1 ==============================
2 LLVM Language Reference Manual
3 ==============================
12 This document is a reference manual for the LLVM assembly language. LLVM
13 is a Static Single Assignment (SSA) based representation that provides
14 type safety, low-level operations, flexibility, and the capability of
15 representing 'all' high-level languages cleanly. It is the common code
16 representation used throughout all phases of the LLVM compilation
22 The LLVM code representation is designed to be used in three different
23 forms: as an in-memory compiler IR, as an on-disk bitcode representation
24 (suitable for fast loading by a Just-In-Time compiler), and as a human
25 readable assembly language representation. This allows LLVM to provide a
26 powerful intermediate representation for efficient compiler
27 transformations and analysis, while providing a natural means to debug
28 and visualize the transformations. The three different forms of LLVM are
29 all equivalent. This document describes the human readable
30 representation and notation.
32 The LLVM representation aims to be light-weight and low-level while
33 being expressive, typed, and extensible at the same time. It aims to be
34 a "universal IR" of sorts, by being at a low enough level that
35 high-level ideas may be cleanly mapped to it (similar to how
36 microprocessors are "universal IR's", allowing many source languages to
37 be mapped to them). By providing type information, LLVM can be used as
38 the target of optimizations: for example, through pointer analysis, it
39 can be proven that a C automatic variable is never accessed outside of
40 the current function, allowing it to be promoted to a simple SSA value
41 instead of a memory location.
48 It is important to note that this document describes 'well formed' LLVM
49 assembly language. There is a difference between what the parser accepts
50 and what is considered 'well formed'. For example, the following
51 instruction is syntactically okay, but not well formed:
57 because the definition of ``%x`` does not dominate all of its uses. The
58 LLVM infrastructure provides a verification pass that may be used to
59 verify that an LLVM module is well formed. This pass is automatically
60 run by the parser after parsing input assembly and by the optimizer
61 before it outputs bitcode. The violations pointed out by the verifier
62 pass indicate bugs in transformation passes or input to the parser.
69 LLVM identifiers come in two basic types: global and local. Global
70 identifiers (functions, global variables) begin with the ``'@'``
71 character. Local identifiers (register names, types) begin with the
72 ``'%'`` character. Additionally, there are three different formats for
73 identifiers, for different purposes:
75 #. Named values are represented as a string of characters with their
76 prefix. For example, ``%foo``, ``@DivisionByZero``,
77 ``%a.really.long.identifier``. The actual regular expression used is
78 '``[%@][a-zA-Z$._][a-zA-Z$._0-9]*``'. Identifiers which require other
79 characters in their names can be surrounded with quotes. Special
80 characters may be escaped using ``"\xx"`` where ``xx`` is the ASCII
81 code for the character in hexadecimal. In this way, any character can
82 be used in a name value, even quotes themselves.
83 #. Unnamed values are represented as an unsigned numeric value with
84 their prefix. For example, ``%12``, ``@2``, ``%44``.
85 #. Constants, which are described in the section Constants_ below.
87 LLVM requires that values start with a prefix for two reasons: Compilers
88 don't need to worry about name clashes with reserved words, and the set
89 of reserved words may be expanded in the future without penalty.
90 Additionally, unnamed identifiers allow a compiler to quickly come up
91 with a temporary variable without having to avoid symbol table
94 Reserved words in LLVM are very similar to reserved words in other
95 languages. There are keywords for different opcodes ('``add``',
96 '``bitcast``', '``ret``', etc...), for primitive type names ('``void``',
97 '``i32``', etc...), and others. These reserved words cannot conflict
98 with variable names, because none of them start with a prefix character
101 Here is an example of LLVM code to multiply the integer variable
108 %result = mul i32 %X, 8
110 After strength reduction:
114 %result = shl i32 %X, 3
120 %0 = add i32 %X, %X ; yields {i32}:%0
121 %1 = add i32 %0, %0 ; yields {i32}:%1
122 %result = add i32 %1, %1
124 This last way of multiplying ``%X`` by 8 illustrates several important
125 lexical features of LLVM:
127 #. Comments are delimited with a '``;``' and go until the end of line.
128 #. Unnamed temporaries are created when the result of a computation is
129 not assigned to a named value.
130 #. Unnamed temporaries are numbered sequentially (using a per-function
131 incrementing counter, starting with 0). Note that basic blocks are
132 included in this numbering. For example, if the entry basic block is not
133 given a label name, then it will get number 0.
135 It also shows a convention that we follow in this document. When
136 demonstrating instructions, we will follow an instruction with a comment
137 that defines the type and name of value produced.
145 LLVM programs are composed of ``Module``'s, each of which is a
146 translation unit of the input programs. Each module consists of
147 functions, global variables, and symbol table entries. Modules may be
148 combined together with the LLVM linker, which merges function (and
149 global variable) definitions, resolves forward declarations, and merges
150 symbol table entries. Here is an example of the "hello world" module:
154 ; Declare the string constant as a global constant.
155 @.str = private unnamed_addr constant [13 x i8] c"hello world\0A\00"
157 ; External declaration of the puts function
158 declare i32 @puts(i8* nocapture) nounwind
160 ; Definition of main function
161 define i32 @main() { ; i32()*
162 ; Convert [13 x i8]* to i8 *...
163 %cast210 = getelementptr [13 x i8]* @.str, i64 0, i64 0
165 ; Call puts function to write out the string to stdout.
166 call i32 @puts(i8* %cast210)
171 !1 = metadata !{i32 42}
174 This example is made up of a :ref:`global variable <globalvars>` named
175 "``.str``", an external declaration of the "``puts``" function, a
176 :ref:`function definition <functionstructure>` for "``main``" and
177 :ref:`named metadata <namedmetadatastructure>` "``foo``".
179 In general, a module is made up of a list of global values (where both
180 functions and global variables are global values). Global values are
181 represented by a pointer to a memory location (in this case, a pointer
182 to an array of char, and a pointer to a function), and have one of the
183 following :ref:`linkage types <linkage>`.
190 All Global Variables and Functions have one of the following types of
194 Global values with "``private``" linkage are only directly
195 accessible by objects in the current module. In particular, linking
196 code into a module with an private global value may cause the
197 private to be renamed as necessary to avoid collisions. Because the
198 symbol is private to the module, all references can be updated. This
199 doesn't show up in any symbol table in the object file.
201 Similar to private, but the value shows as a local symbol
202 (``STB_LOCAL`` in the case of ELF) in the object file. This
203 corresponds to the notion of the '``static``' keyword in C.
204 ``available_externally``
205 Globals with "``available_externally``" linkage are never emitted
206 into the object file corresponding to the LLVM module. They exist to
207 allow inlining and other optimizations to take place given knowledge
208 of the definition of the global, which is known to be somewhere
209 outside the module. Globals with ``available_externally`` linkage
210 are allowed to be discarded at will, and are otherwise the same as
211 ``linkonce_odr``. This linkage type is only allowed on definitions,
214 Globals with "``linkonce``" linkage are merged with other globals of
215 the same name when linkage occurs. This can be used to implement
216 some forms of inline functions, templates, or other code which must
217 be generated in each translation unit that uses it, but where the
218 body may be overridden with a more definitive definition later.
219 Unreferenced ``linkonce`` globals are allowed to be discarded. Note
220 that ``linkonce`` linkage does not actually allow the optimizer to
221 inline the body of this function into callers because it doesn't
222 know if this definition of the function is the definitive definition
223 within the program or whether it will be overridden by a stronger
224 definition. To enable inlining and other optimizations, use
225 "``linkonce_odr``" linkage.
227 "``weak``" linkage has the same merging semantics as ``linkonce``
228 linkage, except that unreferenced globals with ``weak`` linkage may
229 not be discarded. This is used for globals that are declared "weak"
232 "``common``" linkage is most similar to "``weak``" linkage, but they
233 are used for tentative definitions in C, such as "``int X;``" at
234 global scope. Symbols with "``common``" linkage are merged in the
235 same way as ``weak symbols``, and they may not be deleted if
236 unreferenced. ``common`` symbols may not have an explicit section,
237 must have a zero initializer, and may not be marked
238 ':ref:`constant <globalvars>`'. Functions and aliases may not have
241 .. _linkage_appending:
244 "``appending``" linkage may only be applied to global variables of
245 pointer to array type. When two global variables with appending
246 linkage are linked together, the two global arrays are appended
247 together. This is the LLVM, typesafe, equivalent of having the
248 system linker append together "sections" with identical names when
251 The semantics of this linkage follow the ELF object file model: the
252 symbol is weak until linked, if not linked, the symbol becomes null
253 instead of being an undefined reference.
254 ``linkonce_odr``, ``weak_odr``
255 Some languages allow differing globals to be merged, such as two
256 functions with different semantics. Other languages, such as
257 ``C++``, ensure that only equivalent globals are ever merged (the
258 "one definition rule" --- "ODR"). Such languages can use the
259 ``linkonce_odr`` and ``weak_odr`` linkage types to indicate that the
260 global will only be merged with equivalent globals. These linkage
261 types are otherwise the same as their non-``odr`` versions.
263 If none of the above identifiers are used, the global is externally
264 visible, meaning that it participates in linkage and can be used to
265 resolve external symbol references.
267 It is illegal for a function *declaration* to have any linkage type
268 other than ``external`` or ``extern_weak``.
275 LLVM :ref:`functions <functionstructure>`, :ref:`calls <i_call>` and
276 :ref:`invokes <i_invoke>` can all have an optional calling convention
277 specified for the call. The calling convention of any pair of dynamic
278 caller/callee must match, or the behavior of the program is undefined.
279 The following calling conventions are supported by LLVM, and more may be
282 "``ccc``" - The C calling convention
283 This calling convention (the default if no other calling convention
284 is specified) matches the target C calling conventions. This calling
285 convention supports varargs function calls and tolerates some
286 mismatch in the declared prototype and implemented declaration of
287 the function (as does normal C).
288 "``fastcc``" - The fast calling convention
289 This calling convention attempts to make calls as fast as possible
290 (e.g. by passing things in registers). This calling convention
291 allows the target to use whatever tricks it wants to produce fast
292 code for the target, without having to conform to an externally
293 specified ABI (Application Binary Interface). `Tail calls can only
294 be optimized when this, the GHC or the HiPE convention is
295 used. <CodeGenerator.html#id80>`_ This calling convention does not
296 support varargs and requires the prototype of all callees to exactly
297 match the prototype of the function definition.
298 "``coldcc``" - The cold calling convention
299 This calling convention attempts to make code in the caller as
300 efficient as possible under the assumption that the call is not
301 commonly executed. As such, these calls often preserve all registers
302 so that the call does not break any live ranges in the caller side.
303 This calling convention does not support varargs and requires the
304 prototype of all callees to exactly match the prototype of the
305 function definition. Furthermore the inliner doesn't consider such function
307 "``cc 10``" - GHC convention
308 This calling convention has been implemented specifically for use by
309 the `Glasgow Haskell Compiler (GHC) <http://www.haskell.org/ghc>`_.
310 It passes everything in registers, going to extremes to achieve this
311 by disabling callee save registers. This calling convention should
312 not be used lightly but only for specific situations such as an
313 alternative to the *register pinning* performance technique often
314 used when implementing functional programming languages. At the
315 moment only X86 supports this convention and it has the following
318 - On *X86-32* only supports up to 4 bit type parameters. No
319 floating point types are supported.
320 - On *X86-64* only supports up to 10 bit type parameters and 6
321 floating point parameters.
323 This calling convention supports `tail call
324 optimization <CodeGenerator.html#id80>`_ but requires both the
325 caller and callee are using it.
326 "``cc 11``" - The HiPE calling convention
327 This calling convention has been implemented specifically for use by
328 the `High-Performance Erlang
329 (HiPE) <http://www.it.uu.se/research/group/hipe/>`_ compiler, *the*
330 native code compiler of the `Ericsson's Open Source Erlang/OTP
331 system <http://www.erlang.org/download.shtml>`_. It uses more
332 registers for argument passing than the ordinary C calling
333 convention and defines no callee-saved registers. The calling
334 convention properly supports `tail call
335 optimization <CodeGenerator.html#id80>`_ but requires that both the
336 caller and the callee use it. It uses a *register pinning*
337 mechanism, similar to GHC's convention, for keeping frequently
338 accessed runtime components pinned to specific hardware registers.
339 At the moment only X86 supports this convention (both 32 and 64
341 "``webkit_jscc``" - WebKit's JavaScript calling convention
342 This calling convention has been implemented for `WebKit FTL JIT
343 <https://trac.webkit.org/wiki/FTLJIT>`_. It passes arguments on the
344 stack right to left (as cdecl does), and returns a value in the
345 platform's customary return register.
346 "``anyregcc``" - Dynamic calling convention for code patching
347 This is a special convention that supports patching an arbitrary code
348 sequence in place of a call site. This convention forces the call
349 arguments into registers but allows them to be dynamcially
350 allocated. This can currently only be used with calls to
351 llvm.experimental.patchpoint because only this intrinsic records
352 the location of its arguments in a side table. See :doc:`StackMaps`.
353 "``preserve_mostcc``" - The `PreserveMost` calling convention
354 This calling convention attempts to make the code in the caller as little
355 intrusive as possible. This calling convention behaves identical to the `C`
356 calling convention on how arguments and return values are passed, but it
357 uses a different set of caller/callee-saved registers. This alleviates the
358 burden of saving and recovering a large register set before and after the
359 call in the caller. If the arguments are passed in callee-saved registers,
360 then they will be preserved by the callee across the call. This doesn't
361 apply for values returned in callee-saved registers.
363 - On X86-64 the callee preserves all general purpose registers, except for
364 R11. R11 can be used as a scratch register. Floating-point registers
365 (XMMs/YMMs) are not preserved and need to be saved by the caller.
367 The idea behind this convention is to support calls to runtime functions
368 that have a hot path and a cold path. The hot path is usually a small piece
369 of code that doesn't many registers. The cold path might need to call out to
370 another function and therefore only needs to preserve the caller-saved
371 registers, which haven't already been saved by the caller. The
372 `PreserveMost` calling convention is very similar to the `cold` calling
373 convention in terms of caller/callee-saved registers, but they are used for
374 different types of function calls. `coldcc` is for function calls that are
375 rarely executed, whereas `preserve_mostcc` function calls are intended to be
376 on the hot path and definitely executed a lot. Furthermore `preserve_mostcc`
377 doesn't prevent the inliner from inlining the function call.
379 This calling convention will be used by a future version of the ObjectiveC
380 runtime and should therefore still be considered experimental at this time.
381 Although this convention was created to optimize certain runtime calls to
382 the ObjectiveC runtime, it is not limited to this runtime and might be used
383 by other runtimes in the future too. The current implementation only
384 supports X86-64, but the intention is to support more architectures in the
386 "``preserve_allcc``" - The `PreserveAll` calling convention
387 This calling convention attempts to make the code in the caller even less
388 intrusive than the `PreserveMost` calling convention. This calling
389 convention also behaves identical to the `C` calling convention on how
390 arguments and return values are passed, but it uses a different set of
391 caller/callee-saved registers. This removes the burden of saving and
392 recovering a large register set before and after the call in the caller. If
393 the arguments are passed in callee-saved registers, then they will be
394 preserved by the callee across the call. This doesn't apply for values
395 returned in callee-saved registers.
397 - On X86-64 the callee preserves all general purpose registers, except for
398 R11. R11 can be used as a scratch register. Furthermore it also preserves
399 all floating-point registers (XMMs/YMMs).
401 The idea behind this convention is to support calls to runtime functions
402 that don't need to call out to any other functions.
404 This calling convention, like the `PreserveMost` calling convention, will be
405 used by a future version of the ObjectiveC runtime and should be considered
406 experimental at this time.
407 "``cc <n>``" - Numbered convention
408 Any calling convention may be specified by number, allowing
409 target-specific calling conventions to be used. Target specific
410 calling conventions start at 64.
412 More calling conventions can be added/defined on an as-needed basis, to
413 support Pascal conventions or any other well-known target-independent
416 .. _visibilitystyles:
421 All Global Variables and Functions have one of the following visibility
424 "``default``" - Default style
425 On targets that use the ELF object file format, default visibility
426 means that the declaration is visible to other modules and, in
427 shared libraries, means that the declared entity may be overridden.
428 On Darwin, default visibility means that the declaration is visible
429 to other modules. Default visibility corresponds to "external
430 linkage" in the language.
431 "``hidden``" - Hidden style
432 Two declarations of an object with hidden visibility refer to the
433 same object if they are in the same shared object. Usually, hidden
434 visibility indicates that the symbol will not be placed into the
435 dynamic symbol table, so no other module (executable or shared
436 library) can reference it directly.
437 "``protected``" - Protected style
438 On ELF, protected visibility indicates that the symbol will be
439 placed in the dynamic symbol table, but that references within the
440 defining module will bind to the local symbol. That is, the symbol
441 cannot be overridden by another module.
443 A symbol with ``internal`` or ``private`` linkage must have ``default``
451 All Global Variables, Functions and Aliases can have one of the following
455 "``dllimport``" causes the compiler to reference a function or variable via
456 a global pointer to a pointer that is set up by the DLL exporting the
457 symbol. On Microsoft Windows targets, the pointer name is formed by
458 combining ``__imp_`` and the function or variable name.
460 "``dllexport``" causes the compiler to provide a global pointer to a pointer
461 in a DLL, so that it can be referenced with the ``dllimport`` attribute. On
462 Microsoft Windows targets, the pointer name is formed by combining
463 ``__imp_`` and the function or variable name. Since this storage class
464 exists for defining a dll interface, the compiler, assembler and linker know
465 it is externally referenced and must refrain from deleting the symbol.
470 LLVM IR allows you to specify both "identified" and "literal" :ref:`structure
471 types <t_struct>`. Literal types are uniqued structurally, but identified types
472 are never uniqued. An :ref:`opaque structural type <t_opaque>` can also be used
473 to forward declare a type which is not yet available.
475 An example of a identified structure specification is:
479 %mytype = type { %mytype*, i32 }
481 Prior to the LLVM 3.0 release, identified types were structurally uniqued. Only
482 literal types are uniqued in recent versions of LLVM.
489 Global variables define regions of memory allocated at compilation time
492 Global variables definitions must be initialized, may have an explicit section
493 to be placed in, and may have an optional explicit alignment specified.
495 Global variables in other translation units can also be declared, in which
496 case they don't have an initializer.
498 A variable may be defined as ``thread_local``, which means that it will
499 not be shared by threads (each thread will have a separated copy of the
500 variable). Not all targets support thread-local variables. Optionally, a
501 TLS model may be specified:
504 For variables that are only used within the current shared library.
506 For variables in modules that will not be loaded dynamically.
508 For variables defined in the executable and only used within it.
510 The models correspond to the ELF TLS models; see `ELF Handling For
511 Thread-Local Storage <http://people.redhat.com/drepper/tls.pdf>`_ for
512 more information on under which circumstances the different models may
513 be used. The target may choose a different TLS model if the specified
514 model is not supported, or if a better choice of model can be made.
516 A variable may be defined as a global ``constant``, which indicates that
517 the contents of the variable will **never** be modified (enabling better
518 optimization, allowing the global data to be placed in the read-only
519 section of an executable, etc). Note that variables that need runtime
520 initialization cannot be marked ``constant`` as there is a store to the
523 LLVM explicitly allows *declarations* of global variables to be marked
524 constant, even if the final definition of the global is not. This
525 capability can be used to enable slightly better optimization of the
526 program, but requires the language definition to guarantee that
527 optimizations based on the 'constantness' are valid for the translation
528 units that do not include the definition.
530 As SSA values, global variables define pointer values that are in scope
531 (i.e. they dominate) all basic blocks in the program. Global variables
532 always define a pointer to their "content" type because they describe a
533 region of memory, and all memory objects in LLVM are accessed through
536 Global variables can be marked with ``unnamed_addr`` which indicates
537 that the address is not significant, only the content. Constants marked
538 like this can be merged with other constants if they have the same
539 initializer. Note that a constant with significant address *can* be
540 merged with a ``unnamed_addr`` constant, the result being a constant
541 whose address is significant.
543 A global variable may be declared to reside in a target-specific
544 numbered address space. For targets that support them, address spaces
545 may affect how optimizations are performed and/or what target
546 instructions are used to access the variable. The default address space
547 is zero. The address space qualifier must precede any other attributes.
549 LLVM allows an explicit section to be specified for globals. If the
550 target supports it, it will emit globals to the section specified.
552 By default, global initializers are optimized by assuming that global
553 variables defined within the module are not modified from their
554 initial values before the start of the global initializer. This is
555 true even for variables potentially accessible from outside the
556 module, including those with external linkage or appearing in
557 ``@llvm.used`` or dllexported variables. This assumption may be suppressed
558 by marking the variable with ``externally_initialized``.
560 An explicit alignment may be specified for a global, which must be a
561 power of 2. If not present, or if the alignment is set to zero, the
562 alignment of the global is set by the target to whatever it feels
563 convenient. If an explicit alignment is specified, the global is forced
564 to have exactly that alignment. Targets and optimizers are not allowed
565 to over-align the global if the global has an assigned section. In this
566 case, the extra alignment could be observable: for example, code could
567 assume that the globals are densely packed in their section and try to
568 iterate over them as an array, alignment padding would break this
571 Globals can also have a :ref:`DLL storage class <dllstorageclass>`.
575 [@<GlobalVarName> =] [Linkage] [Visibility] [DLLStorageClass] [ThreadLocal]
576 [AddrSpace] [unnamed_addr] [ExternallyInitialized]
577 <global | constant> <Type>
578 [, section "name"] [, align <Alignment>]
580 For example, the following defines a global in a numbered address space
581 with an initializer, section, and alignment:
585 @G = addrspace(5) constant float 1.0, section "foo", align 4
587 The following example just declares a global variable
591 @G = external global i32
593 The following example defines a thread-local global with the
594 ``initialexec`` TLS model:
598 @G = thread_local(initialexec) global i32 0, align 4
600 .. _functionstructure:
605 LLVM function definitions consist of the "``define``" keyword, an
606 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
607 style <visibility>`, an optional :ref:`DLL storage class <dllstorageclass>`,
608 an optional :ref:`calling convention <callingconv>`,
609 an optional ``unnamed_addr`` attribute, a return type, an optional
610 :ref:`parameter attribute <paramattrs>` for the return type, a function
611 name, a (possibly empty) argument list (each with optional :ref:`parameter
612 attributes <paramattrs>`), optional :ref:`function attributes <fnattrs>`,
613 an optional section, an optional alignment, an optional :ref:`garbage
614 collector name <gc>`, an optional :ref:`prefix <prefixdata>`, an opening
615 curly brace, a list of basic blocks, and a closing curly brace.
617 LLVM function declarations consist of the "``declare``" keyword, an
618 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
619 style <visibility>`, an optional :ref:`DLL storage class <dllstorageclass>`,
620 an optional :ref:`calling convention <callingconv>`,
621 an optional ``unnamed_addr`` attribute, a return type, an optional
622 :ref:`parameter attribute <paramattrs>` for the return type, a function
623 name, a possibly empty list of arguments, an optional alignment, an optional
624 :ref:`garbage collector name <gc>` and an optional :ref:`prefix <prefixdata>`.
626 A function definition contains a list of basic blocks, forming the CFG (Control
627 Flow Graph) for the function. Each basic block may optionally start with a label
628 (giving the basic block a symbol table entry), contains a list of instructions,
629 and ends with a :ref:`terminator <terminators>` instruction (such as a branch or
630 function return). If an explicit label is not provided, a block is assigned an
631 implicit numbered label, using the next value from the same counter as used for
632 unnamed temporaries (:ref:`see above<identifiers>`). For example, if a function
633 entry block does not have an explicit label, it will be assigned label "%0",
634 then the first unnamed temporary in that block will be "%1", etc.
636 The first basic block in a function is special in two ways: it is
637 immediately executed on entrance to the function, and it is not allowed
638 to have predecessor basic blocks (i.e. there can not be any branches to
639 the entry block of a function). Because the block can have no
640 predecessors, it also cannot have any :ref:`PHI nodes <i_phi>`.
642 LLVM allows an explicit section to be specified for functions. If the
643 target supports it, it will emit functions to the section specified.
645 An explicit alignment may be specified for a function. If not present,
646 or if the alignment is set to zero, the alignment of the function is set
647 by the target to whatever it feels convenient. If an explicit alignment
648 is specified, the function is forced to have at least that much
649 alignment. All alignments must be a power of 2.
651 If the ``unnamed_addr`` attribute is given, the address is know to not
652 be significant and two identical functions can be merged.
656 define [linkage] [visibility] [DLLStorageClass]
658 <ResultType> @<FunctionName> ([argument list])
659 [unnamed_addr] [fn Attrs] [section "name"] [align N]
660 [gc] [prefix Constant] { ... }
667 Aliases act as "second name" for the aliasee value (which can be either
668 function, global variable, another alias or bitcast of global value).
669 Aliases may have an optional :ref:`linkage type <linkage>`, an optional
670 :ref:`visibility style <visibility>`, and an optional :ref:`DLL storage class
675 @<Name> = [Visibility] [DLLStorageClass] alias [Linkage] <AliaseeTy> @<Aliasee>
677 The linkage must be one of ``private``, ``internal``, ``linkonce``, ``weak``,
678 ``linkonce_odr``, ``weak_odr``, ``external``. Note that some system linkers
679 might not correctly handle dropping a weak symbol that is aliased by a non-weak
682 Alias that are not ``unnamed_addr`` are guaranteed to have the same address as
685 The aliasee must be a definition.
687 Aliases are not allowed to point to aliases with linkages that can be
688 overridden. Since they are only a second name, the possibility of the
689 intermediate alias being overridden cannot be represented in an object file.
