function pointer while still allowing the
1044 function pointer to be called.
1046 To access the data for a given function, a program may bitcast the
1047 function pointer to a pointer to the constant's type and dereference
1048 index -1. This implies that the IR symbol points just past the end of
1049 the prefix data. For instance, take the example of a function annotated
1050 with a single ``i32``,
1052 .. code-block:: llvm
1054 define void @f() prefix i32 123 { ... }
1056 The prefix data can be referenced as,
1058 .. code-block:: llvm
1060 %0 = bitcast *void () @f to *i32
1061 %a = getelementptr inbounds *i32 %0, i32 -1
1064 Prefix data is laid out as if it were an initializer for a global variable
1065 of the prefix data's type. The function will be placed such that the
1066 beginning of the prefix data is aligned. This means that if the size
1067 of the prefix data is not a multiple of the alignment size, the
1068 function's entrypoint will not be aligned. If alignment of the
1069 function's entrypoint is desired, padding must be added to the prefix
1072 A function may have prefix data but no body. This has similar semantics
1073 to the ``available_externally`` linkage in that the data may be used by the
1074 optimizers but will not be emitted in the object file.
1081 The ``prologue`` attribute allows arbitrary code (encoded as bytes) to
1082 be inserted prior to the function body. This can be used for enabling
1083 function hot-patching and instrumentation.
1085 To maintain the semantics of ordinary function calls, the prologue data must
1086 have a particular format. Specifically, it must begin with a sequence of
1087 bytes which decode to a sequence of machine instructions, valid for the
1088 module's target, which transfer control to the point immediately succeeding
1089 the prologue data, without performing any other visible action. This allows
1090 the inliner and other passes to reason about the semantics of the function
1091 definition without needing to reason about the prologue data. Obviously this
1092 makes the format of the prologue data highly target dependent.
1094 A trivial example of valid prologue data for the x86 architecture is ``i8 144``,
1095 which encodes the ``nop`` instruction:
1097 .. code-block:: llvm
1099 define void @f() prologue i8 144 { ... }
1101 Generally prologue data can be formed by encoding a relative branch instruction
1102 which skips the metadata, as in this example of valid prologue data for the
1103 x86_64 architecture, where the first two bytes encode ``jmp .+10``:
1105 .. code-block:: llvm
1107 %0 = type <{ i8, i8, i8* }>
1109 define void @f() prologue %0 <{ i8 235, i8 8, i8* @md}> { ... }
1111 A function may have prologue data but no body. This has similar semantics
1112 to the ``available_externally`` linkage in that the data may be used by the
1113 optimizers but will not be emitted in the object file.
1120 Attribute groups are groups of attributes that are referenced by objects within
1121 the IR. They are important for keeping ``.ll`` files readable, because a lot of
1122 functions will use the same set of attributes. In the degenerative case of a
1123 ``.ll`` file that corresponds to a single ``.c`` file, the single attribute
1124 group will capture the important command line flags used to build that file.
1126 An attribute group is a module-level object. To use an attribute group, an
1127 object references the attribute group's ID (e.g. ``#37``). An object may refer
1128 to more than one attribute group. In that situation, the attributes from the
1129 different groups are merged.
1131 Here is an example of attribute groups for a function that should always be
1132 inlined, has a stack alignment of 4, and which shouldn't use SSE instructions:
1134 .. code-block:: llvm
1136 ; Target-independent attributes:
1137 attributes #0 = { alwaysinline alignstack=4 }
1139 ; Target-dependent attributes:
1140 attributes #1 = { "no-sse" }
1142 ; Function @f has attributes: alwaysinline, alignstack=4, and "no-sse".
1143 define void @f() #0 #1 { ... }
1150 Function attributes are set to communicate additional information about
1151 a function. Function attributes are considered to be part of the
1152 function, not of the function type, so functions with different function
1153 attributes can have the same function type.
1155 Function attributes are simple keywords that follow the type specified.
1156 If multiple attributes are needed, they are space separated. For
1159 .. code-block:: llvm
1161 define void @f() noinline { ... }
1162 define void @f() alwaysinline { ... }
1163 define void @f() alwaysinline optsize { ... }
1164 define void @f() optsize { ... }
1167 This attribute indicates that, when emitting the prologue and
1168 epilogue, the backend should forcibly align the stack pointer.
1169 Specify the desired alignment, which must be a power of two, in
1172 This attribute indicates that the inliner should attempt to inline
1173 this function into callers whenever possible, ignoring any active
1174 inlining size threshold for this caller.
1176 This indicates that the callee function at a call site should be
1177 recognized as a built-in function, even though the function's declaration
1178 uses the ``nobuiltin`` attribute. This is only valid at call sites for
1179 direct calls to functions that are declared with the ``nobuiltin``
1182 This attribute indicates that this function is rarely called. When
1183 computing edge weights, basic blocks post-dominated by a cold
1184 function call are also considered to be cold; and, thus, given low
1187 This attribute indicates that the source code contained a hint that
1188 inlining this function is desirable (such as the "inline" keyword in
1189 C/C++). It is just a hint; it imposes no requirements on the
1192 This attribute indicates that the function should be added to a
1193 jump-instruction table at code-generation time, and that all address-taken
1194 references to this function should be replaced with a reference to the
1195 appropriate jump-instruction-table function pointer. Note that this creates
1196 a new pointer for the original function, which means that code that depends
1197 on function-pointer identity can break. So, any function annotated with
1198 ``jumptable`` must also be ``unnamed_addr``.
1200 This attribute suggests that optimization passes and code generator
1201 passes make choices that keep the code size of this function as small
1202 as possible and perform optimizations that may sacrifice runtime
1203 performance in order to minimize the size of the generated code.
1205 This attribute disables prologue / epilogue emission for the
1206 function. This can have very system-specific consequences.
1208 This indicates that the callee function at a call site is not recognized as
1209 a built-in function. LLVM will retain the original call and not replace it
1210 with equivalent code based on the semantics of the built-in function, unless
1211 the call site uses the ``builtin`` attribute. This is valid at call sites
1212 and on function declarations and definitions.
1214 This attribute indicates that calls to the function cannot be
1215 duplicated. A call to a ``noduplicate`` function may be moved
1216 within its parent function, but may not be duplicated within
1217 its parent function.
1219 A function containing a ``noduplicate`` call may still
1220 be an inlining candidate, provided that the call is not
1221 duplicated by inlining. That implies that the function has
1222 internal linkage and only has one call site, so the original
1223 call is dead after inlining.
1225 This attributes disables implicit floating point instructions.
1227 This attribute indicates that the inliner should never inline this
1228 function in any situation. This attribute may not be used together
1229 with the ``alwaysinline`` attribute.
1231 This attribute suppresses lazy symbol binding for the function. This
1232 may make calls to the function faster, at the cost of extra program
1233 startup time if the function is not called during program startup.
1235 This attribute indicates that the code generator should not use a
1236 red zone, even if the target-specific ABI normally permits it.
1238 This function attribute indicates that the function never returns
1239 normally. This produces undefined behavior at runtime if the
1240 function ever does dynamically return.
1242 This function attribute indicates that the function never raises an
1243 exception. If the function does raise an exception, its runtime
1244 behavior is undefined. However, functions marked nounwind may still
1245 trap or generate asynchronous exceptions. Exception handling schemes
1246 that are recognized by LLVM to handle asynchronous exceptions, such
1247 as SEH, will still provide their implementation defined semantics.
1249 This function attribute indicates that the function is not optimized
1250 by any optimization or code generator passes with the
1251 exception of interprocedural optimization passes.
1252 This attribute cannot be used together with the ``alwaysinline``
1253 attribute; this attribute is also incompatible
1254 with the ``minsize`` attribute and the ``optsize`` attribute.
1256 This attribute requires the ``noinline`` attribute to be specified on
1257 the function as well, so the function is never inlined into any caller.
1258 Only functions with the ``alwaysinline`` attribute are valid
1259 candidates for inlining into the body of this function.
1261 This attribute suggests that optimization passes and code generator
1262 passes make choices that keep the code size of this function low,
1263 and otherwise do optimizations specifically to reduce code size as
1264 long as they do not significantly impact runtime performance.
1266 On a function, this attribute indicates that the function computes its
1267 result (or decides to unwind an exception) based strictly on its arguments,
1268 without dereferencing any pointer arguments or otherwise accessing
1269 any mutable state (e.g. memory, control registers, etc) visible to
1270 caller functions. It does not write through any pointer arguments
1271 (including ``byval`` arguments) and never changes any state visible
1272 to callers. This means that it cannot unwind exceptions by calling
1273 the ``C++`` exception throwing methods.
1275 On an argument, this attribute indicates that the function does not
1276 dereference that pointer argument, even though it may read or write the
1277 memory that the pointer points to if accessed through other pointers.
1279 On a function, this attribute indicates that the function does not write
1280 through any pointer arguments (including ``byval`` arguments) or otherwise
1281 modify any state (e.g. memory, control registers, etc) visible to
1282 caller functions. It may dereference pointer arguments and read
1283 state that may be set in the caller. A readonly function always
1284 returns the same value (or unwinds an exception identically) when
1285 called with the same set of arguments and global state. It cannot
1286 unwind an exception by calling the ``C++`` exception throwing
1289 On an argument, this attribute indicates that the function does not write
1290 through this pointer argument, even though it may write to the memory that
1291 the pointer points to.
1293 This attribute indicates that this function can return twice. The C
1294 ``setjmp`` is an example of such a function. The compiler disables
1295 some optimizations (like tail calls) in the caller of these
1297 ``sanitize_address``
1298 This attribute indicates that AddressSanitizer checks
1299 (dynamic address safety analysis) are enabled for this function.
1301 This attribute indicates that MemorySanitizer checks (dynamic detection
1302 of accesses to uninitialized memory) are enabled for this function.
1304 This attribute indicates that ThreadSanitizer checks
1305 (dynamic thread safety analysis) are enabled for this function.
1307 This attribute indicates that the function should emit a stack
1308 smashing protector. It is in the form of a "canary" --- a random value
1309 placed on the stack before the local variables that's checked upon
1310 return from the function to see if it has been overwritten. A
1311 heuristic is used to determine if a function needs stack protectors
1312 or not. The heuristic used will enable protectors for functions with:
1314 - Character arrays larger than ``ssp-buffer-size`` (default 8).
1315 - Aggregates containing character arrays larger than ``ssp-buffer-size``.
1316 - Calls to alloca() with variable sizes or constant sizes greater than
1317 ``ssp-buffer-size``.
1319 Variables that are identified as requiring a protector will be arranged
1320 on the stack such that they are adjacent to the stack protector guard.
1322 If a function that has an ``ssp`` attribute is inlined into a
1323 function that doesn't have an ``ssp`` attribute, then the resulting
1324 function will have an ``ssp`` attribute.
1326 This attribute indicates that the function should *always* emit a
1327 stack smashing protector. This overrides the ``ssp`` function
1330 Variables that are identified as requiring a protector will be arranged
1331 on the stack such that they are adjacent to the stack protector guard.
1332 The specific layout rules are:
1334 #. Large arrays and structures containing large arrays
1335 (``>= ssp-buffer-size``) are closest to the stack protector.
1336 #. Small arrays and structures containing small arrays
1337 (``< ssp-buffer-size``) are 2nd closest to the protector.
1338 #. Variables that have had their address taken are 3rd closest to the
1341 If a function that has an ``sspreq`` attribute is inlined into a
1342 function that doesn't have an ``sspreq`` attribute or which has an
1343 ``ssp`` or ``sspstrong`` attribute, then the resulting function will have
1344 an ``sspreq`` attribute.
1346 This attribute indicates that the function should emit a stack smashing
1347 protector. This attribute causes a strong heuristic to be used when
1348 determining if a function needs stack protectors. The strong heuristic
1349 will enable protectors for functions with:
1351 - Arrays of any size and type
1352 - Aggregates containing an array of any size and type.
1353 - Calls to alloca().
1354 - Local variables that have had their address taken.
1356 Variables that are identified as requiring a protector will be arranged
1357 on the stack such that they are adjacent to the stack protector guard.
1358 The specific layout rules are:
1360 #. Large arrays and structures containing large arrays
1361 (``>= ssp-buffer-size``) are closest to the stack protector.
1362 #. Small arrays and structures containing small arrays
1363 (``< ssp-buffer-size``) are 2nd closest to the protector.
1364 #. Variables that have had their address taken are 3rd closest to the
1367 This overrides the ``ssp`` function attribute.
1369 If a function that has an ``sspstrong`` attribute is inlined into a
1370 function that doesn't have an ``sspstrong`` attribute, then the
1371 resulting function will have an ``sspstrong`` attribute.
1373 This attribute indicates that the ABI being targeted requires that
1374 an unwind table entry be produce for this function even if we can
1375 show that no exceptions passes by it. This is normally the case for
1376 the ELF x86-64 abi, but it can be disabled for some compilation
1381 Module-Level Inline Assembly
1382 ----------------------------
1384 Modules may contain "module-level inline asm" blocks, which corresponds
1385 to the GCC "file scope inline asm" blocks. These blocks are internally
1386 concatenated by LLVM and treated as a single unit, but may be separated
1387 in the ``.ll`` file if desired. The syntax is very simple:
1389 .. code-block:: llvm
1391 module asm "inline asm code goes here"
1392 module asm "more can go here"
1394 The strings can contain any character by escaping non-printable
1395 characters. The escape sequence used is simply "\\xx" where "xx" is the
1396 two digit hex code for the number.
1398 The inline asm code is simply printed to the machine code .s file when
1399 assembly code is generated.
1401 .. _langref_datalayout:
1406 A module may specify a target specific data layout string that specifies
1407 how data is to be laid out in memory. The syntax for the data layout is
1410 .. code-block:: llvm
1412 target datalayout = "layout specification"
1414 The *layout specification* consists of a list of specifications
1415 separated by the minus sign character ('-'). Each specification starts
1416 with a letter and may include other information after the letter to
1417 define some aspect of the data layout. The specifications accepted are
1421 Specifies that the target lays out data in big-endian form. That is,
1422 the bits with the most significance have the lowest address
1425 Specifies that the target lays out data in little-endian form. That
1426 is, the bits with the least significance have the lowest address
1429 Specifies the natural alignment of the stack in bits. Alignment
1430 promotion of stack variables is limited to the natural stack
1431 alignment to avoid dynamic stack realignment. The stack alignment
1432 must be a multiple of 8-bits. If omitted, the natural stack
1433 alignment defaults to "unspecified", which does not prevent any
1434 alignment promotions.
1435 ``p[n]:<size>:<abi>:<pref>``
1436 This specifies the *size* of a pointer and its ``<abi>`` and
1437 ``<pref>``\erred alignments for address space ``n``. All sizes are in
1438 bits. The address space, ``n`` is optional, and if not specified,
1439 denotes the default address space 0. The value of ``n`` must be
1440 in the range [1,2^23).
1441 ``i<size>:<abi>:<pref>``
1442 This specifies the alignment for an integer type of a given bit
1443 ``<size>``. The value of ``<size>`` must be in the range [1,2^23).
1444 ``v<size>:<abi>:<pref>``
1445 This specifies the alignment for a vector type of a given bit
1447 ``f<size>:<abi>:<pref>``
1448 This specifies the alignment for a floating point type of a given bit
1449 ``<size>``. Only values of ``<size>`` that are supported by the target
1450 will work. 32 (float) and 64 (double) are supported on all targets; 80
1451 or 128 (different flavors of long double) are also supported on some
1454 This specifies the alignment for an object of aggregate type.
1456 If present, specifies that llvm names are mangled in the output. The
1459 * ``e``: ELF mangling: Private symbols get a ``.L`` prefix.
1460 * ``m``: Mips mangling: Private symbols get a ``$`` prefix.
1461 * ``o``: Mach-O mangling: Private symbols get ``L`` prefix. Other
1462 symbols get a ``_`` prefix.
1463 * ``w``: Windows COFF prefix: Similar to Mach-O, but stdcall and fastcall
1464 functions also get a suffix based on the frame size.
1465 ``n<size1>:<size2>:<size3>...``
1466 This specifies a set of native integer widths for the target CPU in
1467 bits. For example, it might contain ``n32`` for 32-bit PowerPC,
1468 ``n32:64`` for PowerPC 64, or ``n8:16:32:64`` for X86-64. Elements of
1469 this set are considered to support most general arithmetic operations
1472 On every specification that takes a ``<abi>:<pref>``, specifying the
1473 ``<pref>`` alignment is optional. If omitted, the preceding ``:``
1474 should be omitted too and ``<pref>`` will be equal to ``<abi>``.
1476 When constructing the data layout for a given target, LLVM starts with a
1477 default set of specifications which are then (possibly) overridden by
1478 the specifications in the ``datalayout`` keyword. The default
1479 specifications are given in this list:
1481 - ``E`` - big endian
1482 - ``p:64:64:64`` - 64-bit pointers with 64-bit alignment.
1483 - ``p[n]:64:64:64`` - Other address spaces are assumed to be the
1484 same as the default address space.
1485 - ``S0`` - natural stack alignment is unspecified
1486 - ``i1:8:8`` - i1 is 8-bit (byte) aligned
1487 - ``i8:8:8`` - i8 is 8-bit (byte) aligned
1488 - ``i16:16:16`` - i16 is 16-bit aligned
1489 - ``i32:32:32`` - i32 is 32-bit aligned
1490 - ``i64:32:64`` - i64 has ABI alignment of 32-bits but preferred
1491 alignment of 64-bits
1492 - ``f16:16:16`` - half is 16-bit aligned
1493 - ``f32:32:32`` - float is 32-bit aligned
1494 - ``f64:64:64`` - double is 64-bit aligned
1495 - ``f128:128:128`` - quad is 128-bit aligned
1496 - ``v64:64:64`` - 64-bit vector is 64-bit aligned
1497 - ``v128:128:128`` - 128-bit vector is 128-bit aligned
1498 - ``a:0:64`` - aggregates are 64-bit aligned
1500 When LLVM is determining the alignment for a given type, it uses the
1503 #. If the type sought is an exact match for one of the specifications,
1504 that specification is used.
1505 #. If no match is found, and the type sought is an integer type, then
1506 the smallest integer type that is larger than the bitwidth of the
1507 sought type is used. If none of the specifications are larger than
1508 the bitwidth then the largest integer type is used. For example,
1509 given the default specifications above, the i7 type will use the
1510 alignment of i8 (next largest) while both i65 and i256 will use the
1511 alignment of i64 (largest specified).
1512 #. If no match is found, and the type sought is a vector type, then the
1513 largest vector type that is smaller than the sought vector type will
1514 be used as a fall back. This happens because <128 x double> can be
1515 implemented in terms of 64 <2 x double>, for example.
1517 The function of the data layout string may not be what you expect.
1518 Notably, this is not a specification from the frontend of what alignment
1519 the code generator should use.
1521 Instead, if specified, the target data layout is required to match what
1522 the ultimate *code generator* expects. This string is used by the
1523 mid-level optimizers to improve code, and this only works if it matches
1524 what the ultimate code generator uses. If you would like to generate IR
1525 that does not embed this target-specific detail into the IR, then you
1526 don't have to specify the string. This will disable some optimizations
1527 that require precise layout information, but this also prevents those
1528 optimizations from introducing target specificity into the IR.
1535 A module may specify a target triple string that describes the target
1536 host. The syntax for the target triple is simply:
1538 .. code-block:: llvm
1540 target triple = "x86_64-apple-macosx10.7.0"
1542 The *target triple* string consists of a series of identifiers delimited
1543 by the minus sign character ('-'). The canonical forms are:
1547 ARCHITECTURE-VENDOR-OPERATING_SYSTEM
1548 ARCHITECTURE-VENDOR-OPERATING_SYSTEM-ENVIRONMENT
1550 This information is passed along to the backend so that it generates
1551 code for the proper architecture. It's possible to override this on the
1552 command line with the ``-mtriple`` command line option.
1554 .. _pointeraliasing:
1556 Pointer Aliasing Rules
1557 ----------------------
1559 Any memory access must be done through a pointer value associated with
1560 an address range of the memory access, otherwise the behavior is
1561 undefined. Pointer values are associated with address ranges according
1562 to the following rules:
1564 - A pointer value is associated with the addresses associated with any
1565 value it is *based* on.
1566 - An address of a global variable is associated with the address range
1567 of the variable's storage.
1568 - The result value of an allocation instruction is associated with the
1569 address range of the allocated storage.
1570 - A null pointer in the default address-space is associated with no
1572 - An integer constant other than zero or a pointer value returned from
1573 a function not defined within LLVM may be associated with address
1574 ranges allocated through mechanisms other than those provided by
1575 LLVM. Such ranges shall not overlap with any ranges of addresses
1576 allocated by mechanisms provided by LLVM.
1578 A pointer value is *based* on another pointer value according to the
1581 - A pointer value formed from a ``getelementptr`` operation is *based*
1582 on the first operand of the ``getelementptr``.
1583 - The result value of a ``bitcast`` is *based* on the operand of the
1585 - A pointer value formed by an ``inttoptr`` is *based* on all pointer
1586 values that contribute (directly or indirectly) to the computation of
1587 the pointer's value.
1588 - The "*based* on" relationship is transitive.
1590 Note that this definition of *"based"* is intentionally similar to the
1591 definition of *"based"* in C99, though it is slightly weaker.
1593 LLVM IR does not associate types with memory. The result type of a
1594 ``load`` merely indicates the size and alignment of the memory from
1595 which to load, as well as the interpretation of the value. The first
1596 operand type of a ``store`` similarly only indicates the size and
1597 alignment of the store.
1599 Consequently, type-based alias analysis, aka TBAA, aka
1600 ``-fstrict-aliasing``, is not applicable to general unadorned LLVM IR.
1601 :ref:`Metadata <metadata>` may be used to encode additional information
1602 which specialized optimization passes may use to implement type-based
1607 Volatile Memory Accesses
1608 ------------------------
1610 Certain memory accesses, such as :ref:`load <i_load>`'s,
1611 :ref:`store <i_store>`'s, and :ref:`llvm.memcpy <int_memcpy>`'s may be
1612 marked ``volatile``. The optimizers must not change the number of
1613 volatile operations or change their order of execution relative to other
1614 volatile operations. The optimizers *may* change the order of volatile
1615 operations relative to non-volatile operations. This is not Java's
1616 "volatile" and has no cross-thread synchronization behavior.
1618 IR-level volatile loads and stores cannot safely be optimized into
1619 llvm.memcpy or llvm.memmove intrinsics even when those intrinsics are
1620 flagged volatile. Likewise, the backend should never split or merge
1621 target-legal volatile load/store instructions.
1623 .. admonition:: Rationale
1625 Platforms may rely on volatile loads and stores of natively supported
1626 data width to be executed as single instruction. For example, in C
1627 this holds for an l-value of volatile primitive type with native
1628 hardware support, but not necessarily for aggregate types. The
1629 frontend upholds these expectations, which are intentionally
1630 unspecified in the IR. The rules above ensure that IR transformation
1631 do not violate the frontend's contract with the language.
1635 Memory Model for Concurrent Operations
1636 --------------------------------------
1638 The LLVM IR does not define any way to start parallel threads of
1639 execution or to register signal handlers. Nonetheless, there are
1640 platform-specific ways to create them, and we define LLVM IR's behavior
1641 in their presence. This model is inspired by the C++0x memory model.
1643 For a more informal introduction to this model, see the :doc:`Atomics`.
1645 We define a *happens-before* partial order as the least partial order
1648 - Is a superset of single-thread program order, and
1649 - When a *synchronizes-with* ``b``, includes an edge from ``a`` to
1650 ``b``. *Synchronizes-with* pairs are introduced by platform-specific
1651 techniques, like pthread locks, thread creation, thread joining,
1652 etc., and by atomic instructions. (See also :ref:`Atomic Memory Ordering
1653 Constraints <ordering>`).
1655 Note that program order does not introduce *happens-before* edges
1656 between a thread and signals executing inside that thread.
1658 Every (defined) read operation (load instructions, memcpy, atomic
1659 loads/read-modify-writes, etc.) R reads a series of bytes written by
1660 (defined) write operations (store instructions, atomic
1661 stores/read-modify-writes, memcpy, etc.). For the purposes of this
1662 section, initialized globals are considered to have a write of the
1663 initializer which is atomic and happens before any other read or write
1664 of the memory in question. For each byte of a read R, R\ :sub:`byte`
1665 may see any write to the same byte, except:
1667 - If write\ :sub:`1` happens before write\ :sub:`2`, and
1668 write\ :sub:`2` happens before R\ :sub:`byte`, then
1669 R\ :sub:`byte` does not see write\ :sub:`1`.
1670 - If R\ :sub:`byte` happens before write\ :sub:`3`, then
1671 R\ :sub:`byte` does not see write\ :sub:`3`.
1673 Given that definition, R\ :sub:`byte` is defined as follows:
1675 - If R is volatile, the result is target-dependent. (Volatile is
1676 supposed to give guarantees which can support ``sig_atomic_t`` in
1677 C/C++, and may be used for accesses to addresses that do not behave
1678 like normal memory. It does not generally provide cross-thread
1680 - Otherwise, if there is no write to the same byte that happens before
1681 R\ :sub:`byte`, R\ :sub:`byte` returns ``undef`` for that byte.
1682 - Otherwise, if R\ :sub:`byte` may see exactly one write,
1683 R\ :sub:`byte` returns the value written by that write.
1684 - Otherwise, if R is atomic, and all the writes R\ :sub:`byte` may
1685 see are atomic, it chooses one of the values written. See the :ref:`Atomic
1686 Memory Ordering Constraints <ordering>` section for additional
1687 constraints on how the choice is made.
1688 - Otherwise R\ :sub:`byte` returns ``undef``.
1690 R returns the value composed of the series of bytes it read. This
1691 implies that some bytes within the value may be ``undef`` **without**
1692 the entire value being ``undef``. Note that this only defines the
1693 semantics of the operation; it doesn't mean that targets will emit more
1694 than one instruction to read the series of bytes.
1696 Note that in cases where none of the atomic intrinsics are used, this
1697 model places only one restriction on IR transformations on top of what
1698 is required for single-threaded execution: introducing a store to a byte
1699 which might not otherwise be stored is not allowed in general.
1700 (Specifically, in the case where another thread might write to and read
1701 from an address, introducing a store can change a load that may see
1702 exactly one write into a load that may see multiple writes.)
1706 Atomic Memory Ordering Constraints
1707 ----------------------------------
1709 Atomic instructions (:ref:`cmpxchg <i_cmpxchg>`,
1710 :ref:`atomicrmw <i_atomicrmw>`, :ref:`fence <i_fence>`,
1711 :ref:`atomic load <i_load>`, and :ref:`atomic store <i_store>`) take
1712 ordering parameters that determine which other atomic instructions on
1713 the same address they *synchronize with*. These semantics are borrowed
1714 from Java and C++0x, but are somewhat more colloquial. If these
1715 descriptions aren't precise enough, check those specs (see spec
1716 references in the :doc:`atomics guide <Atomics>`).
1717 :ref:`fence <i_fence>` instructions treat these orderings somewhat
1718 differently since they don't take an address. See that instruction's
1719 documentation for details.
1721 For a simpler introduction to the ordering constraints, see the
1725 The set of values that can be read is governed by the happens-before
1726 partial order. A value cannot be read unless some operation wrote
1727 it. This is intended to provide a guarantee strong enough to model
1728 Java's non-volatile shared variables. This ordering cannot be
1729 specified for read-modify-write operations; it is not strong enough
1730 to make them atomic in any interesting way.
1732 In addition to the guarantees of ``unordered``, there is a single
1733 total order for modifications by ``monotonic`` operations on each
1734 address. All modification orders must be compatible with the
1735 happens-before order. There is no guarantee that the modification
1736 orders can be combined to a global total order for the whole program
1737 (and this often will not be possible). The read in an atomic
1738 read-modify-write operation (:ref:`cmpxchg <i_cmpxchg>` and
1739 :ref:`atomicrmw <i_atomicrmw>`) reads the value in the modification
1740 order immediately before the value it writes. If one atomic read
1741 happens before another atomic read of the same address, the later
1742 read must see the same value or a later value in the address's
1743 modification order. This disallows reordering of ``monotonic`` (or
1744 stronger) operations on the same address. If an address is written
1745 ``monotonic``-ally by one thread, and other threads ``monotonic``-ally
1746 read that address repeatedly, the other threads must eventually see
1747 the write. This corresponds to the C++0x/C1x
1748 ``memory_order_relaxed``.
1750 In addition to the guarantees of ``monotonic``, a
1751 *synchronizes-with* edge may be formed with a ``release`` operation.
1752 This is intended to model C++'s ``memory_order_acquire``.
1754 In addition to the guarantees of ``monotonic``, if this operation
1755 writes a value which is subsequently read by an ``acquire``
1756 operation, it *synchronizes-with* that operation. (This isn't a
1757 complete description; see the C++0x definition of a release
1758 sequence.) This corresponds to the C++0x/C1x
1759 ``memory_order_release``.
1760 ``acq_rel`` (acquire+release)
1761 Acts as both an ``acquire`` and ``release`` operation on its
1762 address. This corresponds to the C++0x/C1x ``memory_order_acq_rel``.
1763 ``seq_cst`` (sequentially consistent)
1764 In addition to the guarantees of ``acq_rel`` (``acquire`` for an
1765 operation that only reads, ``release`` for an operation that only
1766 writes), there is a global total order on all
1767 sequentially-consistent operations on all addresses, which is
1768 consistent with the *happens-before* partial order and with the
1769 modification orders of all the affected addresses. Each
1770 sequentially-consistent read sees the last preceding write to the
1771 same address in this global order. This corresponds to the C++0x/C1x
1772 ``memory_order_seq_cst`` and Java volatile.
1776 If an atomic operation is marked ``singlethread``, it only *synchronizes
1777 with* or participates in modification and seq\_cst total orderings with
1778 other operations running in the same thread (for example, in signal
1786 LLVM IR floating-point binary ops (:ref:`fadd <i_fadd>`,
1787 :ref:`fsub <i_fsub>`, :ref:`fmul <i_fmul>`, :ref:`fdiv <i_fdiv>`,
1788 :ref:`frem <i_frem>`) have the following flags that can be set to enable
1789 otherwise unsafe floating point operations
1792 No NaNs - Allow optimizations to assume the arguments and result are not
1793 NaN. Such optimizations are required to retain defined behavior over
1794 NaNs, but the value of the result is undefined.
1797 No Infs - Allow optimizations to assume the arguments and result are not
1798 +/-Inf. Such optimizations are required to retain defined behavior over
1799 +/-Inf, but the value of the result is undefined.
1802 No Signed Zeros - Allow optimizations to treat the sign of a zero
1803 argument or result as insignificant.
1806 Allow Reciprocal - Allow optimizations to use the reciprocal of an
1807 argument rather than perform division.
1810 Fast - Allow algebraically equivalent transformations that may
1811 dramatically change results in floating point (e.g. reassociate). This
1812 flag implies all the others.
1816 Use-list Order Directives
1817 -------------------------
1819 Use-list directives encode the in-memory order of each use-list, allowing the
1820 order to be recreated. ``<order-indexes>`` is a comma-separated list of
1821 indexes that are assigned to the referenced value's uses. The referenced
1822 value's use-list is immediately sorted by these indexes.
1824 Use-list directives may appear at function scope or global scope. They are not
1825 instructions, and have no effect on the semantics of the IR. When they're at
1826 function scope, they must appear after the terminator of the final basic block.
1828 If basic blocks have their address taken via ``blockaddress()`` expressions,
1829 ``uselistorder_bb`` can be used to reorder their use-lists from outside their
1836 uselistorder <ty> <value>, { <order-indexes> }
1837 uselistorder_bb @function, %block { <order-indexes> }
1843 define void @foo(i32 %arg1, i32 %arg2) {
1845 ; ... instructions ...
1847 ; ... instructions ...
1849 ; At function scope.
1850 uselistorder i32 %arg1, { 1, 0, 2 }
1851 uselistorder label %bb, { 1, 0 }
1855 uselistorder i32* @global, { 1, 2, 0 }
1856 uselistorder i32 7, { 1, 0 }
1857 uselistorder i32 (i32) @bar, { 1, 0 }
1858 uselistorder_bb @foo, %bb, { 5, 1, 3, 2, 0, 4 }
1865 The LLVM type system is one of the most important features of the
1866 intermediate representation. Being typed enables a number of
1867 optimizations to be performed on the intermediate representation
1868 directly, without having to do extra analyses on the side before the
1869 transformation. A strong type system makes it easier to read the
1870 generated code and enables novel analyses and transformations that are
1871 not feasible to perform on normal three address code representations.