691 .. _namedmetadatastructure:
696 Named metadata is a collection of metadata. :ref:`Metadata
697 nodes <metadata>` (but not metadata strings) are the only valid
698 operands for a named metadata.
702 ; Some unnamed metadata nodes, which are referenced by the named metadata.
703 !0 = metadata !{metadata !"zero"}
704 !1 = metadata !{metadata !"one"}
705 !2 = metadata !{metadata !"two"}
707 !name = !{!0, !1, !2}
714 The return type and each parameter of a function type may have a set of
715 *parameter attributes* associated with them. Parameter attributes are
716 used to communicate additional information about the result or
717 parameters of a function. Parameter attributes are considered to be part
718 of the function, not of the function type, so functions with different
719 parameter attributes can have the same function type.
721 Parameter attributes are simple keywords that follow the type specified.
722 If multiple parameter attributes are needed, they are space separated.
727 declare i32 @printf(i8* noalias nocapture, ...)
728 declare i32 @atoi(i8 zeroext)
729 declare signext i8 @returns_signed_char()
731 Note that any attributes for the function result (``nounwind``,
732 ``readonly``) come immediately after the argument list.
734 Currently, only the following parameter attributes are defined:
737 This indicates to the code generator that the parameter or return
738 value should be zero-extended to the extent required by the target's
739 ABI (which is usually 32-bits, but is 8-bits for a i1 on x86-64) by
740 the caller (for a parameter) or the callee (for a return value).
742 This indicates to the code generator that the parameter or return
743 value should be sign-extended to the extent required by the target's
744 ABI (which is usually 32-bits) by the caller (for a parameter) or
745 the callee (for a return value).
747 This indicates that this parameter or return value should be treated
748 in a special target-dependent fashion during while emitting code for
749 a function call or return (usually, by putting it in a register as
750 opposed to memory, though some targets use it to distinguish between
751 two different kinds of registers). Use of this attribute is
754 This indicates that the pointer parameter should really be passed by
755 value to the function. The attribute implies that a hidden copy of
756 the pointee is made between the caller and the callee, so the callee
757 is unable to modify the value in the caller. This attribute is only
758 valid on LLVM pointer arguments. It is generally used to pass
759 structs and arrays by value, but is also valid on pointers to
760 scalars. The copy is considered to belong to the caller not the
761 callee (for example, ``readonly`` functions should not write to
762 ``byval`` parameters). This is not a valid attribute for return
765 The byval attribute also supports specifying an alignment with the
766 align attribute. It indicates the alignment of the stack slot to
767 form and the known alignment of the pointer specified to the call
768 site. If the alignment is not specified, then the code generator
769 makes a target-specific assumption.
775 The ``inalloca`` argument attribute allows the caller to take the
776 address of outgoing stack arguments. An ``inalloca`` argument must
777 be a pointer to stack memory produced by an ``alloca`` instruction.
778 The alloca, or argument allocation, must also be tagged with the
779 inalloca keyword. Only the past argument may have the ``inalloca``
780 attribute, and that argument is guaranteed to be passed in memory.
782 An argument allocation may be used by a call at most once because
783 the call may deallocate it. The ``inalloca`` attribute cannot be
784 used in conjunction with other attributes that affect argument
785 storage, like ``inreg``, ``nest``, ``sret``, or ``byval``. The
786 ``inalloca`` attribute also disables LLVM's implicit lowering of
787 large aggregate return values, which means that frontend authors
788 must lower them with ``sret`` pointers.
790 When the call site is reached, the argument allocation must have
791 been the most recent stack allocation that is still live, or the
792 results are undefined. It is possible to allocate additional stack
793 space after an argument allocation and before its call site, but it
794 must be cleared off with :ref:`llvm.stackrestore
797 See :doc:`InAlloca` for more information on how to use this
801 This indicates that the pointer parameter specifies the address of a
802 structure that is the return value of the function in the source
803 program. This pointer must be guaranteed by the caller to be valid:
804 loads and stores to the structure may be assumed by the callee
805 not to trap and to be properly aligned. This may only be applied to
806 the first parameter. This is not a valid attribute for return
812 This indicates that pointer values :ref:`based <pointeraliasing>` on
813 the argument or return value do not alias pointer values which are
814 not *based* on it, ignoring certain "irrelevant" dependencies. For a
815 call to the parent function, dependencies between memory references
816 from before or after the call and from those during the call are
817 "irrelevant" to the ``noalias`` keyword for the arguments and return
818 value used in that call. The caller shares the responsibility with
819 the callee for ensuring that these requirements are met. For further
820 details, please see the discussion of the NoAlias response in :ref:`alias
821 analysis <Must, May, or No>`.
823 Note that this definition of ``noalias`` is intentionally similar
824 to the definition of ``restrict`` in C99 for function arguments,
825 though it is slightly weaker.
827 For function return values, C99's ``restrict`` is not meaningful,
828 while LLVM's ``noalias`` is.
830 This indicates that the callee does not make any copies of the
831 pointer that outlive the callee itself. This is not a valid
832 attribute for return values.
837 This indicates that the pointer parameter can be excised using the
838 :ref:`trampoline intrinsics <int_trampoline>`. This is not a valid
839 attribute for return values and can only be applied to one parameter.
842 This indicates that the function always returns the argument as its return
843 value. This is an optimization hint to the code generator when generating
844 the caller, allowing tail call optimization and omission of register saves
845 and restores in some cases; it is not checked or enforced when generating
846 the callee. The parameter and the function return type must be valid
847 operands for the :ref:`bitcast instruction <i_bitcast>`. This is not a
848 valid attribute for return values and can only be applied to one parameter.
852 Garbage Collector Names
853 -----------------------
855 Each function may specify a garbage collector name, which is simply a
860 define void @f() gc "name" { ... }
862 The compiler declares the supported values of *name*. Specifying a
863 collector which will cause the compiler to alter its output in order to
864 support the named garbage collection algorithm.
871 Prefix data is data associated with a function which the code generator
872 will emit immediately before the function body. The purpose of this feature
873 is to allow frontends to associate language-specific runtime metadata with
874 specific functions and make it available through the function pointer while
875 still allowing the function pointer to be called. To access the data for a
876 given function, a program may bitcast the function pointer to a pointer to
877 the constant's type. This implies that the IR symbol points to the start
880 To maintain the semantics of ordinary function calls, the prefix data must
881 have a particular format. Specifically, it must begin with a sequence of
882 bytes which decode to a sequence of machine instructions, valid for the
883 module's target, which transfer control to the point immediately succeeding
884 the prefix data, without performing any other visible action. This allows
885 the inliner and other passes to reason about the semantics of the function
886 definition without needing to reason about the prefix data. Obviously this
887 makes the format of the prefix data highly target dependent.
889 Prefix data is laid out as if it were an initializer for a global variable
890 of the prefix data's type. No padding is automatically placed between the
891 prefix data and the function body. If padding is required, it must be part
894 A trivial example of valid prefix data for the x86 architecture is ``i8 144``,
895 which encodes the ``nop`` instruction:
899 define void @f() prefix i8 144 { ... }
901 Generally prefix data can be formed by encoding a relative branch instruction
902 which skips the metadata, as in this example of valid prefix data for the
903 x86_64 architecture, where the first two bytes encode ``jmp .+10``:
907 %0 = type <{ i8, i8, i8* }>
909 define void @f() prefix %0 <{ i8 235, i8 8, i8* @md}> { ... }
911 A function may have prefix data but no body. This has similar semantics
912 to the ``available_externally`` linkage in that the data may be used by the
913 optimizers but will not be emitted in the object file.
920 Attribute groups are groups of attributes that are referenced by objects within
921 the IR. They are important for keeping ``.ll`` files readable, because a lot of
922 functions will use the same set of attributes. In the degenerative case of a
923 ``.ll`` file that corresponds to a single ``.c`` file, the single attribute
924 group will capture the important command line flags used to build that file.
926 An attribute group is a module-level object. To use an attribute group, an
927 object references the attribute group's ID (e.g. ``#37``). An object may refer
928 to more than one attribute group. In that situation, the attributes from the
929 different groups are merged.
931 Here is an example of attribute groups for a function that should always be
932 inlined, has a stack alignment of 4, and which shouldn't use SSE instructions:
936 ; Target-independent attributes:
937 attributes #0 = { alwaysinline alignstack=4 }
939 ; Target-dependent attributes:
940 attributes #1 = { "no-sse" }
942 ; Function @f has attributes: alwaysinline, alignstack=4, and "no-sse".
943 define void @f() #0 #1 { ... }
950 Function attributes are set to communicate additional information about
951 a function. Function attributes are considered to be part of the
952 function, not of the function type, so functions with different function
953 attributes can have the same function type.
955 Function attributes are simple keywords that follow the type specified.
956 If multiple attributes are needed, they are space separated. For
961 define void @f() noinline { ... }
962 define void @f() alwaysinline { ... }
963 define void @f() alwaysinline optsize { ... }
964 define void @f() optsize { ... }
967 This attribute indicates that, when emitting the prologue and
968 epilogue, the backend should forcibly align the stack pointer.
969 Specify the desired alignment, which must be a power of two, in
972 This attribute indicates that the inliner should attempt to inline
973 this function into callers whenever possible, ignoring any active
974 inlining size threshold for this caller.
976 This indicates that the callee function at a call site should be
977 recognized as a built-in function, even though the function's declaration
978 uses the ``nobuiltin`` attribute. This is only valid at call sites for
979 direct calls to functions which are declared with the ``nobuiltin``
982 This attribute indicates that this function is rarely called. When
983 computing edge weights, basic blocks post-dominated by a cold
984 function call are also considered to be cold; and, thus, given low
987 This attribute indicates that the source code contained a hint that
988 inlining this function is desirable (such as the "inline" keyword in
989 C/C++). It is just a hint; it imposes no requirements on the
992 This attribute suggests that optimization passes and code generator
993 passes make choices that keep the code size of this function as small
994 as possible and perform optimizations that may sacrifice runtime
995 performance in order to minimize the size of the generated code.
997 This attribute disables prologue / epilogue emission for the
998 function. This can have very system-specific consequences.
1000 This indicates that the callee function at a call site is not recognized as
1001 a built-in function. LLVM will retain the original call and not replace it
1002 with equivalent code based on the semantics of the built-in function, unless
1003 the call site uses the ``builtin`` attribute. This is valid at call sites
1004 and on function declarations and definitions.
1006 This attribute indicates that calls to the function cannot be
1007 duplicated. A call to a ``noduplicate`` function may be moved
1008 within its parent function, but may not be duplicated within
1009 its parent function.
1011 A function containing a ``noduplicate`` call may still
1012 be an inlining candidate, provided that the call is not
1013 duplicated by inlining. That implies that the function has
1014 internal linkage and only has one call site, so the original
1015 call is dead after inlining.
1017 This attributes disables implicit floating point instructions.
1019 This attribute indicates that the inliner should never inline this
1020 function in any situation. This attribute may not be used together
1021 with the ``alwaysinline`` attribute.
1023 This attribute suppresses lazy symbol binding for the function. This
1024 may make calls to the function faster, at the cost of extra program
1025 startup time if the function is not called during program startup.
1027 This attribute indicates that the code generator should not use a
1028 red zone, even if the target-specific ABI normally permits it.
1030 This function attribute indicates that the function never returns
1031 normally. This produces undefined behavior at runtime if the
1032 function ever does dynamically return.
1034 This function attribute indicates that the function never returns
1035 with an unwind or exceptional control flow. If the function does
1036 unwind, its runtime behavior is undefined.
1038 This function attribute indicates that the function is not optimized
1039 by any optimization or code generator passes with the
1040 exception of interprocedural optimization passes.
1041 This attribute cannot be used together with the ``alwaysinline``
1042 attribute; this attribute is also incompatible
1043 with the ``minsize`` attribute and the ``optsize`` attribute.
1045 This attribute requires the ``noinline`` attribute to be specified on
1046 the function as well, so the function is never inlined into any caller.
1047 Only functions with the ``alwaysinline`` attribute are valid
1048 candidates for inlining into the body of this function.
1050 This attribute suggests that optimization passes and code generator
1051 passes make choices that keep the code size of this function low,
1052 and otherwise do optimizations specifically to reduce code size as
1053 long as they do not significantly impact runtime performance.
1055 On a function, this attribute indicates that the function computes its
1056 result (or decides to unwind an exception) based strictly on its arguments,
1057 without dereferencing any pointer arguments or otherwise accessing
1058 any mutable state (e.g. memory, control registers, etc) visible to
1059 caller functions. It does not write through any pointer arguments
1060 (including ``byval`` arguments) and never changes any state visible
1061 to callers. This means that it cannot unwind exceptions by calling
1062 the ``C++`` exception throwing methods.
1064 On an argument, this attribute indicates that the function does not
1065 dereference that pointer argument, even though it may read or write the
1066 memory that the pointer points to if accessed through other pointers.
1068 On a function, this attribute indicates that the function does not write
1069 through any pointer arguments (including ``byval`` arguments) or otherwise
1070 modify any state (e.g. memory, control registers, etc) visible to
1071 caller functions. It may dereference pointer arguments and read
1072 state that may be set in the caller. A readonly function always
1073 returns the same value (or unwinds an exception identically) when
1074 called with the same set of arguments and global state. It cannot
1075 unwind an exception by calling the ``C++`` exception throwing
1078 On an argument, this attribute indicates that the function does not write
1079 through this pointer argument, even though it may write to the memory that
1080 the pointer points to.
1082 This attribute indicates that this function can return twice. The C
1083 ``setjmp`` is an example of such a function. The compiler disables
1084 some optimizations (like tail calls) in the caller of these
1086 ``sanitize_address``
1087 This attribute indicates that AddressSanitizer checks
1088 (dynamic address safety analysis) are enabled for this function.
1090 This attribute indicates that MemorySanitizer checks (dynamic detection
1091 of accesses to uninitialized memory) are enabled for this function.
1093 This attribute indicates that ThreadSanitizer checks
1094 (dynamic thread safety analysis) are enabled for this function.
1096 This attribute indicates that the function should emit a stack
1097 smashing protector. It is in the form of a "canary" --- a random value
1098 placed on the stack before the local variables that's checked upon
1099 return from the function to see if it has been overwritten. A
1100 heuristic is used to determine if a function needs stack protectors
1101 or not. The heuristic used will enable protectors for functions with:
1103 - Character arrays larger than ``ssp-buffer-size`` (default 8).
1104 - Aggregates containing character arrays larger than ``ssp-buffer-size``.
1105 - Calls to alloca() with variable sizes or constant sizes greater than
1106 ``ssp-buffer-size``.
1108 Variables that are identified as requiring a protector will be arranged
1109 on the stack such that they are adjacent to the stack protector guard.
1111 If a function that has an ``ssp`` attribute is inlined into a
1112 function that doesn't have an ``ssp`` attribute, then the resulting
1113 function will have an ``ssp`` attribute.
1115 This attribute indicates that the function should *always* emit a
1116 stack smashing protector. This overrides the ``ssp`` function
1119 Variables that are identified as requiring a protector will be arranged
1120 on the stack such that they are adjacent to the stack protector guard.
1121 The specific layout rules are:
1123 #. Large arrays and structures containing large arrays
1124 (``>= ssp-buffer-size``) are closest to the stack protector.
1125 #. Small arrays and structures containing small arrays
1126 (``< ssp-buffer-size``) are 2nd closest to the protector.
1127 #. Variables that have had their address taken are 3rd closest to the
1130 If a function that has an ``sspreq`` attribute is inlined into a
1131 function that doesn't have an ``sspreq`` attribute or which has an
1132 ``ssp`` or ``sspstrong`` attribute, then the resulting function will have
1133 an ``sspreq`` attribute.
1135 This attribute indicates that the function should emit a stack smashing
1136 protector. This attribute causes a strong heuristic to be used when
1137 determining if a function needs stack protectors. The strong heuristic
1138 will enable protectors for functions with:
1140 - Arrays of any size and type
1141 - Aggregates containing an array of any size and type.
1142 - Calls to alloca().
1143 - Local variables that have had their address taken.
1145 Variables that are identified as requiring a protector will be arranged
1146 on the stack such that they are adjacent to the stack protector guard.
1147 The specific layout rules are:
1149 #. Large arrays and structures containing large arrays
1150 (``>= ssp-buffer-size``) are closest to the stack protector.
1151 #. Small arrays and structures containing small arrays
1152 (``< ssp-buffer-size``) are 2nd closest to the protector.
1153 #. Variables that have had their address taken are 3rd closest to the
1156 This overrides the ``ssp`` function attribute.
1158 If a function that has an ``sspstrong`` attribute is inlined into a
1159 function that doesn't have an ``sspstrong`` attribute, then the
1160 resulting function will have an ``sspstrong`` attribute.
1162 This attribute indicates that the ABI being targeted requires that
1163 an unwind table entry be produce for this function even if we can
1164 show that no exceptions passes by it. This is normally the case for
1165 the ELF x86-64 abi, but it can be disabled for some compilation
1170 Module-Level Inline Assembly
1171 ----------------------------
1173 Modules may contain "module-level inline asm" blocks, which corresponds
1174 to the GCC "file scope inline asm" blocks. These blocks are internally
1175 concatenated by LLVM and treated as a single unit, but may be separated
1176 in the ``.ll`` file if desired. The syntax is very simple:
1178 .. code-block:: llvm
1180 module asm "inline asm code goes here"
1181 module asm "more can go here"
1183 The strings can contain any character by escaping non-printable
1184 characters. The escape sequence used is simply "\\xx" where "xx" is the
1185 two digit hex code for the number.
1187 The inline asm code is simply printed to the machine code .s file when
1188 assembly code is generated.
1190 .. _langref_datalayout:
1195 A module may specify a target specific data layout string that specifies
1196 how data is to be laid out in memory. The syntax for the data layout is
1199 .. code-block:: llvm
1201 target datalayout = "layout specification"
1203 The *layout specification* consists of a list of specifications
1204 separated by the minus sign character ('-'). Each specification starts
1205 with a letter and may include other information after the letter to
1206 define some aspect of the data layout. The specifications accepted are
1210 Specifies that the target lays out data in big-endian form. That is,
1211 the bits with the most significance have the lowest address
1214 Specifies that the target lays out data in little-endian form. That
1215 is, the bits with the least significance have the lowest address
1218 Specifies the natural alignment of the stack in bits. Alignment
1219 promotion of stack variables is limited to the natural stack
1220 alignment to avoid dynamic stack realignment. The stack alignment
1221 must be a multiple of 8-bits. If omitted, the natural stack
1222 alignment defaults to "unspecified", which does not prevent any
1223 alignment promotions.
1224 ``p[n]:<size>:<abi>:<pref>``
1225 This specifies the *size* of a pointer and its ``<abi>`` and
1226 ``<pref>``\erred alignments for address space ``n``. All sizes are in
1227 bits. The address space, ``n`` is optional, and if not specified,
1228 denotes the default address space 0. The value of ``n`` must be
1229 in the range [1,2^23).
1230 ``i<size>:<abi>:<pref>``
1231 This specifies the alignment for an integer type of a given bit
1232 ``<size>``. The value of ``<size>`` must be in the range [1,2^23).
1233 ``v<size>:<abi>:<pref>``
1234 This specifies the alignment for a vector type of a given bit
1236 ``f<size>:<abi>:<pref>``
1237 This specifies the alignment for a floating point type of a given bit
1238 ``<size>``. Only values of ``<size>`` that are supported by the target
1239 will work. 32 (float) and 64 (double) are supported on all targets; 80
1240 or 128 (different flavors of long double) are also supported on some
1243 This specifies the alignment for an object of aggregate type.
1245 If present, specifies that llvm names are mangled in the output. The
1248 * ``e``: ELF mangling: Private symbols get a ``.L`` prefix.
1249 * ``m``: Mips mangling: Private symbols get a ``$`` prefix.
1250 * ``o``: Mach-O mangling: Private symbols get ``L`` prefix. Other
1251 symbols get a ``_`` prefix.
1252 * ``w``: Windows COFF prefix: Similar to Mach-O, but stdcall and fastcall
1253 functions also get a suffix based on the frame size.
1254 ``n<size1>:<size2>:<size3>...``
1255 This specifies a set of native integer widths for the target CPU in
1256 bits. For example, it might contain ``n32`` for 32-bit PowerPC,
1257 ``n32:64`` for PowerPC 64, or ``n8:16:32:64`` for X86-64. Elements of
1258 this set are considered to support most general arithmetic operations
1261 On every specification that takes a ``<abi>:<pref>``, specifying the
1262 ``<pref>`` alignment is optional. If omitted, the preceding ``:``
1263 should be omitted too and ``<pref>`` will be equal to ``<abi>``.
1265 When constructing the data layout for a given target, LLVM starts with a
1266 default set of specifications which are then (possibly) overridden by
1267 the specifications in the ``datalayout`` keyword. The default
1268 specifications are given in this list:
1270 - ``E`` - big endian
1271 - ``p:64:64:64`` - 64-bit pointers with 64-bit alignment.
1272 - ``p[n]:64:64:64`` - Other address spaces are assumed to be the
1273 same as the default address space.
1274 - ``S0`` - natural stack alignment is unspecified
1275 - ``i1:8:8`` - i1 is 8-bit (byte) aligned
1276 - ``i8:8:8`` - i8 is 8-bit (byte) aligned
1277 - ``i16:16:16`` - i16 is 16-bit aligned
1278 - ``i32:32:32`` - i32 is 32-bit aligned
1279 - ``i64:32:64`` - i64 has ABI alignment of 32-bits but preferred
1280 alignment of 64-bits
1281 - ``f16:16:16`` - half is 16-bit aligned
1282 - ``f32:32:32`` - float is 32-bit aligned
1283 - ``f64:64:64`` - double is 64-bit aligned
1284 - ``f128:128:128`` - quad is 128-bit aligned
1285 - ``v64:64:64`` - 64-bit vector is 64-bit aligned
1286 - ``v128:128:128`` - 128-bit vector is 128-bit aligned
1287 - ``a:0:64`` - aggregates are 64-bit aligned
1289 When LLVM is determining the alignment for a given type, it uses the
1292 #. If the type sought is an exact match for one of the specifications,
1293 that specification is used.
1294 #. If no match is found, and the type sought is an integer type, then
1295 the smallest integer type that is larger than the bitwidth of the
1296 sought type is used. If none of the specifications are larger than
1297 the bitwidth then the largest integer type is used. For example,
1298 given the default specifications above, the i7 type will use the
1299 alignment of i8 (next largest) while both i65 and i256 will use the
1300 alignment of i64 (largest specified).
1301 #. If no match is found, and the type sought is a vector type, then the
1302 largest vector type that is smaller than the sought vector type will
1303 be used as a fall back. This happens because <128 x double> can be
1304 implemented in terms of 64 <2 x double>, for example.
1306 The function of the data layout string may not be what you expect.
1307 Notably, this is not a specification from the frontend of what alignment
1308 the code generator should use.
1310 Instead, if specified, the target data layout is required to match what
1311 the ultimate *code generator* expects. This string is used by the
1312 mid-level optimizers to improve code, and this only works if it matches
1313 what the ultimate code generator uses. If you would like to generate IR
1314 that does not embed this target-specific detail into the IR, then you
1315 don't have to specify the string. This will disable some optimizations
1316 that require precise layout information, but this also prevents those
1317 optimizations from introducing target specificity into the IR.
1324 A module may specify a target triple string that describes the target
1325 host. The syntax for the target triple is simply:
1327 .. code-block:: llvm
1329 target triple = "x86_64-apple-macosx10.7.0"
1331 The *target triple* string consists of a series of identifiers delimited
1332 by the minus sign character ('-'). The canonical forms are:
1336 ARCHITECTURE-VENDOR-OPERATING_SYSTEM
1337 ARCHITECTURE-VENDOR-OPERATING_SYSTEM-ENVIRONMENT
1339 This information is passed along to the backend so that it generates
1340 code for the proper architecture. It's possible to override this on the
1341 command line with the ``-mtriple`` command line option.
1343 .. _pointeraliasing:
1345 Pointer Aliasing Rules
1346 ----------------------
1348 Any memory access must be done through a pointer value associated with
1349 an address range of the memory access, otherwise the behavior is
1350 undefined. Pointer values are associated with address ranges according
1351 to the following rules:
1353 - A pointer value is associated with the addresses associated with any
1354 value it is *based* on.
1355 - An address of a global variable is associated with the address range
1356 of the variable's storage.
1357 - The result value of an allocation instruction is associated with the
1358 address range of the allocated storage.
1359 - A null pointer in the default address-space is associated with no
1361 - An integer constant other than zero or a pointer value returned from
1362 a function not defined within LLVM may be associated with address
1363 ranges allocated through mechanisms other than those provided by
1364 LLVM. Such ranges shall not overlap with any ranges of addresses
1365 allocated by mechanisms provided by LLVM.
1367 A pointer value is *based* on another pointer value according to the
1370 - A pointer value formed from a ``getelementptr`` operation is *based*
1371 on the first operand of the ``getelementptr``.
1372 - The result value of a ``bitcast`` is *based* on the operand of the
1374 - A pointer value formed by an ``inttoptr`` is *based* on all pointer
1375 values that contribute (directly or indirectly) to the computation of
1376 the pointer's value.
1377 - The "*based* on" relationship is transitive.
1379 Note that this definition of *"based"* is intentionally similar to the
1380 definition of *"based"* in C99, though it is slightly weaker.
1382 LLVM IR does not associate types with memory. The result type of a
1383 ``load`` merely indicates the size and alignment of the memory from
1384 which to load, as well as the interpretation of the value. The first
1385 operand type of a ``store`` similarly only indicates the size and
1386 alignment of the store.
1388 Consequently, type-based alias analysis, aka TBAA, aka
1389 ``-fstrict-aliasing``, is not applicable to general unadorned LLVM IR.
1390 :ref:`Metadata <metadata>` may be used to encode additional information
1391 which specialized optimization passes may use to implement type-based
1396 Volatile Memory Accesses
1397 ------------------------
1399 Certain memory accesses, such as :ref:`load <i_load>`'s,
1400 :ref:`store <i_store>`'s, and :ref:`llvm.memcpy <int_memcpy>`'s may be
1401 marked ``volatile``. The optimizers must not change the number of
1402 volatile operations or change their order of execution relative to other
1403 volatile operations. The optimizers *may* change the order of volatile
1404 operations relative to non-volatile operations. This is not Java's
1405 "volatile" and has no cross-thread synchronization behavior.
1407 IR-level volatile loads and stores cannot safely be optimized into
1408 llvm.memcpy or llvm.memmove intrinsics even when those intrinsics are
1409 flagged volatile. Likewise, the backend should never split or merge
1410 target-legal volatile load/store instructions.
1412 .. admonition:: Rationale
1414 Platforms may rely on volatile loads and stores of natively supported
1415 data width to be executed as single instruction. For example, in C
1416 this holds for an l-value of volatile primitive type with native
1417 hardware support, but not necessarily for aggregate types. The
1418 frontend upholds these expectations, which are intentionally
1419 unspecified in the IR. The rules above ensure that IR transformation
1420 do not violate the frontend's contract with the language.
1424 Memory Model for Concurrent Operations
1425 --------------------------------------
1427 The LLVM IR does not define any way to start parallel threads of
1428 execution or to register signal handlers. Nonetheless, there are
1429 platform-specific ways to create them, and we define LLVM IR's behavior
1430 in their presence. This model is inspired by the C++0x memory model.
1432 For a more informal introduction to this model, see the :doc:`Atomics`.
1434 We define a *happens-before* partial order as the least partial order
1437 - Is a superset of single-thread program order, and
1438 - When a *synchronizes-with* ``b``, includes an edge from ``a`` to
1439 ``b``. *Synchronizes-with* pairs are introduced by platform-specific
1440 techniques, like pthread locks, thread creation, thread joining,
1441 etc., and by atomic instructions. (See also :ref:`Atomic Memory Ordering
1442 Constraints <ordering>`).
1444 Note that program order does not introduce *happens-before* edges
1445 between a thread and signals executing inside that thread.
1447 Every (defined) read operation (load instructions, memcpy, atomic
1448 loads/read-modify-writes, etc.) R reads a series of bytes written by
1449 (defined) write operations (store instructions, atomic
1450 stores/read-modify-writes, memcpy, etc.). For the purposes of this
1451 section, initialized globals are considered to have a write of the
1452 initializer which is atomic and happens before any other read or write
1453 of the memory in question. For each byte of a read R, R\ :sub:`byte`
1454 may see any write to the same byte, except:
1456 - If write\ :sub:`1` happens before write\ :sub:`2`, and
1457 write\ :sub:`2` happens before R\ :sub:`byte`, then
1458 R\ :sub:`byte` does not see write\ :sub:`1`.
1459 - If R\ :sub:`byte` happens before write\ :sub:`3`, then
1460 R\ :sub:`byte` does not see write\ :sub:`3`.
1462 Given that definition, R\ :sub:`byte` is defined as follows:
1464 - If R is volatile, the result is target-dependent. (Volatile is
1465 supposed to give guarantees which can support ``sig_atomic_t`` in
1466 C/C++, and may be used for accesses to addresses which do not behave
1467 like normal memory. It does not generally provide cross-thread
1469 - Otherwise, if there is no write to the same byte that happens before
1470 R\ :sub:`byte`, R\ :sub:`byte` returns ``undef`` for that byte.
1471 - Otherwise, if R\ :sub:`byte` may see exactly one write,
1472 R\ :sub:`byte` returns the value written by that write.
1473 - Otherwise, if R is atomic, and all the writes R\ :sub:`byte` may
1474 see are atomic, it chooses one of the values written. See the :ref:`Atomic
1475 Memory Ordering Constraints <ordering>` section for additional
1476 constraints on how the choice is made.
1477 - Otherwise R\ :sub:`byte` returns ``undef``.
1479 R returns the value composed of the series of bytes it read. This
1480 implies that some bytes within the value may be ``undef`` **without**
1481 the entire value being ``undef``. Note that this only defines the
1482 semantics of the operation; it doesn't mean that targets will emit more
1483 than one instruction to read the series of bytes.
1485 Note that in cases where none of the atomic intrinsics are used, this
1486 model places only one restriction on IR transformations on top of what
1487 is required for single-threaded execution: introducing a store to a byte
1488 which might not otherwise be stored is not allowed in general.
1489 (Specifically, in the case where another thread might write to and read
1490 from an address, introducing a store can change a load that may see
1491 exactly one write into a load that may see multiple writes.)
1495 Atomic Memory Ordering Constraints
1496 ----------------------------------
1498 Atomic instructions (:ref:`cmpxchg <i_cmpxchg>`,
1499 :ref:`atomicrmw <i_atomicrmw>`, :ref:`fence <i_fence>`,
1500 :ref:`atomic load <i_load>`, and :ref:`atomic store <i_store>`) take
1501 ordering parameters that determine which other atomic instructions on
1502 the same address they *synchronize with*. These semantics are borrowed
1503 from Java and C++0x, but are somewhat more colloquial. If these
1504 descriptions aren't precise enough, check those specs (see spec
1505 references in the :doc:`atomics guide <Atomics>`).
1506 :ref:`fence <i_fence>` instructions treat these orderings somewhat
1507 differently since they don't take an address. See that instruction's
1508 documentation for details.
1510 For a simpler introduction to the ordering constraints, see the
1514 The set of values that can be read is governed by the happens-before
1515 partial order. A value cannot be read unless some operation wrote
1516 it. This is intended to provide a guarantee strong enough to model
1517 Java's non-volatile shared variables. This ordering cannot be
1518 specified for read-modify-write operations; it is not strong enough
1519 to make them atomic in any interesting way.
1521 In addition to the guarantees of ``unordered``, there is a single
1522 total order for modifications by ``monotonic`` operations on each
1523 address. All modification orders must be compatible with the
1524 happens-before order. There is no guarantee that the modification
1525 orders can be combined to a global total order for the whole program
1526 (and this often will not be possible). The read in an atomic
1527 read-modify-write operation (:ref:`cmpxchg <i_cmpxchg>` and
1528 :ref:`atomicrmw <i_atomicrmw>`) reads the value in the modification
1529 order immediately before the value it writes. If one atomic read
1530 happens before another atomic read of the same address, the later
1531 read must see the same value or a later value in the address's
1532 modification order. This disallows reordering of ``monotonic`` (or
1533 stronger) operations on the same address. If an address is written
1534 ``monotonic``-ally by one thread, and other threads ``monotonic``-ally
1535 read that address repeatedly, the other threads must eventually see
1536 the write. This corresponds to the C++0x/C1x
1537 ``memory_order_relaxed``.
1539 In addition to the guarantees of ``monotonic``, a
1540 *synchronizes-with* edge may be formed with a ``release`` operation.
1541 This is intended to model C++'s ``memory_order_acquire``.
1543 In addition to the guarantees of ``monotonic``, if this operation
1544 writes a value which is subsequently read by an ``acquire``
1545 operation, it *synchronizes-with* that operation. (This isn't a
1546 complete description; see the C++0x definition of a release
1547 sequence.) This corresponds to the C++0x/C1x
1548 ``memory_order_release``.
1549 ``acq_rel`` (acquire+release)
1550 Acts as both an ``acquire`` and ``release`` operation on its
1551 address. This corresponds to the C++0x/C1x ``memory_order_acq_rel``.
1552 ``seq_cst`` (sequentially consistent)
1553 In addition to the guarantees of ``acq_rel`` (``acquire`` for an
1554 operation which only reads, ``release`` for an operation which only
1555 writes), there is a global total order on all
1556 sequentially-consistent operations on all addresses, which is
1557 consistent with the *happens-before* partial order and with the
1558 modification orders of all the affected addresses. Each
1559 sequentially-consistent read sees the last preceding write to the
1560 same address in this global order. This corresponds to the C++0x/C1x
1561 ``memory_order_seq_cst`` and Java volatile.
1565 If an atomic operation is marked ``singlethread``, it only *synchronizes
1566 with* or participates in modification and seq\_cst total orderings with
1567 other operations running in the same thread (for example, in signal
1575 LLVM IR floating-point binary ops (:ref:`fadd <i_fadd>`,
1576 :ref:`fsub <i_fsub>`, :ref:`fmul <i_fmul>`, :ref:`fdiv <i_fdiv>`,
1577 :ref:`frem <i_frem>`) have the following flags that can set to enable
1578 otherwise unsafe floating point operations
1581 No NaNs - Allow optimizations to assume the arguments and result are not
1582 NaN. Such optimizations are required to retain defined behavior over
1583 NaNs, but the value of the result is undefined.
1586 No Infs - Allow optimizations to assume the arguments and result are not
1587 +/-Inf. Such optimizations are required to retain defined behavior over
1588 +/-Inf, but the value of the result is undefined.
1591 No Signed Zeros - Allow optimizations to treat the sign of a zero
1592 argument or result as insignificant.
1595 Allow Reciprocal - Allow optimizations to use the reciprocal of an
1596 argument rather than perform division.
1599 Fast - Allow algebraically equivalent transformations that may
1600 dramatically change results in floating point (e.g. reassociate). This
1601 flag implies all the others.
1608 The LLVM type system is one of the most important features of the
1609 intermediate representation. Being typed enables a number of
1610 optimizations to be performed on the intermediate representation
1611 directly, without having to do extra analyses on the side before the
1612 transformation. A strong type system makes it easier to read the
1613 generated code and enables novel analyses and transformations that are
1614 not feasible to perform on normal three address code representations.
1624 The void type does not represent any value and has no size.
1642 The function type can be thought of as a function signature. It consists of a
1643 return type and a list of formal parameter types. The return type of a function
1644 type is a void type or first class type --- except for :ref:`label <t_label>`
1645 and :ref:`metadata <t_metadata>` types.
1651 <returntype> (<parameter list>)
1653 ...where '``<parameter list>``' is a comma-separated list of type
1654 specifiers. Optionally, the parameter list may include a type ``...``, which
1655 indicates that the function takes a variable number of arguments. Variable
1656 argument functions can access their arguments with the :ref:`variable argument
1657 handling intrinsic <int_varargs>` functions. '``<returntype>``' is any type
1658 except :ref:`label <t_label>` and :ref:`metadata <t_metadata>`.
1662 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1663 | ``i32 (i32)`` | function taking an ``i32``, returning an ``i32`` |
1664 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1665 | ``float (i16, i32 *) *`` | :ref:`Pointer <t_pointer>` to a function that takes an ``i16`` and a :ref:`pointer <t_pointer>` to ``i32``, returning ``float``. |
1666 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1667 | ``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. |
1668 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1669 | ``{i32, i32} (i32)`` | A function taking an ``i32``, returning a :ref:`structure <t_struct>` containing two ``i32`` values |
1670 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1677 The :ref:`first class <t_firstclass>` types are perhaps the most important.
1678 Values of these types are the only ones which can be produced by
1686 These are the types that are valid in registers from CodeGen's perspective.
1695 The integer type is a very simple type that simply specifies an
1696 arbitrary bit width for the integer type desired. Any bit width from 1
1697 bit to 2\ :sup:`23`\ -1 (about 8 million) can be specified.
1705 The number of bits the integer will occupy is specified by the ``N``
1711 +----------------+------------------------------------------------+
1712 | ``i1`` | a single-bit integer. |
1713 +----------------+------------------------------------------------+
1714 | ``i32`` | a 32-bit integer. |
1715 +----------------+------------------------------------------------+
1716 | ``i1942652`` | a really big integer of over 1 million bits. |
1717 +----------------+------------------------------------------------+
1721 Floating Point Types
1722 """"""""""""""""""""
1731 - 16-bit floating point value
1734 - 32-bit floating point value
1737 - 64-bit floating point value
1740 - 128-bit floating point value (112-bit mantissa)
1743 - 80-bit floating point value (X87)
1746 - 128-bit floating point value (two 64-bits)
1753 The x86_mmx type represents a value held in an MMX register on an x86
1754 machine. The operations allowed on it are quite limited: parameters and
1755 return values, load and store, and bitcast. User-specified MMX
1756 instructions are represented as intrinsic or asm calls with arguments
1757 and/or results of this type. There are no arrays, vectors or constants
1774 The pointer type is used to specify memory locations. Pointers are
1775 commonly used to reference objects in memory.
1777 Pointer types may have an optional address space attribute defining the
1778 numbered address space where the pointed-to object resides. The default
1779 address space is number zero. The semantics of non-zero address spaces
1780 are target-specific.
1782 Note that LLVM does not permit pointers to void (``void*``) nor does it
1783 permit pointers to labels (``label*``). Use ``i8*`` instead.
1793 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1794 | ``[4 x i32]*`` | A :ref:`pointer <t_pointer>` to :ref:`array <t_array>` of four ``i32`` values. |
1795 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1796 | ``i32 (i32*) *`` | A :ref:`pointer <t_pointer>` to a :ref:`function <t_function>` that takes an ``i32*``, returning an ``i32``. |
1797 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1798 | ``i32 addrspace(5)*`` | A :ref:`pointer <t_pointer>` to an ``i32`` value that resides in address space #5. |
1799 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1808 A vector type is a simple derived type that represents a vector of
1809 elements. Vector types are used when multiple primitive data are
1810 operated in parallel using a single instruction (SIMD). A vector type
1811 requires a size (number of elements) and an underlying primitive data
1812 type. Vector types are considered :ref:`first class <t_firstclass>`.
1818 < <# elements> x <elementtype> >
1820 The number of elements is a constant integer value larger than 0;
1821 elementtype may be any integer or floating point type, or a pointer to
1822 these types. Vectors of size zero are not allowed.
1826 +-------------------+--------------------------------------------------+
1827 | ``<4 x i32>`` | Vector of 4 32-bit integer values. |
1828 +-------------------+--------------------------------------------------+
1829 | ``<8 x float>`` | Vector of 8 32-bit floating-point values. |
1830 +-------------------+--------------------------------------------------+
1831 | ``<2 x i64>`` | Vector of 2 64-bit integer values. |
1832 +-------------------+--------------------------------------------------+
1833 | ``<4 x i64*>`` | Vector of 4 pointers to 64-bit integer values. |
1834 +-------------------+--------------------------------------------------+
1843 The label type represents code labels.
1858 The metadata type represents embedded metadata. No derived types may be
1859 created from metadata except for :ref:`function <t_function>` arguments.
1872 Aggregate Types are a subset of derived types that can contain multiple
1873 member types. :ref:`Arrays <t_array>` and :ref:`structs <t_struct>` are
1874 aggregate types. :ref:`Vectors <t_vector>` are not considered to be
1884 The array type is a very simple derived type that arranges elements
1885 sequentially in memory. The array type requires a size (number of
1886 elements) and an underlying data type.
1892 [<# elements> x <elementtype>]
1894 The number of elements is a constant integer value; ``elementtype`` may
1895 be any type with a size.
1899 +------------------+--------------------------------------+
1900 | ``[40 x i32]`` | Array of 40 32-bit integer values. |
1901 +------------------+--------------------------------------+
1902 | ``[41 x i32]`` | Array of 41 32-bit integer values. |
1903 +------------------+--------------------------------------+
1904 | ``[4 x i8]`` | Array of 4 8-bit integer values. |
1905 +------------------+--------------------------------------+
1907 Here are some examples of multidimensional arrays:
1909 +-----------------------------+----------------------------------------------------------+
1910 | ``[3 x [4 x i32]]`` | 3x4 array of 32-bit integer values. |
1911 +-----------------------------+----------------------------------------------------------+
1912 | ``[12 x [10 x float]]`` | 12x10 array of single precision floating point values. |
1913 +-----------------------------+----------------------------------------------------------+
1914 | ``[2 x [3 x [4 x i16]]]`` | 2x3x4 array of 16-bit integer values. |
1915 +-----------------------------+----------------------------------------------------------+
1917 There is no restriction on indexing beyond the end of the array implied
1918 by a static type (though there are restrictions on indexing beyond the
1919 bounds of an allocated object in some cases). This means that
1920 single-dimension 'variable sized array' addressing can be implemented in
1921 LLVM with a zero length array type. An implementation of 'pascal style
1922 arrays' in LLVM could use the type "``{ i32, [0 x float]}``", for
1932 The structure type is used to represent a collection of data members
1933 together in memory. The elements of a structure may be any type that has
1936 Structures in memory are accessed using '``load``' and '``store``' by
1937 getting a pointer to a field with the '``getelementptr``' instruction.
1938 Structures in registers are accessed using the '``extractvalue``' and
1939 '``insertvalue``' instructions.
1941 Structures may optionally be "packed" structures, which indicate that
1942 the alignment of the struct is one byte, and that there is no padding
1943 between the elements. In non-packed structs, padding between field types
1944 is inserted as defined by the DataLayout string in the module, which is
1945 required to match what the underlying code generator expects.
1947 Structures can either be "literal" or "identified". A literal structure
1948 is defined inline with other types (e.g. ``{i32, i32}*``) whereas
1949 identified types are always defined at the top level with a name.
1950 Literal types are uniqued by their contents and can never be recursive
1951 or opaque since there is no way to write one. Identified types can be
1952 recursive, can be opaqued, and are never uniqued.
1958 %T1 = type { <type list> } ; Identified normal struct type
1959 %T2 = type <{ <type list> }> ; Identified packed struct type
1963 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1964 | ``{ i32, i32, i32 }`` | A triple of three ``i32`` values |
1965 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1966 | ``{ 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``. |
1967 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1968 | ``<{ i8, i32 }>`` | A packed struct known to be 5 bytes in size. |
1969 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1973 Opaque Structure Types
1974 """"""""""""""""""""""
1978 Opaque structure types are used to represent named structure types that
1979 do not have a body specified. This corresponds (for example) to the C
1980 notion of a forward declared structure.
1991 +--------------+-------------------+
1992 | ``opaque`` | An opaque type. |
1993 +--------------+-------------------+
2000 LLVM has several different basic types of constants. This section
2001 describes them all and their syntax.
2006 **Boolean constants**
2007 The two strings '``true``' and '``false``' are both valid constants
2009 **Integer constants**
2010 Standard integers (such as '4') are constants of the
2011 :ref:`integer <t_integer>` type. Negative numbers may be used with
2013 **Floating point constants**
2014 Floating point constants use standard decimal notation (e.g.
2015 123.421), exponential notation (e.g. 1.23421e+2), or a more precise
2016 hexadecimal notation (see below). The assembler requires the exact
2017 decimal value of a floating-point constant. For example, the
2018 assembler accepts 1.25 but rejects 1.3 because 1.3 is a repeating
2019 decimal in binary. Floating point constants must have a :ref:`floating
2020 point <t_floating>` type.
2021 **Null pointer constants**
2022 The identifier '``null``' is recognized as a null pointer constant
2023 and must be of :ref:`pointer type <t_pointer>`.
2025 The one non-intuitive notation for constants is the hexadecimal form of
2026 floating point constants. For example, the form
2027 '``double 0x432ff973cafa8000``' is equivalent to (but harder to read
2028 than) '``double 4.5e+15``'. The only time hexadecimal floating point
2029 constants are required (and the only time that they are generated by the
2030 disassembler) is when a floating point constant must be emitted but it
2031 cannot be represented as a decimal floating point number in a reasonable
2032 number of digits. For example, NaN's, infinities, and other special
2033 values are represented in their IEEE hexadecimal format so that assembly
2034 and disassembly do not cause any bits to change in the constants.
2036 When using the hexadecimal form, constants of types half, float, and
2037 double are represented using the 16-digit form shown above (which
2038 matches the IEEE754 representation for double); half and float values
2039 must, however, be exactly representable as IEEE 754 half and single
2040 precision, respectively. Hexadecimal format is always used for long
2041 double, and there are three forms of long double. The 80-bit format used
2042 by x86 is represented as ``0xK`` followed by 20 hexadecimal digits. The
2043 128-bit format used by PowerPC (two adjacent doubles) is represented by
2044 ``0xM`` followed by 32 hexadecimal digits. The IEEE 128-bit format is
2045 represented by ``0xL`` followed by 32 hexadecimal digits. Long doubles
2046 will only work if they match the long double format on your target.
2047 The IEEE 16-bit format (half precision) is represented by ``0xH``
2048 followed by 4 hexadecimal digits. All hexadecimal formats are big-endian
2049 (sign bit at the left).
2051 There are no constants of type x86_mmx.
2053 .. _complexconstants:
2058 Complex constants are a (potentially recursive) combination of simple
2059 constants and smaller complex constants.
2061 **Structure constants**
2062 Structure constants are represented with notation similar to
2063 structure type definitions (a comma separated list of elements,
2064 surrounded by braces (``{}``)). For example:
2065 "``{ i32 4, float 17.0, i32* @G }``", where "``@G``" is declared as
2066 "``@G = external global i32``". Structure constants must have
2067 :ref:`structure type <t_struct>`, and the number and types of elements
2068 must match those specified by the type.
2070 Array constants are represented with notation similar to array type
2071 definitions (a comma separated list of elements, surrounded by
2072 square brackets (``[]``)). For example:
2073 "``[ i32 42, i32 11, i32 74 ]``". Array constants must have
2074 :ref:`array type <t_array>`, and the number and types of elements must
2075 match those specified by the type.
2076 **Vector constants**
2077 Vector constants are represented with notation similar to vector
2078 type definitions (a comma separated list of elements, surrounded by
2079 less-than/greater-than's (``<>``)). For example:
2080 "``< i32 42, i32 11, i32 74, i32 100 >``". Vector constants
2081 must have :ref:`vector type <t_vector>`, and the number and types of
2082 elements must match those specified by the type.
2083 **Zero initialization**
2084 The string '``zeroinitializer``' can be used to zero initialize a
2085 value to zero of *any* type, including scalar and
2086 :ref:`aggregate <t_aggregate>` types. This is often used to avoid
2087 having to print large zero initializers (e.g. for large arrays) and
2088 is always exactly equivalent to using explicit zero initializers.
2090 A metadata node is a structure-like constant with :ref:`metadata
2091 type <t_metadata>`. For example:
2092 "``metadata !{ i32 0, metadata !"test" }``". Unlike other
2093 constants that are meant to be interpreted as part of the
2094 instruction stream, metadata is a place to attach additional
2095 information such as debug info.
2097 Global Variable and Function Addresses
2098 --------------------------------------
2100 The addresses of :ref:`global variables <globalvars>` and
2101 :ref:`functions <functionstructure>` are always implicitly valid
2102 (link-time) constants. These constants are explicitly referenced when
2103 the :ref:`identifier for the global <identifiers>` is used and always have
2104 :ref:`pointer <t_pointer>` type. For example, the following is a legal LLVM
2107 .. code-block:: llvm
2111 @Z = global [2 x i32*] [ i32* @X, i32* @Y ]
2118 The string '``undef``' can be used anywhere a constant is expected, and
2119 indicates that the user of the value may receive an unspecified
2120 bit-pattern. Undefined values may be of any type (other than '``label``'
2121 or '``void``') and be used anywhere a constant is permitted.
2123 Undefined values are useful because they indicate to the compiler that
2124 the program is well defined no matter what value is used. This gives the
2125 compiler more freedom to optimize. Here are some examples of
2126 (potentially surprising) transformations that are valid (in pseudo IR):
2128 .. code-block:: llvm
2138 This is safe because all of the output bits are affected by the undef
2139 bits. Any output bit can have a zero or one depending on the input bits.
2141 .. code-block:: llvm
2152 These logical operations have bits that are not always affected by the
2153 input. For example, if ``%X`` has a zero bit, then the output of the
2154 '``and``' operation will always be a zero for that bit, no matter what
2155 the corresponding bit from the '``undef``' is. As such, it is unsafe to
2156 optimize or assume that the result of the '``and``' is '``undef``'.
2157 However, it is safe to assume that all bits of the '``undef``' could be
2158 0, and optimize the '``and``' to 0. Likewise, it is safe to assume that
2159 all the bits of the '``undef``' operand to the '``or``' could be set,
2160 allowing the '``or``' to be folded to -1.
2162 .. code-block:: llvm
2164 %A = select undef, %X, %Y
2165 %B = select undef, 42, %Y
2166 %C = select %X, %Y, undef
2176 This set of examples shows that undefined '``select``' (and conditional
2177 branch) conditions can go *either way*, but they have to come from one
2178 of the two operands. In the ``%A`` example, if ``%X`` and ``%Y`` were
2179 both known to have a clear low bit, then ``%A`` would have to have a
2180 cleared low bit. However, in the ``%C`` example, the optimizer is
2181 allowed to assume that the '``undef``' operand could be the same as
2182 ``%Y``, allowing the whole '``select``' to be eliminated.
2184 .. code-block:: llvm
2186 %A = xor undef, undef
2203 This example points out that two '``undef``' operands are not
2204 necessarily the same. This can be surprising to people (and also matches
2205 C semantics) where they assume that "``X^X``" is always zero, even if
2206 ``X`` is undefined. This isn't true for a number of reasons, but the
2207 short answer is that an '``undef``' "variable" can arbitrarily change
2208 its value over its "live range". This is true because the variable
2209 doesn't actually *have a live range*. Instead, the value is logically
2210 read from arbitrary registers that happen to be around when needed, so
2211 the value is not necessarily consistent over time. In fact, ``%A`` and
2212 ``%C`` need to have the same semantics or the core LLVM "replace all
2213 uses with" concept would not hold.