1881 The void type does not represent any value and has no size.
1899 The function type can be thought of as a function signature. It consists of a
1900 return type and a list of formal parameter types. The return type of a function
1901 type is a void type or first class type --- except for :ref:`label <t_label>`
1902 and :ref:`metadata <t_metadata>` types.
1908 <returntype> (<parameter list>)
1910 ...where '``<parameter list>``' is a comma-separated list of type
1911 specifiers. Optionally, the parameter list may include a type ``...``, which
1912 indicates that the function takes a variable number of arguments. Variable
1913 argument functions can access their arguments with the :ref:`variable argument
1914 handling intrinsic <int_varargs>` functions. '``<returntype>``' is any type
1915 except :ref:`label <t_label>` and :ref:`metadata <t_metadata>`.
1919 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1920 | ``i32 (i32)`` | function taking an ``i32``, returning an ``i32`` |
1921 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1922 | ``float (i16, i32 *) *`` | :ref:`Pointer <t_pointer>` to a function that takes an ``i16`` and a :ref:`pointer <t_pointer>` to ``i32``, returning ``float``. |
1923 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1924 | ``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. |
1925 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1926 | ``{i32, i32} (i32)`` | A function taking an ``i32``, returning a :ref:`structure <t_struct>` containing two ``i32`` values |
1927 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1934 The :ref:`first class <t_firstclass>` types are perhaps the most important.
1935 Values of these types are the only ones which can be produced by
1943 These are the types that are valid in registers from CodeGen's perspective.
1952 The integer type is a very simple type that simply specifies an
1953 arbitrary bit width for the integer type desired. Any bit width from 1
1954 bit to 2\ :sup:`23`\ -1 (about 8 million) can be specified.
1962 The number of bits the integer will occupy is specified by the ``N``
1968 +----------------+------------------------------------------------+
1969 | ``i1`` | a single-bit integer. |
1970 +----------------+------------------------------------------------+
1971 | ``i32`` | a 32-bit integer. |
1972 +----------------+------------------------------------------------+
1973 | ``i1942652`` | a really big integer of over 1 million bits. |
1974 +----------------+------------------------------------------------+
1978 Floating Point Types
1979 """"""""""""""""""""
1988 - 16-bit floating point value
1991 - 32-bit floating point value
1994 - 64-bit floating point value
1997 - 128-bit floating point value (112-bit mantissa)
2000 - 80-bit floating point value (X87)
2003 - 128-bit floating point value (two 64-bits)
2010 The x86_mmx type represents a value held in an MMX register on an x86
2011 machine. The operations allowed on it are quite limited: parameters and
2012 return values, load and store, and bitcast. User-specified MMX
2013 instructions are represented as intrinsic or asm calls with arguments
2014 and/or results of this type. There are no arrays, vectors or constants
2031 The pointer type is used to specify memory locations. Pointers are
2032 commonly used to reference objects in memory.
2034 Pointer types may have an optional address space attribute defining the
2035 numbered address space where the pointed-to object resides. The default
2036 address space is number zero. The semantics of non-zero address spaces
2037 are target-specific.
2039 Note that LLVM does not permit pointers to void (``void*``) nor does it
2040 permit pointers to labels (``label*``). Use ``i8*`` instead.
2050 +-------------------------+--------------------------------------------------------------------------------------------------------------+
2051 | ``[4 x i32]*`` | A :ref:`pointer <t_pointer>` to :ref:`array <t_array>` of four ``i32`` values. |
2052 +-------------------------+--------------------------------------------------------------------------------------------------------------+
2053 | ``i32 (i32*) *`` | A :ref:`pointer <t_pointer>` to a :ref:`function <t_function>` that takes an ``i32*``, returning an ``i32``. |
2054 +-------------------------+--------------------------------------------------------------------------------------------------------------+
2055 | ``i32 addrspace(5)*`` | A :ref:`pointer <t_pointer>` to an ``i32`` value that resides in address space #5. |
2056 +-------------------------+--------------------------------------------------------------------------------------------------------------+
2065 A vector type is a simple derived type that represents a vector of
2066 elements. Vector types are used when multiple primitive data are
2067 operated in parallel using a single instruction (SIMD). A vector type
2068 requires a size (number of elements) and an underlying primitive data
2069 type. Vector types are considered :ref:`first class <t_firstclass>`.
2075 < <# elements> x <elementtype> >
2077 The number of elements is a constant integer value larger than 0;
2078 elementtype may be any integer, floating point or pointer type. Vectors
2079 of size zero are not allowed.
2083 +-------------------+--------------------------------------------------+
2084 | ``<4 x i32>`` | Vector of 4 32-bit integer values. |
2085 +-------------------+--------------------------------------------------+
2086 | ``<8 x float>`` | Vector of 8 32-bit floating-point values. |
2087 +-------------------+--------------------------------------------------+
2088 | ``<2 x i64>`` | Vector of 2 64-bit integer values. |
2089 +-------------------+--------------------------------------------------+
2090 | ``<4 x i64*>`` | Vector of 4 pointers to 64-bit integer values. |
2091 +-------------------+--------------------------------------------------+
2100 The label type represents code labels.
2115 The metadata type represents embedded metadata. No derived types may be
2116 created from metadata except for :ref:`function <t_function>` arguments.
2129 Aggregate Types are a subset of derived types that can contain multiple
2130 member types. :ref:`Arrays <t_array>` and :ref:`structs <t_struct>` are
2131 aggregate types. :ref:`Vectors <t_vector>` are not considered to be
2141 The array type is a very simple derived type that arranges elements
2142 sequentially in memory. The array type requires a size (number of
2143 elements) and an underlying data type.
2149 [<# elements> x <elementtype>]
2151 The number of elements is a constant integer value; ``elementtype`` may
2152 be any type with a size.
2156 +------------------+--------------------------------------+
2157 | ``[40 x i32]`` | Array of 40 32-bit integer values. |
2158 +------------------+--------------------------------------+
2159 | ``[41 x i32]`` | Array of 41 32-bit integer values. |
2160 +------------------+--------------------------------------+
2161 | ``[4 x i8]`` | Array of 4 8-bit integer values. |
2162 +------------------+--------------------------------------+
2164 Here are some examples of multidimensional arrays:
2166 +-----------------------------+----------------------------------------------------------+
2167 | ``[3 x [4 x i32]]`` | 3x4 array of 32-bit integer values. |
2168 +-----------------------------+----------------------------------------------------------+
2169 | ``[12 x [10 x float]]`` | 12x10 array of single precision floating point values. |
2170 +-----------------------------+----------------------------------------------------------+
2171 | ``[2 x [3 x [4 x i16]]]`` | 2x3x4 array of 16-bit integer values. |
2172 +-----------------------------+----------------------------------------------------------+
2174 There is no restriction on indexing beyond the end of the array implied
2175 by a static type (though there are restrictions on indexing beyond the
2176 bounds of an allocated object in some cases). This means that
2177 single-dimension 'variable sized array' addressing can be implemented in
2178 LLVM with a zero length array type. An implementation of 'pascal style
2179 arrays' in LLVM could use the type "``{ i32, [0 x float]}``", for
2189 The structure type is used to represent a collection of data members
2190 together in memory. The elements of a structure may be any type that has
2193 Structures in memory are accessed using '``load``' and '``store``' by
2194 getting a pointer to a field with the '``getelementptr``' instruction.
2195 Structures in registers are accessed using the '``extractvalue``' and
2196 '``insertvalue``' instructions.
2198 Structures may optionally be "packed" structures, which indicate that
2199 the alignment of the struct is one byte, and that there is no padding
2200 between the elements. In non-packed structs, padding between field types
2201 is inserted as defined by the DataLayout string in the module, which is
2202 required to match what the underlying code generator expects.
2204 Structures can either be "literal" or "identified". A literal structure
2205 is defined inline with other types (e.g. ``{i32, i32}*``) whereas
2206 identified types are always defined at the top level with a name.
2207 Literal types are uniqued by their contents and can never be recursive
2208 or opaque since there is no way to write one. Identified types can be
2209 recursive, can be opaqued, and are never uniqued.
2215 %T1 = type { <type list> } ; Identified normal struct type
2216 %T2 = type <{ <type list> }> ; Identified packed struct type
2220 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2221 | ``{ i32, i32, i32 }`` | A triple of three ``i32`` values |
2222 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2223 | ``{ 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``. |
2224 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2225 | ``<{ i8, i32 }>`` | A packed struct known to be 5 bytes in size. |
2226 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2230 Opaque Structure Types
2231 """"""""""""""""""""""
2235 Opaque structure types are used to represent named structure types that
2236 do not have a body specified. This corresponds (for example) to the C
2237 notion of a forward declared structure.
2248 +--------------+-------------------+
2249 | ``opaque`` | An opaque type. |
2250 +--------------+-------------------+
2257 LLVM has several different basic types of constants. This section
2258 describes them all and their syntax.
2263 **Boolean constants**
2264 The two strings '``true``' and '``false``' are both valid constants
2266 **Integer constants**
2267 Standard integers (such as '4') are constants of the
2268 :ref:`integer <t_integer>` type. Negative numbers may be used with
2270 **Floating point constants**
2271 Floating point constants use standard decimal notation (e.g.
2272 123.421), exponential notation (e.g. 1.23421e+2), or a more precise
2273 hexadecimal notation (see below). The assembler requires the exact
2274 decimal value of a floating-point constant. For example, the
2275 assembler accepts 1.25 but rejects 1.3 because 1.3 is a repeating
2276 decimal in binary. Floating point constants must have a :ref:`floating
2277 point <t_floating>` type.
2278 **Null pointer constants**
2279 The identifier '``null``' is recognized as a null pointer constant
2280 and must be of :ref:`pointer type <t_pointer>`.
2282 The one non-intuitive notation for constants is the hexadecimal form of
2283 floating point constants. For example, the form
2284 '``double 0x432ff973cafa8000``' is equivalent to (but harder to read
2285 than) '``double 4.5e+15``'. The only time hexadecimal floating point
2286 constants are required (and the only time that they are generated by the
2287 disassembler) is when a floating point constant must be emitted but it
2288 cannot be represented as a decimal floating point number in a reasonable
2289 number of digits. For example, NaN's, infinities, and other special
2290 values are represented in their IEEE hexadecimal format so that assembly
2291 and disassembly do not cause any bits to change in the constants.
2293 When using the hexadecimal form, constants of types half, float, and
2294 double are represented using the 16-digit form shown above (which
2295 matches the IEEE754 representation for double); half and float values
2296 must, however, be exactly representable as IEEE 754 half and single
2297 precision, respectively. Hexadecimal format is always used for long
2298 double, and there are three forms of long double. The 80-bit format used
2299 by x86 is represented as ``0xK`` followed by 20 hexadecimal digits. The
2300 128-bit format used by PowerPC (two adjacent doubles) is represented by
2301 ``0xM`` followed by 32 hexadecimal digits. The IEEE 128-bit format is
2302 represented by ``0xL`` followed by 32 hexadecimal digits. Long doubles
2303 will only work if they match the long double format on your target.
2304 The IEEE 16-bit format (half precision) is represented by ``0xH``
2305 followed by 4 hexadecimal digits. All hexadecimal formats are big-endian
2306 (sign bit at the left).
2308 There are no constants of type x86_mmx.
2310 .. _complexconstants:
2315 Complex constants are a (potentially recursive) combination of simple
2316 constants and smaller complex constants.
2318 **Structure constants**
2319 Structure constants are represented with notation similar to
2320 structure type definitions (a comma separated list of elements,
2321 surrounded by braces (``{}``)). For example:
2322 "``{ i32 4, float 17.0, i32* @G }``", where "``@G``" is declared as
2323 "``@G = external global i32``". Structure constants must have
2324 :ref:`structure type <t_struct>`, and the number and types of elements
2325 must match those specified by the type.
2327 Array constants are represented with notation similar to array type
2328 definitions (a comma separated list of elements, surrounded by
2329 square brackets (``[]``)). For example:
2330 "``[ i32 42, i32 11, i32 74 ]``". Array constants must have
2331 :ref:`array type <t_array>`, and the number and types of elements must
2332 match those specified by the type. As a special case, character array
2333 constants may also be represented as a double-quoted string using the ``c``
2334 prefix. For example: "``c"Hello World\0A\00"``".
2335 **Vector constants**
2336 Vector constants are represented with notation similar to vector
2337 type definitions (a comma separated list of elements, surrounded by
2338 less-than/greater-than's (``<>``)). For example:
2339 "``< i32 42, i32 11, i32 74, i32 100 >``". Vector constants
2340 must have :ref:`vector type <t_vector>`, and the number and types of
2341 elements must match those specified by the type.
2342 **Zero initialization**
2343 The string '``zeroinitializer``' can be used to zero initialize a
2344 value to zero of *any* type, including scalar and
2345 :ref:`aggregate <t_aggregate>` types. This is often used to avoid
2346 having to print large zero initializers (e.g. for large arrays) and
2347 is always exactly equivalent to using explicit zero initializers.
2349 A metadata node is a constant tuple without types. For example:
2350 "``!{!0, !{!2, !0}, !"test"}``". Metadata can reference constant values,
2351 for example: "``!{!0, i32 0, i8* @global, i64 (i64)* @function, !"str"}``".
2352 Unlike other typed constants that are meant to be interpreted as part of
2353 the instruction stream, metadata is a place to attach additional
2354 information such as debug info.
2356 Global Variable and Function Addresses
2357 --------------------------------------
2359 The addresses of :ref:`global variables <globalvars>` and
2360 :ref:`functions <functionstructure>` are always implicitly valid
2361 (link-time) constants. These constants are explicitly referenced when
2362 the :ref:`identifier for the global <identifiers>` is used and always have
2363 :ref:`pointer <t_pointer>` type. For example, the following is a legal LLVM
2366 .. code-block:: llvm
2370 @Z = global [2 x i32*] [ i32* @X, i32* @Y ]
2377 The string '``undef``' can be used anywhere a constant is expected, and
2378 indicates that the user of the value may receive an unspecified
2379 bit-pattern. Undefined values may be of any type (other than '``label``'
2380 or '``void``') and be used anywhere a constant is permitted.
2382 Undefined values are useful because they indicate to the compiler that
2383 the program is well defined no matter what value is used. This gives the
2384 compiler more freedom to optimize. Here are some examples of
2385 (potentially surprising) transformations that are valid (in pseudo IR):
2387 .. code-block:: llvm
2397 This is safe because all of the output bits are affected by the undef
2398 bits. Any output bit can have a zero or one depending on the input bits.
2400 .. code-block:: llvm
2411 These logical operations have bits that are not always affected by the
2412 input. For example, if ``%X`` has a zero bit, then the output of the
2413 '``and``' operation will always be a zero for that bit, no matter what
2414 the corresponding bit from the '``undef``' is. As such, it is unsafe to
2415 optimize or assume that the result of the '``and``' is '``undef``'.
2416 However, it is safe to assume that all bits of the '``undef``' could be
2417 0, and optimize the '``and``' to 0. Likewise, it is safe to assume that
2418 all the bits of the '``undef``' operand to the '``or``' could be set,
2419 allowing the '``or``' to be folded to -1.
2421 .. code-block:: llvm
2423 %A = select undef, %X, %Y
2424 %B = select undef, 42, %Y
2425 %C = select %X, %Y, undef
2435 This set of examples shows that undefined '``select``' (and conditional
2436 branch) conditions can go *either way*, but they have to come from one
2437 of the two operands. In the ``%A`` example, if ``%X`` and ``%Y`` were
2438 both known to have a clear low bit, then ``%A`` would have to have a
2439 cleared low bit. However, in the ``%C`` example, the optimizer is
2440 allowed to assume that the '``undef``' operand could be the same as
2441 ``%Y``, allowing the whole '``select``' to be eliminated.
2443 .. code-block:: llvm
2445 %A = xor undef, undef
2462 This example points out that two '``undef``' operands are not
2463 necessarily the same. This can be surprising to people (and also matches
2464 C semantics) where they assume that "``X^X``" is always zero, even if
2465 ``X`` is undefined. This isn't true for a number of reasons, but the
2466 short answer is that an '``undef``' "variable" can arbitrarily change
2467 its value over its "live range". This is true because the variable
2468 doesn't actually *have a live range*. Instead, the value is logically
2469 read from arbitrary registers that happen to be around when needed, so
2470 the value is not necessarily consistent over time. In fact, ``%A`` and
2471 ``%C`` need to have the same semantics or the core LLVM "replace all
2472 uses with" concept would not hold.
2474 .. code-block:: llvm
2482 These examples show the crucial difference between an *undefined value*
2483 and *undefined behavior*. An undefined value (like '``undef``') is
2484 allowed to have an arbitrary bit-pattern. This means that the ``%A``
2485 operation can be constant folded to '``undef``', because the '``undef``'
2486 could be an SNaN, and ``fdiv`` is not (currently) defined on SNaN's.
2487 However, in the second example, we can make a more aggressive
2488 assumption: because the ``undef`` is allowed to be an arbitrary value,
2489 we are allowed to assume that it could be zero. Since a divide by zero
2490 has *undefined behavior*, we are allowed to assume that the operation
2491 does not execute at all. This allows us to delete the divide and all
2492 code after it. Because the undefined operation "can't happen", the
2493 optimizer can assume that it occurs in dead code.
2495 .. code-block:: llvm
2497 a: store undef -> %X
2498 b: store %X -> undef
2503 These examples reiterate the ``fdiv`` example: a store *of* an undefined
2504 value can be assumed to not have any effect; we can assume that the
2505 value is overwritten with bits that happen to match what was already
2506 there. However, a store *to* an undefined location could clobber
2507 arbitrary memory, therefore, it has undefined behavior.
2514 Poison values are similar to :ref:`undef values <undefvalues>`, however
2515 they also represent the fact that an instruction or constant expression
2516 that cannot evoke side effects has nevertheless detected a condition
2517 that results in undefined behavior.
2519 There is currently no way of representing a poison value in the IR; they
2520 only exist when produced by operations such as :ref:`add <i_add>` with
2523 Poison value behavior is defined in terms of value *dependence*:
2525 - Values other than :ref:`phi <i_phi>` nodes depend on their operands.
2526 - :ref:`Phi <i_phi>` nodes depend on the operand corresponding to
2527 their dynamic predecessor basic block.
2528 - Function arguments depend on the corresponding actual argument values
2529 in the dynamic callers of their functions.
2530 - :ref:`Call <i_call>` instructions depend on the :ref:`ret <i_ret>`
2531 instructions that dynamically transfer control back to them.
2532 - :ref:`Invoke <i_invoke>` instructions depend on the
2533 :ref:`ret <i_ret>`, :ref:`resume <i_resume>`, or exception-throwing
2534 call instructions that dynamically transfer control back to them.
2535 - Non-volatile loads and stores depend on the most recent stores to all
2536 of the referenced memory addresses, following the order in the IR
2537 (including loads and stores implied by intrinsics such as
2538 :ref:`@llvm.memcpy <int_memcpy>`.)
2539 - An instruction with externally visible side effects depends on the
2540 most recent preceding instruction with externally visible side
2541 effects, following the order in the IR. (This includes :ref:`volatile
2542 operations <volatile>`.)
2543 - An instruction *control-depends* on a :ref:`terminator
2544 instruction <terminators>` if the terminator instruction has
2545 multiple successors and the instruction is always executed when
2546 control transfers to one of the successors, and may not be executed
2547 when control is transferred to another.
2548 - Additionally, an instruction also *control-depends* on a terminator
2549 instruction if the set of instructions it otherwise depends on would
2550 be different if the terminator had transferred control to a different
2552 - Dependence is transitive.
2554 Poison values have the same behavior as :ref:`undef values <undefvalues>`,
2555 with the additional effect that any instruction that has a *dependence*
2556 on a poison value has undefined behavior.
2558 Here are some examples:
2560 .. code-block:: llvm
2563 %poison = sub nuw i32 0, 1 ; Results in a poison value.
2564 %still_poison = and i32 %poison, 0 ; 0, but also poison.
2565 %poison_yet_again = getelementptr i32* @h, i32 %still_poison
2566 store i32 0, i32* %poison_yet_again ; memory at @h[0] is poisoned
2568 store i32 %poison, i32* @g ; Poison value stored to memory.
2569 %poison2 = load i32* @g ; Poison value loaded back from memory.
2571 store volatile i32 %poison, i32* @g ; External observation; undefined behavior.
2573 %narrowaddr = bitcast i32* @g to i16*
2574 %wideaddr = bitcast i32* @g to i64*
2575 %poison3 = load i16* %narrowaddr ; Returns a poison value.
2576 %poison4 = load i64* %wideaddr ; Returns a poison value.
2578 %cmp = icmp slt i32 %poison, 0 ; Returns a poison value.
2579 br i1 %cmp, label %true, label %end ; Branch to either destination.
2582 store volatile i32 0, i32* @g ; This is control-dependent on %cmp, so
2583 ; it has undefined behavior.
2587 %p = phi i32 [ 0, %entry ], [ 1, %true ]
2588 ; Both edges into this PHI are
2589 ; control-dependent on %cmp, so this
2590 ; always results in a poison value.
2592 store volatile i32 0, i32* @g ; This would depend on the store in %true
2593 ; if %cmp is true, or the store in %entry
2594 ; otherwise, so this is undefined behavior.
2596 br i1 %cmp, label %second_true, label %second_end
2597 ; The same branch again, but this time the
2598 ; true block doesn't have side effects.
2605 store volatile i32 0, i32* @g ; This time, the instruction always depends
2606 ; on the store in %end. Also, it is
2607 ; control-equivalent to %end, so this is
2608 ; well-defined (ignoring earlier undefined
2609 ; behavior in this example).
2613 Addresses of Basic Blocks
2614 -------------------------
2616 ``blockaddress(@function, %block)``
2618 The '``blockaddress``' constant computes the address of the specified
2619 basic block in the specified function, and always has an ``i8*`` type.
2620 Taking the address of the entry block is illegal.
2622 This value only has defined behavior when used as an operand to the
2623 ':ref:`indirectbr <i_indirectbr>`' instruction, or for comparisons
2624 against null. Pointer equality tests between labels addresses results in
2625 undefined behavior --- though, again, comparison against null is ok, and
2626 no label is equal to the null pointer. This may be passed around as an
2627 opaque pointer sized value as long as the bits are not inspected. This
2628 allows ``ptrtoint`` and arithmetic to be performed on these values so
2629 long as the original value is reconstituted before the ``indirectbr``
2632 Finally, some targets may provide defined semantics when using the value
2633 as the operand to an inline assembly, but that is target specific.
2637 Constant Expressions
2638 --------------------
2640 Constant expressions are used to allow expressions involving other
2641 constants to be used as constants. Constant expressions may be of any
2642 :ref:`first class <t_firstclass>` type and may involve any LLVM operation
2643 that does not have side effects (e.g. load and call are not supported).
2644 The following is the syntax for constant expressions:
2646 ``trunc (CST to TYPE)``
2647 Truncate a constant to another type. The bit size of CST must be
2648 larger than the bit size of TYPE. Both types must be integers.
2649 ``zext (CST to TYPE)``
2650 Zero extend a constant to another type. The bit size of CST must be
2651 smaller than the bit size of TYPE. Both types must be integers.
2652 ``sext (CST to TYPE)``
2653 Sign extend a constant to another type. The bit size of CST must be
2654 smaller than the bit size of TYPE. Both types must be integers.
2655 ``fptrunc (CST to TYPE)``
2656 Truncate a floating point constant to another floating point type.
2657 The size of CST must be larger than the size of TYPE. Both types
2658 must be floating point.
2659 ``fpext (CST to TYPE)``
2660 Floating point extend a constant to another type. The size of CST
2661 must be smaller or equal to the size of TYPE. Both types must be
2663 ``fptoui (CST to TYPE)``
2664 Convert a floating point constant to the corresponding unsigned
2665 integer constant. TYPE must be a scalar or vector integer type. CST
2666 must be of scalar or vector floating point type. Both CST and TYPE
2667 must be scalars, or vectors of the same number of elements. If the
2668 value won't fit in the integer type, the results are undefined.
2669 ``fptosi (CST to TYPE)``
2670 Convert a floating point constant to the corresponding signed
2671 integer constant. TYPE must be a scalar or vector integer type. CST
2672 must be of scalar or vector floating point type. Both CST and TYPE
2673 must be scalars, or vectors of the same number of elements. If the
2674 value won't fit in the integer type, the results are undefined.
2675 ``uitofp (CST to TYPE)``
2676 Convert an unsigned integer constant to the corresponding floating
2677 point constant. TYPE must be a scalar or vector floating point type.
2678 CST must be of scalar or vector integer type. Both CST and TYPE must
2679 be scalars, or vectors of the same number of elements. If the value
2680 won't fit in the floating point type, the results are undefined.
2681 ``sitofp (CST to TYPE)``
2682 Convert a signed integer constant to the corresponding floating
2683 point constant. TYPE must be a scalar or vector floating point type.
2684 CST must be of scalar or vector integer type. Both CST and TYPE must
2685 be scalars, or vectors of the same number of elements. If the value
2686 won't fit in the floating point type, the results are undefined.
2687 ``ptrtoint (CST to TYPE)``
2688 Convert a pointer typed constant to the corresponding integer
2689 constant. ``TYPE`` must be an integer type. ``CST`` must be of
2690 pointer type. The ``CST`` value is zero extended, truncated, or
2691 unchanged to make it fit in ``TYPE``.
2692 ``inttoptr (CST to TYPE)``
2693 Convert an integer constant to a pointer constant. TYPE must be a
2694 pointer type. CST must be of integer type. The CST value is zero
2695 extended, truncated, or unchanged to make it fit in a pointer size.
2696 This one is *really* dangerous!
2697 ``bitcast (CST to TYPE)``
2698 Convert a constant, CST, to another TYPE. The constraints of the
2699 operands are the same as those for the :ref:`bitcast
2700 instruction <i_bitcast>`.
2701 ``addrspacecast (CST to TYPE)``
2702 Convert a constant pointer or constant vector of pointer, CST, to another
2703 TYPE in a different address space. The constraints of the operands are the
2704 same as those for the :ref:`addrspacecast instruction <i_addrspacecast>`.
2705 ``getelementptr (CSTPTR, IDX0, IDX1, ...)``, ``getelementptr inbounds (CSTPTR, IDX0, IDX1, ...)``
2706 Perform the :ref:`getelementptr operation <i_getelementptr>` on
2707 constants. As with the :ref:`getelementptr <i_getelementptr>`
2708 instruction, the index list may have zero or more indexes, which are
2709 required to make sense for the type of "CSTPTR".
2710 ``select (COND, VAL1, VAL2)``
2711 Perform the :ref:`select operation <i_select>` on constants.
2712 ``icmp COND (VAL1, VAL2)``
2713 Performs the :ref:`icmp operation <i_icmp>` on constants.
2714 ``fcmp COND (VAL1, VAL2)``
2715 Performs the :ref:`fcmp operation <i_fcmp>` on constants.
2716 ``extractelement (VAL, IDX)``
2717 Perform the :ref:`extractelement operation <i_extractelement>` on
2719 ``insertelement (VAL, ELT, IDX)``
2720 Perform the :ref:`insertelement operation <i_insertelement>` on
2722 ``shufflevector (VEC1, VEC2, IDXMASK)``
2723 Perform the :ref:`shufflevector operation <i_shufflevector>` on
2725 ``extractvalue (VAL, IDX0, IDX1, ...)``
2726 Perform the :ref:`extractvalue operation <i_extractvalue>` on
2727 constants. The index list is interpreted in a similar manner as
2728 indices in a ':ref:`getelementptr <i_getelementptr>`' operation. At
2729 least one index value must be specified.
2730 ``insertvalue (VAL, ELT, IDX0, IDX1, ...)``
2731 Perform the :ref:`insertvalue operation <i_insertvalue>` on constants.
2732 The index list is interpreted in a similar manner as indices in a
2733 ':ref:`getelementptr <i_getelementptr>`' operation. At least one index
2734 value must be specified.
2735 ``OPCODE (LHS, RHS)``
2736 Perform the specified operation of the LHS and RHS constants. OPCODE
2737 may be any of the :ref:`binary <binaryops>` or :ref:`bitwise
2738 binary <bitwiseops>` operations. The constraints on operands are
2739 the same as those for the corresponding instruction (e.g. no bitwise
2740 operations on floating point values are allowed).
2747 Inline Assembler Expressions
2748 ----------------------------
2750 LLVM supports inline assembler expressions (as opposed to :ref:`Module-Level
2751 Inline Assembly <moduleasm>`) through the use of a special value. This
2752 value represents the inline assembler as a string (containing the
2753 instructions to emit), a list of operand constraints (stored as a
2754 string), a flag that indicates whether or not the inline asm expression
2755 has side effects, and a flag indicating whether the function containing
2756 the asm needs to align its stack conservatively. An example inline
2757 assembler expression is:
2759 .. code-block:: llvm
2761 i32 (i32) asm "bswap $0", "=r,r"
2763 Inline assembler expressions may **only** be used as the callee operand
2764 of a :ref:`call <i_call>` or an :ref:`invoke <i_invoke>` instruction.
2765 Thus, typically we have:
2767 .. code-block:: llvm
2769 %X = call i32 asm "bswap $0", "=r,r"(i32 %Y)
2771 Inline asms with side effects not visible in the constraint list must be
2772 marked as having side effects. This is done through the use of the
2773 '``sideeffect``' keyword, like so:
2775 .. code-block:: llvm
2777 call void asm sideeffect "eieio", ""()
2779 In some cases inline asms will contain code that will not work unless
2780 the stack is aligned in some way, such as calls or SSE instructions on
2781 x86, yet will not contain code that does that alignment within the asm.
2782 The compiler should make conservative assumptions about what the asm
2783 might contain and should generate its usual stack alignment code in the
2784 prologue if the '``alignstack``' keyword is present:
2786 .. code-block:: llvm
2788 call void asm alignstack "eieio", ""()
2790 Inline asms also support using non-standard assembly dialects. The
2791 assumed dialect is ATT. When the '``inteldialect``' keyword is present,
2792 the inline asm is using the Intel dialect. Currently, ATT and Intel are
2793 the only supported dialects. An example is:
2795 .. code-block:: llvm
2797 call void asm inteldialect "eieio", ""()
2799 If multiple keywords appear the '``sideeffect``' keyword must come
2800 first, the '``alignstack``' keyword second and the '``inteldialect``'
2806 The call instructions that wrap inline asm nodes may have a
2807 "``!srcloc``" MDNode attached to it that contains a list of constant
2808 integers. If present, the code generator will use the integer as the
2809 location cookie value when report errors through the ``LLVMContext``
2810 error reporting mechanisms. This allows a front-end to correlate backend
2811 errors that occur with inline asm back to the source code that produced
2814 .. code-block:: llvm
2816 call void asm sideeffect "something bad", ""(), !srcloc !42
2818 !42 = !{ i32 1234567 }
2820 It is up to the front-end to make sense of the magic numbers it places
2821 in the IR. If the MDNode contains multiple constants, the code generator
2822 will use the one that corresponds to the line of the asm that the error
2830 LLVM IR allows metadata to be attached to instructions in the program
2831 that can convey extra information about the code to the optimizers and
2832 code generator. One example application of metadata is source-level
2833 debug information. There are two metadata primitives: strings and nodes.
2835 Metadata does not have a type, and is not a value. If referenced from a
2836 ``call`` instruction, it uses the ``metadata`` type.
2838 All metadata are identified in syntax by a exclamation point ('``!``').
2840 Metadata Nodes and Metadata Strings
2841 -----------------------------------
2843 A metadata string is a string surrounded by double quotes. It can
2844 contain any character by escaping non-printable characters with
2845 "``\xx``" where "``xx``" is the two digit hex code. For example:
2848 Metadata nodes are represented with notation similar to structure
2849 constants (a comma separated list of elements, surrounded by braces and
2850 preceded by an exclamation point). Metadata nodes can have any values as
2851 their operand. For example:
2853 .. code-block:: llvm
2855 !{ !"test\00", i32 10}
2857 Metadata nodes that aren't uniqued use the ``distinct`` keyword. For example:
2859 .. code-block:: llvm
2861 !0 = distinct !{!"test\00", i32 10}
2863 ``distinct`` nodes are useful when nodes shouldn't be merged based on their
2864 content. They can also occur when transformations cause uniquing collisions
2865 when metadata operands change.
2867 A :ref:`named metadata <namedmetadatastructure>` is a collection of
2868 metadata nodes, which can be looked up in the module symbol table. For
2871 .. code-block:: llvm
2875 Metadata can be used as function arguments. Here ``llvm.dbg.value``
2876 function is using two metadata arguments:
2878 .. code-block:: llvm
2880 call void @llvm.dbg.value(metadata !24, i64 0, metadata !25)
2882 Metadata can be attached with an instruction. Here metadata ``!21`` is
2883 attached to the ``add`` instruction using the ``!dbg`` identifier:
2885 .. code-block:: llvm
2887 %indvar.next = add i64 %indvar, 1, !dbg !21
2889 More information about specific metadata nodes recognized by the
2890 optimizers and code generator is found below.