2215 .. code-block:: llvm
2223 These examples show the crucial difference between an *undefined value*
2224 and *undefined behavior*. An undefined value (like '``undef``') is
2225 allowed to have an arbitrary bit-pattern. This means that the ``%A``
2226 operation can be constant folded to '``undef``', because the '``undef``'
2227 could be an SNaN, and ``fdiv`` is not (currently) defined on SNaN's.
2228 However, in the second example, we can make a more aggressive
2229 assumption: because the ``undef`` is allowed to be an arbitrary value,
2230 we are allowed to assume that it could be zero. Since a divide by zero
2231 has *undefined behavior*, we are allowed to assume that the operation
2232 does not execute at all. This allows us to delete the divide and all
2233 code after it. Because the undefined operation "can't happen", the
2234 optimizer can assume that it occurs in dead code.
2236 .. code-block:: llvm
2238 a: store undef -> %X
2239 b: store %X -> undef
2244 These examples reiterate the ``fdiv`` example: a store *of* an undefined
2245 value can be assumed to not have any effect; we can assume that the
2246 value is overwritten with bits that happen to match what was already
2247 there. However, a store *to* an undefined location could clobber
2248 arbitrary memory, therefore, it has undefined behavior.
2255 Poison values are similar to :ref:`undef values <undefvalues>`, however
2256 they also represent the fact that an instruction or constant expression
2257 which cannot evoke side effects has nevertheless detected a condition
2258 which results in undefined behavior.
2260 There is currently no way of representing a poison value in the IR; they
2261 only exist when produced by operations such as :ref:`add <i_add>` with
2264 Poison value behavior is defined in terms of value *dependence*:
2266 - Values other than :ref:`phi <i_phi>` nodes depend on their operands.
2267 - :ref:`Phi <i_phi>` nodes depend on the operand corresponding to
2268 their dynamic predecessor basic block.
2269 - Function arguments depend on the corresponding actual argument values
2270 in the dynamic callers of their functions.
2271 - :ref:`Call <i_call>` instructions depend on the :ref:`ret <i_ret>`
2272 instructions that dynamically transfer control back to them.
2273 - :ref:`Invoke <i_invoke>` instructions depend on the
2274 :ref:`ret <i_ret>`, :ref:`resume <i_resume>`, or exception-throwing
2275 call instructions that dynamically transfer control back to them.
2276 - Non-volatile loads and stores depend on the most recent stores to all
2277 of the referenced memory addresses, following the order in the IR
2278 (including loads and stores implied by intrinsics such as
2279 :ref:`@llvm.memcpy <int_memcpy>`.)
2280 - An instruction with externally visible side effects depends on the
2281 most recent preceding instruction with externally visible side
2282 effects, following the order in the IR. (This includes :ref:`volatile
2283 operations <volatile>`.)
2284 - An instruction *control-depends* on a :ref:`terminator
2285 instruction <terminators>` if the terminator instruction has
2286 multiple successors and the instruction is always executed when
2287 control transfers to one of the successors, and may not be executed
2288 when control is transferred to another.
2289 - Additionally, an instruction also *control-depends* on a terminator
2290 instruction if the set of instructions it otherwise depends on would
2291 be different if the terminator had transferred control to a different
2293 - Dependence is transitive.
2295 Poison Values have the same behavior as :ref:`undef values <undefvalues>`,
2296 with the additional affect that any instruction which has a *dependence*
2297 on a poison value has undefined behavior.
2299 Here are some examples:
2301 .. code-block:: llvm
2304 %poison = sub nuw i32 0, 1 ; Results in a poison value.
2305 %still_poison = and i32 %poison, 0 ; 0, but also poison.
2306 %poison_yet_again = getelementptr i32* @h, i32 %still_poison
2307 store i32 0, i32* %poison_yet_again ; memory at @h[0] is poisoned
2309 store i32 %poison, i32* @g ; Poison value stored to memory.
2310 %poison2 = load i32* @g ; Poison value loaded back from memory.
2312 store volatile i32 %poison, i32* @g ; External observation; undefined behavior.
2314 %narrowaddr = bitcast i32* @g to i16*
2315 %wideaddr = bitcast i32* @g to i64*
2316 %poison3 = load i16* %narrowaddr ; Returns a poison value.
2317 %poison4 = load i64* %wideaddr ; Returns a poison value.
2319 %cmp = icmp slt i32 %poison, 0 ; Returns a poison value.
2320 br i1 %cmp, label %true, label %end ; Branch to either destination.
2323 store volatile i32 0, i32* @g ; This is control-dependent on %cmp, so
2324 ; it has undefined behavior.
2328 %p = phi i32 [ 0, %entry ], [ 1, %true ]
2329 ; Both edges into this PHI are
2330 ; control-dependent on %cmp, so this
2331 ; always results in a poison value.
2333 store volatile i32 0, i32* @g ; This would depend on the store in %true
2334 ; if %cmp is true, or the store in %entry
2335 ; otherwise, so this is undefined behavior.
2337 br i1 %cmp, label %second_true, label %second_end
2338 ; The same branch again, but this time the
2339 ; true block doesn't have side effects.
2346 store volatile i32 0, i32* @g ; This time, the instruction always depends
2347 ; on the store in %end. Also, it is
2348 ; control-equivalent to %end, so this is
2349 ; well-defined (ignoring earlier undefined
2350 ; behavior in this example).
2354 Addresses of Basic Blocks
2355 -------------------------
2357 ``blockaddress(@function, %block)``
2359 The '``blockaddress``' constant computes the address of the specified
2360 basic block in the specified function, and always has an ``i8*`` type.
2361 Taking the address of the entry block is illegal.
2363 This value only has defined behavior when used as an operand to the
2364 ':ref:`indirectbr <i_indirectbr>`' instruction, or for comparisons
2365 against null. Pointer equality tests between labels addresses results in
2366 undefined behavior --- though, again, comparison against null is ok, and
2367 no label is equal to the null pointer. This may be passed around as an
2368 opaque pointer sized value as long as the bits are not inspected. This
2369 allows ``ptrtoint`` and arithmetic to be performed on these values so
2370 long as the original value is reconstituted before the ``indirectbr``
2373 Finally, some targets may provide defined semantics when using the value
2374 as the operand to an inline assembly, but that is target specific.
2378 Constant Expressions
2379 --------------------
2381 Constant expressions are used to allow expressions involving other
2382 constants to be used as constants. Constant expressions may be of any
2383 :ref:`first class <t_firstclass>` type and may involve any LLVM operation
2384 that does not have side effects (e.g. load and call are not supported).
2385 The following is the syntax for constant expressions:
2387 ``trunc (CST to TYPE)``
2388 Truncate a constant to another type. The bit size of CST must be
2389 larger than the bit size of TYPE. Both types must be integers.
2390 ``zext (CST to TYPE)``
2391 Zero extend a constant to another type. The bit size of CST must be
2392 smaller than the bit size of TYPE. Both types must be integers.
2393 ``sext (CST to TYPE)``
2394 Sign extend a constant to another type. The bit size of CST must be
2395 smaller than the bit size of TYPE. Both types must be integers.
2396 ``fptrunc (CST to TYPE)``
2397 Truncate a floating point constant to another floating point type.
2398 The size of CST must be larger than the size of TYPE. Both types
2399 must be floating point.
2400 ``fpext (CST to TYPE)``
2401 Floating point extend a constant to another type. The size of CST
2402 must be smaller or equal to the size of TYPE. Both types must be
2404 ``fptoui (CST to TYPE)``
2405 Convert a floating point constant to the corresponding unsigned
2406 integer constant. TYPE must be a scalar or vector integer type. CST
2407 must be of scalar or vector floating point type. Both CST and TYPE
2408 must be scalars, or vectors of the same number of elements. If the
2409 value won't fit in the integer type, the results are undefined.
2410 ``fptosi (CST to TYPE)``
2411 Convert a floating point constant to the corresponding signed
2412 integer constant. TYPE must be a scalar or vector integer type. CST
2413 must be of scalar or vector floating point type. Both CST and TYPE
2414 must be scalars, or vectors of the same number of elements. If the
2415 value won't fit in the integer type, the results are undefined.
2416 ``uitofp (CST to TYPE)``
2417 Convert an unsigned integer constant to the corresponding floating
2418 point constant. TYPE must be a scalar or vector floating point type.
2419 CST must be of scalar or vector integer type. Both CST and TYPE must
2420 be scalars, or vectors of the same number of elements. If the value
2421 won't fit in the floating point type, the results are undefined.
2422 ``sitofp (CST to TYPE)``
2423 Convert a signed integer constant to the corresponding floating
2424 point constant. TYPE must be a scalar or vector floating point type.
2425 CST must be of scalar or vector integer type. Both CST and TYPE must
2426 be scalars, or vectors of the same number of elements. If the value
2427 won't fit in the floating point type, the results are undefined.
2428 ``ptrtoint (CST to TYPE)``
2429 Convert a pointer typed constant to the corresponding integer
2430 constant. ``TYPE`` must be an integer type. ``CST`` must be of
2431 pointer type. The ``CST`` value is zero extended, truncated, or
2432 unchanged to make it fit in ``TYPE``.
2433 ``inttoptr (CST to TYPE)``
2434 Convert an integer constant to a pointer constant. TYPE must be a
2435 pointer type. CST must be of integer type. The CST value is zero
2436 extended, truncated, or unchanged to make it fit in a pointer size.
2437 This one is *really* dangerous!
2438 ``bitcast (CST to TYPE)``
2439 Convert a constant, CST, to another TYPE. The constraints of the
2440 operands are the same as those for the :ref:`bitcast
2441 instruction <i_bitcast>`.
2442 ``addrspacecast (CST to TYPE)``
2443 Convert a constant pointer or constant vector of pointer, CST, to another
2444 TYPE in a different address space. The constraints of the operands are the
2445 same as those for the :ref:`addrspacecast instruction <i_addrspacecast>`.
2446 ``getelementptr (CSTPTR, IDX0, IDX1, ...)``, ``getelementptr inbounds (CSTPTR, IDX0, IDX1, ...)``
2447 Perform the :ref:`getelementptr operation <i_getelementptr>` on
2448 constants. As with the :ref:`getelementptr <i_getelementptr>`
2449 instruction, the index list may have zero or more indexes, which are
2450 required to make sense for the type of "CSTPTR".
2451 ``select (COND, VAL1, VAL2)``
2452 Perform the :ref:`select operation <i_select>` on constants.
2453 ``icmp COND (VAL1, VAL2)``
2454 Performs the :ref:`icmp operation <i_icmp>` on constants.
2455 ``fcmp COND (VAL1, VAL2)``
2456 Performs the :ref:`fcmp operation <i_fcmp>` on constants.
2457 ``extractelement (VAL, IDX)``
2458 Perform the :ref:`extractelement operation <i_extractelement>` on
2460 ``insertelement (VAL, ELT, IDX)``
2461 Perform the :ref:`insertelement operation <i_insertelement>` on
2463 ``shufflevector (VEC1, VEC2, IDXMASK)``
2464 Perform the :ref:`shufflevector operation <i_shufflevector>` on
2466 ``extractvalue (VAL, IDX0, IDX1, ...)``
2467 Perform the :ref:`extractvalue operation <i_extractvalue>` on
2468 constants. The index list is interpreted in a similar manner as
2469 indices in a ':ref:`getelementptr <i_getelementptr>`' operation. At
2470 least one index value must be specified.
2471 ``insertvalue (VAL, ELT, IDX0, IDX1, ...)``
2472 Perform the :ref:`insertvalue operation <i_insertvalue>` on constants.
2473 The index list is interpreted in a similar manner as indices in a
2474 ':ref:`getelementptr <i_getelementptr>`' operation. At least one index
2475 value must be specified.
2476 ``OPCODE (LHS, RHS)``
2477 Perform the specified operation of the LHS and RHS constants. OPCODE
2478 may be any of the :ref:`binary <binaryops>` or :ref:`bitwise
2479 binary <bitwiseops>` operations. The constraints on operands are
2480 the same as those for the corresponding instruction (e.g. no bitwise
2481 operations on floating point values are allowed).
2488 Inline Assembler Expressions
2489 ----------------------------
2491 LLVM supports inline assembler expressions (as opposed to :ref:`Module-Level
2492 Inline Assembly <moduleasm>`) through the use of a special value. This
2493 value represents the inline assembler as a string (containing the
2494 instructions to emit), a list of operand constraints (stored as a
2495 string), a flag that indicates whether or not the inline asm expression
2496 has side effects, and a flag indicating whether the function containing
2497 the asm needs to align its stack conservatively. An example inline
2498 assembler expression is:
2500 .. code-block:: llvm
2502 i32 (i32) asm "bswap $0", "=r,r"
2504 Inline assembler expressions may **only** be used as the callee operand
2505 of a :ref:`call <i_call>` or an :ref:`invoke <i_invoke>` instruction.
2506 Thus, typically we have:
2508 .. code-block:: llvm
2510 %X = call i32 asm "bswap $0", "=r,r"(i32 %Y)
2512 Inline asms with side effects not visible in the constraint list must be
2513 marked as having side effects. This is done through the use of the
2514 '``sideeffect``' keyword, like so:
2516 .. code-block:: llvm
2518 call void asm sideeffect "eieio", ""()
2520 In some cases inline asms will contain code that will not work unless
2521 the stack is aligned in some way, such as calls or SSE instructions on
2522 x86, yet will not contain code that does that alignment within the asm.
2523 The compiler should make conservative assumptions about what the asm
2524 might contain and should generate its usual stack alignment code in the
2525 prologue if the '``alignstack``' keyword is present:
2527 .. code-block:: llvm
2529 call void asm alignstack "eieio", ""()
2531 Inline asms also support using non-standard assembly dialects. The
2532 assumed dialect is ATT. When the '``inteldialect``' keyword is present,
2533 the inline asm is using the Intel dialect. Currently, ATT and Intel are
2534 the only supported dialects. An example is:
2536 .. code-block:: llvm
2538 call void asm inteldialect "eieio", ""()
2540 If multiple keywords appear the '``sideeffect``' keyword must come
2541 first, the '``alignstack``' keyword second and the '``inteldialect``'
2547 The call instructions that wrap inline asm nodes may have a
2548 "``!srcloc``" MDNode attached to it that contains a list of constant
2549 integers. If present, the code generator will use the integer as the
2550 location cookie value when report errors through the ``LLVMContext``
2551 error reporting mechanisms. This allows a front-end to correlate backend
2552 errors that occur with inline asm back to the source code that produced
2555 .. code-block:: llvm
2557 call void asm sideeffect "something bad", ""(), !srcloc !42
2559 !42 = !{ i32 1234567 }
2561 It is up to the front-end to make sense of the magic numbers it places
2562 in the IR. If the MDNode contains multiple constants, the code generator
2563 will use the one that corresponds to the line of the asm that the error
2568 Metadata Nodes and Metadata Strings
2569 -----------------------------------
2571 LLVM IR allows metadata to be attached to instructions in the program
2572 that can convey extra information about the code to the optimizers and
2573 code generator. One example application of metadata is source-level
2574 debug information. There are two metadata primitives: strings and nodes.
2575 All metadata has the ``metadata`` type and is identified in syntax by a
2576 preceding exclamation point ('``!``').
2578 A metadata string is a string surrounded by double quotes. It can
2579 contain any character by escaping non-printable characters with
2580 "``\xx``" where "``xx``" is the two digit hex code. For example:
2583 Metadata nodes are represented with notation similar to structure
2584 constants (a comma separated list of elements, surrounded by braces and
2585 preceded by an exclamation point). Metadata nodes can have any values as
2586 their operand. For example:
2588 .. code-block:: llvm
2590 !{ metadata !"test\00", i32 10}
2592 A :ref:`named metadata <namedmetadatastructure>` is a collection of
2593 metadata nodes, which can be looked up in the module symbol table. For
2596 .. code-block:: llvm
2598 !foo = metadata !{!4, !3}
2600 Metadata can be used as function arguments. Here ``llvm.dbg.value``
2601 function is using two metadata arguments:
2603 .. code-block:: llvm
2605 call void @llvm.dbg.value(metadata !24, i64 0, metadata !25)
2607 Metadata can be attached with an instruction. Here metadata ``!21`` is
2608 attached to the ``add`` instruction using the ``!dbg`` identifier:
2610 .. code-block:: llvm
2612 %indvar.next = add i64 %indvar, 1, !dbg !21
2614 More information about specific metadata nodes recognized by the
2615 optimizers and code generator is found below.
2620 In LLVM IR, memory does not have types, so LLVM's own type system is not
2621 suitable for doing TBAA. Instead, metadata is added to the IR to
2622 describe a type system of a higher level language. This can be used to
2623 implement typical C/C++ TBAA, but it can also be used to implement
2624 custom alias analysis behavior for other languages.
2626 The current metadata format is very simple. TBAA metadata nodes have up
2627 to three fields, e.g.:
2629 .. code-block:: llvm
2631 !0 = metadata !{ metadata !"an example type tree" }
2632 !1 = metadata !{ metadata !"int", metadata !0 }
2633 !2 = metadata !{ metadata !"float", metadata !0 }
2634 !3 = metadata !{ metadata !"const float", metadata !2, i64 1 }
2636 The first field is an identity field. It can be any value, usually a
2637 metadata string, which uniquely identifies the type. The most important
2638 name in the tree is the name of the root node. Two trees with different
2639 root node names are entirely disjoint, even if they have leaves with
2642 The second field identifies the type's parent node in the tree, or is
2643 null or omitted for a root node. A type is considered to alias all of
2644 its descendants and all of its ancestors in the tree. Also, a type is
2645 considered to alias all types in other trees, so that bitcode produced
2646 from multiple front-ends is handled conservatively.
2648 If the third field is present, it's an integer which if equal to 1
2649 indicates that the type is "constant" (meaning
2650 ``pointsToConstantMemory`` should return true; see `other useful
2651 AliasAnalysis methods <AliasAnalysis.html#OtherItfs>`_).
2653 '``tbaa.struct``' Metadata
2654 ^^^^^^^^^^^^^^^^^^^^^^^^^^
2656 The :ref:`llvm.memcpy <int_memcpy>` is often used to implement
2657 aggregate assignment operations in C and similar languages, however it
2658 is defined to copy a contiguous region of memory, which is more than
2659 strictly necessary for aggregate types which contain holes due to
2660 padding. Also, it doesn't contain any TBAA information about the fields
2663 ``!tbaa.struct`` metadata can describe which memory subregions in a
2664 memcpy are padding and what the TBAA tags of the struct are.
2666 The current metadata format is very simple. ``!tbaa.struct`` metadata
2667 nodes are a list of operands which are in conceptual groups of three.
2668 For each group of three, the first operand gives the byte offset of a
2669 field in bytes, the second gives its size in bytes, and the third gives
2672 .. code-block:: llvm
2674 !4 = metadata !{ i64 0, i64 4, metadata !1, i64 8, i64 4, metadata !2 }
2676 This describes a struct with two fields. The first is at offset 0 bytes
2677 with size 4 bytes, and has tbaa tag !1. The second is at offset 8 bytes
2678 and has size 4 bytes and has tbaa tag !2.
2680 Note that the fields need not be contiguous. In this example, there is a
2681 4 byte gap between the two fields. This gap represents padding which
2682 does not carry useful data and need not be preserved.
2684 '``fpmath``' Metadata
2685 ^^^^^^^^^^^^^^^^^^^^^
2687 ``fpmath`` metadata may be attached to any instruction of floating point
2688 type. It can be used to express the maximum acceptable error in the
2689 result of that instruction, in ULPs, thus potentially allowing the
2690 compiler to use a more efficient but less accurate method of computing
2691 it. ULP is defined as follows:
2693 If ``x`` is a real number that lies between two finite consecutive
2694 floating-point numbers ``a`` and ``b``, without being equal to one
2695 of them, then ``ulp(x) = |b - a|``, otherwise ``ulp(x)`` is the
2696 distance between the two non-equal finite floating-point numbers
2697 nearest ``x``. Moreover, ``ulp(NaN)`` is ``NaN``.
2699 The metadata node shall consist of a single positive floating point
2700 number representing the maximum relative error, for example:
2702 .. code-block:: llvm
2704 !0 = metadata !{ float 2.5 } ; maximum acceptable inaccuracy is 2.5 ULPs
2706 '``range``' Metadata
2707 ^^^^^^^^^^^^^^^^^^^^
2709 ``range`` metadata may be attached only to loads of integer types. It
2710 expresses the possible ranges the loaded value is in. The ranges are
2711 represented with a flattened list of integers. The loaded value is known
2712 to be in the union of the ranges defined by each consecutive pair. Each
2713 pair has the following properties:
2715 - The type must match the type loaded by the instruction.
2716 - The pair ``a,b`` represents the range ``[a,b)``.
2717 - Both ``a`` and ``b`` are constants.
2718 - The range is allowed to wrap.
2719 - The range should not represent the full or empty set. That is,
2722 In addition, the pairs must be in signed order of the lower bound and
2723 they must be non-contiguous.
2727 .. code-block:: llvm
2729 %a = load i8* %x, align 1, !range !0 ; Can only be 0 or 1
2730 %b = load i8* %y, align 1, !range !1 ; Can only be 255 (-1), 0 or 1
2731 %c = load i8* %z, align 1, !range !2 ; Can only be 0, 1, 3, 4 or 5
2732 %d = load i8* %z, align 1, !range !3 ; Can only be -2, -1, 3, 4 or 5
2734 !0 = metadata !{ i8 0, i8 2 }
2735 !1 = metadata !{ i8 255, i8 2 }
2736 !2 = metadata !{ i8 0, i8 2, i8 3, i8 6 }
2737 !3 = metadata !{ i8 -2, i8 0, i8 3, i8 6 }
2742 It is sometimes useful to attach information to loop constructs. Currently,
2743 loop metadata is implemented as metadata attached to the branch instruction
2744 in the loop latch block. This type of metadata refer to a metadata node that is
2745 guaranteed to be separate for each loop. The loop identifier metadata is
2746 specified with the name ``llvm.loop``.
2748 The loop identifier metadata is implemented using a metadata that refers to
2749 itself to avoid merging it with any other identifier metadata, e.g.,
2750 during module linkage or function inlining. That is, each loop should refer
2751 to their own identification metadata even if they reside in separate functions.
2752 The following example contains loop identifier metadata for two separate loop
2755 .. code-block:: llvm
2757 !0 = metadata !{ metadata !0 }
2758 !1 = metadata !{ metadata !1 }
2760 The loop identifier metadata can be used to specify additional per-loop
2761 metadata. Any operands after the first operand can be treated as user-defined
2762 metadata. For example the ``llvm.vectorizer.unroll`` metadata is understood
2763 by the loop vectorizer to indicate how many times to unroll the loop:
2765 .. code-block:: llvm
2767 br i1 %exitcond, label %._crit_edge, label %.lr.ph, !llvm.loop !0
2769 !0 = metadata !{ metadata !0, metadata !1 }
2770 !1 = metadata !{ metadata !"llvm.vectorizer.unroll", i32 2 }
2775 Metadata types used to annotate memory accesses with information helpful
2776 for optimizations are prefixed with ``llvm.mem``.
2778 '``llvm.mem.parallel_loop_access``' Metadata
2779 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2781 For a loop to be parallel, in addition to using
2782 the ``llvm.loop`` metadata to mark the loop latch branch instruction,
2783 also all of the memory accessing instructions in the loop body need to be
2784 marked with the ``llvm.mem.parallel_loop_access`` metadata. If there
2785 is at least one memory accessing instruction not marked with the metadata,
2786 the loop must be considered a sequential loop. This causes parallel loops to be
2787 converted to sequential loops due to optimization passes that are unaware of
2788 the parallel semantics and that insert new memory instructions to the loop
2791 Example of a loop that is considered parallel due to its correct use of
2792 both ``llvm.loop`` and ``llvm.mem.parallel_loop_access``
2793 metadata types that refer to the same loop identifier metadata.
2795 .. code-block:: llvm
2799 %val0 = load i32* %arrayidx, !llvm.mem.parallel_loop_access !0
2801 store i32 %val0, i32* %arrayidx1, !llvm.mem.parallel_loop_access !0
2803 br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !0
2807 !0 = metadata !{ metadata !0 }
2809 It is also possible to have nested parallel loops. In that case the
2810 memory accesses refer to a list of loop identifier metadata nodes instead of
2811 the loop identifier metadata node directly:
2813 .. code-block:: llvm
2817 %val1 = load i32* %arrayidx3, !llvm.mem.parallel_loop_access !2
2819 br label %inner.for.body
2823 %val0 = load i32* %arrayidx1, !llvm.mem.parallel_loop_access !0
2825 store i32 %val0, i32* %arrayidx2, !llvm.mem.parallel_loop_access !0
2827 br i1 %exitcond, label %inner.for.end, label %inner.for.body, !llvm.loop !1
2831 store i32 %val1, i32* %arrayidx4, !llvm.mem.parallel_loop_access !2
2833 br i1 %exitcond, label %outer.for.end, label %outer.for.body, !llvm.loop !2
2835 outer.for.end: ; preds = %for.body
2837 !0 = metadata !{ metadata !1, metadata !2 } ; a list of loop identifiers
2838 !1 = metadata !{ metadata !1 } ; an identifier for the inner loop
2839 !2 = metadata !{ metadata !2 } ; an identifier for the outer loop
2841 '``llvm.vectorizer``'
2842 ^^^^^^^^^^^^^^^^^^^^^
2844 Metadata prefixed with ``llvm.vectorizer`` is used to control per-loop
2845 vectorization parameters such as vectorization factor and unroll factor.
2847 ``llvm.vectorizer`` metadata should be used in conjunction with ``llvm.loop``
2848 loop identification metadata.