2892 Specialized Metadata Nodes
2893 ^^^^^^^^^^^^^^^^^^^^^^^^^^
2895 Specialized metadata nodes are custom data structures in metadata (as opposed
2896 to generic tuples). Their fields are labelled, and can be specified in any
2902 ``MDLocation`` nodes represent source debug locations. The ``scope:`` field is
2905 .. code-block:: llvm
2907 !0 = !MDLocation(line: 2900, column: 42, scope: !1, inlinedAt: !2)
2912 In LLVM IR, memory does not have types, so LLVM's own type system is not
2913 suitable for doing TBAA. Instead, metadata is added to the IR to
2914 describe a type system of a higher level language. This can be used to
2915 implement typical C/C++ TBAA, but it can also be used to implement
2916 custom alias analysis behavior for other languages.
2918 The current metadata format is very simple. TBAA metadata nodes have up
2919 to three fields, e.g.:
2921 .. code-block:: llvm
2923 !0 = !{ !"an example type tree" }
2924 !1 = !{ !"int", !0 }
2925 !2 = !{ !"float", !0 }
2926 !3 = !{ !"const float", !2, i64 1 }
2928 The first field is an identity field. It can be any value, usually a
2929 metadata string, which uniquely identifies the type. The most important
2930 name in the tree is the name of the root node. Two trees with different
2931 root node names are entirely disjoint, even if they have leaves with
2934 The second field identifies the type's parent node in the tree, or is
2935 null or omitted for a root node. A type is considered to alias all of
2936 its descendants and all of its ancestors in the tree. Also, a type is
2937 considered to alias all types in other trees, so that bitcode produced
2938 from multiple front-ends is handled conservatively.
2940 If the third field is present, it's an integer which if equal to 1
2941 indicates that the type is "constant" (meaning
2942 ``pointsToConstantMemory`` should return true; see `other useful
2943 AliasAnalysis methods <AliasAnalysis.html#OtherItfs>`_).
2945 '``tbaa.struct``' Metadata
2946 ^^^^^^^^^^^^^^^^^^^^^^^^^^
2948 The :ref:`llvm.memcpy <int_memcpy>` is often used to implement
2949 aggregate assignment operations in C and similar languages, however it
2950 is defined to copy a contiguous region of memory, which is more than
2951 strictly necessary for aggregate types which contain holes due to
2952 padding. Also, it doesn't contain any TBAA information about the fields
2955 ``!tbaa.struct`` metadata can describe which memory subregions in a
2956 memcpy are padding and what the TBAA tags of the struct are.
2958 The current metadata format is very simple. ``!tbaa.struct`` metadata
2959 nodes are a list of operands which are in conceptual groups of three.
2960 For each group of three, the first operand gives the byte offset of a
2961 field in bytes, the second gives its size in bytes, and the third gives
2964 .. code-block:: llvm
2966 !4 = !{ i64 0, i64 4, !1, i64 8, i64 4, !2 }
2968 This describes a struct with two fields. The first is at offset 0 bytes
2969 with size 4 bytes, and has tbaa tag !1. The second is at offset 8 bytes
2970 and has size 4 bytes and has tbaa tag !2.
2972 Note that the fields need not be contiguous. In this example, there is a
2973 4 byte gap between the two fields. This gap represents padding which
2974 does not carry useful data and need not be preserved.
2976 '``noalias``' and '``alias.scope``' Metadata
2977 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2979 ``noalias`` and ``alias.scope`` metadata provide the ability to specify generic
2980 noalias memory-access sets. This means that some collection of memory access
2981 instructions (loads, stores, memory-accessing calls, etc.) that carry
2982 ``noalias`` metadata can specifically be specified not to alias with some other
2983 collection of memory access instructions that carry ``alias.scope`` metadata.
2984 Each type of metadata specifies a list of scopes where each scope has an id and
2985 a domain. When evaluating an aliasing query, if for some some domain, the set
2986 of scopes with that domain in one instruction's ``alias.scope`` list is a
2987 subset of (or equal to) the set of scopes for that domain in another
2988 instruction's ``noalias`` list, then the two memory accesses are assumed not to
2991 The metadata identifying each domain is itself a list containing one or two
2992 entries. The first entry is the name of the domain. Note that if the name is a
2993 string then it can be combined accross functions and translation units. A
2994 self-reference can be used to create globally unique domain names. A
2995 descriptive string may optionally be provided as a second list entry.
2997 The metadata identifying each scope is also itself a list containing two or
2998 three entries. The first entry is the name of the scope. Note that if the name
2999 is a string then it can be combined accross functions and translation units. A
3000 self-reference can be used to create globally unique scope names. A metadata
3001 reference to the scope's domain is the second entry. A descriptive string may
3002 optionally be provided as a third list entry.
3006 .. code-block:: llvm
3008 ; Two scope domains:
3012 ; Some scopes in these domains:
3018 !5 = !{!4} ; A list containing only scope !4
3022 ; These two instructions don't alias:
3023 %0 = load float* %c, align 4, !alias.scope !5
3024 store float %0, float* %arrayidx.i, align 4, !noalias !5
3026 ; These two instructions also don't alias (for domain !1, the set of scopes
3027 ; in the !alias.scope equals that in the !noalias list):
3028 %2 = load float* %c, align 4, !alias.scope !5
3029 store float %2, float* %arrayidx.i2, align 4, !noalias !6
3031 ; These two instructions don't alias (for domain !0, the set of scopes in
3032 ; the !noalias list is not a superset of, or equal to, the scopes in the
3033 ; !alias.scope list):
3034 %2 = load float* %c, align 4, !alias.scope !6
3035 store float %0, float* %arrayidx.i, align 4, !noalias !7
3037 '``fpmath``' Metadata
3038 ^^^^^^^^^^^^^^^^^^^^^
3040 ``fpmath`` metadata may be attached to any instruction of floating point
3041 type. It can be used to express the maximum acceptable error in the
3042 result of that instruction, in ULPs, thus potentially allowing the
3043 compiler to use a more efficient but less accurate method of computing
3044 it. ULP is defined as follows:
3046 If ``x`` is a real number that lies between two finite consecutive
3047 floating-point numbers ``a`` and ``b``, without being equal to one
3048 of them, then ``ulp(x) = |b - a|``, otherwise ``ulp(x)`` is the
3049 distance between the two non-equal finite floating-point numbers
3050 nearest ``x``. Moreover, ``ulp(NaN)`` is ``NaN``.
3052 The metadata node shall consist of a single positive floating point
3053 number representing the maximum relative error, for example:
3055 .. code-block:: llvm
3057 !0 = !{ float 2.5 } ; maximum acceptable inaccuracy is 2.5 ULPs
3061 '``range``' Metadata
3062 ^^^^^^^^^^^^^^^^^^^^
3064 ``range`` metadata may be attached only to ``load``, ``call`` and ``invoke`` of
3065 integer types. It expresses the possible ranges the loaded value or the value
3066 returned by the called function at this call site is in. The ranges are
3067 represented with a flattened list of integers. The loaded value or the value
3068 returned is known to be in the union of the ranges defined by each consecutive
3069 pair. Each pair has the following properties:
3071 - The type must match the type loaded by the instruction.
3072 - The pair ``a,b`` represents the range ``[a,b)``.
3073 - Both ``a`` and ``b`` are constants.
3074 - The range is allowed to wrap.
3075 - The range should not represent the full or empty set. That is,
3078 In addition, the pairs must be in signed order of the lower bound and
3079 they must be non-contiguous.
3083 .. code-block:: llvm
3085 %a = load i8* %x, align 1, !range !0 ; Can only be 0 or 1
3086 %b = load i8* %y, align 1, !range !1 ; Can only be 255 (-1), 0 or 1
3087 %c = call i8 @foo(), !range !2 ; Can only be 0, 1, 3, 4 or 5
3088 %d = invoke i8 @bar() to label %cont
3089 unwind label %lpad, !range !3 ; Can only be -2, -1, 3, 4 or 5
3091 !0 = !{ i8 0, i8 2 }
3092 !1 = !{ i8 255, i8 2 }
3093 !2 = !{ i8 0, i8 2, i8 3, i8 6 }
3094 !3 = !{ i8 -2, i8 0, i8 3, i8 6 }
3099 It is sometimes useful to attach information to loop constructs. Currently,
3100 loop metadata is implemented as metadata attached to the branch instruction
3101 in the loop latch block. This type of metadata refer to a metadata node that is
3102 guaranteed to be separate for each loop. The loop identifier metadata is
3103 specified with the name ``llvm.loop``.
3105 The loop identifier metadata is implemented using a metadata that refers to
3106 itself to avoid merging it with any other identifier metadata, e.g.,
3107 during module linkage or function inlining. That is, each loop should refer
3108 to their own identification metadata even if they reside in separate functions.
3109 The following example contains loop identifier metadata for two separate loop
3112 .. code-block:: llvm
3117 The loop identifier metadata can be used to specify additional
3118 per-loop metadata. Any operands after the first operand can be treated
3119 as user-defined metadata. For example the ``llvm.loop.unroll.count``
3120 suggests an unroll factor to the loop unroller:
3122 .. code-block:: llvm
3124 br i1 %exitcond, label %._crit_edge, label %.lr.ph, !llvm.loop !0
3127 !1 = !{!"llvm.loop.unroll.count", i32 4}
3129 '``llvm.loop.vectorize``' and '``llvm.loop.interleave``'
3130 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3132 Metadata prefixed with ``llvm.loop.vectorize`` or ``llvm.loop.interleave`` are
3133 used to control per-loop vectorization and interleaving parameters such as
3134 vectorization width and interleave count. These metadata should be used in
3135 conjunction with ``llvm.loop`` loop identification metadata. The
3136 ``llvm.loop.vectorize`` and ``llvm.loop.interleave`` metadata are only
3137 optimization hints and the optimizer will only interleave and vectorize loops if
3138 it believes it is safe to do so. The ``llvm.mem.parallel_loop_access`` metadata
3139 which contains information about loop-carried memory dependencies can be helpful
3140 in determining the safety of these transformations.
3142 '``llvm.loop.interleave.count``' Metadata
3143 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3145 This metadata suggests an interleave count to the loop interleaver.
3146 The first operand is the string ``llvm.loop.interleave.count`` and the
3147 second operand is an integer specifying the interleave count. For
3150 .. code-block:: llvm
3152 !0 = !{!"llvm.loop.interleave.count", i32 4}
3154 Note that setting ``llvm.loop.interleave.count`` to 1 disables interleaving
3155 multiple iterations of the loop. If ``llvm.loop.interleave.count`` is set to 0
3156 then the interleave count will be determined automatically.
3158 '``llvm.loop.vectorize.enable``' Metadata
3159 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3161 This metadata selectively enables or disables vectorization for the loop. The
3162 first operand is the string ``llvm.loop.vectorize.enable`` and the second operand
3163 is a bit. If the bit operand value is 1 vectorization is enabled. A value of
3164 0 disables vectorization:
3166 .. code-block:: llvm
3168 !0 = !{!"llvm.loop.vectorize.enable", i1 0}
3169 !1 = !{!"llvm.loop.vectorize.enable", i1 1}
3171 '``llvm.loop.vectorize.width``' Metadata
3172 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3174 This metadata sets the target width of the vectorizer. The first
3175 operand is the string ``llvm.loop.vectorize.width`` and the second
3176 operand is an integer specifying the width. For example:
3178 .. code-block:: llvm
3180 !0 = !{!"llvm.loop.vectorize.width", i32 4}
3182 Note that setting ``llvm.loop.vectorize.width`` to 1 disables
3183 vectorization of the loop. If ``llvm.loop.vectorize.width`` is set to
3184 0 or if the loop does not have this metadata the width will be
3185 determined automatically.
3187 '``llvm.loop.unroll``'
3188 ^^^^^^^^^^^^^^^^^^^^^^
3190 Metadata prefixed with ``llvm.loop.unroll`` are loop unrolling
3191 optimization hints such as the unroll factor. ``llvm.loop.unroll``
3192 metadata should be used in conjunction with ``llvm.loop`` loop
3193 identification metadata. The ``llvm.loop.unroll`` metadata are only
3194 optimization hints and the unrolling will only be performed if the
3195 optimizer believes it is safe to do so.
3197 '``llvm.loop.unroll.count``' Metadata
3198 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3200 This metadata suggests an unroll factor to the loop unroller. The
3201 first operand is the string ``llvm.loop.unroll.count`` and the second
3202 operand is a positive integer specifying the unroll factor. For
3205 .. code-block:: llvm
3207 !0 = !{!"llvm.loop.unroll.count", i32 4}
3209 If the trip count of the loop is less than the unroll count the loop
3210 will be partially unrolled.
3212 '``llvm.loop.unroll.disable``' Metadata
3213 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3215 This metadata either disables loop unrolling. The metadata has a single operand
3216 which is the string ``llvm.loop.unroll.disable``. For example:
3218 .. code-block:: llvm
3220 !0 = !{!"llvm.loop.unroll.disable"}
3222 '``llvm.loop.unroll.full``' Metadata
3223 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3225 This metadata either suggests that the loop should be unrolled fully. The
3226 metadata has a single operand which is the string ``llvm.loop.unroll.disable``.
3229 .. code-block:: llvm
3231 !0 = !{!"llvm.loop.unroll.full"}
3236 Metadata types used to annotate memory accesses with information helpful
3237 for optimizations are prefixed with ``llvm.mem``.
3239 '``llvm.mem.parallel_loop_access``' Metadata
3240 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3242 The ``llvm.mem.parallel_loop_access`` metadata refers to a loop identifier,
3243 or metadata containing a list of loop identifiers for nested loops.
3244 The metadata is attached to memory accessing instructions and denotes that
3245 no loop carried memory dependence exist between it and other instructions denoted
3246 with the same loop identifier.
3248 Precisely, given two instructions ``m1`` and ``m2`` that both have the
3249 ``llvm.mem.parallel_loop_access`` metadata, with ``L1`` and ``L2`` being the
3250 set of loops associated with that metadata, respectively, then there is no loop
3251 carried dependence between ``m1`` and ``m2`` for loops in both ``L1`` and
3254 As a special case, if all memory accessing instructions in a loop have
3255 ``llvm.mem.parallel_loop_access`` metadata that refers to that loop, then the
3256 loop has no loop carried memory dependences and is considered to be a parallel
3259 Note that if not all memory access instructions have such metadata referring to
3260 the loop, then the loop is considered not being trivially parallel. Additional
3261 memory dependence analysis is required to make that determination. As a fail
3262 safe mechanism, this causes loops that were originally parallel to be considered
3263 sequential (if optimization passes that are unaware of the parallel semantics
3264 insert new memory instructions into the loop body).
3266 Example of a loop that is considered parallel due to its correct use of
3267 both ``llvm.loop`` and ``llvm.mem.parallel_loop_access``
3268 metadata types that refer to the same loop identifier metadata.
3270 .. code-block:: llvm
3274 %val0 = load i32* %arrayidx, !llvm.mem.parallel_loop_access !0
3276 store i32 %val0, i32* %arrayidx1, !llvm.mem.parallel_loop_access !0
3278 br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !0
3284 It is also possible to have nested parallel loops. In that case the
3285 memory accesses refer to a list of loop identifier metadata nodes instead of
3286 the loop identifier metadata node directly:
3288 .. code-block:: llvm
3292 %val1 = load i32* %arrayidx3, !llvm.mem.parallel_loop_access !2
3294 br label %inner.for.body
3298 %val0 = load i32* %arrayidx1, !llvm.mem.parallel_loop_access !0
3300 store i32 %val0, i32* %arrayidx2, !llvm.mem.parallel_loop_access !0
3302 br i1 %exitcond, label %inner.for.end, label %inner.for.body, !llvm.loop !1
3306 store i32 %val1, i32* %arrayidx4, !llvm.mem.parallel_loop_access !2
3308 br i1 %exitcond, label %outer.for.end, label %outer.for.body, !llvm.loop !2
3310 outer.for.end: ; preds = %for.body
3312 !0 = !{!1, !2} ; a list of loop identifiers
3313 !1 = !{!1} ; an identifier for the inner loop
3314 !2 = !{!2} ; an identifier for the outer loop
3319 The ``llvm.bitsets`` global metadata is used to implement
3320 :doc:`bitsets <BitSets>`.
3322 Module Flags Metadata
3323 =====================
3325 Information about the module as a whole is difficult to convey to LLVM's
3326 subsystems. The LLVM IR isn't sufficient to transmit this information.
3327 The ``llvm.module.flags`` named metadata exists in order to facilitate
3328 this. These flags are in the form of key / value pairs --- much like a
3329 dictionary --- making it easy for any subsystem who cares about a flag to
3332 The ``llvm.module.flags`` metadata contains a list of metadata triplets.
3333 Each triplet has the following form:
3335 - The first element is a *behavior* flag, which specifies the behavior
3336 when two (or more) modules are merged together, and it encounters two
3337 (or more) metadata with the same ID. The supported behaviors are
3339 - The second element is a metadata string that is a unique ID for the
3340 metadata. Each module may only have one flag entry for each unique ID (not
3341 including entries with the **Require** behavior).
3342 - The third element is the value of the flag.
3344 When two (or more) modules are merged together, the resulting
3345 ``llvm.module.flags`` metadata is the union of the modules' flags. That is, for
3346 each unique metadata ID string, there will be exactly one entry in the merged
3347 modules ``llvm.module.flags`` metadata table, and the value for that entry will
3348 be determined by the merge behavior flag, as described below. The only exception
3349 is that entries with the *Require* behavior are always preserved.
3351 The following behaviors are supported:
3362 Emits an error if two values disagree, otherwise the resulting value
3363 is that of the operands.
3367 Emits a warning if two values disagree. The result value will be the
3368 operand for the flag from the first module being linked.
3372 Adds a requirement that another module flag be present and have a
3373 specified value after linking is performed. The value must be a
3374 metadata pair, where the first element of the pair is the ID of the
3375 module flag to be restricted, and the second element of the pair is
3376 the value the module flag should be restricted to. This behavior can
3377 be used to restrict the allowable results (via triggering of an
3378 error) of linking IDs with the **Override** behavior.
3382 Uses the specified value, regardless of the behavior or value of the
3383 other module. If both modules specify **Override**, but the values
3384 differ, an error will be emitted.
3388 Appends the two values, which are required to be metadata nodes.
3392 Appends the two values, which are required to be metadata
3393 nodes. However, duplicate entries in the second list are dropped
3394 during the append operation.
3396 It is an error for a particular unique flag ID to have multiple behaviors,
3397 except in the case of **Require** (which adds restrictions on another metadata
3398 value) or **Override**.
3400 An example of module flags:
3402 .. code-block:: llvm
3404 !0 = !{ i32 1, !"foo", i32 1 }
3405 !1 = !{ i32 4, !"bar", i32 37 }
3406 !2 = !{ i32 2, !"qux", i32 42 }
3407 !3 = !{ i32 3, !"qux",
3412 !llvm.module.flags = !{ !0, !1, !2, !3 }
3414 - Metadata ``!0`` has the ID ``!"foo"`` and the value '1'. The behavior
3415 if two or more ``!"foo"`` flags are seen is to emit an error if their
3416 values are not equal.
3418 - Metadata ``!1`` has the ID ``!"bar"`` and the value '37'. The
3419 behavior if two or more ``!"bar"`` flags are seen is to use the value
3422 - Metadata ``!2`` has the ID ``!"qux"`` and the value '42'. The
3423 behavior if two or more ``!"qux"`` flags are seen is to emit a
3424 warning if their values are not equal.
3426 - Metadata ``!3`` has the ID ``!"qux"`` and the value:
3432 The behavior is to emit an error if the ``llvm.module.flags`` does not
3433 contain a flag with the ID ``!"foo"`` that has the value '1' after linking is
3436 Objective-C Garbage Collection Module Flags Metadata
3437 ----------------------------------------------------
3439 On the Mach-O platform, Objective-C stores metadata about garbage
3440 collection in a special section called "image info". The metadata
3441 consists of a version number and a bitmask specifying what types of
3442 garbage collection are supported (if any) by the file. If two or more
3443 modules are linked together their garbage collection metadata needs to
3444 be merged rather than appended together.
3446 The Objective-C garbage collection module flags metadata consists of the
3447 following key-value pairs:
3456 * - ``Objective-C Version``
3457 - **[Required]** --- The Objective-C ABI version. Valid values are 1 and 2.
3459 * - ``Objective-C Image Info Version``
3460 - **[Required]** --- The version of the image info section. Currently
3463 * - ``Objective-C Image Info Section``
3464 - **[Required]** --- The section to place the metadata. Valid values are
3465 ``"__OBJC, __image_info, regular"`` for Objective-C ABI version 1, and
3466 ``"__DATA,__objc_imageinfo, regular, no_dead_strip"`` for
3467 Objective-C ABI version 2.
3469 * - ``Objective-C Garbage Collection``
3470 - **[Required]** --- Specifies whether garbage collection is supported or
3471 not. Valid values are 0, for no garbage collection, and 2, for garbage
3472 collection supported.
3474 * - ``Objective-C GC Only``
3475 - **[Optional]** --- Specifies that only garbage collection is supported.
3476 If present, its value must be 6. This flag requires that the
3477 ``Objective-C Garbage Collection`` flag have the value 2.
3479 Some important flag interactions:
3481 - If a module with ``Objective-C Garbage Collection`` set to 0 is
3482 merged with a module with ``Objective-C Garbage Collection`` set to
3483 2, then the resulting module has the
3484 ``Objective-C Garbage Collection`` flag set to 0.
3485 - A module with ``Objective-C Garbage Collection`` set to 0 cannot be
3486 merged with a module with ``Objective-C GC Only`` set to 6.
3488 Automatic Linker Flags Module Flags Metadata
3489 --------------------------------------------
3491 Some targets support embedding flags to the linker inside individual object
3492 files. Typically this is used in conjunction with language extensions which
3493 allow source files to explicitly declare the libraries they depend on, and have
3494 these automatically be transmitted to the linker via object files.
3496 These flags are encoded in the IR using metadata in the module flags section,
3497 using the ``Linker Options`` key. The merge behavior for this flag is required
3498 to be ``AppendUnique``, and the value for the key is expected to be a metadata
3499 node which should be a list of other metadata nodes, each of which should be a
3500 list of metadata strings defining linker options.
3502 For example, the following metadata section specifies two separate sets of
3503 linker options, presumably to link against ``libz`` and the ``Cocoa``
3506 !0 = !{ i32 6, !"Linker Options",
3509 !{ !"-framework", !"Cocoa" } } }
3510 !llvm.module.flags = !{ !0 }
3512 The metadata encoding as lists of lists of options, as opposed to a collapsed
3513 list of options, is chosen so that the IR encoding can use multiple option
3514 strings to specify e.g., a single library, while still having that specifier be
3515 preserved as an atomic element that can be recognized by a target specific
3516 assembly writer or object file emitter.
3518 Each individual option is required to be either a valid option for the target's
3519 linker, or an option that is reserved by the target specific assembly writer or
3520 object file emitter. No other aspect of these options is defined by the IR.
3522 C type width Module Flags Metadata
3523 ----------------------------------
3525 The ARM backend emits a section into each generated object file describing the
3526 options that it was compiled with (in a compiler-independent way) to prevent
3527 linking incompatible objects, and to allow automatic library selection. Some
3528 of these options are not visible at the IR level, namely wchar_t width and enum
3531 To pass this information to the backend, these options are encoded in module
3532 flags metadata, using the following key-value pairs:
3542 - * 0 --- sizeof(wchar_t) == 4
3543 * 1 --- sizeof(wchar_t) == 2
3546 - * 0 --- Enums are at least as large as an ``int``.
3547 * 1 --- Enums are stored in the smallest integer type which can
3548 represent all of its values.
3550 For example, the following metadata section specifies that the module was
3551 compiled with a ``wchar_t`` width of 4 bytes, and the underlying type of an
3552 enum is the smallest type which can represent all of its values::
3554 !llvm.module.flags = !{!0, !1}
3555 !0 = !{i32 1, !"short_wchar", i32 1}
3556 !1 = !{i32 1, !"short_enum", i32 0}
3558 .. _intrinsicglobalvariables:
3560 Intrinsic Global Variables
3561 ==========================
3563 LLVM has a number of "magic" global variables that contain data that
3564 affect code generation or other IR semantics. These are documented here.
3565 All globals of this sort should have a section specified as
3566 "``llvm.metadata``". This section and all globals that start with
3567 "``llvm.``" are reserved for use by LLVM.
3571 The '``llvm.used``' Global Variable
3572 -----------------------------------
3574 The ``@llvm.used`` global is an array which has
3575 :ref:`appending linkage <linkage_appending>`. This array contains a list of
3576 pointers to named global variables, functions and aliases which may optionally
3577 have a pointer cast formed of bitcast or getelementptr. For example, a legal
3580 .. code-block:: llvm
3585 @llvm.used = appending global [2 x i8*] [
3587 i8* bitcast (i32* @Y to i8*)
3588 ], section "llvm.metadata"
3590 If a symbol appears in the ``@llvm.used`` list, then the compiler, assembler,
3591 and linker are required to treat the symbol as if there is a reference to the
3592 symbol that it cannot see (which is why they have to be named). For example, if
3593 a variable has internal linkage and no references other than that from the
3594 ``@llvm.used`` list, it cannot be deleted. This is commonly used to represent
3595 references from inline asms and other things the compiler cannot "see", and
3596 corresponds to "``attribute((used))``" in GNU C.
3598 On some targets, the code generator must emit a directive to the
3599 assembler or object file to prevent the assembler and linker from
3600 molesting the symbol.
3602 .. _gv_llvmcompilerused:
3604 The '``llvm.compiler.used``' Global Variable
3605 --------------------------------------------
3607 The ``@llvm.compiler.used`` directive is the same as the ``@llvm.used``
3608 directive, except that it only prevents the compiler from touching the
3609 symbol. On targets that support it, this allows an intelligent linker to
3610 optimize references to the symbol without being impeded as it would be
3613 This is a rare construct that should only be used in rare circumstances,
3614 and should not be exposed to source languages.
3616 .. _gv_llvmglobalctors:
3618 The '``llvm.global_ctors``' Global Variable
3619 -------------------------------------------
3621 .. code-block:: llvm
3623 %0 = type { i32, void ()*, i8* }
3624 @llvm.global_ctors = appending global [1 x %0] [%0 { i32 65535, void ()* @ctor, i8* @data }]
3626 The ``@llvm.global_ctors`` array contains a list of constructor
3627 functions, priorities, and an optional associated global or function.
3628 The functions referenced by this array will be called in ascending order
3629 of priority (i.e. lowest first) when the module is loaded. The order of
3630 functions with the same priority is not defined.
3632 If the third field is present, non-null, and points to a global variable
3633 or function, the initializer function will only run if the associated
3634 data from the current module is not discarded.
3636 .. _llvmglobaldtors:
3638 The '``llvm.global_dtors``' Global Variable
3639 -------------------------------------------
3641 .. code-block:: llvm
3643 %0 = type { i32, void ()*, i8* }
3644 @llvm.global_dtors = appending global [1 x %0] [%0 { i32 65535, void ()* @dtor, i8* @data }]
3646 The ``@llvm.global_dtors`` array contains a list of destructor
3647 functions, priorities, and an optional associated global or function.
3648 The functions referenced by this array will be called in descending
3649 order of priority (i.e. highest first) when the module is unloaded. The
3650 order of functions with the same priority is not defined.
3652 If the third field is present, non-null, and points to a global variable
3653 or function, the destructor function will only run if the associated
3654 data from the current module is not discarded.
3656 Instruction Reference
3657 =====================
3659 The LLVM instruction set consists of several different classifications
3660 of instructions: :ref:`terminator instructions <terminators>`, :ref:`binary
3661 instructions <binaryops>`, :ref:`bitwise binary
3662 instructions <bitwiseops>`, :ref:`memory instructions <memoryops>`, and
3663 :ref:`other instructions <otherops>`.
3667 Terminator Instructions
3668 -----------------------
3670 As mentioned :ref:`previously <functionstructure>`, every basic block in a
3671 program ends with a "Terminator" instruction, which indicates which
3672 block should be executed after the current block is finished. These
3673 terminator instructions typically yield a '``void``' value: they produce
3674 control flow, not values (the one exception being the
3675 ':ref:`invoke <i_invoke>`' instruction).
3677 The terminator instructions are: ':ref:`ret <i_ret>`',
3678 ':ref:`br <i_br>`', ':ref:`switch <i_switch>`',
3679 ':ref:`indirectbr <i_indirectbr>`', ':ref:`invoke <i_invoke>`',
3680 ':ref:`resume <i_resume>`', and ':ref:`unreachable <i_unreachable>`'.
3684 '``ret``' Instruction
3685 ^^^^^^^^^^^^^^^^^^^^^
3692 ret <type> <value> ; Return a value from a non-void function
3693 ret void ; Return from void function
3698 The '``ret``' instruction is used to return control flow (and optionally
3699 a value) from a function back to the caller.
3701 There are two forms of the '``ret``' instruction: one that returns a
3702 value and then causes control flow, and one that just causes control
3708 The '``ret``' instruction optionally accepts a single argument, the
3709 return value. The type of the return value must be a ':ref:`first
3710 class <t_firstclass>`' type.
3712 A function is not :ref:`well formed <wellformed>` if it it has a non-void
3713 return type and contains a '``ret``' instruction with no return value or
3714 a return value with a type that does not match its type, or if it has a
3715 void return type and contains a '``ret``' instruction with a return
3721 When the '``ret``' instruction is executed, control flow returns back to
3722 the calling function's context. If the caller is a
3723 ":ref:`call <i_call>`" instruction, execution continues at the
3724 instruction after the call. If the caller was an
3725 ":ref:`invoke <i_invoke>`" instruction, execution continues at the
3726 beginning of the "normal" destination block. If the instruction returns
3727 a value, that value shall set the call or invoke instruction's return
3733 .. code-block:: llvm
3735 ret i32 5 ; Return an integer value of 5
3736 ret void ; Return from a void function
3737 ret { i32, i8 } { i32 4, i8 2 } ; Return a struct of values 4 and 2
3741 '``br``' Instruction
3742 ^^^^^^^^^^^^^^^^^^^^
3749 br i1 <cond>, label <iftrue>, label <iffalse>
3750 br label <dest> ; Unconditional branch
3755 The '``br``' instruction is used to cause control flow to transfer to a
3756 different basic block in the current function. There are two forms of
3757 this instruction, corresponding to a conditional branch and an
3758 unconditional branch.
3763 The conditional branch form of the '``br``' instruction takes a single
3764 '``i1``' value and two '``label``' values. The unconditional form of the
3765 '``br``' instruction takes a single '``label``' value as a target.
3770 Upon execution of a conditional '``br``' instruction, the '``i1``'
3771 argument is evaluated. If the value is ``true``, control flows to the
3772 '``iftrue``' ``label`` argument. If "cond" is ``false``, control flows
3773 to the '``iffalse``' ``label`` argument.
3778 .. code-block:: llvm
3781 %cond = icmp eq i32 %a, %b
3782 br i1 %cond, label %IfEqual, label %IfUnequal
3790 '``switch``' Instruction
3791 ^^^^^^^^^^^^^^^^^^^^^^^^
3798 switch <intty> <value>, label <defaultdest> [ <intty> <val>, label <dest> ... ]
3803 The '``switch``' instruction is used to transfer control flow to one of
3804 several different places. It is a generalization of the '``br``'
3805 instruction, allowing a branch to occur to one of many possible
3811 The '``switch``' instruction uses three parameters: an integer
3812 comparison value '``value``', a default '``label``' destination, and an
3813 array of pairs of comparison value constants and '``label``'s. The table
3814 is not allowed to contain duplicate constant entries.
3819 The ``switch`` instruction specifies a table of values and destinations.
3820 When the '``switch``' instruction is executed, this table is searched
3821 for the given value. If the value is found, control flow is transferred
3822 to the corresponding destination; otherwise, control flow is transferred
3823 to the default destination.
3828 Depending on properties of the target machine and the particular
3829 ``switch`` instruction, this instruction may be code generated in
3830 different ways. For example, it could be generated as a series of
3831 chained conditional branches or with a lookup table.
3836 .. code-block:: llvm
3838 ; Emulate a conditional br instruction
3839 %Val = zext i1 %value to i32
3840 switch i32 %Val, label %truedest [ i32 0, label %falsedest ]
3842 ; Emulate an unconditional br instruction
3843 switch i32 0, label %dest [ ]
3845 ; Implement a jump table:
3846 switch i32 %val, label %otherwise [ i32 0, label %onzero
3848 i32 2, label %ontwo ]
3852 '``indirectbr``' Instruction
3853 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3860 indirectbr <somety>* <address>, [ label <dest1>, label <dest2>, ... ]
3865 The '``indirectbr``' instruction implements an indirect branch to a
3866 label within the current function, whose address is specified by
3867 "``address``". Address must be derived from a
3868 :ref:`blockaddress <blockaddress>` constant.
3873 The '``address``' argument is the address of the label to jump to. The
3874 rest of the arguments indicate the full set of possible destinations
3875 that the address may point to. Blocks are allowed to occur multiple
3876 times in the destination list, though this isn't particularly useful.
3878 This destination list is required so that dataflow analysis has an
3879 accurate understanding of the CFG.
3884 Control transfers to the block specified in the address argument. All
3885 possible destination blocks must be listed in the label list, otherwise
3886 this instruction has undefined behavior. This implies that jumps to
3887 labels defined in other functions have undefined behavior as well.
3892 This is typically implemented with a jump through a register.
3897 .. code-block:: llvm
3899 indirectbr i8* %Addr, [ label %bb1, label %bb2, label %bb3 ]
3903 '``invoke``' Instruction
3904 ^^^^^^^^^^^^^^^^^^^^^^^^
3911 <result> = invoke [cconv] [ret attrs] <ptr to function ty> <function ptr val>(<function args>) [fn attrs]
3912 to label <normal label> unwind label <exception label>
3917 The '``invoke``' instruction causes control to transfer to a specified
3918 function, with the possibility of control flow transfer to either the
3919 '``normal``' label or the '``exception``' label. If the callee function
3920 returns with the "``ret``" instruction, control flow will return to the
3921 "normal" label. If the callee (or any indirect callees) returns via the
3922 ":ref:`resume <i_resume>`" instruction or other exception handling
3923 mechanism, control is interrupted and continued at the dynamically
3924 nearest "exception" label.
3926 The '``exception``' label is a `landing
3927 pad <ExceptionHandling.html#overview>`_ for the exception. As such,
3928 '``exception``' label is required to have the
3929 ":ref:`landingpad <i_landingpad>`" instruction, which contains the
3930 information about the behavior of the program after unwinding happens,
3931 as its first non-PHI instruction. The restrictions on the
3932 "``landingpad``" instruction's tightly couples it to the "``invoke``"
3933 instruction, so that the important information contained within the
3934 "``landingpad``" instruction can't be lost through normal code motion.
3939 This instruction requires several arguments:
3941 #. The optional "cconv" marker indicates which :ref:`calling
3942 convention <callingconv>` the call should use. If none is
3943 specified, the call defaults to using C calling conventions.
3944 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
3945 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
3947 #. '``ptr to function ty``': shall be the signature of the pointer to
3948 function value being invoked. In most cases, this is a direct
3949 function invocation, but indirect ``invoke``'s are just as possible,
3950 branching off an arbitrary pointer to function value.
3951 #. '``function ptr val``': An LLVM value containing a pointer to a
3952 function to be invoked.
3953 #. '``function args``': argument list whose types match the function
3954 signature argument types and parameter attributes. All arguments must
3955 be of :ref:`first class <t_firstclass>` type. If the function signature
3956 indicates the function accepts a variable number of arguments, the
3957 extra arguments can be specified.
3958 #. '``normal label``': the label reached when the called function
3959 executes a '``ret``' instruction.
3960 #. '``exception label``': the label reached when a callee returns via
3961 the :ref:`resume <i_resume>` instruction or other exception handling
3963 #. The optional :ref:`function attributes <fnattrs>` list. Only
3964 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
3965 attributes are valid here.
3970 This instruction is designed to operate as a standard '``call``'
3971 instruction in most regards. The primary difference is that it
3972 establishes an association with a label, which is used by the runtime
3973 library to unwind the stack.
3975 This instruction is used in languages with destructors to ensure that
3976 proper cleanup is performed in the case of either a ``longjmp`` or a
3977 thrown exception. Additionally, this is important for implementation of
3978 '``catch``' clauses in high-level languages that support them.
3980 For the purposes of the SSA form, the definition of the value returned
3981 by the '``invoke``' instruction is deemed to occur on the edge from the
3982 current block to the "normal" label. If the callee unwinds then no
3983 return value is available.
3988 .. code-block:: llvm
3990 %retval = invoke i32 @Test(i32 15) to label %Continue
3991 unwind label %TestCleanup ; i32:retval set
3992 %retval = invoke coldcc i32 %Testfnptr(i32 15) to label %Continue
3993 unwind label %TestCleanup ; i32:retval set
3997 '``resume``' Instruction
3998 ^^^^^^^^^^^^^^^^^^^^^^^^
4005 resume <type> <value>
4010 The '``resume``' instruction is a terminator instruction that has no
4016 The '``resume``' instruction requires one argument, which must have the
4017 same type as the result of any '``landingpad``' instruction in the same
4023 The '``resume``' instruction resumes propagation of an existing
4024 (in-flight) exception whose unwinding was interrupted with a
4025 :ref:`landingpad <i_landingpad>` instruction.
4030 .. code-block:: llvm
4032 resume { i8*, i32 } %exn
4036 '``unreachable``' Instruction
4037 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4049 The '``unreachable``' instruction has no defined semantics. This
4050 instruction is used to inform the optimizer that a particular portion of
4051 the code is not reachable. This can be used to indicate that the code
4052 after a no-return function cannot be reached, and other facts.
4057 The '``unreachable``' instruction has no defined semantics.
4064 Binary operators are used to do most of the computation in a program.
4065 They require two operands of the same type, execute an operation on
4066 them, and produce a single value. The operands might represent multiple
4067 data, as is the case with the :ref:`vector <t_vector>` data type. The
4068 result value has the same type as its operands.
4070 There are several different binary operators:
4074 '``add``' Instruction
4075 ^^^^^^^^^^^^^^^^^^^^^
4082 <result> = add <ty> <op1>, <op2> ; yields ty:result
4083 <result> = add nuw <ty> <op1>, <op2> ; yields ty:result
4084 <result> = add nsw <ty> <op1>, <op2> ; yields ty:result
4085 <result> = add nuw nsw <ty> <op1>, <op2> ; yields ty:result
4090 The '``add``' instruction returns the sum of its two operands.
4095 The two arguments to the '``add``' instruction must be
4096 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4097 arguments must have identical types.
4102 The value produced is the integer sum of the two operands.
4104 If the sum has unsigned overflow, the result returned is the
4105 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
4108 Because LLVM integers use a two's complement representation, this
4109 instruction is appropriate for both signed and unsigned integers.
4111 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
4112 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
4113 result value of the ``add`` is a :ref:`poison value <poisonvalues>` if
4114 unsigned and/or signed overflow, respectively, occurs.
4119 .. code-block:: llvm
4121 <result> = add i32 4, %var ; yields i32:result = 4 + %var
4125 '``fadd``' Instruction
4126 ^^^^^^^^^^^^^^^^^^^^^^
4133 <result> = fadd [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
4138 The '``fadd``' instruction returns the sum of its two operands.
4143 The two arguments to the '``fadd``' instruction must be :ref:`floating
4144 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
4145 Both arguments must have identical types.
4150 The value produced is the floating point sum of the two operands. This
4151 instruction can also take any number of :ref:`fast-math flags <fastmath>`,
4152 which are optimization hints to enable otherwise unsafe floating point
4158 .. code-block:: llvm
4160 <result> = fadd float 4.0, %var ; yields float:result = 4.0 + %var
4162 '``sub``' Instruction
4163 ^^^^^^^^^^^^^^^^^^^^^
4170 <result> = sub <ty> <op1>, <op2> ; yields ty:result
4171 <result> = sub nuw <ty> <op1>, <op2> ; yields ty:result
4172 <result> = sub nsw <ty> <op1>, <op2> ; yields ty:result
4173 <result> = sub nuw nsw <ty> <op1>, <op2> ; yields ty:result
4178 The '``sub``' instruction returns the difference of its two operands.
4180 Note that the '``sub``' instruction is used to represent the '``neg``'
4181 instruction present in most other intermediate representations.
4186 The two arguments to the '``sub``' instruction must be
4187 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4188 arguments must have identical types.
4193 The value produced is the integer difference of the two operands.
4195 If the difference has unsigned overflow, the result returned is the
4196 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
4199 Because LLVM integers use a two's complement representation, this
4200 instruction is appropriate for both signed and unsigned integers.
4202 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
4203 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
4204 result value of the ``sub`` is a :ref:`poison value <poisonvalues>` if
4205 unsigned and/or signed overflow, respectively, occurs.
4210 .. code-block:: llvm
4212 <result> = sub i32 4, %var ; yields i32:result = 4 - %var
4213 <result> = sub i32 0, %val ; yields i32:result = -%var
4217 '``fsub``' Instruction
4218 ^^^^^^^^^^^^^^^^^^^^^^
4225 <result> = fsub [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
4230 The '``fsub``' instruction returns the difference of its two operands.
4232 Note that the '``fsub``' instruction is used to represent the '``fneg``'
4233 instruction present in most other intermediate representations.
4238 The two arguments to the '``fsub``' instruction must be :ref:`floating
4239 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
4240 Both arguments must have identical types.
4245 The value produced is the floating point difference of the two operands.
4246 This instruction can also take any number of :ref:`fast-math
4247 flags <fastmath>`, which are optimization hints to enable otherwise
4248 unsafe floating point optimizations:
4253 .. code-block:: llvm
4255 <result> = fsub float 4.0, %var ; yields float:result = 4.0 - %var
4256 <result> = fsub float -0.0, %val ; yields float:result = -%var
4258 '``mul``' Instruction
4259 ^^^^^^^^^^^^^^^^^^^^^
4266 <result> = mul <ty> <op1>, <op2> ; yields ty:result
4267 <result> = mul nuw <ty> <op1>, <op2> ; yields ty:result
4268 <result> = mul nsw <ty> <op1>, <op2> ; yields ty:result
4269 <result> = mul nuw nsw <ty> <op1>, <op2> ; yields ty:result
4274 The '``mul``' instruction returns the product of its two operands.
4279 The two arguments to the '``mul``' instruction must be
4280 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4281 arguments must have identical types.
4286 The value produced is the integer product of the two operands.
4288 If the result of the multiplication has unsigned overflow, the result
4289 returned is the mathematical result modulo 2\ :sup:`n`\ , where n is the
4290 bit width of the result.
4292 Because LLVM integers use a two's complement representation, and the
4293 result is the same width as the operands, this instruction returns the
4294 correct result for both signed and unsigned integers. If a full product
4295 (e.g. ``i32`` * ``i32`` -> ``i64``) is needed, the operands should be
4296 sign-extended or zero-extended as appropriate to the width of the full
4299 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
4300 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
4301 result value of the ``mul`` is a :ref:`poison value <poisonvalues>` if
4302 unsigned and/or signed overflow, respectively, occurs.
4307 .. code-block:: llvm
4309 <result> = mul i32 4, %var ; yields i32:result = 4 * %var
4313 '``fmul``' Instruction
4314 ^^^^^^^^^^^^^^^^^^^^^^
4321 <result> = fmul [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
4326 The '``fmul``' instruction returns the product of its two operands.
4331 The two arguments to the '``fmul``' instruction must be :ref:`floating
4332 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
4333 Both arguments must have identical types.
4338 The value produced is the floating point product of the two operands.
4339 This instruction can also take any number of :ref:`fast-math
4340 flags <fastmath>`, which are optimization hints to enable otherwise
4341 unsafe floating point optimizations:
4346 .. code-block:: llvm
4348 <result> = fmul float 4.0, %var ; yields float:result = 4.0 * %var
4350 '``udiv``' Instruction
4351 ^^^^^^^^^^^^^^^^^^^^^^
4358 <result> = udiv <ty> <op1>, <op2> ; yields ty:result
4359 <result> = udiv exact <ty> <op1>, <op2> ; yields ty:result
4364 The '``udiv``' instruction returns the quotient of its two operands.
4369 The two arguments to the '``udiv``' instruction must be
4370 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4371 arguments must have identical types.
4376 The value produced is the unsigned integer quotient of the two operands.
4378 Note that unsigned integer division and signed integer division are
4379 distinct operations; for signed integer division, use '``sdiv``'.
4381 Division by zero leads to undefined behavior.
4383 If the ``exact`` keyword is present, the result value of the ``udiv`` is
4384 a :ref:`poison value <poisonvalues>` if %op1 is not a multiple of %op2 (as
4385 such, "((a udiv exact b) mul b) == a").
4390 .. code-block:: llvm
4392 <result> = udiv i32 4, %var ; yields i32:result = 4 / %var
4394 '``sdiv``' Instruction
4395 ^^^^^^^^^^^^^^^^^^^^^^
4402 <result> = sdiv <ty> <op1>, <op2> ; yields ty:result
4403 <result> = sdiv exact <ty> <op1>, <op2> ; yields ty:result
4408 The '``sdiv``' instruction returns the quotient of its two operands.
4413 The two arguments to the '``sdiv``' instruction must be
4414 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4415 arguments must have identical types.
4420 The value produced is the signed integer quotient of the two operands
4421 rounded towards zero.
4423 Note that signed integer division and unsigned integer division are
4424 distinct operations; for unsigned integer division, use '``udiv``'.
4426 Division by zero leads to undefined behavior. Overflow also leads to
4427 undefined behavior; this is a rare case, but can occur, for example, by
4428 doing a 32-bit division of -2147483648 by -1.
4430 If the ``exact`` keyword is present, the result value of the ``sdiv`` is
4431 a :ref:`poison value <poisonvalues>` if the result would be rounded.
4436 .. code-block:: llvm
4438 <result> = sdiv i32 4, %var ; yields i32:result = 4 / %var
4442 '``fdiv``' Instruction
4443 ^^^^^^^^^^^^^^^^^^^^^^
4450 <result> = fdiv [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
4455 The '``fdiv``' instruction returns the quotient of its two operands.
4460 The two arguments to the '``fdiv``' instruction must be :ref:`floating
4461 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
4462 Both arguments must have identical types.
4467 The value produced is the floating point quotient of the two operands.
4468 This instruction can also take any number of :ref:`fast-math
4469 flags <fastmath>`, which are optimization hints to enable otherwise
4470 unsafe floating point optimizations:
4475 .. code-block:: llvm
4477 <result> = fdiv float 4.0, %var ; yields float:result = 4.0 / %var
4479 '``urem``' Instruction
4480 ^^^^^^^^^^^^^^^^^^^^^^
4487 <result> = urem <ty> <op1>, <op2> ; yields ty:result
4492 The '``urem``' instruction returns the remainder from the unsigned
4493 division of its two arguments.
4498 The two arguments to the '``urem``' instruction must be
4499 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4500 arguments must have identical types.
4505 This instruction returns the unsigned integer *remainder* of a division.
4506 This instruction always performs an unsigned division to get the
4509 Note that unsigned integer remainder and signed integer remainder are
4510 distinct operations; for signed integer remainder, use '``srem``'.
4512 Taking the remainder of a division by zero leads to undefined behavior.
4517 .. code-block:: llvm
4519 <result> = urem i32 4, %var ; yields i32:result = 4 % %var
4521 '``srem``' Instruction
4522 ^^^^^^^^^^^^^^^^^^^^^^
4529 <result> = srem <ty> <op1>, <op2> ; yields ty:result
4534 The '``srem``' instruction returns the remainder from the signed
4535 division of its two operands. This instruction can also take
4536 :ref:`vector <t_vector>` versions of the values in which case the elements
4542 The two arguments to the '``srem``' instruction must be
4543 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4544 arguments must have identical types.
4549 This instruction returns the *remainder* of a division (where the result
4550 is either zero or has the same sign as the dividend, ``op1``), not the
4551 *modulo* operator (where the result is either zero or has the same sign
4552 as the divisor, ``op2``) of a value. For more information about the
4553 difference, see `The Math
4554 Forum <http://mathforum.org/dr.math/problems/anne.4.28.99.html>`_. For a
4555 table of how this is implemented in various languages, please see
4557 operation <http://en.wikipedia.org/wiki/Modulo_operation>`_.
4559 Note that signed integer remainder and unsigned integer remainder are
4560 distinct operations; for unsigned integer remainder, use '``urem``'.
4562 Taking the remainder of a division by zero leads to undefined behavior.
4563 Overflow also leads to undefined behavior; this is a rare case, but can
4564 occur, for example, by taking the remainder of a 32-bit division of
4565 -2147483648 by -1. (The remainder doesn't actually overflow, but this
4566 rule lets srem be implemented using instructions that return both the
4567 result of the division and the remainder.)
4572 .. code-block:: llvm
4574 <result> = srem i32 4, %var ; yields i32:result = 4 % %var
4578 '``frem``' Instruction
4579 ^^^^^^^^^^^^^^^^^^^^^^
4586 <result> = frem [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
4591 The '``frem``' instruction returns the remainder from the division of
4597 The two arguments to the '``frem``' instruction must be :ref:`floating
4598 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
4599 Both arguments must have identical types.
4604 This instruction returns the *remainder* of a division. The remainder
4605 has the same sign as the dividend. This instruction can also take any
4606 number of :ref:`fast-math flags <fastmath>`, which are optimization hints
4607 to enable otherwise unsafe floating point optimizations:
4612 .. code-block:: llvm
4614 <result> = frem float 4.0, %var ; yields float:result = 4.0 % %var
4618 Bitwise Binary Operations
4619 -------------------------
4621 Bitwise binary operators are used to do various forms of bit-twiddling
4622 in a program. They are generally very efficient instructions and can
4623 commonly be strength reduced from other instructions. They require two
4624 operands of the same type, execute an operation on them, and produce a
4625 single value. The resulting value is the same type as its operands.
4627 '``shl``' Instruction
4628 ^^^^^^^^^^^^^^^^^^^^^
4635 <result> = shl <ty> <op1>, <op2> ; yields ty:result
4636 <result> = shl nuw <ty> <op1>, <op2> ; yields ty:result
4637 <result> = shl nsw <ty> <op1>, <op2> ; yields ty:result
4638 <result> = shl nuw nsw <ty> <op1>, <op2> ; yields ty:result
4643 The '``shl``' instruction returns the first operand shifted to the left
4644 a specified number of bits.
4649 Both arguments to the '``shl``' instruction must be the same
4650 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4651 '``op2``' is treated as an unsigned value.
4656 The value produced is ``op1`` \* 2\ :sup:`op2` mod 2\ :sup:`n`,
4657 where ``n`` is the width of the result. If ``op2`` is (statically or
4658 dynamically) negative or equal to or larger than the number of bits in
4659 ``op1``, the result is undefined. If the arguments are vectors, each
4660 vector element of ``op1`` is shifted by the corresponding shift amount
4663 If the ``nuw`` keyword is present, then the shift produces a :ref:`poison
4664 value <poisonvalues>` if it shifts out any non-zero bits. If the
4665 ``nsw`` keyword is present, then the shift produces a :ref:`poison
4666 value <poisonvalues>` if it shifts out any bits that disagree with the
4667 resultant sign bit. As such, NUW/NSW have the same semantics as they
4668 would if the shift were expressed as a mul instruction with the same
4669 nsw/nuw bits in (mul %op1, (shl 1, %op2)).
4674 .. code-block:: llvm
4676 <result> = shl i32 4, %var ; yields i32: 4 << %var
4677 <result> = shl i32 4, 2 ; yields i32: 16
4678 <result> = shl i32 1, 10 ; yields i32: 1024
4679 <result> = shl i32 1, 32 ; undefined
4680 <result> = shl <2 x i32> < i32 1, i32 1>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 2, i32 4>
4682 '``lshr``' Instruction
4683 ^^^^^^^^^^^^^^^^^^^^^^
4690 <result> = lshr <ty> <op1>, <op2> ; yields ty:result
4691 <result> = lshr exact <ty> <op1>, <op2> ; yields ty:result
4696 The '``lshr``' instruction (logical shift right) returns the first
4697 operand shifted to the right a specified number of bits with zero fill.
4702 Both arguments to the '``lshr``' instruction must be the same
4703 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4704 '``op2``' is treated as an unsigned value.
4709 This instruction always performs a logical shift right operation. The
4710 most significant bits of the result will be filled with zero bits after
4711 the shift. If ``op2`` is (statically or dynamically) equal to or larger
4712 than the number of bits in ``op1``, the result is undefined. If the
4713 arguments are vectors, each vector element of ``op1`` is shifted by the
4714 corresponding shift amount in ``op2``.
4716 If the ``exact`` keyword is present, the result value of the ``lshr`` is
4717 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
4723 .. code-block:: llvm
4725 <result> = lshr i32 4, 1 ; yields i32:result = 2
4726 <result> = lshr i32 4, 2 ; yields i32:result = 1
4727 <result> = lshr i8 4, 3 ; yields i8:result = 0
4728 <result> = lshr i8 -2, 1 ; yields i8:result = 0x7F
4729 <result> = lshr i32 1, 32 ; undefined
4730 <result> = lshr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 0x7FFFFFFF, i32 1>
4732 '``ashr``' Instruction
4733 ^^^^^^^^^^^^^^^^^^^^^^
4740 <result> = ashr <ty> <op1>, <op2> ; yields ty:result
4741 <result> = ashr exact <ty> <op1>, <op2> ; yields ty:result
4746 The '``ashr``' instruction (arithmetic shift right) returns the first
4747 operand shifted to the right a specified number of bits with sign
4753 Both arguments to the '``ashr``' instruction must be the same
4754 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4755 '``op2``' is treated as an unsigned value.
4760 This instruction always performs an arithmetic shift right operation,
4761 The most significant bits of the result will be filled with the sign bit
4762 of ``op1``. If ``op2`` is (statically or dynamically) equal to or larger
4763 than the number of bits in ``op1``, the result is undefined. If the
4764 arguments are vectors, each vector element of ``op1`` is shifted by the
4765 corresponding shift amount in ``op2``.
4767 If the ``exact`` keyword is present, the result value of the ``ashr`` is
4768 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
4774 .. code-block:: llvm
4776 <result> = ashr i32 4, 1 ; yields i32:result = 2
4777 <result> = ashr i32 4, 2 ; yields i32:result = 1
4778 <result> = ashr i8 4, 3 ; yields i8:result = 0
4779 <result> = ashr i8 -2, 1 ; yields i8:result = -1
4780 <result> = ashr i32 1, 32 ; undefined
4781 <result> = ashr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 3> ; yields: result=<2 x i32> < i32 -1, i32 0>
4783 '``and``' Instruction
4784 ^^^^^^^^^^^^^^^^^^^^^
4791 <result> = and <ty> <op1>, <op2> ; yields ty:result
4796 The '``and``' instruction returns the bitwise logical and of its two
4802 The two arguments to the '``and``' instruction must be
4803 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4804 arguments must have identical types.
4809 The truth table used for the '``and``' instruction is:
4826 .. code-block:: llvm
4828 <result> = and i32 4, %var ; yields i32:result = 4 & %var
4829 <result> = and i32 15, 40 ; yields i32:result = 8
4830 <result> = and i32 4, 8 ; yields i32:result = 0
4832 '``or``' Instruction
4833 ^^^^^^^^^^^^^^^^^^^^
4840 <result> = or <ty> <op1>, <op2> ; yields ty:result
4845 The '``or``' instruction returns the bitwise logical inclusive or of its
4851 The two arguments to the '``or``' instruction must be
4852 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4853 arguments must have identical types.
4858 The truth table used for the '``or``' instruction is:
4877 <result> = or i32 4, %var ; yields i32:result = 4 | %var
4878 <result> = or i32 15, 40 ; yields i32:result = 47
4879 <result> = or i32 4, 8 ; yields i32:result = 12
4881 '``xor``' Instruction
4882 ^^^^^^^^^^^^^^^^^^^^^
4889 <result> = xor <ty> <op1>, <op2> ; yields ty:result
4894 The '``xor``' instruction returns the bitwise logical exclusive or of
4895 its two operands. The ``xor`` is used to implement the "one's
4896 complement" operation, which is the "~" operator in C.
4901 The two arguments to the '``xor``' instruction must be
4902 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4903 arguments must have identical types.
4908 The truth table used for the '``xor``' instruction is:
4925 .. code-block:: llvm
4927 <result> = xor i32 4, %var ; yields i32:result = 4 ^ %var
4928 <result> = xor i32 15, 40 ; yields i32:result = 39
4929 <result> = xor i32 4, 8 ; yields i32:result = 12
4930 <result> = xor i32 %V, -1 ; yields i32:result = ~%V
4935 LLVM supports several instructions to represent vector operations in a
4936 target-independent manner. These instructions cover the element-access
4937 and vector-specific operations needed to process vectors effectively.
4938 While LLVM does directly support these vector operations, many
4939 sophisticated algorithms will want to use target-specific intrinsics to
4940 take full advantage of a specific target.
4942 .. _i_extractelement:
4944 '``extractelement``' Instruction
4945 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4952 <result> = extractelement <n x <ty>> <val>, <ty2> <idx> ; yields <ty>
4957 The '``extractelement``' instruction extracts a single scalar element
4958 from a vector at a specified index.
4963 The first operand of an '``extractelement``' instruction is a value of
4964 :ref:`vector <t_vector>` type. The second operand is an index indicating
4965 the position from which to extract the element. The index may be a
4966 variable of any integer type.
4971 The result is a scalar of the same type as the element type of ``val``.
4972 Its value is the value at position ``idx`` of ``val``. If ``idx``
4973 exceeds the length of ``val``, the results are undefined.
4978 .. code-block:: llvm
4980 <result> = extractelement <4 x i32> %vec, i32 0 ; yields i32
4982 .. _i_insertelement:
4984 '``insertelement``' Instruction
4985 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4992 <result> = insertelement <n x <ty>> <val>, <ty> <elt>, <ty2> <idx> ; yields <n x <ty>>
4997 The '``insertelement``' instruction inserts a scalar element into a
4998 vector at a specified index.
5003 The first operand of an '``insertelement``' instruction is a value of
5004 :ref:`vector <t_vector>` type. The second operand is a scalar value whose
5005 type must equal the element type of the first operand. The third operand
5006 is an index indicating the position at which to insert the value. The
5007 index may be a variable of any integer type.
5012 The result is a vector of the same type as ``val``. Its element values
5013 are those of ``val`` except at position ``idx``, where it gets the value
5014 ``elt``. If ``idx`` exceeds the length of ``val``, the results are
5020 .. code-block:: llvm
5022 <result> = insertelement <4 x i32> %vec, i32 1, i32 0 ; yields <4 x i32>
5024 .. _i_shufflevector:
5026 '``shufflevector``' Instruction
5027 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5034 <result> = shufflevector <n x <ty>> <v1>, <n x <ty>> <v2>, <m x i32> <mask> ; yields <m x <ty>>
5039 The '``shufflevector``' instruction constructs a permutation of elements
5040 from two input vectors, returning a vector with the same element type as
5041 the input and length that is the same as the shuffle mask.
5046 The first two operands of a '``shufflevector``' instruction are vectors
5047 with the same type. The third argument is a shuffle mask whose element
5048 type is always 'i32'. The result of the instruction is a vector whose
5049 length is the same as the shuffle mask and whose element type is the
5050 same as the element type of the first two operands.
5052 The shuffle mask operand is required to be a constant vector with either
5053 constant integer or undef values.
5058 The elements of the two input vectors are numbered from left to right
5059 across both of the vectors. The shuffle mask operand specifies, for each
5060 element of the result vector, which element of the two input vectors the
5061 result element gets. The element selector may be undef (meaning "don't
5062 care") and the second operand may be undef if performing a shuffle from
5068 .. code-block:: llvm
5070 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
5071 <4 x i32> <i32 0, i32 4, i32 1, i32 5> ; yields <4 x i32>
5072 <result> = shufflevector <4 x i32> %v1, <4 x i32> undef,
5073 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32> - Identity shuffle.
5074 <result> = shufflevector <8 x i32> %v1, <8 x i32> undef,
5075 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32>
5076 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
5077 <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7 > ; yields <8 x i32>
5079 Aggregate Operations
5080 --------------------
5082 LLVM supports several instructions for working with
5083 :ref:`aggregate <t_aggregate>` values.
5087 '``extractvalue``' Instruction
5088 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5095 <result> = extractvalue <aggregate type> <val>, <idx>{, <idx>}*
5100 The '``extractvalue``' instruction extracts the value of a member field
5101 from an :ref:`aggregate <t_aggregate>` value.
5106 The first operand of an '``extractvalue``' instruction is a value of
5107 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The operands are
5108 constant indices to specify which value to extract in a similar manner
5109 as indices in a '``getelementptr``' instruction.
5111 The major differences to ``getelementptr`` indexing are:
5113 - Since the value being indexed is not a pointer, the first index is
5114 omitted and assumed to be zero.
5115 - At least one index must be specified.
5116 - Not only struct indices but also array indices must be in bounds.
5121 The result is the value at the position in the aggregate specified by
5127 .. code-block:: llvm
5129 <result> = extractvalue {i32, float} %agg, 0 ; yields i32
5133 '``insertvalue``' Instruction
5134 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5141 <result> = insertvalue <aggregate type> <val>, <ty> <elt>, <idx>{, <idx>}* ; yields <aggregate type>
5146 The '``insertvalue``' instruction inserts a value into a member field in
5147 an :ref:`aggregate <t_aggregate>` value.
5152 The first operand of an '``insertvalue``' instruction is a value of
5153 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The second operand is
5154 a first-class value to insert. The following operands are constant
5155 indices indicating the position at which to insert the value in a
5156 similar manner as indices in a '``extractvalue``' instruction. The value
5157 to insert must have the same type as the value identified by the
5163 The result is an aggregate of the same type as ``val``. Its value is
5164 that of ``val`` except that the value at the position specified by the
5165 indices is that of ``elt``.
5170 .. code-block:: llvm
5172 %agg1 = insertvalue {i32, float} undef, i32 1, 0 ; yields {i32 1, float undef}
5173 %agg2 = insertvalue {i32, float} %agg1, float %val, 1 ; yields {i32 1, float %val}
5174 %agg3 = insertvalue {i32, {float}} undef, float %val, 1, 0 ; yields {i32 undef, {float %val}}
5178 Memory Access and Addressing Operations
5179 ---------------------------------------
5181 A key design point of an SSA-based representation is how it represents
5182 memory. In LLVM, no memory locations are in SSA form, which makes things
5183 very simple. This section describes how to read, write, and allocate
5188 '``alloca``' Instruction
5189 ^^^^^^^^^^^^^^^^^^^^^^^^
5196 <result> = alloca [inalloca] <type> [, <ty> <NumElements>] [, align <alignment>] ; yields type*:result
5201 The '``alloca``' instruction allocates memory on the stack frame of the
5202 currently executing function, to be automatically released when this
5203 function returns to its caller. The object is always allocated in the
5204 generic address space (address space zero).
5209 The '``alloca``' instruction allocates ``sizeof(<type>)*NumElements``
5210 bytes of memory on the runtime stack, returning a pointer of the
5211 appropriate type to the program. If "NumElements" is specified, it is
5212 the number of elements allocated, otherwise "NumElements" is defaulted
5213 to be one. If a constant alignment is specified, the value result of the
5214 allocation is guaranteed to be aligned to at least that boundary. The
5215 alignment may not be greater than ``1 << 29``. If not specified, or if
5216 zero, the target can choose to align the allocation on any convenient
5217 boundary compatible with the type.
5219 '``type``' may be any sized type.
5224 Memory is allocated; a pointer is returned. The operation is undefined
5225 if there is insufficient stack space for the allocation. '``alloca``'d
5226 memory is automatically released when the function returns. The
5227 '``alloca``' instruction is commonly used to represent automatic
5228 variables that must have an address available. When the function returns
5229 (either with the ``ret`` or ``resume`` instructions), the memory is
5230 reclaimed. Allocating zero bytes is legal, but the result is undefined.
5231 The order in which memory is allocated (ie., which way the stack grows)
5237 .. code-block:: llvm
5239 %ptr = alloca i32 ; yields i32*:ptr
5240 %ptr = alloca i32, i32 4 ; yields i32*:ptr
5241 %ptr = alloca i32, i32 4, align 1024 ; yields i32*:ptr
5242 %ptr = alloca i32, align 1024 ; yields i32*:ptr
5246 '``load``' Instruction
5247 ^^^^^^^^^^^^^^^^^^^^^^
5254 <result> = load [volatile] <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>][, !invariant.load !<index>][, !nonnull !<index>]
5255 <result> = load atomic [volatile] <ty>* <pointer> [singlethread] <ordering>, align <alignment>
5256 !<index> = !{ i32 1 }
5261 The '``load``' instruction is used to read from memory.
5266 The argument to the ``load`` instruction specifies the memory address
5267 from which to load. The pointer must point to a :ref:`first
5268 class <t_firstclass>` type. If the ``load`` is marked as ``volatile``,
5269 then the optimizer is not allowed to modify the number or order of
5270 execution of this ``load`` with other :ref:`volatile
5271 operations <volatile>`.