2850 '``llvm.vectorizer.unroll``' Metadata
2851 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2853 This metadata instructs the loop vectorizer to unroll the specified
2854 loop exactly ``N`` times.
2856 The first operand is the string ``llvm.vectorizer.unroll`` and the second
2857 operand is an integer specifying the unroll factor. For example:
2859 .. code-block:: llvm
2861 !0 = metadata !{ metadata !"llvm.vectorizer.unroll", i32 4 }
2863 Note that setting ``llvm.vectorizer.unroll`` to 1 disables unrolling of the
2866 If ``llvm.vectorizer.unroll`` is set to 0 then the amount of unrolling will be
2867 determined automatically.
2869 '``llvm.vectorizer.width``' Metadata
2870 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2872 This metadata sets the target width of the vectorizer to ``N``. Without
2873 this metadata, the vectorizer will choose a width automatically.
2874 Regardless of this metadata, the vectorizer will only vectorize loops if
2875 it believes it is valid to do so.
2877 The first operand is the string ``llvm.vectorizer.width`` and the second
2878 operand is an integer specifying the width. For example:
2880 .. code-block:: llvm
2882 !0 = metadata !{ metadata !"llvm.vectorizer.width", i32 4 }
2884 Note that setting ``llvm.vectorizer.width`` to 1 disables vectorization of the
2887 If ``llvm.vectorizer.width`` is set to 0 then the width will be determined
2890 Module Flags Metadata
2891 =====================
2893 Information about the module as a whole is difficult to convey to LLVM's
2894 subsystems. The LLVM IR isn't sufficient to transmit this information.
2895 The ``llvm.module.flags`` named metadata exists in order to facilitate
2896 this. These flags are in the form of key / value pairs --- much like a
2897 dictionary --- making it easy for any subsystem who cares about a flag to
2900 The ``llvm.module.flags`` metadata contains a list of metadata triplets.
2901 Each triplet has the following form:
2903 - The first element is a *behavior* flag, which specifies the behavior
2904 when two (or more) modules are merged together, and it encounters two
2905 (or more) metadata with the same ID. The supported behaviors are
2907 - The second element is a metadata string that is a unique ID for the
2908 metadata. Each module may only have one flag entry for each unique ID (not
2909 including entries with the **Require** behavior).
2910 - The third element is the value of the flag.
2912 When two (or more) modules are merged together, the resulting
2913 ``llvm.module.flags`` metadata is the union of the modules' flags. That is, for
2914 each unique metadata ID string, there will be exactly one entry in the merged
2915 modules ``llvm.module.flags`` metadata table, and the value for that entry will
2916 be determined by the merge behavior flag, as described below. The only exception
2917 is that entries with the *Require* behavior are always preserved.
2919 The following behaviors are supported:
2930 Emits an error if two values disagree, otherwise the resulting value
2931 is that of the operands.
2935 Emits a warning if two values disagree. The result value will be the
2936 operand for the flag from the first module being linked.
2940 Adds a requirement that another module flag be present and have a
2941 specified value after linking is performed. The value must be a
2942 metadata pair, where the first element of the pair is the ID of the
2943 module flag to be restricted, and the second element of the pair is
2944 the value the module flag should be restricted to. This behavior can
2945 be used to restrict the allowable results (via triggering of an
2946 error) of linking IDs with the **Override** behavior.
2950 Uses the specified value, regardless of the behavior or value of the
2951 other module. If both modules specify **Override**, but the values
2952 differ, an error will be emitted.
2956 Appends the two values, which are required to be metadata nodes.
2960 Appends the two values, which are required to be metadata
2961 nodes. However, duplicate entries in the second list are dropped
2962 during the append operation.
2964 It is an error for a particular unique flag ID to have multiple behaviors,
2965 except in the case of **Require** (which adds restrictions on another metadata
2966 value) or **Override**.
2968 An example of module flags:
2970 .. code-block:: llvm
2972 !0 = metadata !{ i32 1, metadata !"foo", i32 1 }
2973 !1 = metadata !{ i32 4, metadata !"bar", i32 37 }
2974 !2 = metadata !{ i32 2, metadata !"qux", i32 42 }
2975 !3 = metadata !{ i32 3, metadata !"qux",
2977 metadata !"foo", i32 1
2980 !llvm.module.flags = !{ !0, !1, !2, !3 }
2982 - Metadata ``!0`` has the ID ``!"foo"`` and the value '1'. The behavior
2983 if two or more ``!"foo"`` flags are seen is to emit an error if their
2984 values are not equal.
2986 - Metadata ``!1`` has the ID ``!"bar"`` and the value '37'. The
2987 behavior if two or more ``!"bar"`` flags are seen is to use the value
2990 - Metadata ``!2`` has the ID ``!"qux"`` and the value '42'. The
2991 behavior if two or more ``!"qux"`` flags are seen is to emit a
2992 warning if their values are not equal.
2994 - Metadata ``!3`` has the ID ``!"qux"`` and the value:
2998 metadata !{ metadata !"foo", i32 1 }
3000 The behavior is to emit an error if the ``llvm.module.flags`` does not
3001 contain a flag with the ID ``!"foo"`` that has the value '1' after linking is
3004 Objective-C Garbage Collection Module Flags Metadata
3005 ----------------------------------------------------
3007 On the Mach-O platform, Objective-C stores metadata about garbage
3008 collection in a special section called "image info". The metadata
3009 consists of a version number and a bitmask specifying what types of
3010 garbage collection are supported (if any) by the file. If two or more
3011 modules are linked together their garbage collection metadata needs to
3012 be merged rather than appended together.
3014 The Objective-C garbage collection module flags metadata consists of the
3015 following key-value pairs:
3024 * - ``Objective-C Version``
3025 - **[Required]** --- The Objective-C ABI version. Valid values are 1 and 2.
3027 * - ``Objective-C Image Info Version``
3028 - **[Required]** --- The version of the image info section. Currently
3031 * - ``Objective-C Image Info Section``
3032 - **[Required]** --- The section to place the metadata. Valid values are
3033 ``"__OBJC, __image_info, regular"`` for Objective-C ABI version 1, and
3034 ``"__DATA,__objc_imageinfo, regular, no_dead_strip"`` for
3035 Objective-C ABI version 2.
3037 * - ``Objective-C Garbage Collection``
3038 - **[Required]** --- Specifies whether garbage collection is supported or
3039 not. Valid values are 0, for no garbage collection, and 2, for garbage
3040 collection supported.
3042 * - ``Objective-C GC Only``
3043 - **[Optional]** --- Specifies that only garbage collection is supported.
3044 If present, its value must be 6. This flag requires that the
3045 ``Objective-C Garbage Collection`` flag have the value 2.
3047 Some important flag interactions:
3049 - If a module with ``Objective-C Garbage Collection`` set to 0 is
3050 merged with a module with ``Objective-C Garbage Collection`` set to
3051 2, then the resulting module has the
3052 ``Objective-C Garbage Collection`` flag set to 0.
3053 - A module with ``Objective-C Garbage Collection`` set to 0 cannot be
3054 merged with a module with ``Objective-C GC Only`` set to 6.
3056 Automatic Linker Flags Module Flags Metadata
3057 --------------------------------------------
3059 Some targets support embedding flags to the linker inside individual object
3060 files. Typically this is used in conjunction with language extensions which
3061 allow source files to explicitly declare the libraries they depend on, and have
3062 these automatically be transmitted to the linker via object files.
3064 These flags are encoded in the IR using metadata in the module flags section,
3065 using the ``Linker Options`` key. The merge behavior for this flag is required
3066 to be ``AppendUnique``, and the value for the key is expected to be a metadata
3067 node which should be a list of other metadata nodes, each of which should be a
3068 list of metadata strings defining linker options.
3070 For example, the following metadata section specifies two separate sets of
3071 linker options, presumably to link against ``libz`` and the ``Cocoa``
3074 !0 = metadata !{ i32 6, metadata !"Linker Options",
3076 metadata !{ metadata !"-lz" },
3077 metadata !{ metadata !"-framework", metadata !"Cocoa" } } }
3078 !llvm.module.flags = !{ !0 }
3080 The metadata encoding as lists of lists of options, as opposed to a collapsed
3081 list of options, is chosen so that the IR encoding can use multiple option
3082 strings to specify e.g., a single library, while still having that specifier be
3083 preserved as an atomic element that can be recognized by a target specific
3084 assembly writer or object file emitter.
3086 Each individual option is required to be either a valid option for the target's
3087 linker, or an option that is reserved by the target specific assembly writer or
3088 object file emitter. No other aspect of these options is defined by the IR.
3090 .. _intrinsicglobalvariables:
3092 Intrinsic Global Variables
3093 ==========================
3095 LLVM has a number of "magic" global variables that contain data that
3096 affect code generation or other IR semantics. These are documented here.
3097 All globals of this sort should have a section specified as
3098 "``llvm.metadata``". This section and all globals that start with
3099 "``llvm.``" are reserved for use by LLVM.
3103 The '``llvm.used``' Global Variable
3104 -----------------------------------
3106 The ``@llvm.used`` global is an array which has
3107 :ref:`appending linkage <linkage_appending>`. This array contains a list of
3108 pointers to named global variables, functions and aliases which may optionally
3109 have a pointer cast formed of bitcast or getelementptr. For example, a legal
3112 .. code-block:: llvm
3117 @llvm.used = appending global [2 x i8*] [
3119 i8* bitcast (i32* @Y to i8*)
3120 ], section "llvm.metadata"
3122 If a symbol appears in the ``@llvm.used`` list, then the compiler, assembler,
3123 and linker are required to treat the symbol as if there is a reference to the
3124 symbol that it cannot see (which is why they have to be named). For example, if
3125 a variable has internal linkage and no references other than that from the
3126 ``@llvm.used`` list, it cannot be deleted. This is commonly used to represent
3127 references from inline asms and other things the compiler cannot "see", and
3128 corresponds to "``attribute((used))``" in GNU C.
3130 On some targets, the code generator must emit a directive to the
3131 assembler or object file to prevent the assembler and linker from
3132 molesting the symbol.
3134 .. _gv_llvmcompilerused:
3136 The '``llvm.compiler.used``' Global Variable
3137 --------------------------------------------
3139 The ``@llvm.compiler.used`` directive is the same as the ``@llvm.used``
3140 directive, except that it only prevents the compiler from touching the
3141 symbol. On targets that support it, this allows an intelligent linker to
3142 optimize references to the symbol without being impeded as it would be
3145 This is a rare construct that should only be used in rare circumstances,
3146 and should not be exposed to source languages.
3148 .. _gv_llvmglobalctors:
3150 The '``llvm.global_ctors``' Global Variable
3151 -------------------------------------------
3153 .. code-block:: llvm
3155 %0 = type { i32, void ()* }
3156 @llvm.global_ctors = appending global [1 x %0] [%0 { i32 65535, void ()* @ctor }]
3158 The ``@llvm.global_ctors`` array contains a list of constructor
3159 functions and associated priorities. The functions referenced by this
3160 array will be called in ascending order of priority (i.e. lowest first)
3161 when the module is loaded. The order of functions with the same priority
3164 .. _llvmglobaldtors:
3166 The '``llvm.global_dtors``' Global Variable
3167 -------------------------------------------
3169 .. code-block:: llvm
3171 %0 = type { i32, void ()* }
3172 @llvm.global_dtors = appending global [1 x %0] [%0 { i32 65535, void ()* @dtor }]
3174 The ``@llvm.global_dtors`` array contains a list of destructor functions
3175 and associated priorities. The functions referenced by this array will
3176 be called in descending order of priority (i.e. highest first) when the
3177 module is loaded. The order of functions with the same priority is not
3180 Instruction Reference
3181 =====================
3183 The LLVM instruction set consists of several different classifications
3184 of instructions: :ref:`terminator instructions <terminators>`, :ref:`binary
3185 instructions <binaryops>`, :ref:`bitwise binary
3186 instructions <bitwiseops>`, :ref:`memory instructions <memoryops>`, and
3187 :ref:`other instructions <otherops>`.
3191 Terminator Instructions
3192 -----------------------
3194 As mentioned :ref:`previously <functionstructure>`, every basic block in a
3195 program ends with a "Terminator" instruction, which indicates which
3196 block should be executed after the current block is finished. These
3197 terminator instructions typically yield a '``void``' value: they produce
3198 control flow, not values (the one exception being the
3199 ':ref:`invoke <i_invoke>`' instruction).
3201 The terminator instructions are: ':ref:`ret <i_ret>`',
3202 ':ref:`br <i_br>`', ':ref:`switch <i_switch>`',
3203 ':ref:`indirectbr <i_indirectbr>`', ':ref:`invoke <i_invoke>`',
3204 ':ref:`resume <i_resume>`', and ':ref:`unreachable <i_unreachable>`'.
3208 '``ret``' Instruction
3209 ^^^^^^^^^^^^^^^^^^^^^
3216 ret <type> <value> ; Return a value from a non-void function
3217 ret void ; Return from void function
3222 The '``ret``' instruction is used to return control flow (and optionally
3223 a value) from a function back to the caller.
3225 There are two forms of the '``ret``' instruction: one that returns a
3226 value and then causes control flow, and one that just causes control
3232 The '``ret``' instruction optionally accepts a single argument, the
3233 return value. The type of the return value must be a ':ref:`first
3234 class <t_firstclass>`' type.
3236 A function is not :ref:`well formed <wellformed>` if it it has a non-void
3237 return type and contains a '``ret``' instruction with no return value or
3238 a return value with a type that does not match its type, or if it has a
3239 void return type and contains a '``ret``' instruction with a return
3245 When the '``ret``' instruction is executed, control flow returns back to
3246 the calling function's context. If the caller is a
3247 ":ref:`call <i_call>`" instruction, execution continues at the
3248 instruction after the call. If the caller was an
3249 ":ref:`invoke <i_invoke>`" instruction, execution continues at the
3250 beginning of the "normal" destination block. If the instruction returns
3251 a value, that value shall set the call or invoke instruction's return
3257 .. code-block:: llvm
3259 ret i32 5 ; Return an integer value of 5
3260 ret void ; Return from a void function
3261 ret { i32, i8 } { i32 4, i8 2 } ; Return a struct of values 4 and 2
3265 '``br``' Instruction
3266 ^^^^^^^^^^^^^^^^^^^^
3273 br i1 <cond>, label <iftrue>, label <iffalse>
3274 br label <dest> ; Unconditional branch
3279 The '``br``' instruction is used to cause control flow to transfer to a
3280 different basic block in the current function. There are two forms of
3281 this instruction, corresponding to a conditional branch and an
3282 unconditional branch.
3287 The conditional branch form of the '``br``' instruction takes a single
3288 '``i1``' value and two '``label``' values. The unconditional form of the
3289 '``br``' instruction takes a single '``label``' value as a target.
3294 Upon execution of a conditional '``br``' instruction, the '``i1``'
3295 argument is evaluated. If the value is ``true``, control flows to the
3296 '``iftrue``' ``label`` argument. If "cond" is ``false``, control flows
3297 to the '``iffalse``' ``label`` argument.
3302 .. code-block:: llvm
3305 %cond = icmp eq i32 %a, %b
3306 br i1 %cond, label %IfEqual, label %IfUnequal
3314 '``switch``' Instruction
3315 ^^^^^^^^^^^^^^^^^^^^^^^^
3322 switch <intty> <value>, label <defaultdest> [ <intty> <val>, label <dest> ... ]
3327 The '``switch``' instruction is used to transfer control flow to one of
3328 several different places. It is a generalization of the '``br``'
3329 instruction, allowing a branch to occur to one of many possible
3335 The '``switch``' instruction uses three parameters: an integer
3336 comparison value '``value``', a default '``label``' destination, and an
3337 array of pairs of comparison value constants and '``label``'s. The table
3338 is not allowed to contain duplicate constant entries.
3343 The ``switch`` instruction specifies a table of values and destinations.
3344 When the '``switch``' instruction is executed, this table is searched
3345 for the given value. If the value is found, control flow is transferred
3346 to the corresponding destination; otherwise, control flow is transferred
3347 to the default destination.
3352 Depending on properties of the target machine and the particular
3353 ``switch`` instruction, this instruction may be code generated in
3354 different ways. For example, it could be generated as a series of
3355 chained conditional branches or with a lookup table.
3360 .. code-block:: llvm
3362 ; Emulate a conditional br instruction
3363 %Val = zext i1 %value to i32
3364 switch i32 %Val, label %truedest [ i32 0, label %falsedest ]
3366 ; Emulate an unconditional br instruction
3367 switch i32 0, label %dest [ ]
3369 ; Implement a jump table:
3370 switch i32 %val, label %otherwise [ i32 0, label %onzero
3372 i32 2, label %ontwo ]
3376 '``indirectbr``' Instruction
3377 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3384 indirectbr <somety>* <address>, [ label <dest1>, label <dest2>, ... ]
3389 The '``indirectbr``' instruction implements an indirect branch to a
3390 label within the current function, whose address is specified by
3391 "``address``". Address must be derived from a
3392 :ref:`blockaddress <blockaddress>` constant.
3397 The '``address``' argument is the address of the label to jump to. The
3398 rest of the arguments indicate the full set of possible destinations
3399 that the address may point to. Blocks are allowed to occur multiple
3400 times in the destination list, though this isn't particularly useful.
3402 This destination list is required so that dataflow analysis has an
3403 accurate understanding of the CFG.
3408 Control transfers to the block specified in the address argument. All
3409 possible destination blocks must be listed in the label list, otherwise
3410 this instruction has undefined behavior. This implies that jumps to
3411 labels defined in other functions have undefined behavior as well.
3416 This is typically implemented with a jump through a register.
3421 .. code-block:: llvm
3423 indirectbr i8* %Addr, [ label %bb1, label %bb2, label %bb3 ]
3427 '``invoke``' Instruction
3428 ^^^^^^^^^^^^^^^^^^^^^^^^
3435 <result> = invoke [cconv] [ret attrs] <ptr to function ty> <function ptr val>(<function args>) [fn attrs]
3436 to label <normal label> unwind label <exception label>
3441 The '``invoke``' instruction causes control to transfer to a specified
3442 function, with the possibility of control flow transfer to either the
3443 '``normal``' label or the '``exception``' label. If the callee function
3444 returns with the "``ret``" instruction, control flow will return to the
3445 "normal" label. If the callee (or any indirect callees) returns via the
3446 ":ref:`resume <i_resume>`" instruction or other exception handling
3447 mechanism, control is interrupted and continued at the dynamically
3448 nearest "exception" label.
3450 The '``exception``' label is a `landing
3451 pad <ExceptionHandling.html#overview>`_ for the exception. As such,
3452 '``exception``' label is required to have the
3453 ":ref:`landingpad <i_landingpad>`" instruction, which contains the
3454 information about the behavior of the program after unwinding happens,
3455 as its first non-PHI instruction. The restrictions on the
3456 "``landingpad``" instruction's tightly couples it to the "``invoke``"
3457 instruction, so that the important information contained within the
3458 "``landingpad``" instruction can't be lost through normal code motion.
3463 This instruction requires several arguments:
3465 #. The optional "cconv" marker indicates which :ref:`calling
3466 convention <callingconv>` the call should use. If none is
3467 specified, the call defaults to using C calling conventions.
3468 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
3469 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
3471 #. '``ptr to function ty``': shall be the signature of the pointer to
3472 function value being invoked. In most cases, this is a direct
3473 function invocation, but indirect ``invoke``'s are just as possible,
3474 branching off an arbitrary pointer to function value.
3475 #. '``function ptr val``': An LLVM value containing a pointer to a
3476 function to be invoked.
3477 #. '``function args``': argument list whose types match the function
3478 signature argument types and parameter attributes. All arguments must
3479 be of :ref:`first class <t_firstclass>` type. If the function signature
3480 indicates the function accepts a variable number of arguments, the
3481 extra arguments can be specified.
3482 #. '``normal label``': the label reached when the called function
3483 executes a '``ret``' instruction.
3484 #. '``exception label``': the label reached when a callee returns via
3485 the :ref:`resume <i_resume>` instruction or other exception handling
3487 #. The optional :ref:`function attributes <fnattrs>` list. Only
3488 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
3489 attributes are valid here.
3494 This instruction is designed to operate as a standard '``call``'
3495 instruction in most regards. The primary difference is that it
3496 establishes an association with a label, which is used by the runtime
3497 library to unwind the stack.
3499 This instruction is used in languages with destructors to ensure that
3500 proper cleanup is performed in the case of either a ``longjmp`` or a
3501 thrown exception. Additionally, this is important for implementation of
3502 '``catch``' clauses in high-level languages that support them.
3504 For the purposes of the SSA form, the definition of the value returned
3505 by the '``invoke``' instruction is deemed to occur on the edge from the
3506 current block to the "normal" label. If the callee unwinds then no
3507 return value is available.
3512 .. code-block:: llvm
3514 %retval = invoke i32 @Test(i32 15) to label %Continue
3515 unwind label %TestCleanup ; {i32}:retval set
3516 %retval = invoke coldcc i32 %Testfnptr(i32 15) to label %Continue
3517 unwind label %TestCleanup ; {i32}:retval set
3521 '``resume``' Instruction
3522 ^^^^^^^^^^^^^^^^^^^^^^^^
3529 resume <type> <value>
3534 The '``resume``' instruction is a terminator instruction that has no
3540 The '``resume``' instruction requires one argument, which must have the
3541 same type as the result of any '``landingpad``' instruction in the same
3547 The '``resume``' instruction resumes propagation of an existing
3548 (in-flight) exception whose unwinding was interrupted with a
3549 :ref:`landingpad <i_landingpad>` instruction.
3554 .. code-block:: llvm
3556 resume { i8*, i32 } %exn
3560 '``unreachable``' Instruction
3561 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3573 The '``unreachable``' instruction has no defined semantics. This
3574 instruction is used to inform the optimizer that a particular portion of
3575 the code is not reachable. This can be used to indicate that the code
3576 after a no-return function cannot be reached, and other facts.
3581 The '``unreachable``' instruction has no defined semantics.
3588 Binary operators are used to do most of the computation in a program.
3589 They require two operands of the same type, execute an operation on
3590 them, and produce a single value. The operands might represent multiple
3591 data, as is the case with the :ref:`vector <t_vector>` data type. The
3592 result value has the same type as its operands.
3594 There are several different binary operators:
3598 '``add``' Instruction
3599 ^^^^^^^^^^^^^^^^^^^^^
3606 <result> = add <ty> <op1>, <op2> ; yields {ty}:result
3607 <result> = add nuw <ty> <op1>, <op2> ; yields {ty}:result
3608 <result> = add nsw <ty> <op1>, <op2> ; yields {ty}:result
3609 <result> = add nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3614 The '``add``' instruction returns the sum of its two operands.
3619 The two arguments to the '``add``' instruction must be
3620 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3621 arguments must have identical types.
3626 The value produced is the integer sum of the two operands.
3628 If the sum has unsigned overflow, the result returned is the
3629 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
3632 Because LLVM integers use a two's complement representation, this
3633 instruction is appropriate for both signed and unsigned integers.
3635 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3636 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3637 result value of the ``add`` is a :ref:`poison value <poisonvalues>` if
3638 unsigned and/or signed overflow, respectively, occurs.
3643 .. code-block:: llvm
3645 <result> = add i32 4, %var ; yields {i32}:result = 4 + %var
3649 '``fadd``' Instruction
3650 ^^^^^^^^^^^^^^^^^^^^^^
3657 <result> = fadd [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3662 The '``fadd``' instruction returns the sum of its two operands.
3667 The two arguments to the '``fadd``' instruction must be :ref:`floating
3668 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3669 Both arguments must have identical types.
3674 The value produced is the floating point sum of the two operands. This
3675 instruction can also take any number of :ref:`fast-math flags <fastmath>`,
3676 which are optimization hints to enable otherwise unsafe floating point
3682 .. code-block:: llvm
3684 <result> = fadd float 4.0, %var ; yields {float}:result = 4.0 + %var
3686 '``sub``' Instruction
3687 ^^^^^^^^^^^^^^^^^^^^^
3694 <result> = sub <ty> <op1>, <op2> ; yields {ty}:result
3695 <result> = sub nuw <ty> <op1>, <op2> ; yields {ty}:result
3696 <result> = sub nsw <ty> <op1>, <op2> ; yields {ty}:result
3697 <result> = sub nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3702 The '``sub``' instruction returns the difference of its two operands.
3704 Note that the '``sub``' instruction is used to represent the '``neg``'
3705 instruction present in most other intermediate representations.
3710 The two arguments to the '``sub``' instruction must be
3711 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3712 arguments must have identical types.
3717 The value produced is the integer difference of the two operands.
3719 If the difference has unsigned overflow, the result returned is the
3720 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
3723 Because LLVM integers use a two's complement representation, this
3724 instruction is appropriate for both signed and unsigned integers.
3726 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3727 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3728 result value of the ``sub`` is a :ref:`poison value <poisonvalues>` if
3729 unsigned and/or signed overflow, respectively, occurs.
3734 .. code-block:: llvm
3736 <result> = sub i32 4, %var ; yields {i32}:result = 4 - %var
3737 <result> = sub i32 0, %val ; yields {i32}:result = -%var
3741 '``fsub``' Instruction
3742 ^^^^^^^^^^^^^^^^^^^^^^
3749 <result> = fsub [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3754 The '``fsub``' instruction returns the difference of its two operands.
3756 Note that the '``fsub``' instruction is used to represent the '``fneg``'
3757 instruction present in most other intermediate representations.
3762 The two arguments to the '``fsub``' instruction must be :ref:`floating
3763 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3764 Both arguments must have identical types.