5273 If the ``load`` is marked as ``atomic``, it takes an extra
5274 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
5275 ``release`` and ``acq_rel`` orderings are not valid on ``load``
5276 instructions. Atomic loads produce :ref:`defined <memmodel>` results
5277 when they may see multiple atomic stores. The type of the pointee must
5278 be an integer type whose bit width is a power of two greater than or
5279 equal to eight and less than or equal to a target-specific size limit.
5280 ``align`` must be explicitly specified on atomic loads, and the load has
5281 undefined behavior if the alignment is not set to a value which is at
5282 least the size in bytes of the pointee. ``!nontemporal`` does not have
5283 any defined semantics for atomic loads.
5285 The optional constant ``align`` argument specifies the alignment of the
5286 operation (that is, the alignment of the memory address). A value of 0
5287 or an omitted ``align`` argument means that the operation has the ABI
5288 alignment for the target. It is the responsibility of the code emitter
5289 to ensure that the alignment information is correct. Overestimating the
5290 alignment results in undefined behavior. Underestimating the alignment
5291 may produce less efficient code. An alignment of 1 is always safe. The
5292 maximum possible alignment is ``1 << 29``.
5294 The optional ``!nontemporal`` metadata must reference a single
5295 metadata name ``<index>`` corresponding to a metadata node with one
5296 ``i32`` entry of value 1. The existence of the ``!nontemporal``
5297 metadata on the instruction tells the optimizer and code generator
5298 that this load is not expected to be reused in the cache. The code
5299 generator may select special instructions to save cache bandwidth, such
5300 as the ``MOVNT`` instruction on x86.
5302 The optional ``!invariant.load`` metadata must reference a single
5303 metadata name ``<index>`` corresponding to a metadata node with no
5304 entries. The existence of the ``!invariant.load`` metadata on the
5305 instruction tells the optimizer and code generator that the address
5306 operand to this load points to memory which can be assumed unchanged.
5307 Being invariant does not imply that a location is dereferenceable,
5308 but it does imply that once the location is known dereferenceable
5309 its value is henceforth unchanging.
5311 The optional ``!nonnull`` metadata must reference a single
5312 metadata name ``<index>`` corresponding to a metadata node with no
5313 entries. The existence of the ``!nonnull`` metadata on the
5314 instruction tells the optimizer that the value loaded is known to
5315 never be null. This is analogous to the ''nonnull'' attribute
5316 on parameters and return values. This metadata can only be applied
5317 to loads of a pointer type.
5322 The location of memory pointed to is loaded. If the value being loaded
5323 is of scalar type then the number of bytes read does not exceed the
5324 minimum number of bytes needed to hold all bits of the type. For
5325 example, loading an ``i24`` reads at most three bytes. When loading a
5326 value of a type like ``i20`` with a size that is not an integral number
5327 of bytes, the result is undefined if the value was not originally
5328 written using a store of the same type.
5333 .. code-block:: llvm
5335 %ptr = alloca i32 ; yields i32*:ptr
5336 store i32 3, i32* %ptr ; yields void
5337 %val = load i32* %ptr ; yields i32:val = i32 3
5341 '``store``' Instruction
5342 ^^^^^^^^^^^^^^^^^^^^^^^
5349 store [volatile] <ty> <value>, <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>] ; yields void
5350 store atomic [volatile] <ty> <value>, <ty>* <pointer> [singlethread] <ordering>, align <alignment> ; yields void
5355 The '``store``' instruction is used to write to memory.
5360 There are two arguments to the ``store`` instruction: a value to store
5361 and an address at which to store it. The type of the ``<pointer>``
5362 operand must be a pointer to the :ref:`first class <t_firstclass>` type of
5363 the ``<value>`` operand. If the ``store`` is marked as ``volatile``,
5364 then the optimizer is not allowed to modify the number or order of
5365 execution of this ``store`` with other :ref:`volatile
5366 operations <volatile>`.
5368 If the ``store`` is marked as ``atomic``, it takes an extra
5369 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
5370 ``acquire`` and ``acq_rel`` orderings aren't valid on ``store``
5371 instructions. Atomic loads produce :ref:`defined <memmodel>` results
5372 when they may see multiple atomic stores. The type of the pointee must
5373 be an integer type whose bit width is a power of two greater than or
5374 equal to eight and less than or equal to a target-specific size limit.
5375 ``align`` must be explicitly specified on atomic stores, and the store
5376 has undefined behavior if the alignment is not set to a value which is
5377 at least the size in bytes of the pointee. ``!nontemporal`` does not
5378 have any defined semantics for atomic stores.
5380 The optional constant ``align`` argument specifies the alignment of the
5381 operation (that is, the alignment of the memory address). A value of 0
5382 or an omitted ``align`` argument means that the operation has the ABI
5383 alignment for the target. It is the responsibility of the code emitter
5384 to ensure that the alignment information is correct. Overestimating the
5385 alignment results in undefined behavior. Underestimating the
5386 alignment may produce less efficient code. An alignment of 1 is always
5387 safe. The maximum possible alignment is ``1 << 29``.
5389 The optional ``!nontemporal`` metadata must reference a single metadata
5390 name ``<index>`` corresponding to a metadata node with one ``i32`` entry of
5391 value 1. The existence of the ``!nontemporal`` metadata on the instruction
5392 tells the optimizer and code generator that this load is not expected to
5393 be reused in the cache. The code generator may select special
5394 instructions to save cache bandwidth, such as the MOVNT instruction on
5400 The contents of memory are updated to contain ``<value>`` at the
5401 location specified by the ``<pointer>`` operand. If ``<value>`` is
5402 of scalar type then the number of bytes written does not exceed the
5403 minimum number of bytes needed to hold all bits of the type. For
5404 example, storing an ``i24`` writes at most three bytes. When writing a
5405 value of a type like ``i20`` with a size that is not an integral number
5406 of bytes, it is unspecified what happens to the extra bits that do not
5407 belong to the type, but they will typically be overwritten.
5412 .. code-block:: llvm
5414 %ptr = alloca i32 ; yields i32*:ptr
5415 store i32 3, i32* %ptr ; yields void
5416 %val = load i32* %ptr ; yields i32:val = i32 3
5420 '``fence``' Instruction
5421 ^^^^^^^^^^^^^^^^^^^^^^^
5428 fence [singlethread] <ordering> ; yields void
5433 The '``fence``' instruction is used to introduce happens-before edges
5439 '``fence``' instructions take an :ref:`ordering <ordering>` argument which
5440 defines what *synchronizes-with* edges they add. They can only be given
5441 ``acquire``, ``release``, ``acq_rel``, and ``seq_cst`` orderings.
5446 A fence A which has (at least) ``release`` ordering semantics
5447 *synchronizes with* a fence B with (at least) ``acquire`` ordering
5448 semantics if and only if there exist atomic operations X and Y, both
5449 operating on some atomic object M, such that A is sequenced before X, X
5450 modifies M (either directly or through some side effect of a sequence
5451 headed by X), Y is sequenced before B, and Y observes M. This provides a
5452 *happens-before* dependency between A and B. Rather than an explicit
5453 ``fence``, one (but not both) of the atomic operations X or Y might
5454 provide a ``release`` or ``acquire`` (resp.) ordering constraint and
5455 still *synchronize-with* the explicit ``fence`` and establish the
5456 *happens-before* edge.
5458 A ``fence`` which has ``seq_cst`` ordering, in addition to having both
5459 ``acquire`` and ``release`` semantics specified above, participates in
5460 the global program order of other ``seq_cst`` operations and/or fences.
5462 The optional ":ref:`singlethread <singlethread>`" argument specifies
5463 that the fence only synchronizes with other fences in the same thread.
5464 (This is useful for interacting with signal handlers.)
5469 .. code-block:: llvm
5471 fence acquire ; yields void
5472 fence singlethread seq_cst ; yields void
5476 '``cmpxchg``' Instruction
5477 ^^^^^^^^^^^^^^^^^^^^^^^^^
5484 cmpxchg [weak] [volatile] <ty>* <pointer>, <ty> <cmp>, <ty> <new> [singlethread] <success ordering> <failure ordering> ; yields { ty, i1 }
5489 The '``cmpxchg``' instruction is used to atomically modify memory. It
5490 loads a value in memory and compares it to a given value. If they are
5491 equal, it tries to store a new value into the memory.
5496 There are three arguments to the '``cmpxchg``' instruction: an address
5497 to operate on, a value to compare to the value currently be at that
5498 address, and a new value to place at that address if the compared values
5499 are equal. The type of '<cmp>' must be an integer type whose bit width
5500 is a power of two greater than or equal to eight and less than or equal
5501 to a target-specific size limit. '<cmp>' and '<new>' must have the same
5502 type, and the type of '<pointer>' must be a pointer to that type. If the
5503 ``cmpxchg`` is marked as ``volatile``, then the optimizer is not allowed
5504 to modify the number or order of execution of this ``cmpxchg`` with
5505 other :ref:`volatile operations <volatile>`.
5507 The success and failure :ref:`ordering <ordering>` arguments specify how this
5508 ``cmpxchg`` synchronizes with other atomic operations. Both ordering parameters
5509 must be at least ``monotonic``, the ordering constraint on failure must be no
5510 stronger than that on success, and the failure ordering cannot be either
5511 ``release`` or ``acq_rel``.
5513 The optional "``singlethread``" argument declares that the ``cmpxchg``
5514 is only atomic with respect to code (usually signal handlers) running in
5515 the same thread as the ``cmpxchg``. Otherwise the cmpxchg is atomic with
5516 respect to all other code in the system.
5518 The pointer passed into cmpxchg must have alignment greater than or
5519 equal to the size in memory of the operand.
5524 The contents of memory at the location specified by the '``<pointer>``' operand
5525 is read and compared to '``<cmp>``'; if the read value is the equal, the
5526 '``<new>``' is written. The original value at the location is returned, together
5527 with a flag indicating success (true) or failure (false).
5529 If the cmpxchg operation is marked as ``weak`` then a spurious failure is
5530 permitted: the operation may not write ``<new>`` even if the comparison
5533 If the cmpxchg operation is strong (the default), the i1 value is 1 if and only
5534 if the value loaded equals ``cmp``.
5536 A successful ``cmpxchg`` is a read-modify-write instruction for the purpose of
5537 identifying release sequences. A failed ``cmpxchg`` is equivalent to an atomic
5538 load with an ordering parameter determined the second ordering parameter.
5543 .. code-block:: llvm
5546 %orig = atomic load i32* %ptr unordered ; yields i32
5550 %cmp = phi i32 [ %orig, %entry ], [%old, %loop]
5551 %squared = mul i32 %cmp, %cmp
5552 %val_success = cmpxchg i32* %ptr, i32 %cmp, i32 %squared acq_rel monotonic ; yields { i32, i1 }
5553 %value_loaded = extractvalue { i32, i1 } %val_success, 0
5554 %success = extractvalue { i32, i1 } %val_success, 1
5555 br i1 %success, label %done, label %loop
5562 '``atomicrmw``' Instruction
5563 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
5570 atomicrmw [volatile] <operation> <ty>* <pointer>, <ty> <value> [singlethread] <ordering> ; yields ty
5575 The '``atomicrmw``' instruction is used to atomically modify memory.
5580 There are three arguments to the '``atomicrmw``' instruction: an
5581 operation to apply, an address whose value to modify, an argument to the
5582 operation. The operation must be one of the following keywords:
5596 The type of '<value>' must be an integer type whose bit width is a power
5597 of two greater than or equal to eight and less than or equal to a
5598 target-specific size limit. The type of the '``<pointer>``' operand must
5599 be a pointer to that type. If the ``atomicrmw`` is marked as
5600 ``volatile``, then the optimizer is not allowed to modify the number or
5601 order of execution of this ``atomicrmw`` with other :ref:`volatile
5602 operations <volatile>`.
5607 The contents of memory at the location specified by the '``<pointer>``'
5608 operand are atomically read, modified, and written back. The original
5609 value at the location is returned. The modification is specified by the
5612 - xchg: ``*ptr = val``
5613 - add: ``*ptr = *ptr + val``
5614 - sub: ``*ptr = *ptr - val``
5615 - and: ``*ptr = *ptr & val``
5616 - nand: ``*ptr = ~(*ptr & val)``
5617 - or: ``*ptr = *ptr | val``
5618 - xor: ``*ptr = *ptr ^ val``
5619 - max: ``*ptr = *ptr > val ? *ptr : val`` (using a signed comparison)
5620 - min: ``*ptr = *ptr < val ? *ptr : val`` (using a signed comparison)
5621 - umax: ``*ptr = *ptr > val ? *ptr : val`` (using an unsigned
5623 - umin: ``*ptr = *ptr < val ? *ptr : val`` (using an unsigned
5629 .. code-block:: llvm
5631 %old = atomicrmw add i32* %ptr, i32 1 acquire ; yields i32
5633 .. _i_getelementptr:
5635 '``getelementptr``' Instruction
5636 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5643 <result> = getelementptr <pty>* <ptrval>{, <ty> <idx>}*
5644 <result> = getelementptr inbounds <pty>* <ptrval>{, <ty> <idx>}*
5645 <result> = getelementptr <ptr vector> ptrval, <vector index type> idx
5650 The '``getelementptr``' instruction is used to get the address of a
5651 subelement of an :ref:`aggregate <t_aggregate>` data structure. It performs
5652 address calculation only and does not access memory.
5657 The first argument is always a pointer or a vector of pointers, and
5658 forms the basis of the calculation. The remaining arguments are indices
5659 that indicate which of the elements of the aggregate object are indexed.
5660 The interpretation of each index is dependent on the type being indexed
5661 into. The first index always indexes the pointer value given as the
5662 first argument, the second index indexes a value of the type pointed to
5663 (not necessarily the value directly pointed to, since the first index
5664 can be non-zero), etc. The first type indexed into must be a pointer
5665 value, subsequent types can be arrays, vectors, and structs. Note that
5666 subsequent types being indexed into can never be pointers, since that
5667 would require loading the pointer before continuing calculation.
5669 The type of each index argument depends on the type it is indexing into.
5670 When indexing into a (optionally packed) structure, only ``i32`` integer
5671 **constants** are allowed (when using a vector of indices they must all
5672 be the **same** ``i32`` integer constant). When indexing into an array,
5673 pointer or vector, integers of any width are allowed, and they are not
5674 required to be constant. These integers are treated as signed values
5677 For example, let's consider a C code fragment and how it gets compiled
5693 int *foo(struct ST *s) {
5694 return &s[1].Z.B[5][13];
5697 The LLVM code generated by Clang is:
5699 .. code-block:: llvm
5701 %struct.RT = type { i8, [10 x [20 x i32]], i8 }
5702 %struct.ST = type { i32, double, %struct.RT }
5704 define i32* @foo(%struct.ST* %s) nounwind uwtable readnone optsize ssp {
5706 %arrayidx = getelementptr inbounds %struct.ST* %s, i64 1, i32 2, i32 1, i64 5, i64 13
5713 In the example above, the first index is indexing into the
5714 '``%struct.ST*``' type, which is a pointer, yielding a '``%struct.ST``'
5715 = '``{ i32, double, %struct.RT }``' type, a structure. The second index
5716 indexes into the third element of the structure, yielding a
5717 '``%struct.RT``' = '``{ i8 , [10 x [20 x i32]], i8 }``' type, another
5718 structure. The third index indexes into the second element of the
5719 structure, yielding a '``[10 x [20 x i32]]``' type, an array. The two
5720 dimensions of the array are subscripted into, yielding an '``i32``'
5721 type. The '``getelementptr``' instruction returns a pointer to this
5722 element, thus computing a value of '``i32*``' type.
5724 Note that it is perfectly legal to index partially through a structure,
5725 returning a pointer to an inner element. Because of this, the LLVM code
5726 for the given testcase is equivalent to:
5728 .. code-block:: llvm
5730 define i32* @foo(%struct.ST* %s) {
5731 %t1 = getelementptr %struct.ST* %s, i32 1 ; yields %struct.ST*:%t1
5732 %t2 = getelementptr %struct.ST* %t1, i32 0, i32 2 ; yields %struct.RT*:%t2
5733 %t3 = getelementptr %struct.RT* %t2, i32 0, i32 1 ; yields [10 x [20 x i32]]*:%t3
5734 %t4 = getelementptr [10 x [20 x i32]]* %t3, i32 0, i32 5 ; yields [20 x i32]*:%t4
5735 %t5 = getelementptr [20 x i32]* %t4, i32 0, i32 13 ; yields i32*:%t5
5739 If the ``inbounds`` keyword is present, the result value of the
5740 ``getelementptr`` is a :ref:`poison value <poisonvalues>` if the base
5741 pointer is not an *in bounds* address of an allocated object, or if any
5742 of the addresses that would be formed by successive addition of the
5743 offsets implied by the indices to the base address with infinitely
5744 precise signed arithmetic are not an *in bounds* address of that
5745 allocated object. The *in bounds* addresses for an allocated object are
5746 all the addresses that point into the object, plus the address one byte
5747 past the end. In cases where the base is a vector of pointers the
5748 ``inbounds`` keyword applies to each of the computations element-wise.
5750 If the ``inbounds`` keyword is not present, the offsets are added to the
5751 base address with silently-wrapping two's complement arithmetic. If the
5752 offsets have a different width from the pointer, they are sign-extended
5753 or truncated to the width of the pointer. The result value of the
5754 ``getelementptr`` may be outside the object pointed to by the base
5755 pointer. The result value may not necessarily be used to access memory
5756 though, even if it happens to point into allocated storage. See the
5757 :ref:`Pointer Aliasing Rules <pointeraliasing>` section for more
5760 The getelementptr instruction is often confusing. For some more insight
5761 into how it works, see :doc:`the getelementptr FAQ <GetElementPtr>`.
5766 .. code-block:: llvm
5768 ; yields [12 x i8]*:aptr
5769 %aptr = getelementptr {i32, [12 x i8]}* %saptr, i64 0, i32 1
5771 %vptr = getelementptr {i32, <2 x i8>}* %svptr, i64 0, i32 1, i32 1
5773 %eptr = getelementptr [12 x i8]* %aptr, i64 0, i32 1
5775 %iptr = getelementptr [10 x i32]* @arr, i16 0, i16 0
5777 In cases where the pointer argument is a vector of pointers, each index
5778 must be a vector with the same number of elements. For example:
5780 .. code-block:: llvm
5782 %A = getelementptr <4 x i8*> %ptrs, <4 x i64> %offsets,
5784 Conversion Operations
5785 ---------------------
5787 The instructions in this category are the conversion instructions
5788 (casting) which all take a single operand and a type. They perform
5789 various bit conversions on the operand.
5791 '``trunc .. to``' Instruction
5792 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5799 <result> = trunc <ty> <value> to <ty2> ; yields ty2
5804 The '``trunc``' instruction truncates its operand to the type ``ty2``.
5809 The '``trunc``' instruction takes a value to trunc, and a type to trunc
5810 it to. Both types must be of :ref:`integer <t_integer>` types, or vectors
5811 of the same number of integers. The bit size of the ``value`` must be
5812 larger than the bit size of the destination type, ``ty2``. Equal sized
5813 types are not allowed.
5818 The '``trunc``' instruction truncates the high order bits in ``value``
5819 and converts the remaining bits to ``ty2``. Since the source size must
5820 be larger than the destination size, ``trunc`` cannot be a *no-op cast*.
5821 It will always truncate bits.
5826 .. code-block:: llvm
5828 %X = trunc i32 257 to i8 ; yields i8:1
5829 %Y = trunc i32 123 to i1 ; yields i1:true
5830 %Z = trunc i32 122 to i1 ; yields i1:false
5831 %W = trunc <2 x i16> <i16 8, i16 7> to <2 x i8> ; yields <i8 8, i8 7>
5833 '``zext .. to``' Instruction
5834 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5841 <result> = zext <ty> <value> to <ty2> ; yields ty2
5846 The '``zext``' instruction zero extends its operand to type ``ty2``.
5851 The '``zext``' instruction takes a value to cast, and a type to cast it
5852 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
5853 the same number of integers. The bit size of the ``value`` must be
5854 smaller than the bit size of the destination type, ``ty2``.
5859 The ``zext`` fills the high order bits of the ``value`` with zero bits
5860 until it reaches the size of the destination type, ``ty2``.
5862 When zero extending from i1, the result will always be either 0 or 1.
5867 .. code-block:: llvm
5869 %X = zext i32 257 to i64 ; yields i64:257
5870 %Y = zext i1 true to i32 ; yields i32:1
5871 %Z = zext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
5873 '``sext .. to``' Instruction
5874 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5881 <result> = sext <ty> <value> to <ty2> ; yields ty2
5886 The '``sext``' sign extends ``value`` to the type ``ty2``.
5891 The '``sext``' instruction takes a value to cast, and a type to cast it
5892 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
5893 the same number of integers. The bit size of the ``value`` must be
5894 smaller than the bit size of the destination type, ``ty2``.
5899 The '``sext``' instruction performs a sign extension by copying the sign
5900 bit (highest order bit) of the ``value`` until it reaches the bit size
5901 of the type ``ty2``.
5903 When sign extending from i1, the extension always results in -1 or 0.
5908 .. code-block:: llvm
5910 %X = sext i8 -1 to i16 ; yields i16 :65535
5911 %Y = sext i1 true to i32 ; yields i32:-1
5912 %Z = sext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
5914 '``fptrunc .. to``' Instruction
5915 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5922 <result> = fptrunc <ty> <value> to <ty2> ; yields ty2
5927 The '``fptrunc``' instruction truncates ``value`` to type ``ty2``.
5932 The '``fptrunc``' instruction takes a :ref:`floating point <t_floating>`
5933 value to cast and a :ref:`floating point <t_floating>` type to cast it to.
5934 The size of ``value`` must be larger than the size of ``ty2``. This
5935 implies that ``fptrunc`` cannot be used to make a *no-op cast*.
5940 The '``fptrunc``' instruction truncates a ``value`` from a larger
5941 :ref:`floating point <t_floating>` type to a smaller :ref:`floating
5942 point <t_floating>` type. If the value cannot fit within the
5943 destination type, ``ty2``, then the results are undefined.
5948 .. code-block:: llvm
5950 %X = fptrunc double 123.0 to float ; yields float:123.0
5951 %Y = fptrunc double 1.0E+300 to float ; yields undefined
5953 '``fpext .. to``' Instruction
5954 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5961 <result> = fpext <ty> <value> to <ty2> ; yields ty2
5966 The '``fpext``' extends a floating point ``value`` to a larger floating
5972 The '``fpext``' instruction takes a :ref:`floating point <t_floating>`
5973 ``value`` to cast, and a :ref:`floating point <t_floating>` type to cast it
5974 to. The source type must be smaller than the destination type.
5979 The '``fpext``' instruction extends the ``value`` from a smaller
5980 :ref:`floating point <t_floating>` type to a larger :ref:`floating
5981 point <t_floating>` type. The ``fpext`` cannot be used to make a
5982 *no-op cast* because it always changes bits. Use ``bitcast`` to make a
5983 *no-op cast* for a floating point cast.
5988 .. code-block:: llvm
5990 %X = fpext float 3.125 to double ; yields double:3.125000e+00
5991 %Y = fpext double %X to fp128 ; yields fp128:0xL00000000000000004000900000000000
5993 '``fptoui .. to``' Instruction
5994 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6001 <result> = fptoui <ty> <value> to <ty2> ; yields ty2
6006 The '``fptoui``' converts a floating point ``value`` to its unsigned
6007 integer equivalent of type ``ty2``.
6012 The '``fptoui``' instruction takes a value to cast, which must be a
6013 scalar or vector :ref:`floating point <t_floating>` value, and a type to
6014 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
6015 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
6016 type with the same number of elements as ``ty``
6021 The '``fptoui``' instruction converts its :ref:`floating
6022 point <t_floating>` operand into the nearest (rounding towards zero)
6023 unsigned integer value. If the value cannot fit in ``ty2``, the results
6029 .. code-block:: llvm
6031 %X = fptoui double 123.0 to i32 ; yields i32:123
6032 %Y = fptoui float 1.0E+300 to i1 ; yields undefined:1
6033 %Z = fptoui float 1.04E+17 to i8 ; yields undefined:1
6035 '``fptosi .. to``' Instruction
6036 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6043 <result> = fptosi <ty> <value> to <ty2> ; yields ty2
6048 The '``fptosi``' instruction converts :ref:`floating point <t_floating>`
6049 ``value`` to type ``ty2``.
6054 The '``fptosi``' instruction takes a value to cast, which must be a
6055 scalar or vector :ref:`floating point <t_floating>` value, and a type to
6056 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
6057 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
6058 type with the same number of elements as ``ty``
6063 The '``fptosi``' instruction converts its :ref:`floating
6064 point <t_floating>` operand into the nearest (rounding towards zero)
6065 signed integer value. If the value cannot fit in ``ty2``, the results
6071 .. code-block:: llvm
6073 %X = fptosi double -123.0 to i32 ; yields i32:-123
6074 %Y = fptosi float 1.0E-247 to i1 ; yields undefined:1
6075 %Z = fptosi float 1.04E+17 to i8 ; yields undefined:1
6077 '``uitofp .. to``' Instruction
6078 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6085 <result> = uitofp <ty> <value> to <ty2> ; yields ty2
6090 The '``uitofp``' instruction regards ``value`` as an unsigned integer
6091 and converts that value to the ``ty2`` type.
6096 The '``uitofp``' instruction takes a value to cast, which must be a
6097 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
6098 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
6099 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
6100 type with the same number of elements as ``ty``
6105 The '``uitofp``' instruction interprets its operand as an unsigned
6106 integer quantity and converts it to the corresponding floating point
6107 value. If the value cannot fit in the floating point value, the results
6113 .. code-block:: llvm
6115 %X = uitofp i32 257 to float ; yields float:257.0
6116 %Y = uitofp i8 -1 to double ; yields double:255.0
6118 '``sitofp .. to``' Instruction
6119 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6126 <result> = sitofp <ty> <value> to <ty2> ; yields ty2
6131 The '``sitofp``' instruction regards ``value`` as a signed integer and
6132 converts that value to the ``ty2`` type.
6137 The '``sitofp``' instruction takes a value to cast, which must be a
6138 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
6139 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
6140 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
6141 type with the same number of elements as ``ty``
6146 The '``sitofp``' instruction interprets its operand as a signed integer
6147 quantity and converts it to the corresponding floating point value. If
6148 the value cannot fit in the floating point value, the results are
6154 .. code-block:: llvm
6156 %X = sitofp i32 257 to float ; yields float:257.0
6157 %Y = sitofp i8 -1 to double ; yields double:-1.0
6161 '``ptrtoint .. to``' Instruction
6162 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6169 <result> = ptrtoint <ty> <value> to <ty2> ; yields ty2
6174 The '``ptrtoint``' instruction converts the pointer or a vector of
6175 pointers ``value`` to the integer (or vector of integers) type ``ty2``.
6180 The '``ptrtoint``' instruction takes a ``value`` to cast, which must be
6181 a a value of type :ref:`pointer <t_pointer>` or a vector of pointers, and a
6182 type to cast it to ``ty2``, which must be an :ref:`integer <t_integer>` or
6183 a vector of integers type.
6188 The '``ptrtoint``' instruction converts ``value`` to integer type
6189 ``ty2`` by interpreting the pointer value as an integer and either
6190 truncating or zero extending that value to the size of the integer type.
6191 If ``value`` is smaller than ``ty2`` then a zero extension is done. If
6192 ``value`` is larger than ``ty2`` then a truncation is done. If they are
6193 the same size, then nothing is done (*no-op cast*) other than a type
6199 .. code-block:: llvm
6201 %X = ptrtoint i32* %P to i8 ; yields truncation on 32-bit architecture
6202 %Y = ptrtoint i32* %P to i64 ; yields zero extension on 32-bit architecture
6203 %Z = ptrtoint <4 x i32*> %P to <4 x i64>; yields vector zero extension for a vector of addresses on 32-bit architecture
6207 '``inttoptr .. to``' Instruction
6208 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6215 <result> = inttoptr <ty> <value> to <ty2> ; yields ty2
6220 The '``inttoptr``' instruction converts an integer ``value`` to a
6221 pointer type, ``ty2``.
6226 The '``inttoptr``' instruction takes an :ref:`integer <t_integer>` value to
6227 cast, and a type to cast it to, which must be a :ref:`pointer <t_pointer>`
6233 The '``inttoptr``' instruction converts ``value`` to type ``ty2`` by
6234 applying either a zero extension or a truncation depending on the size
6235 of the integer ``value``. If ``value`` is larger than the size of a
6236 pointer then a truncation is done. If ``value`` is smaller than the size
6237 of a pointer then a zero extension is done. If they are the same size,
6238 nothing is done (*no-op cast*).
6243 .. code-block:: llvm
6245 %X = inttoptr i32 255 to i32* ; yields zero extension on 64-bit architecture
6246 %Y = inttoptr i32 255 to i32* ; yields no-op on 32-bit architecture
6247 %Z = inttoptr i64 0 to i32* ; yields truncation on 32-bit architecture
6248 %Z = inttoptr <4 x i32> %G to <4 x i8*>; yields truncation of vector G to four pointers
6252 '``bitcast .. to``' Instruction
6253 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6260 <result> = bitcast <ty> <value> to <ty2> ; yields ty2
6265 The '``bitcast``' instruction converts ``value`` to type ``ty2`` without
6271 The '``bitcast``' instruction takes a value to cast, which must be a
6272 non-aggregate first class value, and a type to cast it to, which must
6273 also be a non-aggregate :ref:`first class <t_firstclass>` type. The
6274 bit sizes of ``value`` and the destination type, ``ty2``, must be
6275 identical. If the source type is a pointer, the destination type must
6276 also be a pointer of the same size. This instruction supports bitwise
6277 conversion of vectors to integers and to vectors of other types (as
6278 long as they have the same size).
6283 The '``bitcast``' instruction converts ``value`` to type ``ty2``. It
6284 is always a *no-op cast* because no bits change with this
6285 conversion. The conversion is done as if the ``value`` had been stored
6286 to memory and read back as type ``ty2``. Pointer (or vector of
6287 pointers) types may only be converted to other pointer (or vector of
6288 pointers) types with the same address space through this instruction.
6289 To convert pointers to other types, use the :ref:`inttoptr <i_inttoptr>`
6290 or :ref:`ptrtoint <i_ptrtoint>` instructions first.
6295 .. code-block:: llvm
6297 %X = bitcast i8 255 to i8 ; yields i8 :-1
6298 %Y = bitcast i32* %x to sint* ; yields sint*:%x
6299 %Z = bitcast <2 x int> %V to i64; ; yields i64: %V
6300 %Z = bitcast <2 x i32*> %V to <2 x i64*> ; yields <2 x i64*>
6302 .. _i_addrspacecast:
6304 '``addrspacecast .. to``' Instruction
6305 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6312 <result> = addrspacecast <pty> <ptrval> to <pty2> ; yields pty2
6317 The '``addrspacecast``' instruction converts ``ptrval`` from ``pty`` in
6318 address space ``n`` to type ``pty2`` in address space ``m``.
6323 The '``addrspacecast``' instruction takes a pointer or vector of pointer value
6324 to cast and a pointer type to cast it to, which must have a different
6330 The '``addrspacecast``' instruction converts the pointer value
6331 ``ptrval`` to type ``pty2``. It can be a *no-op cast* or a complex
6332 value modification, depending on the target and the address space
6333 pair. Pointer conversions within the same address space must be
6334 performed with the ``bitcast`` instruction. Note that if the address space
6335 conversion is legal then both result and operand refer to the same memory
6341 .. code-block:: llvm
6343 %X = addrspacecast i32* %x to i32 addrspace(1)* ; yields i32 addrspace(1)*:%x
6344 %Y = addrspacecast i32 addrspace(1)* %y to i64 addrspace(2)* ; yields i64 addrspace(2)*:%y
6345 %Z = addrspacecast <4 x i32*> %z to <4 x float addrspace(3)*> ; yields <4 x float addrspace(3)*>:%z
6352 The instructions in this category are the "miscellaneous" instructions,
6353 which defy better classification.
6357 '``icmp``' Instruction
6358 ^^^^^^^^^^^^^^^^^^^^^^
6365 <result> = icmp <cond> <ty> <op1>, <op2> ; yields i1 or <N x i1>:result
6370 The '``icmp``' instruction returns a boolean value or a vector of
6371 boolean values based on comparison of its two integer, integer vector,
6372 pointer, or pointer vector operands.