3769 The value produced is the floating point difference of the two operands.
3770 This instruction can also take any number of :ref:`fast-math
3771 flags <fastmath>`, which are optimization hints to enable otherwise
3772 unsafe floating point optimizations:
3777 .. code-block:: llvm
3779 <result> = fsub float 4.0, %var ; yields {float}:result = 4.0 - %var
3780 <result> = fsub float -0.0, %val ; yields {float}:result = -%var
3782 '``mul``' Instruction
3783 ^^^^^^^^^^^^^^^^^^^^^
3790 <result> = mul <ty> <op1>, <op2> ; yields {ty}:result
3791 <result> = mul nuw <ty> <op1>, <op2> ; yields {ty}:result
3792 <result> = mul nsw <ty> <op1>, <op2> ; yields {ty}:result
3793 <result> = mul nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3798 The '``mul``' instruction returns the product of its two operands.
3803 The two arguments to the '``mul``' instruction must be
3804 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3805 arguments must have identical types.
3810 The value produced is the integer product of the two operands.
3812 If the result of the multiplication has unsigned overflow, the result
3813 returned is the mathematical result modulo 2\ :sup:`n`\ , where n is the
3814 bit width of the result.
3816 Because LLVM integers use a two's complement representation, and the
3817 result is the same width as the operands, this instruction returns the
3818 correct result for both signed and unsigned integers. If a full product
3819 (e.g. ``i32`` * ``i32`` -> ``i64``) is needed, the operands should be
3820 sign-extended or zero-extended as appropriate to the width of the full
3823 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3824 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3825 result value of the ``mul`` is a :ref:`poison value <poisonvalues>` if
3826 unsigned and/or signed overflow, respectively, occurs.
3831 .. code-block:: llvm
3833 <result> = mul i32 4, %var ; yields {i32}:result = 4 * %var
3837 '``fmul``' Instruction
3838 ^^^^^^^^^^^^^^^^^^^^^^
3845 <result> = fmul [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3850 The '``fmul``' instruction returns the product of its two operands.
3855 The two arguments to the '``fmul``' instruction must be :ref:`floating
3856 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3857 Both arguments must have identical types.
3862 The value produced is the floating point product of the two operands.
3863 This instruction can also take any number of :ref:`fast-math
3864 flags <fastmath>`, which are optimization hints to enable otherwise
3865 unsafe floating point optimizations:
3870 .. code-block:: llvm
3872 <result> = fmul float 4.0, %var ; yields {float}:result = 4.0 * %var
3874 '``udiv``' Instruction
3875 ^^^^^^^^^^^^^^^^^^^^^^
3882 <result> = udiv <ty> <op1>, <op2> ; yields {ty}:result
3883 <result> = udiv exact <ty> <op1>, <op2> ; yields {ty}:result
3888 The '``udiv``' instruction returns the quotient of its two operands.
3893 The two arguments to the '``udiv``' instruction must be
3894 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3895 arguments must have identical types.
3900 The value produced is the unsigned integer quotient of the two operands.
3902 Note that unsigned integer division and signed integer division are
3903 distinct operations; for signed integer division, use '``sdiv``'.
3905 Division by zero leads to undefined behavior.
3907 If the ``exact`` keyword is present, the result value of the ``udiv`` is
3908 a :ref:`poison value <poisonvalues>` if %op1 is not a multiple of %op2 (as
3909 such, "((a udiv exact b) mul b) == a").
3914 .. code-block:: llvm
3916 <result> = udiv i32 4, %var ; yields {i32}:result = 4 / %var
3918 '``sdiv``' Instruction
3919 ^^^^^^^^^^^^^^^^^^^^^^
3926 <result> = sdiv <ty> <op1>, <op2> ; yields {ty}:result
3927 <result> = sdiv exact <ty> <op1>, <op2> ; yields {ty}:result
3932 The '``sdiv``' instruction returns the quotient of its two operands.
3937 The two arguments to the '``sdiv``' instruction must be
3938 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3939 arguments must have identical types.
3944 The value produced is the signed integer quotient of the two operands
3945 rounded towards zero.
3947 Note that signed integer division and unsigned integer division are
3948 distinct operations; for unsigned integer division, use '``udiv``'.
3950 Division by zero leads to undefined behavior. Overflow also leads to
3951 undefined behavior; this is a rare case, but can occur, for example, by
3952 doing a 32-bit division of -2147483648 by -1.
3954 If the ``exact`` keyword is present, the result value of the ``sdiv`` is
3955 a :ref:`poison value <poisonvalues>` if the result would be rounded.
3960 .. code-block:: llvm
3962 <result> = sdiv i32 4, %var ; yields {i32}:result = 4 / %var
3966 '``fdiv``' Instruction
3967 ^^^^^^^^^^^^^^^^^^^^^^
3974 <result> = fdiv [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3979 The '``fdiv``' instruction returns the quotient of its two operands.
3984 The two arguments to the '``fdiv``' instruction must be :ref:`floating
3985 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3986 Both arguments must have identical types.
3991 The value produced is the floating point quotient of the two operands.
3992 This instruction can also take any number of :ref:`fast-math
3993 flags <fastmath>`, which are optimization hints to enable otherwise
3994 unsafe floating point optimizations:
3999 .. code-block:: llvm
4001 <result> = fdiv float 4.0, %var ; yields {float}:result = 4.0 / %var
4003 '``urem``' Instruction
4004 ^^^^^^^^^^^^^^^^^^^^^^
4011 <result> = urem <ty> <op1>, <op2> ; yields {ty}:result
4016 The '``urem``' instruction returns the remainder from the unsigned
4017 division of its two arguments.
4022 The two arguments to the '``urem``' instruction must be
4023 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4024 arguments must have identical types.
4029 This instruction returns the unsigned integer *remainder* of a division.
4030 This instruction always performs an unsigned division to get the
4033 Note that unsigned integer remainder and signed integer remainder are
4034 distinct operations; for signed integer remainder, use '``srem``'.
4036 Taking the remainder of a division by zero leads to undefined behavior.
4041 .. code-block:: llvm
4043 <result> = urem i32 4, %var ; yields {i32}:result = 4 % %var
4045 '``srem``' Instruction
4046 ^^^^^^^^^^^^^^^^^^^^^^
4053 <result> = srem <ty> <op1>, <op2> ; yields {ty}:result
4058 The '``srem``' instruction returns the remainder from the signed
4059 division of its two operands. This instruction can also take
4060 :ref:`vector <t_vector>` versions of the values in which case the elements
4066 The two arguments to the '``srem``' instruction must be
4067 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4068 arguments must have identical types.
4073 This instruction returns the *remainder* of a division (where the result
4074 is either zero or has the same sign as the dividend, ``op1``), not the
4075 *modulo* operator (where the result is either zero or has the same sign
4076 as the divisor, ``op2``) of a value. For more information about the
4077 difference, see `The Math
4078 Forum <http://mathforum.org/dr.math/problems/anne.4.28.99.html>`_. For a
4079 table of how this is implemented in various languages, please see
4081 operation <http://en.wikipedia.org/wiki/Modulo_operation>`_.
4083 Note that signed integer remainder and unsigned integer remainder are
4084 distinct operations; for unsigned integer remainder, use '``urem``'.
4086 Taking the remainder of a division by zero leads to undefined behavior.
4087 Overflow also leads to undefined behavior; this is a rare case, but can
4088 occur, for example, by taking the remainder of a 32-bit division of
4089 -2147483648 by -1. (The remainder doesn't actually overflow, but this
4090 rule lets srem be implemented using instructions that return both the
4091 result of the division and the remainder.)
4096 .. code-block:: llvm
4098 <result> = srem i32 4, %var ; yields {i32}:result = 4 % %var
4102 '``frem``' Instruction
4103 ^^^^^^^^^^^^^^^^^^^^^^
4110 <result> = frem [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
4115 The '``frem``' instruction returns the remainder from the division of
4121 The two arguments to the '``frem``' instruction must be :ref:`floating
4122 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
4123 Both arguments must have identical types.
4128 This instruction returns the *remainder* of a division. The remainder
4129 has the same sign as the dividend. This instruction can also take any
4130 number of :ref:`fast-math flags <fastmath>`, which are optimization hints
4131 to enable otherwise unsafe floating point optimizations:
4136 .. code-block:: llvm
4138 <result> = frem float 4.0, %var ; yields {float}:result = 4.0 % %var
4142 Bitwise Binary Operations
4143 -------------------------
4145 Bitwise binary operators are used to do various forms of bit-twiddling
4146 in a program. They are generally very efficient instructions and can
4147 commonly be strength reduced from other instructions. They require two
4148 operands of the same type, execute an operation on them, and produce a
4149 single value. The resulting value is the same type as its operands.
4151 '``shl``' Instruction
4152 ^^^^^^^^^^^^^^^^^^^^^
4159 <result> = shl <ty> <op1>, <op2> ; yields {ty}:result
4160 <result> = shl nuw <ty> <op1>, <op2> ; yields {ty}:result
4161 <result> = shl nsw <ty> <op1>, <op2> ; yields {ty}:result
4162 <result> = shl nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
4167 The '``shl``' instruction returns the first operand shifted to the left
4168 a specified number of bits.
4173 Both arguments to the '``shl``' instruction must be the same
4174 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4175 '``op2``' is treated as an unsigned value.
4180 The value produced is ``op1`` \* 2\ :sup:`op2` mod 2\ :sup:`n`,
4181 where ``n`` is the width of the result. If ``op2`` is (statically or
4182 dynamically) negative or equal to or larger than the number of bits in
4183 ``op1``, the result is undefined. If the arguments are vectors, each
4184 vector element of ``op1`` is shifted by the corresponding shift amount
4187 If the ``nuw`` keyword is present, then the shift produces a :ref:`poison
4188 value <poisonvalues>` if it shifts out any non-zero bits. If the
4189 ``nsw`` keyword is present, then the shift produces a :ref:`poison
4190 value <poisonvalues>` if it shifts out any bits that disagree with the
4191 resultant sign bit. As such, NUW/NSW have the same semantics as they
4192 would if the shift were expressed as a mul instruction with the same
4193 nsw/nuw bits in (mul %op1, (shl 1, %op2)).
4198 .. code-block:: llvm
4200 <result> = shl i32 4, %var ; yields {i32}: 4 << %var
4201 <result> = shl i32 4, 2 ; yields {i32}: 16
4202 <result> = shl i32 1, 10 ; yields {i32}: 1024
4203 <result> = shl i32 1, 32 ; undefined
4204 <result> = shl <2 x i32> < i32 1, i32 1>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 2, i32 4>
4206 '``lshr``' Instruction
4207 ^^^^^^^^^^^^^^^^^^^^^^
4214 <result> = lshr <ty> <op1>, <op2> ; yields {ty}:result
4215 <result> = lshr exact <ty> <op1>, <op2> ; yields {ty}:result
4220 The '``lshr``' instruction (logical shift right) returns the first
4221 operand shifted to the right a specified number of bits with zero fill.
4226 Both arguments to the '``lshr``' instruction must be the same
4227 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4228 '``op2``' is treated as an unsigned value.
4233 This instruction always performs a logical shift right operation. The
4234 most significant bits of the result will be filled with zero bits after
4235 the shift. If ``op2`` is (statically or dynamically) equal to or larger
4236 than the number of bits in ``op1``, the result is undefined. If the
4237 arguments are vectors, each vector element of ``op1`` is shifted by the
4238 corresponding shift amount in ``op2``.
4240 If the ``exact`` keyword is present, the result value of the ``lshr`` is
4241 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
4247 .. code-block:: llvm
4249 <result> = lshr i32 4, 1 ; yields {i32}:result = 2
4250 <result> = lshr i32 4, 2 ; yields {i32}:result = 1
4251 <result> = lshr i8 4, 3 ; yields {i8}:result = 0
4252 <result> = lshr i8 -2, 1 ; yields {i8}:result = 0x7F
4253 <result> = lshr i32 1, 32 ; undefined
4254 <result> = lshr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 0x7FFFFFFF, i32 1>
4256 '``ashr``' Instruction
4257 ^^^^^^^^^^^^^^^^^^^^^^
4264 <result> = ashr <ty> <op1>, <op2> ; yields {ty}:result
4265 <result> = ashr exact <ty> <op1>, <op2> ; yields {ty}:result
4270 The '``ashr``' instruction (arithmetic shift right) returns the first
4271 operand shifted to the right a specified number of bits with sign
4277 Both arguments to the '``ashr``' instruction must be the same
4278 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4279 '``op2``' is treated as an unsigned value.
4284 This instruction always performs an arithmetic shift right operation,
4285 The most significant bits of the result will be filled with the sign bit
4286 of ``op1``. If ``op2`` is (statically or dynamically) equal to or larger
4287 than the number of bits in ``op1``, the result is undefined. If the
4288 arguments are vectors, each vector element of ``op1`` is shifted by the
4289 corresponding shift amount in ``op2``.
4291 If the ``exact`` keyword is present, the result value of the ``ashr`` is
4292 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
4298 .. code-block:: llvm
4300 <result> = ashr i32 4, 1 ; yields {i32}:result = 2
4301 <result> = ashr i32 4, 2 ; yields {i32}:result = 1
4302 <result> = ashr i8 4, 3 ; yields {i8}:result = 0
4303 <result> = ashr i8 -2, 1 ; yields {i8}:result = -1
4304 <result> = ashr i32 1, 32 ; undefined
4305 <result> = ashr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 3> ; yields: result=<2 x i32> < i32 -1, i32 0>
4307 '``and``' Instruction
4308 ^^^^^^^^^^^^^^^^^^^^^
4315 <result> = and <ty> <op1>, <op2> ; yields {ty}:result
4320 The '``and``' instruction returns the bitwise logical and of its two
4326 The two arguments to the '``and``' instruction must be
4327 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4328 arguments must have identical types.
4333 The truth table used for the '``and``' instruction is:
4350 .. code-block:: llvm
4352 <result> = and i32 4, %var ; yields {i32}:result = 4 & %var
4353 <result> = and i32 15, 40 ; yields {i32}:result = 8
4354 <result> = and i32 4, 8 ; yields {i32}:result = 0
4356 '``or``' Instruction
4357 ^^^^^^^^^^^^^^^^^^^^
4364 <result> = or <ty> <op1>, <op2> ; yields {ty}:result
4369 The '``or``' instruction returns the bitwise logical inclusive or of its
4375 The two arguments to the '``or``' instruction must be
4376 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4377 arguments must have identical types.
4382 The truth table used for the '``or``' instruction is:
4401 <result> = or i32 4, %var ; yields {i32}:result = 4 | %var
4402 <result> = or i32 15, 40 ; yields {i32}:result = 47
4403 <result> = or i32 4, 8 ; yields {i32}:result = 12
4405 '``xor``' Instruction
4406 ^^^^^^^^^^^^^^^^^^^^^
4413 <result> = xor <ty> <op1>, <op2> ; yields {ty}:result
4418 The '``xor``' instruction returns the bitwise logical exclusive or of
4419 its two operands. The ``xor`` is used to implement the "one's
4420 complement" operation, which is the "~" operator in C.
4425 The two arguments to the '``xor``' instruction must be
4426 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4427 arguments must have identical types.
4432 The truth table used for the '``xor``' instruction is:
4449 .. code-block:: llvm
4451 <result> = xor i32 4, %var ; yields {i32}:result = 4 ^ %var
4452 <result> = xor i32 15, 40 ; yields {i32}:result = 39
4453 <result> = xor i32 4, 8 ; yields {i32}:result = 12
4454 <result> = xor i32 %V, -1 ; yields {i32}:result = ~%V
4459 LLVM supports several instructions to represent vector operations in a
4460 target-independent manner. These instructions cover the element-access
4461 and vector-specific operations needed to process vectors effectively.
4462 While LLVM does directly support these vector operations, many
4463 sophisticated algorithms will want to use target-specific intrinsics to
4464 take full advantage of a specific target.
4466 .. _i_extractelement:
4468 '``extractelement``' Instruction
4469 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4476 <result> = extractelement <n x <ty>> <val>, <ty2> <idx> ; yields <ty>
4481 The '``extractelement``' instruction extracts a single scalar element
4482 from a vector at a specified index.
4487 The first operand of an '``extractelement``' instruction is a value of
4488 :ref:`vector <t_vector>` type. The second operand is an index indicating
4489 the position from which to extract the element. The index may be a
4490 variable of any integer type.
4495 The result is a scalar of the same type as the element type of ``val``.
4496 Its value is the value at position ``idx`` of ``val``. If ``idx``
4497 exceeds the length of ``val``, the results are undefined.
4502 .. code-block:: llvm
4504 <result> = extractelement <4 x i32> %vec, i32 0 ; yields i32
4506 .. _i_insertelement:
4508 '``insertelement``' Instruction
4509 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4516 <result> = insertelement <n x <ty>> <val>, <ty> <elt>, <ty2> <idx> ; yields <n x <ty>>
4521 The '``insertelement``' instruction inserts a scalar element into a
4522 vector at a specified index.
4527 The first operand of an '``insertelement``' instruction is a value of
4528 :ref:`vector <t_vector>` type. The second operand is a scalar value whose
4529 type must equal the element type of the first operand. The third operand
4530 is an index indicating the position at which to insert the value. The
4531 index may be a variable of any integer type.
4536 The result is a vector of the same type as ``val``. Its element values
4537 are those of ``val`` except at position ``idx``, where it gets the value
4538 ``elt``. If ``idx`` exceeds the length of ``val``, the results are
4544 .. code-block:: llvm
4546 <result> = insertelement <4 x i32> %vec, i32 1, i32 0 ; yields <4 x i32>
4548 .. _i_shufflevector:
4550 '``shufflevector``' Instruction
4551 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4558 <result> = shufflevector <n x <ty>> <v1>, <n x <ty>> <v2>, <m x i32> <mask> ; yields <m x <ty>>
4563 The '``shufflevector``' instruction constructs a permutation of elements
4564 from two input vectors, returning a vector with the same element type as
4565 the input and length that is the same as the shuffle mask.
4570 The first two operands of a '``shufflevector``' instruction are vectors
4571 with the same type. The third argument is a shuffle mask whose element
4572 type is always 'i32'. The result of the instruction is a vector whose
4573 length is the same as the shuffle mask and whose element type is the
4574 same as the element type of the first two operands.
4576 The shuffle mask operand is required to be a constant vector with either
4577 constant integer or undef values.
4582 The elements of the two input vectors are numbered from left to right
4583 across both of the vectors. The shuffle mask operand specifies, for each
4584 element of the result vector, which element of the two input vectors the
4585 result element gets. The element selector may be undef (meaning "don't
4586 care") and the second operand may be undef if performing a shuffle from
4592 .. code-block:: llvm
4594 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
4595 <4 x i32> <i32 0, i32 4, i32 1, i32 5> ; yields <4 x i32>
4596 <result> = shufflevector <4 x i32> %v1, <4 x i32> undef,
4597 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32> - Identity shuffle.
4598 <result> = shufflevector <8 x i32> %v1, <8 x i32> undef,
4599 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32>
4600 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
4601 <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7 > ; yields <8 x i32>
4603 Aggregate Operations
4604 --------------------
4606 LLVM supports several instructions for working with
4607 :ref:`aggregate <t_aggregate>` values.
4611 '``extractvalue``' Instruction
4612 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4619 <result> = extractvalue <aggregate type> <val>, <idx>{, <idx>}*
4624 The '``extractvalue``' instruction extracts the value of a member field
4625 from an :ref:`aggregate <t_aggregate>` value.
4630 The first operand of an '``extractvalue``' instruction is a value of
4631 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The operands are
4632 constant indices to specify which value to extract in a similar manner
4633 as indices in a '``getelementptr``' instruction.
4635 The major differences to ``getelementptr`` indexing are:
4637 - Since the value being indexed is not a pointer, the first index is
4638 omitted and assumed to be zero.
4639 - At least one index must be specified.
4640 - Not only struct indices but also array indices must be in bounds.
4645 The result is the value at the position in the aggregate specified by
4651 .. code-block:: llvm
4653 <result> = extractvalue {i32, float} %agg, 0 ; yields i32
4657 '``insertvalue``' Instruction
4658 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4665 <result> = insertvalue <aggregate type> <val>, <ty> <elt>, <idx>{, <idx>}* ; yields <aggregate type>
4670 The '``insertvalue``' instruction inserts a value into a member field in
4671 an :ref:`aggregate <t_aggregate>` value.
4676 The first operand of an '``insertvalue``' instruction is a value of
4677 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The second operand is
4678 a first-class value to insert. The following operands are constant
4679 indices indicating the position at which to insert the value in a
4680 similar manner as indices in a '``extractvalue``' instruction. The value
4681 to insert must have the same type as the value identified by the
4687 The result is an aggregate of the same type as ``val``. Its value is
4688 that of ``val`` except that the value at the position specified by the
4689 indices is that of ``elt``.
4694 .. code-block:: llvm
4696 %agg1 = insertvalue {i32, float} undef, i32 1, 0 ; yields {i32 1, float undef}
4697 %agg2 = insertvalue {i32, float} %agg1, float %val, 1 ; yields {i32 1, float %val}
4698 %agg3 = insertvalue {i32, {float}} %agg1, float %val, 1, 0 ; yields {i32 1, float %val}
4702 Memory Access and Addressing Operations
4703 ---------------------------------------
4705 A key design point of an SSA-based representation is how it represents
4706 memory. In LLVM, no memory locations are in SSA form, which makes things
4707 very simple. This section describes how to read, write, and allocate
4712 '``alloca``' Instruction
4713 ^^^^^^^^^^^^^^^^^^^^^^^^
4720 <result> = alloca [inalloca] <type> [, <ty> <NumElements>] [, align <alignment>] ; yields {type*}:result
4725 The '``alloca``' instruction allocates memory on the stack frame of the
4726 currently executing function, to be automatically released when this
4727 function returns to its caller. The object is always allocated in the
4728 generic address space (address space zero).
4733 The '``alloca``' instruction allocates ``sizeof(<type>)*NumElements``
4734 bytes of memory on the runtime stack, returning a pointer of the
4735 appropriate type to the program. If "NumElements" is specified, it is
4736 the number of elements allocated, otherwise "NumElements" is defaulted
4737 to be one. If a constant alignment is specified, the value result of the
4738 allocation is guaranteed to be aligned to at least that boundary. If not
4739 specified, or if zero, the target can choose to align the allocation on
4740 any convenient boundary compatible with the type.
4742 '``type``' may be any sized type.
4747 Memory is allocated; a pointer is returned. The operation is undefined
4748 if there is insufficient stack space for the allocation. '``alloca``'d
4749 memory is automatically released when the function returns. The
4750 '``alloca``' instruction is commonly used to represent automatic
4751 variables that must have an address available. When the function returns
4752 (either with the ``ret`` or ``resume`` instructions), the memory is
4753 reclaimed. Allocating zero bytes is legal, but the result is undefined.
4754 The order in which memory is allocated (ie., which way the stack grows)
4760 .. code-block:: llvm
4762 %ptr = alloca i32 ; yields {i32*}:ptr
4763 %ptr = alloca i32, i32 4 ; yields {i32*}:ptr
4764 %ptr = alloca i32, i32 4, align 1024 ; yields {i32*}:ptr
4765 %ptr = alloca i32, align 1024 ; yields {i32*}:ptr
4769 '``load``' Instruction
4770 ^^^^^^^^^^^^^^^^^^^^^^
4777 <result> = load [volatile] <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>][, !invariant.load !<index>]
4778 <result> = load atomic [volatile] <ty>* <pointer> [singlethread] <ordering>, align <alignment>
4779 !<index> = !{ i32 1 }
4784 The '``load``' instruction is used to read from memory.
4789 The argument to the ``load`` instruction specifies the memory address
4790 from which to load. The pointer must point to a :ref:`first
4791 class <t_firstclass>` type. If the ``load`` is marked as ``volatile``,
4792 then the optimizer is not allowed to modify the number or order of
4793 execution of this ``load`` with other :ref:`volatile
4794 operations <volatile>`.
4796 If the ``load`` is marked as ``atomic``, it takes an extra
4797 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
4798 ``release`` and ``acq_rel`` orderings are not valid on ``load``
4799 instructions. Atomic loads produce :ref:`defined <memmodel>` results
4800 when they may see multiple atomic stores. The type of the pointee must
4801 be an integer type whose bit width is a power of two greater than or
4802 equal to eight and less than or equal to a target-specific size limit.
4803 ``align`` must be explicitly specified on atomic loads, and the load has
4804 undefined behavior if the alignment is not set to a value which is at
4805 least the size in bytes of the pointee. ``!nontemporal`` does not have
4806 any defined semantics for atomic loads.
4808 The optional constant ``align`` argument specifies the alignment of the
4809 operation (that is, the alignment of the memory address). A value of 0
4810 or an omitted ``align`` argument means that the operation has the ABI
4811 alignment for the target. It is the responsibility of the code emitter
4812 to ensure that the alignment information is correct. Overestimating the
4813 alignment results in undefined behavior. Underestimating the alignment
4814 may produce less efficient code. An alignment of 1 is always safe.
4816 The optional ``!nontemporal`` metadata must reference a single
4817 metadata name ``<index>`` corresponding to a metadata node with one
4818 ``i32`` entry of value 1. The existence of the ``!nontemporal``
4819 metadata on the instruction tells the optimizer and code generator
4820 that this load is not expected to be reused in the cache. The code
4821 generator may select special instructions to save cache bandwidth, such
4822 as the ``MOVNT`` instruction on x86.
4824 The optional ``!invariant.load`` metadata must reference a single
4825 metadata name ``<index>`` corresponding to a metadata node with no
4826 entries. The existence of the ``!invariant.load`` metadata on the
4827 instruction tells the optimizer and code generator that this load
4828 address points to memory which does not change value during program
4829 execution. The optimizer may then move this load around, for example, by
4830 hoisting it out of loops using loop invariant code motion.