6377 The '``icmp``' instruction takes three operands. The first operand is
6378 the condition code indicating the kind of comparison to perform. It is
6379 not a value, just a keyword. The possible condition code are:
6382 #. ``ne``: not equal
6383 #. ``ugt``: unsigned greater than
6384 #. ``uge``: unsigned greater or equal
6385 #. ``ult``: unsigned less than
6386 #. ``ule``: unsigned less or equal
6387 #. ``sgt``: signed greater than
6388 #. ``sge``: signed greater or equal
6389 #. ``slt``: signed less than
6390 #. ``sle``: signed less or equal
6392 The remaining two arguments must be :ref:`integer <t_integer>` or
6393 :ref:`pointer <t_pointer>` or integer :ref:`vector <t_vector>` typed. They
6394 must also be identical types.
6399 The '``icmp``' compares ``op1`` and ``op2`` according to the condition
6400 code given as ``cond``. The comparison performed always yields either an
6401 :ref:`i1 <t_integer>` or vector of ``i1`` result, as follows:
6403 #. ``eq``: yields ``true`` if the operands are equal, ``false``
6404 otherwise. No sign interpretation is necessary or performed.
6405 #. ``ne``: yields ``true`` if the operands are unequal, ``false``
6406 otherwise. No sign interpretation is necessary or performed.
6407 #. ``ugt``: interprets the operands as unsigned values and yields
6408 ``true`` if ``op1`` is greater than ``op2``.
6409 #. ``uge``: interprets the operands as unsigned values and yields
6410 ``true`` if ``op1`` is greater than or equal to ``op2``.
6411 #. ``ult``: interprets the operands as unsigned values and yields
6412 ``true`` if ``op1`` is less than ``op2``.
6413 #. ``ule``: interprets the operands as unsigned values and yields
6414 ``true`` if ``op1`` is less than or equal to ``op2``.
6415 #. ``sgt``: interprets the operands as signed values and yields ``true``
6416 if ``op1`` is greater than ``op2``.
6417 #. ``sge``: interprets the operands as signed values and yields ``true``
6418 if ``op1`` is greater than or equal to ``op2``.
6419 #. ``slt``: interprets the operands as signed values and yields ``true``
6420 if ``op1`` is less than ``op2``.
6421 #. ``sle``: interprets the operands as signed values and yields ``true``
6422 if ``op1`` is less than or equal to ``op2``.
6424 If the operands are :ref:`pointer <t_pointer>` typed, the pointer values
6425 are compared as if they were integers.
6427 If the operands are integer vectors, then they are compared element by
6428 element. The result is an ``i1`` vector with the same number of elements
6429 as the values being compared. Otherwise, the result is an ``i1``.
6434 .. code-block:: llvm
6436 <result> = icmp eq i32 4, 5 ; yields: result=false
6437 <result> = icmp ne float* %X, %X ; yields: result=false
6438 <result> = icmp ult i16 4, 5 ; yields: result=true
6439 <result> = icmp sgt i16 4, 5 ; yields: result=false
6440 <result> = icmp ule i16 -4, 5 ; yields: result=false
6441 <result> = icmp sge i16 4, 5 ; yields: result=false
6443 Note that the code generator does not yet support vector types with the
6444 ``icmp`` instruction.
6448 '``fcmp``' Instruction
6449 ^^^^^^^^^^^^^^^^^^^^^^
6456 <result> = fcmp <cond> <ty> <op1>, <op2> ; yields i1 or <N x i1>:result
6461 The '``fcmp``' instruction returns a boolean value or vector of boolean
6462 values based on comparison of its operands.
6464 If the operands are floating point scalars, then the result type is a
6465 boolean (:ref:`i1 <t_integer>`).
6467 If the operands are floating point vectors, then the result type is a
6468 vector of boolean with the same number of elements as the operands being
6474 The '``fcmp``' instruction takes three operands. The first operand is
6475 the condition code indicating the kind of comparison to perform. It is
6476 not a value, just a keyword. The possible condition code are:
6478 #. ``false``: no comparison, always returns false
6479 #. ``oeq``: ordered and equal
6480 #. ``ogt``: ordered and greater than
6481 #. ``oge``: ordered and greater than or equal
6482 #. ``olt``: ordered and less than
6483 #. ``ole``: ordered and less than or equal
6484 #. ``one``: ordered and not equal
6485 #. ``ord``: ordered (no nans)
6486 #. ``ueq``: unordered or equal
6487 #. ``ugt``: unordered or greater than
6488 #. ``uge``: unordered or greater than or equal
6489 #. ``ult``: unordered or less than
6490 #. ``ule``: unordered or less than or equal
6491 #. ``une``: unordered or not equal
6492 #. ``uno``: unordered (either nans)
6493 #. ``true``: no comparison, always returns true
6495 *Ordered* means that neither operand is a QNAN while *unordered* means
6496 that either operand may be a QNAN.
6498 Each of ``val1`` and ``val2`` arguments must be either a :ref:`floating
6499 point <t_floating>` type or a :ref:`vector <t_vector>` of floating point
6500 type. They must have identical types.
6505 The '``fcmp``' instruction compares ``op1`` and ``op2`` according to the
6506 condition code given as ``cond``. If the operands are vectors, then the
6507 vectors are compared element by element. Each comparison performed
6508 always yields an :ref:`i1 <t_integer>` result, as follows:
6510 #. ``false``: always yields ``false``, regardless of operands.
6511 #. ``oeq``: yields ``true`` if both operands are not a QNAN and ``op1``
6512 is equal to ``op2``.
6513 #. ``ogt``: yields ``true`` if both operands are not a QNAN and ``op1``
6514 is greater than ``op2``.
6515 #. ``oge``: yields ``true`` if both operands are not a QNAN and ``op1``
6516 is greater than or equal to ``op2``.
6517 #. ``olt``: yields ``true`` if both operands are not a QNAN and ``op1``
6518 is less than ``op2``.
6519 #. ``ole``: yields ``true`` if both operands are not a QNAN and ``op1``
6520 is less than or equal to ``op2``.
6521 #. ``one``: yields ``true`` if both operands are not a QNAN and ``op1``
6522 is not equal to ``op2``.
6523 #. ``ord``: yields ``true`` if both operands are not a QNAN.
6524 #. ``ueq``: yields ``true`` if either operand is a QNAN or ``op1`` is
6526 #. ``ugt``: yields ``true`` if either operand is a QNAN or ``op1`` is
6527 greater than ``op2``.
6528 #. ``uge``: yields ``true`` if either operand is a QNAN or ``op1`` is
6529 greater than or equal to ``op2``.
6530 #. ``ult``: yields ``true`` if either operand is a QNAN or ``op1`` is
6532 #. ``ule``: yields ``true`` if either operand is a QNAN or ``op1`` is
6533 less than or equal to ``op2``.
6534 #. ``une``: yields ``true`` if either operand is a QNAN or ``op1`` is
6535 not equal to ``op2``.
6536 #. ``uno``: yields ``true`` if either operand is a QNAN.
6537 #. ``true``: always yields ``true``, regardless of operands.
6542 .. code-block:: llvm
6544 <result> = fcmp oeq float 4.0, 5.0 ; yields: result=false
6545 <result> = fcmp one float 4.0, 5.0 ; yields: result=true
6546 <result> = fcmp olt float 4.0, 5.0 ; yields: result=true
6547 <result> = fcmp ueq double 1.0, 2.0 ; yields: result=false
6549 Note that the code generator does not yet support vector types with the
6550 ``fcmp`` instruction.
6554 '``phi``' Instruction
6555 ^^^^^^^^^^^^^^^^^^^^^
6562 <result> = phi <ty> [ <val0>, <label0>], ...
6567 The '``phi``' instruction is used to implement the φ node in the SSA
6568 graph representing the function.
6573 The type of the incoming values is specified with the first type field.
6574 After this, the '``phi``' instruction takes a list of pairs as
6575 arguments, with one pair for each predecessor basic block of the current
6576 block. Only values of :ref:`first class <t_firstclass>` type may be used as
6577 the value arguments to the PHI node. Only labels may be used as the
6580 There must be no non-phi instructions between the start of a basic block
6581 and the PHI instructions: i.e. PHI instructions must be first in a basic
6584 For the purposes of the SSA form, the use of each incoming value is
6585 deemed to occur on the edge from the corresponding predecessor block to
6586 the current block (but after any definition of an '``invoke``'
6587 instruction's return value on the same edge).
6592 At runtime, the '``phi``' instruction logically takes on the value
6593 specified by the pair corresponding to the predecessor basic block that
6594 executed just prior to the current block.
6599 .. code-block:: llvm
6601 Loop: ; Infinite loop that counts from 0 on up...
6602 %indvar = phi i32 [ 0, %LoopHeader ], [ %nextindvar, %Loop ]
6603 %nextindvar = add i32 %indvar, 1
6608 '``select``' Instruction
6609 ^^^^^^^^^^^^^^^^^^^^^^^^
6616 <result> = select selty <cond>, <ty> <val1>, <ty> <val2> ; yields ty
6618 selty is either i1 or {<N x i1>}
6623 The '``select``' instruction is used to choose one value based on a
6624 condition, without IR-level branching.
6629 The '``select``' instruction requires an 'i1' value or a vector of 'i1'
6630 values indicating the condition, and two values of the same :ref:`first
6631 class <t_firstclass>` type. If the val1/val2 are vectors and the
6632 condition is a scalar, then entire vectors are selected, not individual
6638 If the condition is an i1 and it evaluates to 1, the instruction returns
6639 the first value argument; otherwise, it returns the second value
6642 If the condition is a vector of i1, then the value arguments must be
6643 vectors of the same size, and the selection is done element by element.
6648 .. code-block:: llvm
6650 %X = select i1 true, i8 17, i8 42 ; yields i8:17
6654 '``call``' Instruction
6655 ^^^^^^^^^^^^^^^^^^^^^^
6662 <result> = [tail | musttail] call [cconv] [ret attrs] <ty> [<fnty>*] <fnptrval>(<function args>) [fn attrs]
6667 The '``call``' instruction represents a simple function call.
6672 This instruction requires several arguments:
6674 #. The optional ``tail`` and ``musttail`` markers indicate that the optimizers
6675 should perform tail call optimization. The ``tail`` marker is a hint that
6676 `can be ignored <CodeGenerator.html#sibcallopt>`_. The ``musttail`` marker
6677 means that the call must be tail call optimized in order for the program to
6678 be correct. The ``musttail`` marker provides these guarantees:
6680 #. The call will not cause unbounded stack growth if it is part of a
6681 recursive cycle in the call graph.
6682 #. Arguments with the :ref:`inalloca <attr_inalloca>` attribute are
6685 Both markers imply that the callee does not access allocas or varargs from
6686 the caller. Calls marked ``musttail`` must obey the following additional
6689 - The call must immediately precede a :ref:`ret <i_ret>` instruction,
6690 or a pointer bitcast followed by a ret instruction.
6691 - The ret instruction must return the (possibly bitcasted) value
6692 produced by the call or void.
6693 - The caller and callee prototypes must match. Pointer types of
6694 parameters or return types may differ in pointee type, but not
6696 - The calling conventions of the caller and callee must match.
6697 - All ABI-impacting function attributes, such as sret, byval, inreg,
6698 returned, and inalloca, must match.
6699 - The callee must be varargs iff the caller is varargs. Bitcasting a
6700 non-varargs function to the appropriate varargs type is legal so
6701 long as the non-varargs prefixes obey the other rules.
6703 Tail call optimization for calls marked ``tail`` is guaranteed to occur if
6704 the following conditions are met:
6706 - Caller and callee both have the calling convention ``fastcc``.
6707 - The call is in tail position (ret immediately follows call and ret
6708 uses value of call or is void).
6709 - Option ``-tailcallopt`` is enabled, or
6710 ``llvm::GuaranteedTailCallOpt`` is ``true``.
6711 - `Platform-specific constraints are
6712 met. <CodeGenerator.html#tailcallopt>`_
6714 #. The optional "cconv" marker indicates which :ref:`calling
6715 convention <callingconv>` the call should use. If none is
6716 specified, the call defaults to using C calling conventions. The
6717 calling convention of the call must match the calling convention of
6718 the target function, or else the behavior is undefined.
6719 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
6720 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
6722 #. '``ty``': the type of the call instruction itself which is also the
6723 type of the return value. Functions that return no value are marked
6725 #. '``fnty``': shall be the signature of the pointer to function value
6726 being invoked. The argument types must match the types implied by
6727 this signature. This type can be omitted if the function is not
6728 varargs and if the function type does not return a pointer to a
6730 #. '``fnptrval``': An LLVM value containing a pointer to a function to
6731 be invoked. In most cases, this is a direct function invocation, but
6732 indirect ``call``'s are just as possible, calling an arbitrary pointer
6734 #. '``function args``': argument list whose types match the function
6735 signature argument types and parameter attributes. All arguments must
6736 be of :ref:`first class <t_firstclass>` type. If the function signature
6737 indicates the function accepts a variable number of arguments, the
6738 extra arguments can be specified.
6739 #. The optional :ref:`function attributes <fnattrs>` list. Only
6740 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
6741 attributes are valid here.
6746 The '``call``' instruction is used to cause control flow to transfer to
6747 a specified function, with its incoming arguments bound to the specified
6748 values. Upon a '``ret``' instruction in the called function, control
6749 flow continues with the instruction after the function call, and the
6750 return value of the function is bound to the result argument.
6755 .. code-block:: llvm
6757 %retval = call i32 @test(i32 %argc)
6758 call i32 (i8*, ...)* @printf(i8* %msg, i32 12, i8 42) ; yields i32
6759 %X = tail call i32 @foo() ; yields i32
6760 %Y = tail call fastcc i32 @foo() ; yields i32
6761 call void %foo(i8 97 signext)
6763 %struct.A = type { i32, i8 }
6764 %r = call %struct.A @foo() ; yields { i32, i8 }
6765 %gr = extractvalue %struct.A %r, 0 ; yields i32
6766 %gr1 = extractvalue %struct.A %r, 1 ; yields i8
6767 %Z = call void @foo() noreturn ; indicates that %foo never returns normally
6768 %ZZ = call zeroext i32 @bar() ; Return value is %zero extended
6770 llvm treats calls to some functions with names and arguments that match
6771 the standard C99 library as being the C99 library functions, and may
6772 perform optimizations or generate code for them under that assumption.
6773 This is something we'd like to change in the future to provide better
6774 support for freestanding environments and non-C-based languages.
6778 '``va_arg``' Instruction
6779 ^^^^^^^^^^^^^^^^^^^^^^^^
6786 <resultval> = va_arg <va_list*> <arglist>, <argty>
6791 The '``va_arg``' instruction is used to access arguments passed through
6792 the "variable argument" area of a function call. It is used to implement
6793 the ``va_arg`` macro in C.
6798 This instruction takes a ``va_list*`` value and the type of the
6799 argument. It returns a value of the specified argument type and
6800 increments the ``va_list`` to point to the next argument. The actual
6801 type of ``va_list`` is target specific.
6806 The '``va_arg``' instruction loads an argument of the specified type
6807 from the specified ``va_list`` and causes the ``va_list`` to point to
6808 the next argument. For more information, see the variable argument
6809 handling :ref:`Intrinsic Functions <int_varargs>`.
6811 It is legal for this instruction to be called in a function which does
6812 not take a variable number of arguments, for example, the ``vfprintf``
6815 ``va_arg`` is an LLVM instruction instead of an :ref:`intrinsic
6816 function <intrinsics>` because it takes a type as an argument.
6821 See the :ref:`variable argument processing <int_varargs>` section.
6823 Note that the code generator does not yet fully support va\_arg on many
6824 targets. Also, it does not currently support va\_arg with aggregate
6825 types on any target.
6829 '``landingpad``' Instruction
6830 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6837 <resultval> = landingpad <resultty> personality <type> <pers_fn> <clause>+
6838 <resultval> = landingpad <resultty> personality <type> <pers_fn> cleanup <clause>*
6840 <clause> := catch <type> <value>
6841 <clause> := filter <array constant type> <array constant>
6846 The '``landingpad``' instruction is used by `LLVM's exception handling
6847 system <ExceptionHandling.html#overview>`_ to specify that a basic block
6848 is a landing pad --- one where the exception lands, and corresponds to the
6849 code found in the ``catch`` portion of a ``try``/``catch`` sequence. It
6850 defines values supplied by the personality function (``pers_fn``) upon
6851 re-entry to the function. The ``resultval`` has the type ``resultty``.
6856 This instruction takes a ``pers_fn`` value. This is the personality
6857 function associated with the unwinding mechanism. The optional
6858 ``cleanup`` flag indicates that the landing pad block is a cleanup.
6860 A ``clause`` begins with the clause type --- ``catch`` or ``filter`` --- and
6861 contains the global variable representing the "type" that may be caught
6862 or filtered respectively. Unlike the ``catch`` clause, the ``filter``
6863 clause takes an array constant as its argument. Use
6864 "``[0 x i8**] undef``" for a filter which cannot throw. The
6865 '``landingpad``' instruction must contain *at least* one ``clause`` or
6866 the ``cleanup`` flag.
6871 The '``landingpad``' instruction defines the values which are set by the
6872 personality function (``pers_fn``) upon re-entry to the function, and
6873 therefore the "result type" of the ``landingpad`` instruction. As with
6874 calling conventions, how the personality function results are
6875 represented in LLVM IR is target specific.
6877 The clauses are applied in order from top to bottom. If two
6878 ``landingpad`` instructions are merged together through inlining, the
6879 clauses from the calling function are appended to the list of clauses.
6880 When the call stack is being unwound due to an exception being thrown,
6881 the exception is compared against each ``clause`` in turn. If it doesn't
6882 match any of the clauses, and the ``cleanup`` flag is not set, then
6883 unwinding continues further up the call stack.
6885 The ``landingpad`` instruction has several restrictions:
6887 - A landing pad block is a basic block which is the unwind destination
6888 of an '``invoke``' instruction.
6889 - A landing pad block must have a '``landingpad``' instruction as its
6890 first non-PHI instruction.
6891 - There can be only one '``landingpad``' instruction within the landing
6893 - A basic block that is not a landing pad block may not include a
6894 '``landingpad``' instruction.
6895 - All '``landingpad``' instructions in a function must have the same
6896 personality function.
6901 .. code-block:: llvm
6903 ;; A landing pad which can catch an integer.
6904 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6906 ;; A landing pad that is a cleanup.
6907 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6909 ;; A landing pad which can catch an integer and can only throw a double.
6910 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6912 filter [1 x i8**] [@_ZTId]
6919 LLVM supports the notion of an "intrinsic function". These functions
6920 have well known names and semantics and are required to follow certain
6921 restrictions. Overall, these intrinsics represent an extension mechanism
6922 for the LLVM language that does not require changing all of the
6923 transformations in LLVM when adding to the language (or the bitcode
6924 reader/writer, the parser, etc...).
6926 Intrinsic function names must all start with an "``llvm.``" prefix. This
6927 prefix is reserved in LLVM for intrinsic names; thus, function names may
6928 not begin with this prefix. Intrinsic functions must always be external
6929 functions: you cannot define the body of intrinsic functions. Intrinsic
6930 functions may only be used in call or invoke instructions: it is illegal
6931 to take the address of an intrinsic function. Additionally, because
6932 intrinsic functions are part of the LLVM language, it is required if any
6933 are added that they be documented here.
6935 Some intrinsic functions can be overloaded, i.e., the intrinsic
6936 represents a family of functions that perform the same operation but on
6937 different data types. Because LLVM can represent over 8 million
6938 different integer types, overloading is used commonly to allow an
6939 intrinsic function to operate on any integer type. One or more of the
6940 argument types or the result type can be overloaded to accept any
6941 integer type. Argument types may also be defined as exactly matching a
6942 previous argument's type or the result type. This allows an intrinsic
6943 function which accepts multiple arguments, but needs all of them to be
6944 of the same type, to only be overloaded with respect to a single
6945 argument or the result.
6947 Overloaded intrinsics will have the names of its overloaded argument
6948 types encoded into its function name, each preceded by a period. Only
6949 those types which are overloaded result in a name suffix. Arguments
6950 whose type is matched against another type do not. For example, the
6951 ``llvm.ctpop`` function can take an integer of any width and returns an
6952 integer of exactly the same integer width. This leads to a family of
6953 functions such as ``i8 @llvm.ctpop.i8(i8 %val)`` and
6954 ``i29 @llvm.ctpop.i29(i29 %val)``. Only one type, the return type, is
6955 overloaded, and only one type suffix is required. Because the argument's
6956 type is matched against the return type, it does not require its own
6959 To learn how to add an intrinsic function, please see the `Extending
6960 LLVM Guide <ExtendingLLVM.html>`_.
6964 Variable Argument Handling Intrinsics
6965 -------------------------------------
6967 Variable argument support is defined in LLVM with the
6968 :ref:`va_arg <i_va_arg>` instruction and these three intrinsic
6969 functions. These functions are related to the similarly named macros
6970 defined in the ``<stdarg.h>`` header file.
6972 All of these functions operate on arguments that use a target-specific
6973 value type "``va_list``". The LLVM assembly language reference manual
6974 does not define what this type is, so all transformations should be
6975 prepared to handle these functions regardless of the type used.
6977 This example shows how the :ref:`va_arg <i_va_arg>` instruction and the
6978 variable argument handling intrinsic functions are used.
6980 .. code-block:: llvm
6982 ; This struct is different for every platform. For most platforms,
6983 ; it is merely an i8*.
6984 %struct.va_list = type { i8* }
6986 ; For Unix x86_64 platforms, va_list is the following struct:
6987 ; %struct.va_list = type { i32, i32, i8*, i8* }
6989 define i32 @test(i32 %X, ...) {
6990 ; Initialize variable argument processing
6991 %ap = alloca %struct.va_list
6992 %ap2 = bitcast %struct.va_list* %ap to i8*
6993 call void @llvm.va_start(i8* %ap2)
6995 ; Read a single integer argument
6996 %tmp = va_arg i8* %ap2, i32
6998 ; Demonstrate usage of llvm.va_copy and llvm.va_end
7000 %aq2 = bitcast i8** %aq to i8*
7001 call void @llvm.va_copy(i8* %aq2, i8* %ap2)
7002 call void @llvm.va_end(i8* %aq2)
7004 ; Stop processing of arguments.
7005 call void @llvm.va_end(i8* %ap2)
7009 declare void @llvm.va_start(i8*)
7010 declare void @llvm.va_copy(i8*, i8*)
7011 declare void @llvm.va_end(i8*)
7015 '``llvm.va_start``' Intrinsic
7016 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7023 declare void @llvm.va_start(i8* <arglist>)
7028 The '``llvm.va_start``' intrinsic initializes ``*<arglist>`` for
7029 subsequent use by ``va_arg``.
7034 The argument is a pointer to a ``va_list`` element to initialize.
7039 The '``llvm.va_start``' intrinsic works just like the ``va_start`` macro
7040 available in C. In a target-dependent way, it initializes the
7041 ``va_list`` element to which the argument points, so that the next call
7042 to ``va_arg`` will produce the first variable argument passed to the
7043 function. Unlike the C ``va_start`` macro, this intrinsic does not need
7044 to know the last argument of the function as the compiler can figure
7047 '``llvm.va_end``' Intrinsic
7048 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7055 declare void @llvm.va_end(i8* <arglist>)
7060 The '``llvm.va_end``' intrinsic destroys ``*<arglist>``, which has been
7061 initialized previously with ``llvm.va_start`` or ``llvm.va_copy``.
7066 The argument is a pointer to a ``va_list`` to destroy.
7071 The '``llvm.va_end``' intrinsic works just like the ``va_end`` macro
7072 available in C. In a target-dependent way, it destroys the ``va_list``
7073 element to which the argument points. Calls to
7074 :ref:`llvm.va_start <int_va_start>` and
7075 :ref:`llvm.va_copy <int_va_copy>` must be matched exactly with calls to
7080 '``llvm.va_copy``' Intrinsic
7081 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7088 declare void @llvm.va_copy(i8* <destarglist>, i8* <srcarglist>)
7093 The '``llvm.va_copy``' intrinsic copies the current argument position
7094 from the source argument list to the destination argument list.
7099 The first argument is a pointer to a ``va_list`` element to initialize.
7100 The second argument is a pointer to a ``va_list`` element to copy from.
7105 The '``llvm.va_copy``' intrinsic works just like the ``va_copy`` macro
7106 available in C. In a target-dependent way, it copies the source
7107 ``va_list`` element into the destination ``va_list`` element. This
7108 intrinsic is necessary because the `` llvm.va_start`` intrinsic may be
7109 arbitrarily complex and require, for example, memory allocation.
7111 Accurate Garbage Collection Intrinsics
7112 --------------------------------------
7114 LLVM's support for `Accurate Garbage Collection <GarbageCollection.html>`_
7115 (GC) requires the frontend to generate code containing appropriate intrinsic
7116 calls and select an appropriate GC strategy which knows how to lower these
7117 intrinsics in a manner which is appropriate for the target collector.
7119 These intrinsics allow identification of :ref:`GC roots on the
7120 stack <int_gcroot>`, as well as garbage collector implementations that
7121 require :ref:`read <int_gcread>` and :ref:`write <int_gcwrite>` barriers.
7122 Frontends for type-safe garbage collected languages should generate
7123 these intrinsics to make use of the LLVM garbage collectors. For more
7124 details, see `Garbage Collection with LLVM <GarbageCollection.html>`_.
7126 Experimental Statepoint Intrinsics
7127 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7129 LLVM provides an second experimental set of intrinsics for describing garbage
7130 collection safepoints in compiled code. These intrinsics are an alternative
7131 to the ``llvm.gcroot`` intrinsics, but are compatible with the ones for
7132 :ref:`read <int_gcread>` and :ref:`write <int_gcwrite>` barriers. The
7133 differences in approach are covered in the `Garbage Collection with LLVM
7134 <GarbageCollection.html>`_ documentation. The intrinsics themselves are
7135 described in :doc:`Statepoints`.
7139 '``llvm.gcroot``' Intrinsic
7140 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7147 declare void @llvm.gcroot(i8** %ptrloc, i8* %metadata)
7152 The '``llvm.gcroot``' intrinsic declares the existence of a GC root to
7153 the code generator, and allows some metadata to be associated with it.
7158 The first argument specifies the address of a stack object that contains
7159 the root pointer. The second pointer (which must be either a constant or
7160 a global value address) contains the meta-data to be associated with the
7166 At runtime, a call to this intrinsic stores a null pointer into the
7167 "ptrloc" location. At compile-time, the code generator generates
7168 information to allow the runtime to find the pointer at GC safe points.
7169 The '``llvm.gcroot``' intrinsic may only be used in a function which
7170 :ref:`specifies a GC algorithm <gc>`.
7174 '``llvm.gcread``' Intrinsic
7175 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7182 declare i8* @llvm.gcread(i8* %ObjPtr, i8** %Ptr)
7187 The '``llvm.gcread``' intrinsic identifies reads of references from heap
7188 locations, allowing garbage collector implementations that require read
7194 The second argument is the address to read from, which should be an
7195 address allocated from the garbage collector. The first object is a
7196 pointer to the start of the referenced object, if needed by the language
7197 runtime (otherwise null).
7202 The '``llvm.gcread``' intrinsic has the same semantics as a load
7203 instruction, but may be replaced with substantially more complex code by
7204 the garbage collector runtime, as needed. The '``llvm.gcread``'
7205 intrinsic may only be used in a function which :ref:`specifies a GC
7210 '``llvm.gcwrite``' Intrinsic
7211 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7218 declare void @llvm.gcwrite(i8* %P1, i8* %Obj, i8** %P2)
7223 The '``llvm.gcwrite``' intrinsic identifies writes of references to heap
7224 locations, allowing garbage collector implementations that require write
7225 barriers (such as generational or reference counting collectors).
7230 The first argument is the reference to store, the second is the start of
7231 the object to store it to, and the third is the address of the field of
7232 Obj to store to. If the runtime does not require a pointer to the
7233 object, Obj may be null.
7238 The '``llvm.gcwrite``' intrinsic has the same semantics as a store
7239 instruction, but may be replaced with substantially more complex code by
7240 the garbage collector runtime, as needed. The '``llvm.gcwrite``'
7241 intrinsic may only be used in a function which :ref:`specifies a GC
7244 Code Generator Intrinsics
7245 -------------------------
7247 These intrinsics are provided by LLVM to expose special features that
7248 may only be implemented with code generator support.
7250 '``llvm.returnaddress``' Intrinsic
7251 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7258 declare i8 *@llvm.returnaddress(i32 <level>)
7263 The '``llvm.returnaddress``' intrinsic attempts to compute a
7264 target-specific value indicating the return address of the current
7265 function or one of its callers.
7270 The argument to this intrinsic indicates which function to return the
7271 address for. Zero indicates the calling function, one indicates its
7272 caller, etc. The argument is **required** to be a constant integer
7278 The '``llvm.returnaddress``' intrinsic either returns a pointer
7279 indicating the return address of the specified call frame, or zero if it
7280 cannot be identified. The value returned by this intrinsic is likely to
7281 be incorrect or 0 for arguments other than zero, so it should only be
7282 used for debugging purposes.
7284 Note that calling this intrinsic does not prevent function inlining or
7285 other aggressive transformations, so the value returned may not be that
7286 of the obvious source-language caller.
7288 '``llvm.frameaddress``' Intrinsic
7289 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7296 declare i8* @llvm.frameaddress(i32 <level>)
7301 The '``llvm.frameaddress``' intrinsic attempts to return the
7302 target-specific frame pointer value for the specified stack frame.
7307 The argument to this intrinsic indicates which function to return the
7308 frame pointer for. Zero indicates the calling function, one indicates
7309 its caller, etc. The argument is **required** to be a constant integer
7315 The '``llvm.frameaddress``' intrinsic either returns a pointer
7316 indicating the frame address of the specified call frame, or zero if it
7317 cannot be identified. The value returned by this intrinsic is likely to
7318 be incorrect or 0 for arguments other than zero, so it should only be
7319 used for debugging purposes.
7321 Note that calling this intrinsic does not prevent function inlining or
7322 other aggressive transformations, so the value returned may not be that
7323 of the obvious source-language caller.
7325 '``llvm.frameallocate``' and '``llvm.framerecover``' Intrinsics
7326 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7333 declare i8* @llvm.frameallocate(i32 %size)
7334 declare i8* @llvm.framerecover(i8* %func, i8* %fp)
7339 The '``llvm.frameallocate``' intrinsic allocates stack memory at some fixed
7340 offset from the frame pointer, and the '``llvm.framerecover``'
7341 intrinsic applies that offset to a live frame pointer to recover the address of
7342 the allocation. The offset is computed during frame layout of the caller of
7343 ``llvm.frameallocate``.
7348 The ``size`` argument to '``llvm.frameallocate``' must be a constant integer
7349 indicating the amount of stack memory to allocate. As with allocas, allocating
7350 zero bytes is legal, but the result is undefined.
7352 The ``func`` argument to '``llvm.framerecover``' must be a constant
7353 bitcasted pointer to a function defined in the current module. The code
7354 generator cannot determine the frame allocation offset of functions defined in
7357 The ``fp`` argument to '``llvm.framerecover``' must be a frame
7358 pointer of a call frame that is currently live. The return value of
7359 '``llvm.frameaddress``' is one way to produce such a value, but most platforms
7360 also expose the frame pointer through stack unwinding mechanisms.
7365 These intrinsics allow a group of functions to access one stack memory
7366 allocation in an ancestor stack frame. The memory returned from
7367 '``llvm.frameallocate``' may be allocated prior to stack realignment, so the
7368 memory is only aligned to the ABI-required stack alignment. Each function may
7369 only call '``llvm.frameallocate``' one or zero times from the function entry
7370 block. The frame allocation intrinsic inhibits inlining, as any frame
7371 allocations in the inlined function frame are likely to be at a different
7372 offset from the one used by '``llvm.framerecover``' called with the
7375 .. _int_read_register:
7376 .. _int_write_register:
7378 '``llvm.read_register``' and '``llvm.write_register``' Intrinsics
7379 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7386 declare i32 @llvm.read_register.i32(metadata)
7387 declare i64 @llvm.read_register.i64(metadata)
7388 declare void @llvm.write_register.i32(metadata, i32 @value)
7389 declare void @llvm.write_register.i64(metadata, i64 @value)
7395 The '``llvm.read_register``' and '``llvm.write_register``' intrinsics
7396 provides access to the named register. The register must be valid on
7397 the architecture being compiled to. The type needs to be compatible
7398 with the register being read.
7403 The '``llvm.read_register``' intrinsic returns the current value of the
7404 register, where possible. The '``llvm.write_register``' intrinsic sets
7405 the current value of the register, where possible.
7407 This is useful to implement named register global variables that need
7408 to always be mapped to a specific register, as is common practice on
7409 bare-metal programs including OS kernels.
7411 The compiler doesn't check for register availability or use of the used
7412 register in surrounding code, including inline assembly. Because of that,
7413 allocatable registers are not supported.
7415 Warning: So far it only works with the stack pointer on selected
7416 architectures (ARM, AArch64, PowerPC and x86_64). Significant amount of
7417 work is needed to support other registers and even more so, allocatable
7422 '``llvm.stacksave``' Intrinsic
7423 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7430 declare i8* @llvm.stacksave()
7435 The '``llvm.stacksave``' intrinsic is used to remember the current state
7436 of the function stack, for use with
7437 :ref:`llvm.stackrestore <int_stackrestore>`. This is useful for
7438 implementing language features like scoped automatic variable sized
7444 This intrinsic returns a opaque pointer value that can be passed to
7445 :ref:`llvm.stackrestore <int_stackrestore>`. When an
7446 ``llvm.stackrestore`` intrinsic is executed with a value saved from
7447 ``llvm.stacksave``, it effectively restores the state of the stack to
7448 the state it was in when the ``llvm.stacksave`` intrinsic executed. In
7449 practice, this pops any :ref:`alloca <i_alloca>` blocks from the stack that
7450 were allocated after the ``llvm.stacksave`` was executed.
7452 .. _int_stackrestore:
7454 '``llvm.stackrestore``' Intrinsic
7455 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7462 declare void @llvm.stackrestore(i8* %ptr)
7467 The '``llvm.stackrestore``' intrinsic is used to restore the state of
7468 the function stack to the state it was in when the corresponding
7469 :ref:`llvm.stacksave <int_stacksave>` intrinsic executed. This is
7470 useful for implementing language features like scoped automatic variable
7471 sized arrays in C99.
7476 See the description for :ref:`llvm.stacksave <int_stacksave>`.
7478 '``llvm.prefetch``' Intrinsic
7479 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7486 declare void @llvm.prefetch(i8* <address>, i32 <rw>, i32 <locality>, i32 <cache type>)
7491 The '``llvm.prefetch``' intrinsic is a hint to the code generator to
7492 insert a prefetch instruction if supported; otherwise, it is a noop.
7493 Prefetches have no effect on the behavior of the program but can change
7494 its performance characteristics.
7499 ``address`` is the address to be prefetched, ``rw`` is the specifier
7500 determining if the fetch should be for a read (0) or write (1), and
7501 ``locality`` is a temporal locality specifier ranging from (0) - no
7502 locality, to (3) - extremely local keep in cache. The ``cache type``
7503 specifies whether the prefetch is performed on the data (1) or
7504 instruction (0) cache. The ``rw``, ``locality`` and ``cache type``
7505 arguments must be constant integers.
7510 This intrinsic does not modify the behavior of the program. In
7511 particular, prefetches cannot trap and do not produce a value. On
7512 targets that support this intrinsic, the prefetch can provide hints to
7513 the processor cache for better performance.
7515 '``llvm.pcmarker``' Intrinsic
7516 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7523 declare void @llvm.pcmarker(i32 <id>)
7528 The '``llvm.pcmarker``' intrinsic is a method to export a Program
7529 Counter (PC) in a region of code to simulators and other tools. The
7530 method is target specific, but it is expected that the marker will use
7531 exported symbols to transmit the PC of the marker. The marker makes no
7532 guarantees that it will remain with any specific instruction after
7533 optimizations. It is possible that the presence of a marker will inhibit
7534 optimizations. The intended use is to be inserted after optimizations to
7535 allow correlations of simulation runs.
7540 ``id`` is a numerical id identifying the marker.
7545 This intrinsic does not modify the behavior of the program. Backends
7546 that do not support this intrinsic may ignore it.
7548 '``llvm.readcyclecounter``' Intrinsic
7549 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7556 declare i64 @llvm.readcyclecounter()
7561 The '``llvm.readcyclecounter``' intrinsic provides access to the cycle
7562 counter register (or similar low latency, high accuracy clocks) on those
7563 targets that support it. On X86, it should map to RDTSC. On Alpha, it
7564 should map to RPCC. As the backing counters overflow quickly (on the
7565 order of 9 seconds on alpha), this should only be used for small
7571 When directly supported, reading the cycle counter should not modify any
7572 memory. Implementations are allowed to either return a application
7573 specific value or a system wide value. On backends without support, this
7574 is lowered to a constant 0.
7576 Note that runtime support may be conditional on the privilege-level code is
7577 running at and the host platform.
7579 '``llvm.clear_cache``' Intrinsic
7580 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7587 declare void @llvm.clear_cache(i8*, i8*)
7592 The '``llvm.clear_cache``' intrinsic ensures visibility of modifications
7593 in the specified range to the execution unit of the processor. On
7594 targets with non-unified instruction and data cache, the implementation
7595 flushes the instruction cache.
7600 On platforms with coherent instruction and data caches (e.g. x86), this
7601 intrinsic is a nop. On platforms with non-coherent instruction and data
7602 cache (e.g. ARM, MIPS), the intrinsic is lowered either to appropriate
7603 instructions or a system call, if cache flushing requires special
7606 The default behavior is to emit a call to ``__clear_cache`` from the run
7609 This instrinsic does *not* empty the instruction pipeline. Modifications
7610 of the current function are outside the scope of the intrinsic.
7612 '``llvm.instrprof_increment``' Intrinsic
7613 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7620 declare void @llvm.instrprof_increment(i8* <name>, i64 <hash>,
7621 i32 <num-counters>, i32 <index>)
7626 The '``llvm.instrprof_increment``' intrinsic can be emitted by a
7627 frontend for use with instrumentation based profiling. These will be
7628 lowered by the ``-instrprof`` pass to generate execution counts of a
7634 The first argument is a pointer to a global variable containing the
7635 name of the entity being instrumented. This should generally be the
7636 (mangled) function name for a set of counters.
7638 The second argument is a hash value that can be used by the consumer
7639 of the profile data to detect changes to the instrumented source, and
7640 the third is the number of counters associated with ``name``. It is an
7641 error if ``hash`` or ``num-counters`` differ between two instances of
7642 ``instrprof_increment`` that refer to the same name.
7644 The last argument refers to which of the counters for ``name`` should
7645 be incremented. It should be a value between 0 and ``num-counters``.
7650 This intrinsic represents an increment of a profiling counter. It will
7651 cause the ``-instrprof`` pass to generate the appropriate data
7652 structures and the code to increment the appropriate value, in a
7653 format that can be written out by a compiler runtime and consumed via
7654 the ``llvm-profdata`` tool.
7656 Standard C Library Intrinsics
7657 -----------------------------
7659 LLVM provides intrinsics for a few important standard C library
7660 functions. These intrinsics allow source-language front-ends to pass
7661 information about the alignment of the pointer arguments to the code
7662 generator, providing opportunity for more efficient code generation.
7666 '``llvm.memcpy``' Intrinsic
7667 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7672 This is an overloaded intrinsic. You can use ``llvm.memcpy`` on any
7673 integer bit width and for different address spaces. Not all targets
7674 support all bit widths however.
7678 declare void @llvm.memcpy.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
7679 i32 <len>, i32 <align>, i1 <isvolatile>)
7680 declare void @llvm.memcpy.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
7681 i64 <len>, i32 <align>, i1 <isvolatile>)
7686 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
7687 source location to the destination location.
7689 Note that, unlike the standard libc function, the ``llvm.memcpy.*``
7690 intrinsics do not return a value, takes extra alignment/isvolatile
7691 arguments and the pointers can be in specified address spaces.
7696 The first argument is a pointer to the destination, the second is a
7697 pointer to the source. The third argument is an integer argument
7698 specifying the number of bytes to copy, the fourth argument is the
7699 alignment of the source and destination locations, and the fifth is a
7700 boolean indicating a volatile access.
7702 If the call to this intrinsic has an alignment value that is not 0 or 1,
7703 then the caller guarantees that both the source and destination pointers
7704 are aligned to that boundary.
7706 If the ``isvolatile`` parameter is ``true``, the ``llvm.memcpy`` call is
7707 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
7708 very cleanly specified and it is unwise to depend on it.
7713 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
7714 source location to the destination location, which are not allowed to
7715 overlap. It copies "len" bytes of memory over. If the argument is known
7716 to be aligned to some boundary, this can be specified as the fourth
7717 argument, otherwise it should be set to 0 or 1 (both meaning no alignment).
7719 '``llvm.memmove``' Intrinsic
7720 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7725 This is an overloaded intrinsic. You can use llvm.memmove on any integer
7726 bit width and for different address space. Not all targets support all
7731 declare void @llvm.memmove.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
7732 i32 <len>, i32 <align>, i1 <isvolatile>)
7733 declare void @llvm.memmove.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
7734 i64 <len>, i32 <align>, i1 <isvolatile>)
7739 The '``llvm.memmove.*``' intrinsics move a block of memory from the
7740 source location to the destination location. It is similar to the
7741 '``llvm.memcpy``' intrinsic but allows the two memory locations to
7744 Note that, unlike the standard libc function, the ``llvm.memmove.*``
7745 intrinsics do not return a value, takes extra alignment/isvolatile
7746 arguments and the pointers can be in specified address spaces.
7751 The first argument is a pointer to the destination, the second is a
7752 pointer to the source. The third argument is an integer argument
7753 specifying the number of bytes to copy, the fourth argument is the
7754 alignment of the source and destination locations, and the fifth is a
7755 boolean indicating a volatile access.
7757 If the call to this intrinsic has an alignment value that is not 0 or 1,
7758 then the caller guarantees that the source and destination pointers are
7759 aligned to that boundary.
7761 If the ``isvolatile`` parameter is ``true``, the ``llvm.memmove`` call
7762 is a :ref:`volatile operation <volatile>`. The detailed access behavior is
7763 not very cleanly specified and it is unwise to depend on it.
7768 The '``llvm.memmove.*``' intrinsics copy a block of memory from the
7769 source location to the destination location, which may overlap. It
7770 copies "len" bytes of memory over. If the argument is known to be
7771 aligned to some boundary, this can be specified as the fourth argument,
7772 otherwise it should be set to 0 or 1 (both meaning no alignment).
7774 '``llvm.memset.*``' Intrinsics
7775 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7780 This is an overloaded intrinsic. You can use llvm.memset on any integer
7781 bit width and for different address spaces. However, not all targets
7782 support all bit widths.
7786 declare void @llvm.memset.p0i8.i32(i8* <dest>, i8 <val>,
7787 i32 <len>, i32 <align>, i1 <isvolatile>)
7788 declare void @llvm.memset.p0i8.i64(i8* <dest>, i8 <val>,
7789 i64 <len>, i32 <align>, i1 <isvolatile>)
7794 The '``llvm.memset.*``' intrinsics fill a block of memory with a
7795 particular byte value.
7797 Note that, unlike the standard libc function, the ``llvm.memset``
7798 intrinsic does not return a value and takes extra alignment/volatile
7799 arguments. Also, the destination can be in an arbitrary address space.
7804 The first argument is a pointer to the destination to fill, the second
7805 is the byte value with which to fill it, the third argument is an
7806 integer argument specifying the number of bytes to fill, and the fourth
7807 argument is the known alignment of the destination location.
7809 If the call to this intrinsic has an alignment value that is not 0 or 1,
7810 then the caller guarantees that the destination pointer is aligned to
7813 If the ``isvolatile`` parameter is ``true``, the ``llvm.memset`` call is
7814 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
7815 very cleanly specified and it is unwise to depend on it.
7820 The '``llvm.memset.*``' intrinsics fill "len" bytes of memory starting
7821 at the destination location. If the argument is known to be aligned to
7822 some boundary, this can be specified as the fourth argument, otherwise
7823 it should be set to 0 or 1 (both meaning no alignment).
7825 '``llvm.sqrt.*``' Intrinsic
7826 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7831 This is an overloaded intrinsic. You can use ``llvm.sqrt`` on any
7832 floating point or vector of floating point type. Not all targets support
7837 declare float @llvm.sqrt.f32(float %Val)
7838 declare double @llvm.sqrt.f64(double %Val)
7839 declare x86_fp80 @llvm.sqrt.f80(x86_fp80 %Val)
7840 declare fp128 @llvm.sqrt.f128(fp128 %Val)
7841 declare ppc_fp128 @llvm.sqrt.ppcf128(ppc_fp128 %Val)
7846 The '``llvm.sqrt``' intrinsics return the sqrt of the specified operand,
7847 returning the same value as the libm '``sqrt``' functions would. Unlike
7848 ``sqrt`` in libm, however, ``llvm.sqrt`` has undefined behavior for
7849 negative numbers other than -0.0 (which allows for better optimization,
7850 because there is no need to worry about errno being set).
7851 ``llvm.sqrt(-0.0)`` is defined to return -0.0 like IEEE sqrt.
7856 The argument and return value are floating point numbers of the same
7862 This function returns the sqrt of the specified operand if it is a
7863 nonnegative floating point number.
7865 '``llvm.powi.*``' Intrinsic
7866 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7871 This is an overloaded intrinsic. You can use ``llvm.powi`` on any
7872 floating point or vector of floating point type. Not all targets support
7877 declare float @llvm.powi.f32(float %Val, i32 %power)
7878 declare double @llvm.powi.f64(double %Val, i32 %power)
7879 declare x86_fp80 @llvm.powi.f80(x86_fp80 %Val, i32 %power)
7880 declare fp128 @llvm.powi.f128(fp128 %Val, i32 %power)
7881 declare ppc_fp128 @llvm.powi.ppcf128(ppc_fp128 %Val, i32 %power)
7886 The '``llvm.powi.*``' intrinsics return the first operand raised to the
7887 specified (positive or negative) power. The order of evaluation of
7888 multiplications is not defined. When a vector of floating point type is
7889 used, the second argument remains a scalar integer value.
7894 The second argument is an integer power, and the first is a value to
7895 raise to that power.
7900 This function returns the first value raised to the second power with an
7901 unspecified sequence of rounding operations.
7903 '``llvm.sin.*``' Intrinsic
7904 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7909 This is an overloaded intrinsic. You can use ``llvm.sin`` on any
7910 floating point or vector of floating point type. Not all targets support
7915 declare float @llvm.sin.f32(float %Val)
7916 declare double @llvm.sin.f64(double %Val)
7917 declare x86_fp80 @llvm.sin.f80(x86_fp80 %Val)
7918 declare fp128 @llvm.sin.f128(fp128 %Val)
7919 declare ppc_fp128 @llvm.sin.ppcf128(ppc_fp128 %Val)
7924 The '``llvm.sin.*``' intrinsics return the sine of the operand.
7929 The argument and return value are floating point numbers of the same
7935 This function returns the sine of the specified operand, returning the
7936 same values as the libm ``sin`` functions would, and handles error
7937 conditions in the same way.
7939 '``llvm.cos.*``' Intrinsic
7940 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7945 This is an overloaded intrinsic. You can use ``llvm.cos`` on any
7946 floating point or vector of floating point type. Not all targets support
7951 declare float @llvm.cos.f32(float %Val)
7952 declare double @llvm.cos.f64(double %Val)
7953 declare x86_fp80 @llvm.cos.f80(x86_fp80 %Val)
7954 declare fp128 @llvm.cos.f128(fp128 %Val)
7955 declare ppc_fp128 @llvm.cos.ppcf128(ppc_fp128 %Val)
7960 The '``llvm.cos.*``' intrinsics return the cosine of the operand.
7965 The argument and return value are floating point numbers of the same
7971 This function returns the cosine of the specified operand, returning the
7972 same values as the libm ``cos`` functions would, and handles error
7973 conditions in the same way.
7975 '``llvm.pow.*``' Intrinsic
7976 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7981 This is an overloaded intrinsic. You can use ``llvm.pow`` on any
7982 floating point or vector of floating point type. Not all targets support
7987 declare float @llvm.pow.f32(float %Val, float %Power)
7988 declare double @llvm.pow.f64(double %Val, double %Power)
7989 declare x86_fp80 @llvm.pow.f80(x86_fp80 %Val, x86_fp80 %Power)
7990 declare fp128 @llvm.pow.f128(fp128 %Val, fp128 %Power)
7991 declare ppc_fp128 @llvm.pow.ppcf128(ppc_fp128 %Val, ppc_fp128 Power)
7996 The '``llvm.pow.*``' intrinsics return the first operand raised to the
7997 specified (positive or negative) power.
8002 The second argument is a floating point power, and the first is a value
8003 to raise to that power.
8008 This function returns the first value raised to the second power,
8009 returning the same values as the libm ``pow`` functions would, and
8010 handles error conditions in the same way.
8012 '``llvm.exp.*``' Intrinsic
8013 ^^^^^^^^^^^^^^^^^^^^^^^^^^
8018 This is an overloaded intrinsic. You can use ``llvm.exp`` on any
8019 floating point or vector of floating point type. Not all targets support
8024 declare float @llvm.exp.f32(float %Val)
8025 declare double @llvm.exp.f64(double %Val)
8026 declare x86_fp80 @llvm.exp.f80(x86_fp80 %Val)
8027 declare fp128 @llvm.exp.f128(fp128 %Val)
8028 declare ppc_fp128 @llvm.exp.ppcf128(ppc_fp128 %Val)
8033 The '``llvm.exp.*``' intrinsics perform the exp function.
8038 The argument and return value are floating point numbers of the same
8044 This function returns the same values as the libm ``exp`` functions
8045 would, and handles error conditions in the same way.
8047 '``llvm.exp2.*``' Intrinsic
8048 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8053 This is an overloaded intrinsic. You can use ``llvm.exp2`` on any
8054 floating point or vector of floating point type. Not all targets support
8059 declare float @llvm.exp2.f32(float %Val)
8060 declare double @llvm.exp2.f64(double %Val)
8061 declare x86_fp80 @llvm.exp2.f80(x86_fp80 %Val)
8062 declare fp128 @llvm.exp2.f128(fp128 %Val)
8063 declare ppc_fp128 @llvm.exp2.ppcf128(ppc_fp128 %Val)
8068 The '``llvm.exp2.*``' intrinsics perform the exp2 function.
8073 The argument and return value are floating point numbers of the same
8079 This function returns the same values as the libm ``exp2`` functions
8080 would, and handles error conditions in the same way.
8082 '``llvm.log.*``' Intrinsic
8083 ^^^^^^^^^^^^^^^^^^^^^^^^^^
8088 This is an overloaded intrinsic. You can use ``llvm.log`` on any
8089 floating point or vector of floating point type. Not all targets support
8094 declare float @llvm.log.f32(float %Val)
8095 declare double @llvm.log.f64(double %Val)
8096 declare x86_fp80 @llvm.log.f80(x86_fp80 %Val)
8097 declare fp128 @llvm.log.f128(fp128 %Val)
8098 declare ppc_fp128 @llvm.log.ppcf128(ppc_fp128 %Val)
8103 The '``llvm.log.*``' intrinsics perform the log function.
8108 The argument and return value are floating point numbers of the same
8114 This function returns the same values as the libm ``log`` functions
8115 would, and handles error conditions in the same way.
8117 '``llvm.log10.*``' Intrinsic
8118 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8123 This is an overloaded intrinsic. You can use ``llvm.log10`` on any
8124 floating point or vector of floating point type. Not all targets support
8129 declare float @llvm.log10.f32(float %Val)
8130 declare double @llvm.log10.f64(double %Val)
8131 declare x86_fp80 @llvm.log10.f80(x86_fp80 %Val)
8132 declare fp128 @llvm.log10.f128(fp128 %Val)
8133 declare ppc_fp128 @llvm.log10.ppcf128(ppc_fp128 %Val)
8138 The '``llvm.log10.*``' intrinsics perform the log10 function.
8143 The argument and return value are floating point numbers of the same
8149 This function returns the same values as the libm ``log10`` functions
8150 would, and handles error conditions in the same way.
8152 '``llvm.log2.*``' Intrinsic
8153 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8158 This is an overloaded intrinsic. You can use ``llvm.log2`` on any
8159 floating point or vector of floating point type. Not all targets support
8164 declare float @llvm.log2.f32(float %Val)
8165 declare double @llvm.log2.f64(double %Val)
8166 declare x86_fp80 @llvm.log2.f80(x86_fp80 %Val)
8167 declare fp128 @llvm.log2.f128(fp128 %Val)
8168 declare ppc_fp128 @llvm.log2.ppcf128(ppc_fp128 %Val)
8173 The '``llvm.log2.*``' intrinsics perform the log2 function.
8178 The argument and return value are floating point numbers of the same
8184 This function returns the same values as the libm ``log2`` functions
8185 would, and handles error conditions in the same way.
8187 '``llvm.fma.*``' Intrinsic
8188 ^^^^^^^^^^^^^^^^^^^^^^^^^^
8193 This is an overloaded intrinsic. You can use ``llvm.fma`` on any
8194 floating point or vector of floating point type. Not all targets support
8199 declare float @llvm.fma.f32(float %a, float %b, float %c)
8200 declare double @llvm.fma.f64(double %a, double %b, double %c)
8201 declare x86_fp80 @llvm.fma.f80(x86_fp80 %a, x86_fp80 %b, x86_fp80 %c)
8202 declare fp128 @llvm.fma.f128(fp128 %a, fp128 %b, fp128 %c)
8203 declare ppc_fp128 @llvm.fma.ppcf128(ppc_fp128 %a, ppc_fp128 %b, ppc_fp128 %c)
8208 The '``llvm.fma.*``' intrinsics perform the fused multiply-add
8214 The argument and return value are floating point numbers of the same
8220 This function returns the same values as the libm ``fma`` functions
8221 would, and does not set errno.
8223 '``llvm.fabs.*``' Intrinsic
8224 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8229 This is an overloaded intrinsic. You can use ``llvm.fabs`` on any
8230 floating point or vector of floating point type. Not all targets support
8235 declare float @llvm.fabs.f32(float %Val)
8236 declare double @llvm.fabs.f64(double %Val)
8237 declare x86_fp80 @llvm.fabs.f80(x86_fp80 %Val)
8238 declare fp128 @llvm.fabs.f128(fp128 %Val)
8239 declare ppc_fp128 @llvm.fabs.ppcf128(ppc_fp128 %Val)
8244 The '``llvm.fabs.*``' intrinsics return the absolute value of the
8250 The argument and return value are floating point numbers of the same
8256 This function returns the same values as the libm ``fabs`` functions
8257 would, and handles error conditions in the same way.
8259 '``llvm.minnum.*``' Intrinsic
8260 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8265 This is an overloaded intrinsic. You can use ``llvm.minnum`` on any
8266 floating point or vector of floating point type. Not all targets support
8271 declare float @llvm.minnum.f32(float %Val0, float %Val1)
8272 declare double @llvm.minnum.f64(double %Val0, double %Val1)
8273 declare x86_fp80 @llvm.minnum.f80(x86_fp80 %Val0, x86_fp80 %Val1)
8274 declare fp128 @llvm.minnum.f128(fp128 %Val0, fp128 %Val1)
8275 declare ppc_fp128 @llvm.minnum.ppcf128(ppc_fp128 %Val0, ppc_fp128 %Val1)
8280 The '``llvm.minnum.*``' intrinsics return the minimum of the two
8287 The arguments and return value are floating point numbers of the same
8293 Follows the IEEE-754 semantics for minNum, which also match for libm's
8296 If either operand is a NaN, returns the other non-NaN operand. Returns
8297 NaN only if both operands are NaN. If the operands compare equal,
8298 returns a value that compares equal to both operands. This means that
8299 fmin(+/-0.0, +/-0.0) could return either -0.0 or 0.0.
8301 '``llvm.maxnum.*``' Intrinsic
8302 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8307 This is an overloaded intrinsic. You can use ``llvm.maxnum`` on any
8308 floating point or vector of floating point type. Not all targets support
8313 declare float @llvm.maxnum.f32(float %Val0, float %Val1l)
8314 declare double @llvm.maxnum.f64(double %Val0, double %Val1)
8315 declare x86_fp80 @llvm.maxnum.f80(x86_fp80 %Val0, x86_fp80 %Val1)
8316 declare fp128 @llvm.maxnum.f128(fp128 %Val0, fp128 %Val1)
8317 declare ppc_fp128 @llvm.maxnum.ppcf128(ppc_fp128 %Val0, ppc_fp128 %Val1)
8322 The '``llvm.maxnum.*``' intrinsics return the maximum of the two
8329 The arguments and return value are floating point numbers of the same
8334 Follows the IEEE-754 semantics for maxNum, which also match for libm's
8337 If either operand is a NaN, returns the other non-NaN operand. Returns
8338 NaN only if both operands are NaN. If the operands compare equal,
8339 returns a value that compares equal to both operands. This means that
8340 fmax(+/-0.0, +/-0.0) could return either -0.0 or 0.0.
8342 '``llvm.copysign.*``' Intrinsic
8343 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8348 This is an overloaded intrinsic. You can use ``llvm.copysign`` on any
8349 floating point or vector of floating point type. Not all targets support
8354 declare float @llvm.copysign.f32(float %Mag, float %Sgn)
8355 declare double @llvm.copysign.f64(double %Mag, double %Sgn)
8356 declare x86_fp80 @llvm.copysign.f80(x86_fp80 %Mag, x86_fp80 %Sgn)
8357 declare fp128 @llvm.copysign.f128(fp128 %Mag, fp128 %Sgn)
8358 declare ppc_fp128 @llvm.copysign.ppcf128(ppc_fp128 %Mag, ppc_fp128 %Sgn)
8363 The '``llvm.copysign.*``' intrinsics return a value with the magnitude of the
8364 first operand and the sign of the second operand.
8369 The arguments and return value are floating point numbers of the same
8375 This function returns the same values as the libm ``copysign``
8376 functions would, and handles error conditions in the same way.
8378 '``llvm.floor.*``' Intrinsic
8379 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8384 This is an overloaded intrinsic. You can use ``llvm.floor`` on any
8385 floating point or vector of floating point type. Not all targets support
8390 declare float @llvm.floor.f32(float %Val)
8391 declare double @llvm.floor.f64(double %Val)
8392 declare x86_fp80 @llvm.floor.f80(x86_fp80 %Val)
8393 declare fp128 @llvm.floor.f128(fp128 %Val)
8394 declare ppc_fp128 @llvm.floor.ppcf128(ppc_fp128 %Val)
8399 The '``llvm.floor.*``' intrinsics return the floor of the operand.
8404 The argument and return value are floating point numbers of the same
8410 This function returns the same values as the libm ``floor`` functions
8411 would, and handles error conditions in the same way.
8413 '``llvm.ceil.*``' Intrinsic
8414 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8419 This is an overloaded intrinsic. You can use ``llvm.ceil`` on any
8420 floating point or vector of floating point type. Not all targets support
8425 declare float @llvm.ceil.f32(float %Val)
8426 declare double @llvm.ceil.f64(double %Val)
8427 declare x86_fp80 @llvm.ceil.f80(x86_fp80 %Val)
8428 declare fp128 @llvm.ceil.f128(fp128 %Val)
8429 declare ppc_fp128 @llvm.ceil.ppcf128(ppc_fp128 %Val)
8434 The '``llvm.ceil.*``' intrinsics return the ceiling of the operand.
8439 The argument and return value are floating point numbers of the same
8445 This function returns the same values as the libm ``ceil`` functions
8446 would, and handles error conditions in the same way.
8448 '``llvm.trunc.*``' Intrinsic
8449 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8454 This is an overloaded intrinsic. You can use ``llvm.trunc`` on any
8455 floating point or vector of floating point type. Not all targets support
8460 declare float @llvm.trunc.f32(float %Val)
8461 declare double @llvm.trunc.f64(double %Val)
8462 declare x86_fp80 @llvm.trunc.f80(x86_fp80 %Val)
8463 declare fp128 @llvm.trunc.f128(fp128 %Val)
8464 declare ppc_fp128 @llvm.trunc.ppcf128(ppc_fp128 %Val)
8469 The '``llvm.trunc.*``' intrinsics returns the operand rounded to the
8470 nearest integer not larger in magnitude than the operand.
8475 The argument and return value are floating point numbers of the same
8481 This function returns the same values as the libm ``trunc`` functions
8482 would, and handles error conditions in the same way.
8484 '``llvm.rint.*``' Intrinsic
8485 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8490 This is an overloaded intrinsic. You can use ``llvm.rint`` on any
8491 floating point or vector of floating point type. Not all targets support
8496 declare float @llvm.rint.f32(float %Val)
8497 declare double @llvm.rint.f64(double %Val)
8498 declare x86_fp80 @llvm.rint.f80(x86_fp80 %Val)
8499 declare fp128 @llvm.rint.f128(fp128 %Val)
8500 declare ppc_fp128 @llvm.rint.ppcf128(ppc_fp128 %Val)
8505 The '``llvm.rint.*``' intrinsics returns the operand rounded to the
8506 nearest integer. It may raise an inexact floating-point exception if the
8507 operand isn't an integer.
8512 The argument and return value are floating point numbers of the same
8518 This function returns the same values as the libm ``rint`` functions
8519 would, and handles error conditions in the same way.
8521 '``llvm.nearbyint.*``' Intrinsic
8522 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8527 This is an overloaded intrinsic. You can use ``llvm.nearbyint`` on any
8528 floating point or vector of floating point type. Not all targets support
8533 declare float @llvm.nearbyint.f32(float %Val)
8534 declare double @llvm.nearbyint.f64(double %Val)
8535 declare x86_fp80 @llvm.nearbyint.f80(x86_fp80 %Val)
8536 declare fp128 @llvm.nearbyint.f128(fp128 %Val)
8537 declare ppc_fp128 @llvm.nearbyint.ppcf128(ppc_fp128 %Val)
8542 The '``llvm.nearbyint.*``' intrinsics returns the operand rounded to the
8548 The argument and return value are floating point numbers of the same
8554 This function returns the same values as the libm ``nearbyint``
8555 functions would, and handles error conditions in the same way.
8557 '``llvm.round.*``' Intrinsic
8558 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8563 This is an overloaded intrinsic. You can use ``llvm.round`` on any
8564 floating point or vector of floating point type. Not all targets support
8569 declare float @llvm.round.f32(float %Val)
8570 declare double @llvm.round.f64(double %Val)
8571 declare x86_fp80 @llvm.round.f80(x86_fp80 %Val)
8572 declare fp128 @llvm.round.f128(fp128 %Val)
8573 declare ppc_fp128 @llvm.round.ppcf128(ppc_fp128 %Val)
8578 The '``llvm.round.*``' intrinsics returns the operand rounded to the
8584 The argument and return value are floating point numbers of the same
8590 This function returns the same values as the libm ``round``
8591 functions would, and handles error conditions in the same way.
8593 Bit Manipulation Intrinsics
8594 ---------------------------
8596 LLVM provides intrinsics for a few important bit manipulation
8597 operations. These allow efficient code generation for some algorithms.
8599 '``llvm.bswap.*``' Intrinsics
8600 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8605 This is an overloaded intrinsic function. You can use bswap on any
8606 integer type that is an even number of bytes (i.e. BitWidth % 16 == 0).
8610 declare i16 @llvm.bswap.i16(i16 <id>)
8611 declare i32 @llvm.bswap.i32(i32 <id>)
8612 declare i64 @llvm.bswap.i64(i64 <id>)
8617 The '``llvm.bswap``' family of intrinsics is used to byte swap integer
8618 values with an even number of bytes (positive multiple of 16 bits).
8619 These are useful for performing operations on data that is not in the
8620 target's native byte order.
8625 The ``llvm.bswap.i16`` intrinsic returns an i16 value that has the high
8626 and low byte of the input i16 swapped. Similarly, the ``llvm.bswap.i32``
8627 intrinsic returns an i32 value that has the four bytes of the input i32
8628 swapped, so that if the input bytes are numbered 0, 1, 2, 3 then the
8629 returned i32 will have its bytes in 3, 2, 1, 0 order. The
8630 ``llvm.bswap.i48``, ``llvm.bswap.i64`` and other intrinsics extend this
8631 concept to additional even-byte lengths (6 bytes, 8 bytes and more,
8634 '``llvm.ctpop.*``' Intrinsic
8635 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8640 This is an overloaded intrinsic. You can use llvm.ctpop on any integer
8641 bit width, or on any vector with integer elements. Not all targets
8642 support all bit widths or vector types, however.
8646 declare i8 @llvm.ctpop.i8(i8 <src>)
8647 declare i16 @llvm.ctpop.i16(i16 <src>)
8648 declare i32 @llvm.ctpop.i32(i32 <src>)
8649 declare i64 @llvm.ctpop.i64(i64 <src>)
8650 declare i256 @llvm.ctpop.i256(i256 <src>)
8651 declare <2 x i32> @llvm.ctpop.v2i32(<2 x i32> <src>)
8656 The '``llvm.ctpop``' family of intrinsics counts the number of bits set
8662 The only argument is the value to be counted. The argument may be of any
8663 integer type, or a vector with integer elements. The return type must
8664 match the argument type.