4835 The location of memory pointed to is loaded. If the value being loaded
4836 is of scalar type then the number of bytes read does not exceed the
4837 minimum number of bytes needed to hold all bits of the type. For
4838 example, loading an ``i24`` reads at most three bytes. When loading a
4839 value of a type like ``i20`` with a size that is not an integral number
4840 of bytes, the result is undefined if the value was not originally
4841 written using a store of the same type.
4846 .. code-block:: llvm
4848 %ptr = alloca i32 ; yields {i32*}:ptr
4849 store i32 3, i32* %ptr ; yields {void}
4850 %val = load i32* %ptr ; yields {i32}:val = i32 3
4854 '``store``' Instruction
4855 ^^^^^^^^^^^^^^^^^^^^^^^
4862 store [volatile] <ty> <value>, <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>] ; yields {void}
4863 store atomic [volatile] <ty> <value>, <ty>* <pointer> [singlethread] <ordering>, align <alignment> ; yields {void}
4868 The '``store``' instruction is used to write to memory.
4873 There are two arguments to the ``store`` instruction: a value to store
4874 and an address at which to store it. The type of the ``<pointer>``
4875 operand must be a pointer to the :ref:`first class <t_firstclass>` type of
4876 the ``<value>`` operand. If the ``store`` is marked as ``volatile``,
4877 then the optimizer is not allowed to modify the number or order of
4878 execution of this ``store`` with other :ref:`volatile
4879 operations <volatile>`.
4881 If the ``store`` is marked as ``atomic``, it takes an extra
4882 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
4883 ``acquire`` and ``acq_rel`` orderings aren't valid on ``store``
4884 instructions. Atomic loads produce :ref:`defined <memmodel>` results
4885 when they may see multiple atomic stores. The type of the pointee must
4886 be an integer type whose bit width is a power of two greater than or
4887 equal to eight and less than or equal to a target-specific size limit.
4888 ``align`` must be explicitly specified on atomic stores, and the store
4889 has undefined behavior if the alignment is not set to a value which is
4890 at least the size in bytes of the pointee. ``!nontemporal`` does not
4891 have any defined semantics for atomic stores.
4893 The optional constant ``align`` argument specifies the alignment of the
4894 operation (that is, the alignment of the memory address). A value of 0
4895 or an omitted ``align`` argument means that the operation has the ABI
4896 alignment for the target. It is the responsibility of the code emitter
4897 to ensure that the alignment information is correct. Overestimating the
4898 alignment results in undefined behavior. Underestimating the
4899 alignment may produce less efficient code. An alignment of 1 is always
4902 The optional ``!nontemporal`` metadata must reference a single metadata
4903 name ``<index>`` corresponding to a metadata node with one ``i32`` entry of
4904 value 1. The existence of the ``!nontemporal`` metadata on the instruction
4905 tells the optimizer and code generator that this load is not expected to
4906 be reused in the cache. The code generator may select special
4907 instructions to save cache bandwidth, such as the MOVNT instruction on
4913 The contents of memory are updated to contain ``<value>`` at the
4914 location specified by the ``<pointer>`` operand. If ``<value>`` is
4915 of scalar type then the number of bytes written does not exceed the
4916 minimum number of bytes needed to hold all bits of the type. For
4917 example, storing an ``i24`` writes at most three bytes. When writing a
4918 value of a type like ``i20`` with a size that is not an integral number
4919 of bytes, it is unspecified what happens to the extra bits that do not
4920 belong to the type, but they will typically be overwritten.
4925 .. code-block:: llvm
4927 %ptr = alloca i32 ; yields {i32*}:ptr
4928 store i32 3, i32* %ptr ; yields {void}
4929 %val = load i32* %ptr ; yields {i32}:val = i32 3
4933 '``fence``' Instruction
4934 ^^^^^^^^^^^^^^^^^^^^^^^
4941 fence [singlethread] <ordering> ; yields {void}
4946 The '``fence``' instruction is used to introduce happens-before edges
4952 '``fence``' instructions take an :ref:`ordering <ordering>` argument which
4953 defines what *synchronizes-with* edges they add. They can only be given
4954 ``acquire``, ``release``, ``acq_rel``, and ``seq_cst`` orderings.
4959 A fence A which has (at least) ``release`` ordering semantics
4960 *synchronizes with* a fence B with (at least) ``acquire`` ordering
4961 semantics if and only if there exist atomic operations X and Y, both
4962 operating on some atomic object M, such that A is sequenced before X, X
4963 modifies M (either directly or through some side effect of a sequence
4964 headed by X), Y is sequenced before B, and Y observes M. This provides a
4965 *happens-before* dependency between A and B. Rather than an explicit
4966 ``fence``, one (but not both) of the atomic operations X or Y might
4967 provide a ``release`` or ``acquire`` (resp.) ordering constraint and
4968 still *synchronize-with* the explicit ``fence`` and establish the
4969 *happens-before* edge.
4971 A ``fence`` which has ``seq_cst`` ordering, in addition to having both
4972 ``acquire`` and ``release`` semantics specified above, participates in
4973 the global program order of other ``seq_cst`` operations and/or fences.
4975 The optional ":ref:`singlethread <singlethread>`" argument specifies
4976 that the fence only synchronizes with other fences in the same thread.
4977 (This is useful for interacting with signal handlers.)
4982 .. code-block:: llvm
4984 fence acquire ; yields {void}
4985 fence singlethread seq_cst ; yields {void}
4989 '``cmpxchg``' Instruction
4990 ^^^^^^^^^^^^^^^^^^^^^^^^^
4997 cmpxchg [volatile] <ty>* <pointer>, <ty> <cmp>, <ty> <new> [singlethread] <success ordering> <failure ordering> ; yields {ty}
5002 The '``cmpxchg``' instruction is used to atomically modify memory. It
5003 loads a value in memory and compares it to a given value. If they are
5004 equal, it stores a new value into the memory.
5009 There are three arguments to the '``cmpxchg``' instruction: an address
5010 to operate on, a value to compare to the value currently be at that
5011 address, and a new value to place at that address if the compared values
5012 are equal. The type of '<cmp>' must be an integer type whose bit width
5013 is a power of two greater than or equal to eight and less than or equal
5014 to a target-specific size limit. '<cmp>' and '<new>' must have the same
5015 type, and the type of '<pointer>' must be a pointer to that type. If the
5016 ``cmpxchg`` is marked as ``volatile``, then the optimizer is not allowed
5017 to modify the number or order of execution of this ``cmpxchg`` with
5018 other :ref:`volatile operations <volatile>`.
5020 The success and failure :ref:`ordering <ordering>` arguments specify how this
5021 ``cmpxchg`` synchronizes with other atomic operations. The both ordering
5022 parameters must be at least ``monotonic``, the ordering constraint on failure
5023 must be no stronger than that on success, and the failure ordering cannot be
5024 either ``release`` or ``acq_rel``.
5026 The optional "``singlethread``" argument declares that the ``cmpxchg``
5027 is only atomic with respect to code (usually signal handlers) running in
5028 the same thread as the ``cmpxchg``. Otherwise the cmpxchg is atomic with
5029 respect to all other code in the system.
5031 The pointer passed into cmpxchg must have alignment greater than or
5032 equal to the size in memory of the operand.
5037 The contents of memory at the location specified by the '``<pointer>``'
5038 operand is read and compared to '``<cmp>``'; if the read value is the
5039 equal, '``<new>``' is written. The original value at the location is
5042 A successful ``cmpxchg`` is a read-modify-write instruction for the purpose of
5043 identifying release sequences. A failed ``cmpxchg`` is equivalent to an atomic
5044 load with an ordering parameter determined the second ordering parameter.
5049 .. code-block:: llvm
5052 %orig = atomic load i32* %ptr unordered ; yields {i32}
5056 %cmp = phi i32 [ %orig, %entry ], [%old, %loop]
5057 %squared = mul i32 %cmp, %cmp
5058 %old = cmpxchg i32* %ptr, i32 %cmp, i32 %squared acq_rel monotonic ; yields {i32}
5059 %success = icmp eq i32 %cmp, %old
5060 br i1 %success, label %done, label %loop
5067 '``atomicrmw``' Instruction
5068 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
5075 atomicrmw [volatile] <operation> <ty>* <pointer>, <ty> <value> [singlethread] <ordering> ; yields {ty}
5080 The '``atomicrmw``' instruction is used to atomically modify memory.
5085 There are three arguments to the '``atomicrmw``' instruction: an
5086 operation to apply, an address whose value to modify, an argument to the
5087 operation. The operation must be one of the following keywords:
5101 The type of '<value>' must be an integer type whose bit width is a power
5102 of two greater than or equal to eight and less than or equal to a
5103 target-specific size limit. The type of the '``<pointer>``' operand must
5104 be a pointer to that type. If the ``atomicrmw`` is marked as
5105 ``volatile``, then the optimizer is not allowed to modify the number or
5106 order of execution of this ``atomicrmw`` with other :ref:`volatile
5107 operations <volatile>`.
5112 The contents of memory at the location specified by the '``<pointer>``'
5113 operand are atomically read, modified, and written back. The original
5114 value at the location is returned. The modification is specified by the
5117 - xchg: ``*ptr = val``
5118 - add: ``*ptr = *ptr + val``
5119 - sub: ``*ptr = *ptr - val``
5120 - and: ``*ptr = *ptr & val``
5121 - nand: ``*ptr = ~(*ptr & val)``
5122 - or: ``*ptr = *ptr | val``
5123 - xor: ``*ptr = *ptr ^ val``
5124 - max: ``*ptr = *ptr > val ? *ptr : val`` (using a signed comparison)
5125 - min: ``*ptr = *ptr < val ? *ptr : val`` (using a signed comparison)
5126 - umax: ``*ptr = *ptr > val ? *ptr : val`` (using an unsigned
5128 - umin: ``*ptr = *ptr < val ? *ptr : val`` (using an unsigned
5134 .. code-block:: llvm
5136 %old = atomicrmw add i32* %ptr, i32 1 acquire ; yields {i32}
5138 .. _i_getelementptr:
5140 '``getelementptr``' Instruction
5141 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5148 <result> = getelementptr <pty>* <ptrval>{, <ty> <idx>}*
5149 <result> = getelementptr inbounds <pty>* <ptrval>{, <ty> <idx>}*
5150 <result> = getelementptr <ptr vector> ptrval, <vector index type> idx
5155 The '``getelementptr``' instruction is used to get the address of a
5156 subelement of an :ref:`aggregate <t_aggregate>` data structure. It performs
5157 address calculation only and does not access memory.
5162 The first argument is always a pointer or a vector of pointers, and
5163 forms the basis of the calculation. The remaining arguments are indices
5164 that indicate which of the elements of the aggregate object are indexed.
5165 The interpretation of each index is dependent on the type being indexed
5166 into. The first index always indexes the pointer value given as the
5167 first argument, the second index indexes a value of the type pointed to
5168 (not necessarily the value directly pointed to, since the first index
5169 can be non-zero), etc. The first type indexed into must be a pointer
5170 value, subsequent types can be arrays, vectors, and structs. Note that
5171 subsequent types being indexed into can never be pointers, since that
5172 would require loading the pointer before continuing calculation.
5174 The type of each index argument depends on the type it is indexing into.
5175 When indexing into a (optionally packed) structure, only ``i32`` integer
5176 **constants** are allowed (when using a vector of indices they must all
5177 be the **same** ``i32`` integer constant). When indexing into an array,
5178 pointer or vector, integers of any width are allowed, and they are not
5179 required to be constant. These integers are treated as signed values
5182 For example, let's consider a C code fragment and how it gets compiled
5198 int *foo(struct ST *s) {
5199 return &s[1].Z.B[5][13];
5202 The LLVM code generated by Clang is:
5204 .. code-block:: llvm
5206 %struct.RT = type { i8, [10 x [20 x i32]], i8 }
5207 %struct.ST = type { i32, double, %struct.RT }
5209 define i32* @foo(%struct.ST* %s) nounwind uwtable readnone optsize ssp {
5211 %arrayidx = getelementptr inbounds %struct.ST* %s, i64 1, i32 2, i32 1, i64 5, i64 13
5218 In the example above, the first index is indexing into the
5219 '``%struct.ST*``' type, which is a pointer, yielding a '``%struct.ST``'
5220 = '``{ i32, double, %struct.RT }``' type, a structure. The second index
5221 indexes into the third element of the structure, yielding a
5222 '``%struct.RT``' = '``{ i8 , [10 x [20 x i32]], i8 }``' type, another
5223 structure. The third index indexes into the second element of the
5224 structure, yielding a '``[10 x [20 x i32]]``' type, an array. The two
5225 dimensions of the array are subscripted into, yielding an '``i32``'
5226 type. The '``getelementptr``' instruction returns a pointer to this
5227 element, thus computing a value of '``i32*``' type.
5229 Note that it is perfectly legal to index partially through a structure,
5230 returning a pointer to an inner element. Because of this, the LLVM code
5231 for the given testcase is equivalent to:
5233 .. code-block:: llvm
5235 define i32* @foo(%struct.ST* %s) {
5236 %t1 = getelementptr %struct.ST* %s, i32 1 ; yields %struct.ST*:%t1
5237 %t2 = getelementptr %struct.ST* %t1, i32 0, i32 2 ; yields %struct.RT*:%t2
5238 %t3 = getelementptr %struct.RT* %t2, i32 0, i32 1 ; yields [10 x [20 x i32]]*:%t3
5239 %t4 = getelementptr [10 x [20 x i32]]* %t3, i32 0, i32 5 ; yields [20 x i32]*:%t4
5240 %t5 = getelementptr [20 x i32]* %t4, i32 0, i32 13 ; yields i32*:%t5
5244 If the ``inbounds`` keyword is present, the result value of the
5245 ``getelementptr`` is a :ref:`poison value <poisonvalues>` if the base
5246 pointer is not an *in bounds* address of an allocated object, or if any
5247 of the addresses that would be formed by successive addition of the
5248 offsets implied by the indices to the base address with infinitely
5249 precise signed arithmetic are not an *in bounds* address of that
5250 allocated object. The *in bounds* addresses for an allocated object are
5251 all the addresses that point into the object, plus the address one byte
5252 past the end. In cases where the base is a vector of pointers the
5253 ``inbounds`` keyword applies to each of the computations element-wise.
5255 If the ``inbounds`` keyword is not present, the offsets are added to the
5256 base address with silently-wrapping two's complement arithmetic. If the
5257 offsets have a different width from the pointer, they are sign-extended
5258 or truncated to the width of the pointer. The result value of the
5259 ``getelementptr`` may be outside the object pointed to by the base
5260 pointer. The result value may not necessarily be used to access memory
5261 though, even if it happens to point into allocated storage. See the
5262 :ref:`Pointer Aliasing Rules <pointeraliasing>` section for more
5265 The getelementptr instruction is often confusing. For some more insight
5266 into how it works, see :doc:`the getelementptr FAQ <GetElementPtr>`.
5271 .. code-block:: llvm
5273 ; yields [12 x i8]*:aptr
5274 %aptr = getelementptr {i32, [12 x i8]}* %saptr, i64 0, i32 1
5276 %vptr = getelementptr {i32, <2 x i8>}* %svptr, i64 0, i32 1, i32 1
5278 %eptr = getelementptr [12 x i8]* %aptr, i64 0, i32 1
5280 %iptr = getelementptr [10 x i32]* @arr, i16 0, i16 0
5282 In cases where the pointer argument is a vector of pointers, each index
5283 must be a vector with the same number of elements. For example:
5285 .. code-block:: llvm
5287 %A = getelementptr <4 x i8*> %ptrs, <4 x i64> %offsets,
5289 Conversion Operations
5290 ---------------------
5292 The instructions in this category are the conversion instructions
5293 (casting) which all take a single operand and a type. They perform
5294 various bit conversions on the operand.
5296 '``trunc .. to``' Instruction
5297 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5304 <result> = trunc <ty> <value> to <ty2> ; yields ty2
5309 The '``trunc``' instruction truncates its operand to the type ``ty2``.
5314 The '``trunc``' instruction takes a value to trunc, and a type to trunc
5315 it to. Both types must be of :ref:`integer <t_integer>` types, or vectors
5316 of the same number of integers. The bit size of the ``value`` must be
5317 larger than the bit size of the destination type, ``ty2``. Equal sized
5318 types are not allowed.
5323 The '``trunc``' instruction truncates the high order bits in ``value``
5324 and converts the remaining bits to ``ty2``. Since the source size must
5325 be larger than the destination size, ``trunc`` cannot be a *no-op cast*.
5326 It will always truncate bits.
5331 .. code-block:: llvm
5333 %X = trunc i32 257 to i8 ; yields i8:1
5334 %Y = trunc i32 123 to i1 ; yields i1:true
5335 %Z = trunc i32 122 to i1 ; yields i1:false
5336 %W = trunc <2 x i16> <i16 8, i16 7> to <2 x i8> ; yields <i8 8, i8 7>
5338 '``zext .. to``' Instruction
5339 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5346 <result> = zext <ty> <value> to <ty2> ; yields ty2
5351 The '``zext``' instruction zero extends its operand to type ``ty2``.
5356 The '``zext``' instruction takes a value to cast, and a type to cast it
5357 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
5358 the same number of integers. The bit size of the ``value`` must be
5359 smaller than the bit size of the destination type, ``ty2``.
5364 The ``zext`` fills the high order bits of the ``value`` with zero bits
5365 until it reaches the size of the destination type, ``ty2``.
5367 When zero extending from i1, the result will always be either 0 or 1.
5372 .. code-block:: llvm
5374 %X = zext i32 257 to i64 ; yields i64:257
5375 %Y = zext i1 true to i32 ; yields i32:1
5376 %Z = zext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
5378 '``sext .. to``' Instruction
5379 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5386 <result> = sext <ty> <value> to <ty2> ; yields ty2
5391 The '``sext``' sign extends ``value`` to the type ``ty2``.
5396 The '``sext``' instruction takes a value to cast, and a type to cast it
5397 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
5398 the same number of integers. The bit size of the ``value`` must be
5399 smaller than the bit size of the destination type, ``ty2``.
5404 The '``sext``' instruction performs a sign extension by copying the sign
5405 bit (highest order bit) of the ``value`` until it reaches the bit size
5406 of the type ``ty2``.
5408 When sign extending from i1, the extension always results in -1 or 0.
5413 .. code-block:: llvm
5415 %X = sext i8 -1 to i16 ; yields i16 :65535
5416 %Y = sext i1 true to i32 ; yields i32:-1
5417 %Z = sext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
5419 '``fptrunc .. to``' Instruction
5420 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5427 <result> = fptrunc <ty> <value> to <ty2> ; yields ty2
5432 The '``fptrunc``' instruction truncates ``value`` to type ``ty2``.
5437 The '``fptrunc``' instruction takes a :ref:`floating point <t_floating>`
5438 value to cast and a :ref:`floating point <t_floating>` type to cast it to.
5439 The size of ``value`` must be larger than the size of ``ty2``. This
5440 implies that ``fptrunc`` cannot be used to make a *no-op cast*.
5445 The '``fptrunc``' instruction truncates a ``value`` from a larger
5446 :ref:`floating point <t_floating>` type to a smaller :ref:`floating
5447 point <t_floating>` type. If the value cannot fit within the
5448 destination type, ``ty2``, then the results are undefined.
5453 .. code-block:: llvm
5455 %X = fptrunc double 123.0 to float ; yields float:123.0
5456 %Y = fptrunc double 1.0E+300 to float ; yields undefined
5458 '``fpext .. to``' Instruction
5459 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5466 <result> = fpext <ty> <value> to <ty2> ; yields ty2
5471 The '``fpext``' extends a floating point ``value`` to a larger floating
5477 The '``fpext``' instruction takes a :ref:`floating point <t_floating>`
5478 ``value`` to cast, and a :ref:`floating point <t_floating>` type to cast it
5479 to. The source type must be smaller than the destination type.
5484 The '``fpext``' instruction extends the ``value`` from a smaller
5485 :ref:`floating point <t_floating>` type to a larger :ref:`floating
5486 point <t_floating>` type. The ``fpext`` cannot be used to make a
5487 *no-op cast* because it always changes bits. Use ``bitcast`` to make a
5488 *no-op cast* for a floating point cast.
5493 .. code-block:: llvm
5495 %X = fpext float 3.125 to double ; yields double:3.125000e+00
5496 %Y = fpext double %X to fp128 ; yields fp128:0xL00000000000000004000900000000000
5498 '``fptoui .. to``' Instruction
5499 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5506 <result> = fptoui <ty> <value> to <ty2> ; yields ty2
5511 The '``fptoui``' converts a floating point ``value`` to its unsigned
5512 integer equivalent of type ``ty2``.
5517 The '``fptoui``' instruction takes a value to cast, which must be a
5518 scalar or vector :ref:`floating point <t_floating>` value, and a type to
5519 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
5520 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
5521 type with the same number of elements as ``ty``
5526 The '``fptoui``' instruction converts its :ref:`floating
5527 point <t_floating>` operand into the nearest (rounding towards zero)
5528 unsigned integer value. If the value cannot fit in ``ty2``, the results
5534 .. code-block:: llvm
5536 %X = fptoui double 123.0 to i32 ; yields i32:123
5537 %Y = fptoui float 1.0E+300 to i1 ; yields undefined:1
5538 %Z = fptoui float 1.04E+17 to i8 ; yields undefined:1
5540 '``fptosi .. to``' Instruction
5541 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5548 <result> = fptosi <ty> <value> to <ty2> ; yields ty2
5553 The '``fptosi``' instruction converts :ref:`floating point <t_floating>`
5554 ``value`` to type ``ty2``.
5559 The '``fptosi``' instruction takes a value to cast, which must be a
5560 scalar or vector :ref:`floating point <t_floating>` value, and a type to
5561 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
5562 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
5563 type with the same number of elements as ``ty``
5568 The '``fptosi``' instruction converts its :ref:`floating
5569 point <t_floating>` operand into the nearest (rounding towards zero)
5570 signed integer value. If the value cannot fit in ``ty2``, the results
5576 .. code-block:: llvm
5578 %X = fptosi double -123.0 to i32 ; yields i32:-123
5579 %Y = fptosi float 1.0E-247 to i1 ; yields undefined:1
5580 %Z = fptosi float 1.04E+17 to i8 ; yields undefined:1
5582 '``uitofp .. to``' Instruction
5583 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5590 <result> = uitofp <ty> <value> to <ty2> ; yields ty2
5595 The '``uitofp``' instruction regards ``value`` as an unsigned integer
5596 and converts that value to the ``ty2`` type.
5601 The '``uitofp``' instruction takes a value to cast, which must be a
5602 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
5603 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
5604 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
5605 type with the same number of elements as ``ty``
5610 The '``uitofp``' instruction interprets its operand as an unsigned
5611 integer quantity and converts it to the corresponding floating point
5612 value. If the value cannot fit in the floating point value, the results
5618 .. code-block:: llvm
5620 %X = uitofp i32 257 to float ; yields float:257.0
5621 %Y = uitofp i8 -1 to double ; yields double:255.0
5623 '``sitofp .. to``' Instruction
5624 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5631 <result> = sitofp <ty> <value> to <ty2> ; yields ty2
5636 The '``sitofp``' instruction regards ``value`` as a signed integer and
5637 converts that value to the ``ty2`` type.
5642 The '``sitofp``' instruction takes a value to cast, which must be a
5643 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
5644 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
5645 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
5646 type with the same number of elements as ``ty``
5651 The '``sitofp``' instruction interprets its operand as a signed integer
5652 quantity and converts it to the corresponding floating point value. If
5653 the value cannot fit in the floating point value, the results are
5659 .. code-block:: llvm
5661 %X = sitofp i32 257 to float ; yields float:257.0
5662 %Y = sitofp i8 -1 to double ; yields double:-1.0
5666 '``ptrtoint .. to``' Instruction
5667 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5674 <result> = ptrtoint <ty> <value> to <ty2> ; yields ty2
5679 The '``ptrtoint``' instruction converts the pointer or a vector of
5680 pointers ``value`` to the integer (or vector of integers) type ``ty2``.
5685 The '``ptrtoint``' instruction takes a ``value`` to cast, which must be
5686 a a value of type :ref:`pointer <t_pointer>` or a vector of pointers, and a
5687 type to cast it to ``ty2``, which must be an :ref:`integer <t_integer>` or
5688 a vector of integers type.
5693 The '``ptrtoint``' instruction converts ``value`` to integer type
5694 ``ty2`` by interpreting the pointer value as an integer and either
5695 truncating or zero extending that value to the size of the integer type.
5696 If ``value`` is smaller than ``ty2`` then a zero extension is done. If
5697 ``value`` is larger than ``ty2`` then a truncation is done. If they are
5698 the same size, then nothing is done (*no-op cast*) other than a type
5704 .. code-block:: llvm
5706 %X = ptrtoint i32* %P to i8 ; yields truncation on 32-bit architecture
5707 %Y = ptrtoint i32* %P to i64 ; yields zero extension on 32-bit architecture
5708 %Z = ptrtoint <4 x i32*> %P to <4 x i64>; yields vector zero extension for a vector of addresses on 32-bit architecture
5712 '``inttoptr .. to``' Instruction
5713 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5720 <result> = inttoptr <ty> <value> to <ty2> ; yields ty2
5725 The '``inttoptr``' instruction converts an integer ``value`` to a
5726 pointer type, ``ty2``.
5731 The '``inttoptr``' instruction takes an :ref:`integer <t_integer>` value to
5732 cast, and a type to cast it to, which must be a :ref:`pointer <t_pointer>`
5738 The '``inttoptr``' instruction converts ``value`` to type ``ty2`` by
5739 applying either a zero extension or a truncation depending on the size
5740 of the integer ``value``. If ``value`` is larger than the size of a
5741 pointer then a truncation is done. If ``value`` is smaller than the size
5742 of a pointer then a zero extension is done. If they are the same size,
5743 nothing is done (*no-op cast*).