8669 The '``llvm.ctpop``' intrinsic counts the 1's in a variable, or within
8670 each element of a vector.
8672 '``llvm.ctlz.*``' Intrinsic
8673 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8678 This is an overloaded intrinsic. You can use ``llvm.ctlz`` on any
8679 integer bit width, or any vector whose elements are integers. Not all
8680 targets support all bit widths or vector types, however.
8684 declare i8 @llvm.ctlz.i8 (i8 <src>, i1 <is_zero_undef>)
8685 declare i16 @llvm.ctlz.i16 (i16 <src>, i1 <is_zero_undef>)
8686 declare i32 @llvm.ctlz.i32 (i32 <src>, i1 <is_zero_undef>)
8687 declare i64 @llvm.ctlz.i64 (i64 <src>, i1 <is_zero_undef>)
8688 declare i256 @llvm.ctlz.i256(i256 <src>, i1 <is_zero_undef>)
8689 declase <2 x i32> @llvm.ctlz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
8694 The '``llvm.ctlz``' family of intrinsic functions counts the number of
8695 leading zeros in a variable.
8700 The first argument is the value to be counted. This argument may be of
8701 any integer type, or a vector with integer element type. The return
8702 type must match the first argument type.
8704 The second argument must be a constant and is a flag to indicate whether
8705 the intrinsic should ensure that a zero as the first argument produces a
8706 defined result. Historically some architectures did not provide a
8707 defined result for zero values as efficiently, and many algorithms are
8708 now predicated on avoiding zero-value inputs.
8713 The '``llvm.ctlz``' intrinsic counts the leading (most significant)
8714 zeros in a variable, or within each element of the vector. If
8715 ``src == 0`` then the result is the size in bits of the type of ``src``
8716 if ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
8717 ``llvm.ctlz(i32 2) = 30``.
8719 '``llvm.cttz.*``' Intrinsic
8720 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8725 This is an overloaded intrinsic. You can use ``llvm.cttz`` on any
8726 integer bit width, or any vector of integer elements. Not all targets
8727 support all bit widths or vector types, however.
8731 declare i8 @llvm.cttz.i8 (i8 <src>, i1 <is_zero_undef>)
8732 declare i16 @llvm.cttz.i16 (i16 <src>, i1 <is_zero_undef>)
8733 declare i32 @llvm.cttz.i32 (i32 <src>, i1 <is_zero_undef>)
8734 declare i64 @llvm.cttz.i64 (i64 <src>, i1 <is_zero_undef>)
8735 declare i256 @llvm.cttz.i256(i256 <src>, i1 <is_zero_undef>)
8736 declase <2 x i32> @llvm.cttz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
8741 The '``llvm.cttz``' family of intrinsic functions counts the number of
8747 The first argument is the value to be counted. This argument may be of
8748 any integer type, or a vector with integer element type. The return
8749 type must match the first argument type.
8751 The second argument must be a constant and is a flag to indicate whether
8752 the intrinsic should ensure that a zero as the first argument produces a
8753 defined result. Historically some architectures did not provide a
8754 defined result for zero values as efficiently, and many algorithms are
8755 now predicated on avoiding zero-value inputs.
8760 The '``llvm.cttz``' intrinsic counts the trailing (least significant)
8761 zeros in a variable, or within each element of a vector. If ``src == 0``
8762 then the result is the size in bits of the type of ``src`` if
8763 ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
8764 ``llvm.cttz(2) = 1``.
8766 Arithmetic with Overflow Intrinsics
8767 -----------------------------------
8769 LLVM provides intrinsics for some arithmetic with overflow operations.
8771 '``llvm.sadd.with.overflow.*``' Intrinsics
8772 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8777 This is an overloaded intrinsic. You can use ``llvm.sadd.with.overflow``
8778 on any integer bit width.
8782 declare {i16, i1} @llvm.sadd.with.overflow.i16(i16 %a, i16 %b)
8783 declare {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
8784 declare {i64, i1} @llvm.sadd.with.overflow.i64(i64 %a, i64 %b)
8789 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
8790 a signed addition of the two arguments, and indicate whether an overflow
8791 occurred during the signed summation.
8796 The arguments (%a and %b) and the first element of the result structure
8797 may be of integer types of any bit width, but they must have the same
8798 bit width. The second element of the result structure must be of type
8799 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
8805 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
8806 a signed addition of the two variables. They return a structure --- the
8807 first element of which is the signed summation, and the second element
8808 of which is a bit specifying if the signed summation resulted in an
8814 .. code-block:: llvm
8816 %res = call {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
8817 %sum = extractvalue {i32, i1} %res, 0
8818 %obit = extractvalue {i32, i1} %res, 1
8819 br i1 %obit, label %overflow, label %normal
8821 '``llvm.uadd.with.overflow.*``' Intrinsics
8822 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8827 This is an overloaded intrinsic. You can use ``llvm.uadd.with.overflow``
8828 on any integer bit width.
8832 declare {i16, i1} @llvm.uadd.with.overflow.i16(i16 %a, i16 %b)
8833 declare {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
8834 declare {i64, i1} @llvm.uadd.with.overflow.i64(i64 %a, i64 %b)
8839 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
8840 an unsigned addition of the two arguments, and indicate whether a carry
8841 occurred during the unsigned summation.
8846 The arguments (%a and %b) and the first element of the result structure
8847 may be of integer types of any bit width, but they must have the same
8848 bit width. The second element of the result structure must be of type
8849 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
8855 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
8856 an unsigned addition of the two arguments. They return a structure --- the
8857 first element of which is the sum, and the second element of which is a
8858 bit specifying if the unsigned summation resulted in a carry.
8863 .. code-block:: llvm
8865 %res = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
8866 %sum = extractvalue {i32, i1} %res, 0
8867 %obit = extractvalue {i32, i1} %res, 1
8868 br i1 %obit, label %carry, label %normal
8870 '``llvm.ssub.with.overflow.*``' Intrinsics
8871 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8876 This is an overloaded intrinsic. You can use ``llvm.ssub.with.overflow``
8877 on any integer bit width.
8881 declare {i16, i1} @llvm.ssub.with.overflow.i16(i16 %a, i16 %b)
8882 declare {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
8883 declare {i64, i1} @llvm.ssub.with.overflow.i64(i64 %a, i64 %b)
8888 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
8889 a signed subtraction of the two arguments, and indicate whether an
8890 overflow occurred during the signed subtraction.
8895 The arguments (%a and %b) and the first element of the result structure
8896 may be of integer types of any bit width, but they must have the same
8897 bit width. The second element of the result structure must be of type
8898 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
8904 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
8905 a signed subtraction of the two arguments. They return a structure --- the
8906 first element of which is the subtraction, and the second element of
8907 which is a bit specifying if the signed subtraction resulted in an
8913 .. code-block:: llvm
8915 %res = call {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
8916 %sum = extractvalue {i32, i1} %res, 0
8917 %obit = extractvalue {i32, i1} %res, 1
8918 br i1 %obit, label %overflow, label %normal
8920 '``llvm.usub.with.overflow.*``' Intrinsics
8921 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8926 This is an overloaded intrinsic. You can use ``llvm.usub.with.overflow``
8927 on any integer bit width.
8931 declare {i16, i1} @llvm.usub.with.overflow.i16(i16 %a, i16 %b)
8932 declare {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
8933 declare {i64, i1} @llvm.usub.with.overflow.i64(i64 %a, i64 %b)
8938 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
8939 an unsigned subtraction of the two arguments, and indicate whether an
8940 overflow occurred during the unsigned subtraction.
8945 The arguments (%a and %b) and the first element of the result structure
8946 may be of integer types of any bit width, but they must have the same
8947 bit width. The second element of the result structure must be of type
8948 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
8954 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
8955 an unsigned subtraction of the two arguments. They return a structure ---
8956 the first element of which is the subtraction, and the second element of
8957 which is a bit specifying if the unsigned subtraction resulted in an
8963 .. code-block:: llvm
8965 %res = call {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
8966 %sum = extractvalue {i32, i1} %res, 0
8967 %obit = extractvalue {i32, i1} %res, 1
8968 br i1 %obit, label %overflow, label %normal
8970 '``llvm.smul.with.overflow.*``' Intrinsics
8971 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8976 This is an overloaded intrinsic. You can use ``llvm.smul.with.overflow``
8977 on any integer bit width.
8981 declare {i16, i1} @llvm.smul.with.overflow.i16(i16 %a, i16 %b)
8982 declare {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
8983 declare {i64, i1} @llvm.smul.with.overflow.i64(i64 %a, i64 %b)
8988 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
8989 a signed multiplication of the two arguments, and indicate whether an
8990 overflow occurred during the signed multiplication.
8995 The arguments (%a and %b) and the first element of the result structure
8996 may be of integer types of any bit width, but they must have the same
8997 bit width. The second element of the result structure must be of type
8998 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
9004 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
9005 a signed multiplication of the two arguments. They return a structure ---
9006 the first element of which is the multiplication, and the second element
9007 of which is a bit specifying if the signed multiplication resulted in an
9013 .. code-block:: llvm
9015 %res = call {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
9016 %sum = extractvalue {i32, i1} %res, 0
9017 %obit = extractvalue {i32, i1} %res, 1
9018 br i1 %obit, label %overflow, label %normal
9020 '``llvm.umul.with.overflow.*``' Intrinsics
9021 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9026 This is an overloaded intrinsic. You can use ``llvm.umul.with.overflow``
9027 on any integer bit width.
9031 declare {i16, i1} @llvm.umul.with.overflow.i16(i16 %a, i16 %b)
9032 declare {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
9033 declare {i64, i1} @llvm.umul.with.overflow.i64(i64 %a, i64 %b)
9038 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
9039 a unsigned multiplication of the two arguments, and indicate whether an
9040 overflow occurred during the unsigned multiplication.
9045 The arguments (%a and %b) and the first element of the result structure
9046 may be of integer types of any bit width, but they must have the same
9047 bit width. The second element of the result structure must be of type
9048 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
9054 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
9055 an unsigned multiplication of the two arguments. They return a structure ---
9056 the first element of which is the multiplication, and the second
9057 element of which is a bit specifying if the unsigned multiplication
9058 resulted in an overflow.
9063 .. code-block:: llvm
9065 %res = call {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
9066 %sum = extractvalue {i32, i1} %res, 0
9067 %obit = extractvalue {i32, i1} %res, 1
9068 br i1 %obit, label %overflow, label %normal
9070 Specialised Arithmetic Intrinsics
9071 ---------------------------------
9073 '``llvm.fmuladd.*``' Intrinsic
9074 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9081 declare float @llvm.fmuladd.f32(float %a, float %b, float %c)
9082 declare double @llvm.fmuladd.f64(double %a, double %b, double %c)
9087 The '``llvm.fmuladd.*``' intrinsic functions represent multiply-add
9088 expressions that can be fused if the code generator determines that (a) the
9089 target instruction set has support for a fused operation, and (b) that the
9090 fused operation is more efficient than the equivalent, separate pair of mul
9091 and add instructions.
9096 The '``llvm.fmuladd.*``' intrinsics each take three arguments: two
9097 multiplicands, a and b, and an addend c.
9106 %0 = call float @llvm.fmuladd.f32(%a, %b, %c)
9108 is equivalent to the expression a \* b + c, except that rounding will
9109 not be performed between the multiplication and addition steps if the
9110 code generator fuses the operations. Fusion is not guaranteed, even if
9111 the target platform supports it. If a fused multiply-add is required the
9112 corresponding llvm.fma.\* intrinsic function should be used
9113 instead. This never sets errno, just as '``llvm.fma.*``'.
9118 .. code-block:: llvm
9120 %r2 = call float @llvm.fmuladd.f32(float %a, float %b, float %c) ; yields float:r2 = (a * b) + c
9122 Half Precision Floating Point Intrinsics
9123 ----------------------------------------
9125 For most target platforms, half precision floating point is a
9126 storage-only format. This means that it is a dense encoding (in memory)
9127 but does not support computation in the format.
9129 This means that code must first load the half-precision floating point
9130 value as an i16, then convert it to float with
9131 :ref:`llvm.convert.from.fp16 <int_convert_from_fp16>`. Computation can
9132 then be performed on the float value (including extending to double
9133 etc). To store the value back to memory, it is first converted to float
9134 if needed, then converted to i16 with
9135 :ref:`llvm.convert.to.fp16 <int_convert_to_fp16>`, then storing as an
9138 .. _int_convert_to_fp16:
9140 '``llvm.convert.to.fp16``' Intrinsic
9141 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9148 declare i16 @llvm.convert.to.fp16.f32(float %a)
9149 declare i16 @llvm.convert.to.fp16.f64(double %a)
9154 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion from a
9155 conventional floating point type to half precision floating point format.
9160 The intrinsic function contains single argument - the value to be
9166 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion from a
9167 conventional floating point format to half precision floating point format. The
9168 return value is an ``i16`` which contains the converted number.
9173 .. code-block:: llvm
9175 %res = call i16 @llvm.convert.to.fp16.f32(float %a)
9176 store i16 %res, i16* @x, align 2
9178 .. _int_convert_from_fp16:
9180 '``llvm.convert.from.fp16``' Intrinsic
9181 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9188 declare float @llvm.convert.from.fp16.f32(i16 %a)
9189 declare double @llvm.convert.from.fp16.f64(i16 %a)
9194 The '``llvm.convert.from.fp16``' intrinsic function performs a
9195 conversion from half precision floating point format to single precision
9196 floating point format.
9201 The intrinsic function contains single argument - the value to be
9207 The '``llvm.convert.from.fp16``' intrinsic function performs a
9208 conversion from half single precision floating point format to single
9209 precision floating point format. The input half-float value is
9210 represented by an ``i16`` value.
9215 .. code-block:: llvm
9217 %a = load i16* @x, align 2
9218 %res = call float @llvm.convert.from.fp16(i16 %a)
9223 The LLVM debugger intrinsics (which all start with ``llvm.dbg.``
9224 prefix), are described in the `LLVM Source Level
9225 Debugging <SourceLevelDebugging.html#format_common_intrinsics>`_
9228 Exception Handling Intrinsics
9229 -----------------------------
9231 The LLVM exception handling intrinsics (which all start with
9232 ``llvm.eh.`` prefix), are described in the `LLVM Exception
9233 Handling <ExceptionHandling.html#format_common_intrinsics>`_ document.
9237 Trampoline Intrinsics
9238 ---------------------
9240 These intrinsics make it possible to excise one parameter, marked with
9241 the :ref:`nest <nest>` attribute, from a function. The result is a
9242 callable function pointer lacking the nest parameter - the caller does
9243 not need to provide a value for it. Instead, the value to use is stored
9244 in advance in a "trampoline", a block of memory usually allocated on the
9245 stack, which also contains code to splice the nest value into the
9246 argument list. This is used to implement the GCC nested function address
9249 For example, if the function is ``i32 f(i8* nest %c, i32 %x, i32 %y)``
9250 then the resulting function pointer has signature ``i32 (i32, i32)*``.
9251 It can be created as follows:
9253 .. code-block:: llvm
9255 %tramp = alloca [10 x i8], align 4 ; size and alignment only correct for X86
9256 %tramp1 = getelementptr [10 x i8]* %tramp, i32 0, i32 0
9257 call i8* @llvm.init.trampoline(i8* %tramp1, i8* bitcast (i32 (i8*, i32, i32)* @f to i8*), i8* %nval)
9258 %p = call i8* @llvm.adjust.trampoline(i8* %tramp1)
9259 %fp = bitcast i8* %p to i32 (i32, i32)*
9261 The call ``%val = call i32 %fp(i32 %x, i32 %y)`` is then equivalent to
9262 ``%val = call i32 %f(i8* %nval, i32 %x, i32 %y)``.
9266 '``llvm.init.trampoline``' Intrinsic
9267 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9274 declare void @llvm.init.trampoline(i8* <tramp>, i8* <func>, i8* <nval>)
9279 This fills the memory pointed to by ``tramp`` with executable code,
9280 turning it into a trampoline.
9285 The ``llvm.init.trampoline`` intrinsic takes three arguments, all
9286 pointers. The ``tramp`` argument must point to a sufficiently large and
9287 sufficiently aligned block of memory; this memory is written to by the
9288 intrinsic. Note that the size and the alignment are target-specific -
9289 LLVM currently provides no portable way of determining them, so a
9290 front-end that generates this intrinsic needs to have some
9291 target-specific knowledge. The ``func`` argument must hold a function
9292 bitcast to an ``i8*``.
9297 The block of memory pointed to by ``tramp`` is filled with target
9298 dependent code, turning it into a function. Then ``tramp`` needs to be
9299 passed to :ref:`llvm.adjust.trampoline <int_at>` to get a pointer which can
9300 be :ref:`bitcast (to a new function) and called <int_trampoline>`. The new
9301 function's signature is the same as that of ``func`` with any arguments
9302 marked with the ``nest`` attribute removed. At most one such ``nest``
9303 argument is allowed, and it must be of pointer type. Calling the new
9304 function is equivalent to calling ``func`` with the same argument list,
9305 but with ``nval`` used for the missing ``nest`` argument. If, after
9306 calling ``llvm.init.trampoline``, the memory pointed to by ``tramp`` is
9307 modified, then the effect of any later call to the returned function
9308 pointer is undefined.
9312 '``llvm.adjust.trampoline``' Intrinsic
9313 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9320 declare i8* @llvm.adjust.trampoline(i8* <tramp>)
9325 This performs any required machine-specific adjustment to the address of
9326 a trampoline (passed as ``tramp``).
9331 ``tramp`` must point to a block of memory which already has trampoline
9332 code filled in by a previous call to
9333 :ref:`llvm.init.trampoline <int_it>`.
9338 On some architectures the address of the code to be executed needs to be
9339 different than the address where the trampoline is actually stored. This
9340 intrinsic returns the executable address corresponding to ``tramp``
9341 after performing the required machine specific adjustments. The pointer
9342 returned can then be :ref:`bitcast and executed <int_trampoline>`.
9344 Masked Vector Load and Store Intrinsics
9345 ---------------------------------------
9347 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.
9351 '``llvm.masked.load.*``' Intrinsics
9352 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9356 This is an overloaded intrinsic. The loaded data is a vector of any integer or floating point data type.
9360 declare <16 x float> @llvm.masked.load.v16f32 (<16 x float>* <ptr>, i32 <alignment>, <16 x i1> <mask>, <16 x float> <passthru>)
9361 declare <2 x double> @llvm.masked.load.v2f64 (<2 x double>* <ptr>, i32 <alignment>, <2 x i1> <mask>, <2 x double> <passthru>)
9366 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 in the passthru operand.
9372 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 'i1' 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 passthru operand are the same vector types.
9378 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.
9379 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.
9384 %res = call <16 x float> @llvm.masked.load.v16f32 (<16 x float>* %ptr, i32 4, <16 x i1>%mask, <16 x float> %passthru)
9386 ;; The result of the two following instructions is identical aside from potential memory access exception
9387 %loadlal = load <16 x float>* %ptr, align 4
9388 %res = select <16 x i1> %mask, <16 x float> %loadlal, <16 x float> %passthru
9392 '``llvm.masked.store.*``' Intrinsics
9393 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9397 This is an overloaded intrinsic. The data stored in memory is a vector of any integer or floating point data type.
9401 declare void @llvm.masked.store.v8i32 (<8 x i32> <value>, <8 x i32> * <ptr>, i32 <alignment>, <8 x i1> <mask>)
9402 declare void @llvm.masked.store.v16f32(<16 x i32> <value>, <16 x i32>* <ptr>, i32 <alignment>, <16 x i1> <mask>)
9407 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.
9412 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.
9418 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.
9419 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.
9423 call void @llvm.masked.store.v16f32(<16 x float> %value, <16 x float>* %ptr, i32 4, <16 x i1> %mask)
9425 ;; The result of the following instructions is identical aside from potential data races and memory access exceptions
9426 %oldval = load <16 x float>* %ptr, align 4
9427 %res = select <16 x i1> %mask, <16 x float> %value, <16 x float> %oldval
9428 store <16 x float> %res, <16 x float>* %ptr, align 4
9434 This class of intrinsics provides information about the lifetime of
9435 memory objects and ranges where variables are immutable.
9439 '``llvm.lifetime.start``' Intrinsic
9440 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9447 declare void @llvm.lifetime.start(i64 <size>, i8* nocapture <ptr>)
9452 The '``llvm.lifetime.start``' intrinsic specifies the start of a memory
9458 The first argument is a constant integer representing the size of the
9459 object, or -1 if it is variable sized. The second argument is a pointer
9465 This intrinsic indicates that before this point in the code, the value
9466 of the memory pointed to by ``ptr`` is dead. This means that it is known
9467 to never be used and has an undefined value. A load from the pointer
9468 that precedes this intrinsic can be replaced with ``'undef'``.
9472 '``llvm.lifetime.end``' Intrinsic
9473 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9480 declare void @llvm.lifetime.end(i64 <size>, i8* nocapture <ptr>)
9485 The '``llvm.lifetime.end``' intrinsic specifies the end of a memory
9491 The first argument is a constant integer representing the size of the
9492 object, or -1 if it is variable sized. The second argument is a pointer
9498 This intrinsic indicates that after this point in the code, the value of
9499 the memory pointed to by ``ptr`` is dead. This means that it is known to
9500 never be used and has an undefined value. Any stores into the memory
9501 object following this intrinsic may be removed as dead.
9503 '``llvm.invariant.start``' Intrinsic
9504 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9511 declare {}* @llvm.invariant.start(i64 <size>, i8* nocapture <ptr>)
9516 The '``llvm.invariant.start``' intrinsic specifies that the contents of
9517 a memory object will not change.
9522 The first argument is a constant integer representing the size of the
9523 object, or -1 if it is variable sized. The second argument is a pointer
9529 This intrinsic indicates that until an ``llvm.invariant.end`` that uses
9530 the return value, the referenced memory location is constant and
9533 '``llvm.invariant.end``' Intrinsic
9534 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9541 declare void @llvm.invariant.end({}* <start>, i64 <size>, i8* nocapture <ptr>)
9546 The '``llvm.invariant.end``' intrinsic specifies that the contents of a
9547 memory object are mutable.
9552 The first argument is the matching ``llvm.invariant.start`` intrinsic.
9553 The second argument is a constant integer representing the size of the
9554 object, or -1 if it is variable sized and the third argument is a
9555 pointer to the object.
9560 This intrinsic indicates that the memory is mutable again.
9565 This class of intrinsics is designed to be generic and has no specific
9568 '``llvm.var.annotation``' Intrinsic
9569 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9576 declare void @llvm.var.annotation(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
9581 The '``llvm.var.annotation``' intrinsic.
9586 The first argument is a pointer to a value, the second is a pointer to a
9587 global string, the third is a pointer to a global string which is the
9588 source file name, and the last argument is the line number.
9593 This intrinsic allows annotation of local variables with arbitrary
9594 strings. This can be useful for special purpose optimizations that want
9595 to look for these annotations. These have no other defined use; they are
9596 ignored by code generation and optimization.
9598 '``llvm.ptr.annotation.*``' Intrinsic
9599 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9604 This is an overloaded intrinsic. You can use '``llvm.ptr.annotation``' on a
9605 pointer to an integer of any width. *NOTE* you must specify an address space for
9606 the pointer. The identifier for the default address space is the integer
9611 declare i8* @llvm.ptr.annotation.p<address space>i8(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
9612 declare i16* @llvm.ptr.annotation.p<address space>i16(i16* <val>, i8* <str>, i8* <str>, i32 <int>)
9613 declare i32* @llvm.ptr.annotation.p<address space>i32(i32* <val>, i8* <str>, i8* <str>, i32 <int>)
9614 declare i64* @llvm.ptr.annotation.p<address space>i64(i64* <val>, i8* <str>, i8* <str>, i32 <int>)
9615 declare i256* @llvm.ptr.annotation.p<address space>i256(i256* <val>, i8* <str>, i8* <str>, i32 <int>)
9620 The '``llvm.ptr.annotation``' intrinsic.
9625 The first argument is a pointer to an integer value of arbitrary bitwidth
9626 (result of some expression), the second is a pointer to a global string, the
9627 third is a pointer to a global string which is the source file name, and the
9628 last argument is the line number. It returns the value of the first argument.
9633 This intrinsic allows annotation of a pointer to an integer with arbitrary
9634 strings. This can be useful for special purpose optimizations that want to look
9635 for these annotations. These have no other defined use; they are ignored by code
9636 generation and optimization.
9638 '``llvm.annotation.*``' Intrinsic
9639 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9644 This is an overloaded intrinsic. You can use '``llvm.annotation``' on
9645 any integer bit width.
9649 declare i8 @llvm.annotation.i8(i8 <val>, i8* <str>, i8* <str>, i32 <int>)
9650 declare i16 @llvm.annotation.i16(i16 <val>, i8* <str>, i8* <str>, i32 <int>)
9651 declare i32 @llvm.annotation.i32(i32 <val>, i8* <str>, i8* <str>, i32 <int>)
9652 declare i64 @llvm.annotation.i64(i64 <val>, i8* <str>, i8* <str>, i32 <int>)
9653 declare i256 @llvm.annotation.i256(i256 <val>, i8* <str>, i8* <str>, i32 <int>)
9658 The '``llvm.annotation``' intrinsic.
9663 The first argument is an integer value (result of some expression), the
9664 second is a pointer to a global string, the third is a pointer to a
9665 global string which is the source file name, and the last argument is
9666 the line number. It returns the value of the first argument.
9671 This intrinsic allows annotations to be put on arbitrary expressions
9672 with arbitrary strings. This can be useful for special purpose
9673 optimizations that want to look for these annotations. These have no
9674 other defined use; they are ignored by code generation and optimization.
9676 '``llvm.trap``' Intrinsic
9677 ^^^^^^^^^^^^^^^^^^^^^^^^^
9684 declare void @llvm.trap() noreturn nounwind
9689 The '``llvm.trap``' intrinsic.
9699 This intrinsic is lowered to the target dependent trap instruction. If
9700 the target does not have a trap instruction, this intrinsic will be
9701 lowered to a call of the ``abort()`` function.
9703 '``llvm.debugtrap``' Intrinsic
9704 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9711 declare void @llvm.debugtrap() nounwind
9716 The '``llvm.debugtrap``' intrinsic.
9726 This intrinsic is lowered to code which is intended to cause an
9727 execution trap with the intention of requesting the attention of a
9730 '``llvm.stackprotector``' Intrinsic
9731 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9738 declare void @llvm.stackprotector(i8* <guard>, i8** <slot>)
9743 The ``llvm.stackprotector`` intrinsic takes the ``guard`` and stores it
9744 onto the stack at ``slot``. The stack slot is adjusted to ensure that it
9745 is placed on the stack before local variables.
9750 The ``llvm.stackprotector`` intrinsic requires two pointer arguments.
9751 The first argument is the value loaded from the stack guard
9752 ``@__stack_chk_guard``. The second variable is an ``alloca`` that has
9753 enough space to hold the value of the guard.
9758 This intrinsic causes the prologue/epilogue inserter to force the position of
9759 the ``AllocaInst`` stack slot to be before local variables on the stack. This is
9760 to ensure that if a local variable on the stack is overwritten, it will destroy
9761 the value of the guard. When the function exits, the guard on the stack is
9762 checked against the original guard by ``llvm.stackprotectorcheck``. If they are
9763 different, then ``llvm.stackprotectorcheck`` causes the program to abort by
9764 calling the ``__stack_chk_fail()`` function.
9766 '``llvm.stackprotectorcheck``' Intrinsic
9767 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9774 declare void @llvm.stackprotectorcheck(i8** <guard>)
9779 The ``llvm.stackprotectorcheck`` intrinsic compares ``guard`` against an already
9780 created stack protector and if they are not equal calls the
9781 ``__stack_chk_fail()`` function.
9786 The ``llvm.stackprotectorcheck`` intrinsic requires one pointer argument, the
9787 the variable ``@__stack_chk_guard``.
9792 This intrinsic is provided to perform the stack protector check by comparing
9793 ``guard`` with the stack slot created by ``llvm.stackprotector`` and if the
9794 values do not match call the ``__stack_chk_fail()`` function.
9796 The reason to provide this as an IR level intrinsic instead of implementing it
9797 via other IR operations is that in order to perform this operation at the IR
9798 level without an intrinsic, one would need to create additional basic blocks to
9799 handle the success/failure cases. This makes it difficult to stop the stack
9800 protector check from disrupting sibling tail calls in Codegen. With this
9801 intrinsic, we are able to generate the stack protector basic blocks late in
9802 codegen after the tail call decision has occurred.
9804 '``llvm.objectsize``' Intrinsic
9805 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9812 declare i32 @llvm.objectsize.i32(i8* <object>, i1 <min>)
9813 declare i64 @llvm.objectsize.i64(i8* <object>, i1 <min>)
9818 The ``llvm.objectsize`` intrinsic is designed to provide information to
9819 the optimizers to determine at compile time whether a) an operation
9820 (like memcpy) will overflow a buffer that corresponds to an object, or
9821 b) that a runtime check for overflow isn't necessary. An object in this
9822 context means an allocation of a specific class, structure, array, or
9828 The ``llvm.objectsize`` intrinsic takes two arguments. The first
9829 argument is a pointer to or into the ``object``. The second argument is
9830 a boolean and determines whether ``llvm.objectsize`` returns 0 (if true)
9831 or -1 (if false) when the object size is unknown. The second argument
9832 only accepts constants.
9837 The ``llvm.objectsize`` intrinsic is lowered to a constant representing
9838 the size of the object concerned. If the size cannot be determined at
9839 compile time, ``llvm.objectsize`` returns ``i32/i64 -1 or 0`` (depending
9840 on the ``min`` argument).
9842 '``llvm.expect``' Intrinsic
9843 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9848 This is an overloaded intrinsic. You can use ``llvm.expect`` on any
9853 declare i1 @llvm.expect.i1(i1 <val>, i1 <expected_val>)
9854 declare i32 @llvm.expect.i32(i32 <val>, i32 <expected_val>)
9855 declare i64 @llvm.expect.i64(i64 <val>, i64 <expected_val>)
9860 The ``llvm.expect`` intrinsic provides information about expected (the
9861 most probable) value of ``val``, which can be used by optimizers.
9866 The ``llvm.expect`` intrinsic takes two arguments. The first argument is
9867 a value. The second argument is an expected value, this needs to be a
9868 constant value, variables are not allowed.
9873 This intrinsic is lowered to the ``val``.
9875 '``llvm.assume``' Intrinsic
9876 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9883 declare void @llvm.assume(i1 %cond)
9888 The ``llvm.assume`` allows the optimizer to assume that the provided
9889 condition is true. This information can then be used in simplifying other parts
9895 The condition which the optimizer may assume is always true.
9900 The intrinsic allows the optimizer to assume that the provided condition is
9901 always true whenever the control flow reaches the intrinsic call. No code is
9902 generated for this intrinsic, and instructions that contribute only to the
9903 provided condition are not used for code generation. If the condition is
9904 violated during execution, the behavior is undefined.
9906 Note that the optimizer might limit the transformations performed on values
9907 used by the ``llvm.assume`` intrinsic in order to preserve the instructions
9908 only used to form the intrinsic's input argument. This might prove undesirable
9909 if the extra information provided by the ``llvm.assume`` intrinsic does not cause
9910 sufficient overall improvement in code quality. For this reason,
9911 ``llvm.assume`` should not be used to document basic mathematical invariants
9912 that the optimizer can otherwise deduce or facts that are of little use to the
9917 '``llvm.bitset.test``' Intrinsic
9918 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9925 declare i1 @llvm.bitset.test(i8* %ptr, metadata %bitset) nounwind readnone
9931 The first argument is a pointer to be tested. The second argument is a
9932 metadata string containing the name of a :doc:`bitset <BitSets>`.
9937 The ``llvm.bitset.test`` intrinsic tests whether the given pointer is a
9938 member of the given bitset.
9940 '``llvm.donothing``' Intrinsic
9941 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9948 declare void @llvm.donothing() nounwind readnone
9953 The ``llvm.donothing`` intrinsic doesn't perform any operation. It's one of only
9954 two intrinsics (besides ``llvm.experimental.patchpoint``) that can be called
9955 with an invoke instruction.
9965 This intrinsic does nothing, and it's removed by optimizers and ignored
9968 Stack Map Intrinsics
9969 --------------------
9971 LLVM provides experimental intrinsics to support runtime patching
9972 mechanisms commonly desired in dynamic language JITs. These intrinsics
9973 are described in :doc:`StackMaps`.