5748 .. code-block:: llvm
5750 %X = inttoptr i32 255 to i32* ; yields zero extension on 64-bit architecture
5751 %Y = inttoptr i32 255 to i32* ; yields no-op on 32-bit architecture
5752 %Z = inttoptr i64 0 to i32* ; yields truncation on 32-bit architecture
5753 %Z = inttoptr <4 x i32> %G to <4 x i8*>; yields truncation of vector G to four pointers
5757 '``bitcast .. to``' Instruction
5758 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5765 <result> = bitcast <ty> <value> to <ty2> ; yields ty2
5770 The '``bitcast``' instruction converts ``value`` to type ``ty2`` without
5776 The '``bitcast``' instruction takes a value to cast, which must be a
5777 non-aggregate first class value, and a type to cast it to, which must
5778 also be a non-aggregate :ref:`first class <t_firstclass>` type. The
5779 bit sizes of ``value`` and the destination type, ``ty2``, must be
5780 identical. If the source type is a pointer, the destination type must
5781 also be a pointer of the same size. This instruction supports bitwise
5782 conversion of vectors to integers and to vectors of other types (as
5783 long as they have the same size).
5788 The '``bitcast``' instruction converts ``value`` to type ``ty2``. It
5789 is always a *no-op cast* because no bits change with this
5790 conversion. The conversion is done as if the ``value`` had been stored
5791 to memory and read back as type ``ty2``. Pointer (or vector of
5792 pointers) types may only be converted to other pointer (or vector of
5793 pointers) types with the same address space through this instruction.
5794 To convert pointers to other types, use the :ref:`inttoptr <i_inttoptr>`
5795 or :ref:`ptrtoint <i_ptrtoint>` instructions first.
5800 .. code-block:: llvm
5802 %X = bitcast i8 255 to i8 ; yields i8 :-1
5803 %Y = bitcast i32* %x to sint* ; yields sint*:%x
5804 %Z = bitcast <2 x int> %V to i64; ; yields i64: %V
5805 %Z = bitcast <2 x i32*> %V to <2 x i64*> ; yields <2 x i64*>
5807 .. _i_addrspacecast:
5809 '``addrspacecast .. to``' Instruction
5810 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5817 <result> = addrspacecast <pty> <ptrval> to <pty2> ; yields pty2
5822 The '``addrspacecast``' instruction converts ``ptrval`` from ``pty`` in
5823 address space ``n`` to type ``pty2`` in address space ``m``.
5828 The '``addrspacecast``' instruction takes a pointer or vector of pointer value
5829 to cast and a pointer type to cast it to, which must have a different
5835 The '``addrspacecast``' instruction converts the pointer value
5836 ``ptrval`` to type ``pty2``. It can be a *no-op cast* or a complex
5837 value modification, depending on the target and the address space
5838 pair. Pointer conversions within the same address space must be
5839 performed with the ``bitcast`` instruction. Note that if the address space
5840 conversion is legal then both result and operand refer to the same memory
5846 .. code-block:: llvm
5848 %X = addrspacecast i32* %x to i32 addrspace(1)* ; yields i32 addrspace(1)*:%x
5849 %Y = addrspacecast i32 addrspace(1)* %y to i64 addrspace(2)* ; yields i64 addrspace(2)*:%y
5850 %Z = addrspacecast <4 x i32*> %z to <4 x float addrspace(3)*> ; yields <4 x float addrspace(3)*>:%z
5857 The instructions in this category are the "miscellaneous" instructions,
5858 which defy better classification.
5862 '``icmp``' Instruction
5863 ^^^^^^^^^^^^^^^^^^^^^^
5870 <result> = icmp <cond> <ty> <op1>, <op2> ; yields {i1} or {<N x i1>}:result
5875 The '``icmp``' instruction returns a boolean value or a vector of
5876 boolean values based on comparison of its two integer, integer vector,
5877 pointer, or pointer vector operands.
5882 The '``icmp``' instruction takes three operands. The first operand is
5883 the condition code indicating the kind of comparison to perform. It is
5884 not a value, just a keyword. The possible condition code are:
5887 #. ``ne``: not equal
5888 #. ``ugt``: unsigned greater than
5889 #. ``uge``: unsigned greater or equal
5890 #. ``ult``: unsigned less than
5891 #. ``ule``: unsigned less or equal
5892 #. ``sgt``: signed greater than
5893 #. ``sge``: signed greater or equal
5894 #. ``slt``: signed less than
5895 #. ``sle``: signed less or equal
5897 The remaining two arguments must be :ref:`integer <t_integer>` or
5898 :ref:`pointer <t_pointer>` or integer :ref:`vector <t_vector>` typed. They
5899 must also be identical types.
5904 The '``icmp``' compares ``op1`` and ``op2`` according to the condition
5905 code given as ``cond``. The comparison performed always yields either an
5906 :ref:`i1 <t_integer>` or vector of ``i1`` result, as follows:
5908 #. ``eq``: yields ``true`` if the operands are equal, ``false``
5909 otherwise. No sign interpretation is necessary or performed.
5910 #. ``ne``: yields ``true`` if the operands are unequal, ``false``
5911 otherwise. No sign interpretation is necessary or performed.
5912 #. ``ugt``: interprets the operands as unsigned values and yields
5913 ``true`` if ``op1`` is greater than ``op2``.
5914 #. ``uge``: interprets the operands as unsigned values and yields
5915 ``true`` if ``op1`` is greater than or equal to ``op2``.
5916 #. ``ult``: interprets the operands as unsigned values and yields
5917 ``true`` if ``op1`` is less than ``op2``.
5918 #. ``ule``: interprets the operands as unsigned values and yields
5919 ``true`` if ``op1`` is less than or equal to ``op2``.
5920 #. ``sgt``: interprets the operands as signed values and yields ``true``
5921 if ``op1`` is greater than ``op2``.
5922 #. ``sge``: interprets the operands as signed values and yields ``true``
5923 if ``op1`` is greater than or equal to ``op2``.
5924 #. ``slt``: interprets the operands as signed values and yields ``true``
5925 if ``op1`` is less than ``op2``.
5926 #. ``sle``: interprets the operands as signed values and yields ``true``
5927 if ``op1`` is less than or equal to ``op2``.
5929 If the operands are :ref:`pointer <t_pointer>` typed, the pointer values
5930 are compared as if they were integers.
5932 If the operands are integer vectors, then they are compared element by
5933 element. The result is an ``i1`` vector with the same number of elements
5934 as the values being compared. Otherwise, the result is an ``i1``.
5939 .. code-block:: llvm
5941 <result> = icmp eq i32 4, 5 ; yields: result=false
5942 <result> = icmp ne float* %X, %X ; yields: result=false
5943 <result> = icmp ult i16 4, 5 ; yields: result=true
5944 <result> = icmp sgt i16 4, 5 ; yields: result=false
5945 <result> = icmp ule i16 -4, 5 ; yields: result=false
5946 <result> = icmp sge i16 4, 5 ; yields: result=false
5948 Note that the code generator does not yet support vector types with the
5949 ``icmp`` instruction.
5953 '``fcmp``' Instruction
5954 ^^^^^^^^^^^^^^^^^^^^^^
5961 <result> = fcmp <cond> <ty> <op1>, <op2> ; yields {i1} or {<N x i1>}:result
5966 The '``fcmp``' instruction returns a boolean value or vector of boolean
5967 values based on comparison of its operands.
5969 If the operands are floating point scalars, then the result type is a
5970 boolean (:ref:`i1 <t_integer>`).
5972 If the operands are floating point vectors, then the result type is a
5973 vector of boolean with the same number of elements as the operands being
5979 The '``fcmp``' instruction takes three operands. The first operand is
5980 the condition code indicating the kind of comparison to perform. It is
5981 not a value, just a keyword. The possible condition code are:
5983 #. ``false``: no comparison, always returns false
5984 #. ``oeq``: ordered and equal
5985 #. ``ogt``: ordered and greater than
5986 #. ``oge``: ordered and greater than or equal
5987 #. ``olt``: ordered and less than
5988 #. ``ole``: ordered and less than or equal
5989 #. ``one``: ordered and not equal
5990 #. ``ord``: ordered (no nans)
5991 #. ``ueq``: unordered or equal
5992 #. ``ugt``: unordered or greater than
5993 #. ``uge``: unordered or greater than or equal
5994 #. ``ult``: unordered or less than
5995 #. ``ule``: unordered or less than or equal
5996 #. ``une``: unordered or not equal
5997 #. ``uno``: unordered (either nans)
5998 #. ``true``: no comparison, always returns true
6000 *Ordered* means that neither operand is a QNAN while *unordered* means
6001 that either operand may be a QNAN.
6003 Each of ``val1`` and ``val2`` arguments must be either a :ref:`floating
6004 point <t_floating>` type or a :ref:`vector <t_vector>` of floating point
6005 type. They must have identical types.
6010 The '``fcmp``' instruction compares ``op1`` and ``op2`` according to the
6011 condition code given as ``cond``. If the operands are vectors, then the
6012 vectors are compared element by element. Each comparison performed
6013 always yields an :ref:`i1 <t_integer>` result, as follows:
6015 #. ``false``: always yields ``false``, regardless of operands.
6016 #. ``oeq``: yields ``true`` if both operands are not a QNAN and ``op1``
6017 is equal to ``op2``.
6018 #. ``ogt``: yields ``true`` if both operands are not a QNAN and ``op1``
6019 is greater than ``op2``.
6020 #. ``oge``: yields ``true`` if both operands are not a QNAN and ``op1``
6021 is greater than or equal to ``op2``.
6022 #. ``olt``: yields ``true`` if both operands are not a QNAN and ``op1``
6023 is less than ``op2``.
6024 #. ``ole``: yields ``true`` if both operands are not a QNAN and ``op1``
6025 is less than or equal to ``op2``.
6026 #. ``one``: yields ``true`` if both operands are not a QNAN and ``op1``
6027 is not equal to ``op2``.
6028 #. ``ord``: yields ``true`` if both operands are not a QNAN.
6029 #. ``ueq``: yields ``true`` if either operand is a QNAN or ``op1`` is
6031 #. ``ugt``: yields ``true`` if either operand is a QNAN or ``op1`` is
6032 greater than ``op2``.
6033 #. ``uge``: yields ``true`` if either operand is a QNAN or ``op1`` is
6034 greater than or equal to ``op2``.
6035 #. ``ult``: yields ``true`` if either operand is a QNAN or ``op1`` is
6037 #. ``ule``: yields ``true`` if either operand is a QNAN or ``op1`` is
6038 less than or equal to ``op2``.
6039 #. ``une``: yields ``true`` if either operand is a QNAN or ``op1`` is
6040 not equal to ``op2``.
6041 #. ``uno``: yields ``true`` if either operand is a QNAN.
6042 #. ``true``: always yields ``true``, regardless of operands.
6047 .. code-block:: llvm
6049 <result> = fcmp oeq float 4.0, 5.0 ; yields: result=false
6050 <result> = fcmp one float 4.0, 5.0 ; yields: result=true
6051 <result> = fcmp olt float 4.0, 5.0 ; yields: result=true
6052 <result> = fcmp ueq double 1.0, 2.0 ; yields: result=false
6054 Note that the code generator does not yet support vector types with the
6055 ``fcmp`` instruction.
6059 '``phi``' Instruction
6060 ^^^^^^^^^^^^^^^^^^^^^
6067 <result> = phi <ty> [ <val0>, <label0>], ...
6072 The '``phi``' instruction is used to implement the φ node in the SSA
6073 graph representing the function.
6078 The type of the incoming values is specified with the first type field.
6079 After this, the '``phi``' instruction takes a list of pairs as
6080 arguments, with one pair for each predecessor basic block of the current
6081 block. Only values of :ref:`first class <t_firstclass>` type may be used as
6082 the value arguments to the PHI node. Only labels may be used as the
6085 There must be no non-phi instructions between the start of a basic block
6086 and the PHI instructions: i.e. PHI instructions must be first in a basic
6089 For the purposes of the SSA form, the use of each incoming value is
6090 deemed to occur on the edge from the corresponding predecessor block to
6091 the current block (but after any definition of an '``invoke``'
6092 instruction's return value on the same edge).
6097 At runtime, the '``phi``' instruction logically takes on the value
6098 specified by the pair corresponding to the predecessor basic block that
6099 executed just prior to the current block.
6104 .. code-block:: llvm
6106 Loop: ; Infinite loop that counts from 0 on up...
6107 %indvar = phi i32 [ 0, %LoopHeader ], [ %nextindvar, %Loop ]
6108 %nextindvar = add i32 %indvar, 1
6113 '``select``' Instruction
6114 ^^^^^^^^^^^^^^^^^^^^^^^^
6121 <result> = select selty <cond>, <ty> <val1>, <ty> <val2> ; yields ty
6123 selty is either i1 or {<N x i1>}
6128 The '``select``' instruction is used to choose one value based on a
6129 condition, without IR-level branching.
6134 The '``select``' instruction requires an 'i1' value or a vector of 'i1'
6135 values indicating the condition, and two values of the same :ref:`first
6136 class <t_firstclass>` type. If the val1/val2 are vectors and the
6137 condition is a scalar, then entire vectors are selected, not individual
6143 If the condition is an i1 and it evaluates to 1, the instruction returns
6144 the first value argument; otherwise, it returns the second value
6147 If the condition is a vector of i1, then the value arguments must be
6148 vectors of the same size, and the selection is done element by element.
6153 .. code-block:: llvm
6155 %X = select i1 true, i8 17, i8 42 ; yields i8:17
6159 '``call``' Instruction
6160 ^^^^^^^^^^^^^^^^^^^^^^
6167 <result> = [tail | musttail] call [cconv] [ret attrs] <ty> [<fnty>*] <fnptrval>(<function args>) [fn attrs]
6172 The '``call``' instruction represents a simple function call.
6177 This instruction requires several arguments:
6179 #. The optional ``tail`` and ``musttail`` markers indicate that the optimizers
6180 should perform tail call optimization. The ``tail`` marker is a hint that
6181 `can be ignored <CodeGenerator.html#sibcallopt>`_. The ``musttail`` marker
6182 means that the call must be tail call optimized in order for the program to
6183 be correct. The ``musttail`` marker provides these guarantees:
6185 #. The call will not cause unbounded stack growth if it is part of a
6186 recursive cycle in the call graph.
6187 #. Arguments with the :ref:`inalloca <attr_inalloca>` attribute are
6190 Both markers imply that the callee does not access allocas or varargs from
6191 the caller. Calls marked ``musttail`` must obey the following additional
6194 - The call must immediately precede a :ref:`ret <i_ret>` instruction,
6195 or a pointer bitcast followed by a ret instruction.
6196 - The ret instruction must return the (possibly bitcasted) value
6197 produced by the call or void.
6198 - The caller and callee prototypes must match. Pointer types of
6199 parameters or return types may differ in pointee type, but not
6201 - The calling conventions of the caller and callee must match.
6202 - All ABI-impacting function attributes, such as sret, byval, inreg,
6203 returned, and inalloca, must match.
6205 Tail call optimization for calls marked ``tail`` is guaranteed to occur if
6206 the following conditions are met:
6208 - Caller and callee both have the calling convention ``fastcc``.
6209 - The call is in tail position (ret immediately follows call and ret
6210 uses value of call or is void).
6211 - Option ``-tailcallopt`` is enabled, or
6212 ``llvm::GuaranteedTailCallOpt`` is ``true``.
6213 - `Platform specific constraints are
6214 met. <CodeGenerator.html#tailcallopt>`_
6216 #. The optional "cconv" marker indicates which :ref:`calling
6217 convention <callingconv>` the call should use. If none is
6218 specified, the call defaults to using C calling conventions. The
6219 calling convention of the call must match the calling convention of
6220 the target function, or else the behavior is undefined.
6221 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
6222 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
6224 #. '``ty``': the type of the call instruction itself which is also the
6225 type of the return value. Functions that return no value are marked
6227 #. '``fnty``': shall be the signature of the pointer to function value
6228 being invoked. The argument types must match the types implied by
6229 this signature. This type can be omitted if the function is not
6230 varargs and if the function type does not return a pointer to a
6232 #. '``fnptrval``': An LLVM value containing a pointer to a function to
6233 be invoked. In most cases, this is a direct function invocation, but
6234 indirect ``call``'s are just as possible, calling an arbitrary pointer
6236 #. '``function args``': argument list whose types match the function
6237 signature argument types and parameter attributes. All arguments must
6238 be of :ref:`first class <t_firstclass>` type. If the function signature
6239 indicates the function accepts a variable number of arguments, the
6240 extra arguments can be specified.
6241 #. The optional :ref:`function attributes <fnattrs>` list. Only
6242 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
6243 attributes are valid here.
6248 The '``call``' instruction is used to cause control flow to transfer to
6249 a specified function, with its incoming arguments bound to the specified
6250 values. Upon a '``ret``' instruction in the called function, control
6251 flow continues with the instruction after the function call, and the
6252 return value of the function is bound to the result argument.
6257 .. code-block:: llvm
6259 %retval = call i32 @test(i32 %argc)
6260 call i32 (i8*, ...)* @printf(i8* %msg, i32 12, i8 42) ; yields i32
6261 %X = tail call i32 @foo() ; yields i32
6262 %Y = tail call fastcc i32 @foo() ; yields i32
6263 call void %foo(i8 97 signext)
6265 %struct.A = type { i32, i8 }
6266 %r = call %struct.A @foo() ; yields { 32, i8 }
6267 %gr = extractvalue %struct.A %r, 0 ; yields i32
6268 %gr1 = extractvalue %struct.A %r, 1 ; yields i8
6269 %Z = call void @foo() noreturn ; indicates that %foo never returns normally
6270 %ZZ = call zeroext i32 @bar() ; Return value is %zero extended
6272 llvm treats calls to some functions with names and arguments that match
6273 the standard C99 library as being the C99 library functions, and may
6274 perform optimizations or generate code for them under that assumption.
6275 This is something we'd like to change in the future to provide better
6276 support for freestanding environments and non-C-based languages.
6280 '``va_arg``' Instruction
6281 ^^^^^^^^^^^^^^^^^^^^^^^^
6288 <resultval> = va_arg <va_list*> <arglist>, <argty>
6293 The '``va_arg``' instruction is used to access arguments passed through
6294 the "variable argument" area of a function call. It is used to implement
6295 the ``va_arg`` macro in C.
6300 This instruction takes a ``va_list*`` value and the type of the
6301 argument. It returns a value of the specified argument type and
6302 increments the ``va_list`` to point to the next argument. The actual
6303 type of ``va_list`` is target specific.
6308 The '``va_arg``' instruction loads an argument of the specified type
6309 from the specified ``va_list`` and causes the ``va_list`` to point to
6310 the next argument. For more information, see the variable argument
6311 handling :ref:`Intrinsic Functions <int_varargs>`.
6313 It is legal for this instruction to be called in a function which does
6314 not take a variable number of arguments, for example, the ``vfprintf``
6317 ``va_arg`` is an LLVM instruction instead of an :ref:`intrinsic
6318 function <intrinsics>` because it takes a type as an argument.
6323 See the :ref:`variable argument processing <int_varargs>` section.
6325 Note that the code generator does not yet fully support va\_arg on many
6326 targets. Also, it does not currently support va\_arg with aggregate
6327 types on any target.
6331 '``landingpad``' Instruction
6332 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6339 <resultval> = landingpad <resultty> personality <type> <pers_fn> <clause>+
6340 <resultval> = landingpad <resultty> personality <type> <pers_fn> cleanup <clause>*
6342 <clause> := catch <type> <value>
6343 <clause> := filter <array constant type> <array constant>
6348 The '``landingpad``' instruction is used by `LLVM's exception handling
6349 system <ExceptionHandling.html#overview>`_ to specify that a basic block
6350 is a landing pad --- one where the exception lands, and corresponds to the
6351 code found in the ``catch`` portion of a ``try``/``catch`` sequence. It
6352 defines values supplied by the personality function (``pers_fn``) upon
6353 re-entry to the function. The ``resultval`` has the type ``resultty``.
6358 This instruction takes a ``pers_fn`` value. This is the personality
6359 function associated with the unwinding mechanism. The optional
6360 ``cleanup`` flag indicates that the landing pad block is a cleanup.
6362 A ``clause`` begins with the clause type --- ``catch`` or ``filter`` --- and
6363 contains the global variable representing the "type" that may be caught
6364 or filtered respectively. Unlike the ``catch`` clause, the ``filter``
6365 clause takes an array constant as its argument. Use
6366 "``[0 x i8**] undef``" for a filter which cannot throw. The
6367 '``landingpad``' instruction must contain *at least* one ``clause`` or
6368 the ``cleanup`` flag.
6373 The '``landingpad``' instruction defines the values which are set by the
6374 personality function (``pers_fn``) upon re-entry to the function, and
6375 therefore the "result type" of the ``landingpad`` instruction. As with
6376 calling conventions, how the personality function results are
6377 represented in LLVM IR is target specific.
6379 The clauses are applied in order from top to bottom. If two
6380 ``landingpad`` instructions are merged together through inlining, the
6381 clauses from the calling function are appended to the list of clauses.
6382 When the call stack is being unwound due to an exception being thrown,
6383 the exception is compared against each ``clause`` in turn. If it doesn't
6384 match any of the clauses, and the ``cleanup`` flag is not set, then
6385 unwinding continues further up the call stack.
6387 The ``landingpad`` instruction has several restrictions:
6389 - A landing pad block is a basic block which is the unwind destination
6390 of an '``invoke``' instruction.
6391 - A landing pad block must have a '``landingpad``' instruction as its
6392 first non-PHI instruction.
6393 - There can be only one '``landingpad``' instruction within the landing
6395 - A basic block that is not a landing pad block may not include a
6396 '``landingpad``' instruction.
6397 - All '``landingpad``' instructions in a function must have the same
6398 personality function.
6403 .. code-block:: llvm
6405 ;; A landing pad which can catch an integer.
6406 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6408 ;; A landing pad that is a cleanup.
6409 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6411 ;; A landing pad which can catch an integer and can only throw a double.
6412 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6414 filter [1 x i8**] [@_ZTId]
6421 LLVM supports the notion of an "intrinsic function". These functions
6422 have well known names and semantics and are required to follow certain
6423 restrictions. Overall, these intrinsics represent an extension mechanism
6424 for the LLVM language that does not require changing all of the
6425 transformations in LLVM when adding to the language (or the bitcode
6426 reader/writer, the parser, etc...).
6428 Intrinsic function names must all start with an "``llvm.``" prefix. This
6429 prefix is reserved in LLVM for intrinsic names; thus, function names may
6430 not begin with this prefix. Intrinsic functions must always be external
6431 functions: you cannot define the body of intrinsic functions. Intrinsic
6432 functions may only be used in call or invoke instructions: it is illegal
6433 to take the address of an intrinsic function. Additionally, because
6434 intrinsic functions are part of the LLVM language, it is required if any
6435 are added that they be documented here.
6437 Some intrinsic functions can be overloaded, i.e., the intrinsic
6438 represents a family of functions that perform the same operation but on
6439 different data types. Because LLVM can represent over 8 million
6440 different integer types, overloading is used commonly to allow an
6441 intrinsic function to operate on any integer type. One or more of the
6442 argument types or the result type can be overloaded to accept any
6443 integer type. Argument types may also be defined as exactly matching a
6444 previous argument's type or the result type. This allows an intrinsic
6445 function which accepts multiple arguments, but needs all of them to be
6446 of the same type, to only be overloaded with respect to a single
6447 argument or the result.
6449 Overloaded intrinsics will have the names of its overloaded argument
6450 types encoded into its function name, each preceded by a period. Only
6451 those types which are overloaded result in a name suffix. Arguments
6452 whose type is matched against another type do not. For example, the
6453 ``llvm.ctpop`` function can take an integer of any width and returns an
6454 integer of exactly the same integer width. This leads to a family of
6455 functions such as ``i8 @llvm.ctpop.i8(i8 %val)`` and
6456 ``i29 @llvm.ctpop.i29(i29 %val)``. Only one type, the return type, is
6457 overloaded, and only one type suffix is required. Because the argument's
6458 type is matched against the return type, it does not require its own
6461 To learn how to add an intrinsic function, please see the `Extending
6462 LLVM Guide <ExtendingLLVM.html>`_.
6466 Variable Argument Handling Intrinsics
6467 -------------------------------------
6469 Variable argument support is defined in LLVM with the
6470 :ref:`va_arg <i_va_arg>` instruction and these three intrinsic
6471 functions. These functions are related to the similarly named macros
6472 defined in the ``<stdarg.h>`` header file.
6474 All of these functions operate on arguments that use a target-specific
6475 value type "``va_list``". The LLVM assembly language reference manual
6476 does not define what this type is, so all transformations should be
6477 prepared to handle these functions regardless of the type used.
6479 This example shows how the :ref:`va_arg <i_va_arg>` instruction and the
6480 variable argument handling intrinsic functions are used.
6482 .. code-block:: llvm
6484 define i32 @test(i32 %X, ...) {
6485 ; Initialize variable argument processing
6487 %ap2 = bitcast i8** %ap to i8*
6488 call void @llvm.va_start(i8* %ap2)
6490 ; Read a single integer argument
6491 %tmp = va_arg i8** %ap, i32
6493 ; Demonstrate usage of llvm.va_copy and llvm.va_end
6495 %aq2 = bitcast i8** %aq to i8*
6496 call void @llvm.va_copy(i8* %aq2, i8* %ap2)
6497 call void @llvm.va_end(i8* %aq2)
6499 ; Stop processing of arguments.
6500 call void @llvm.va_end(i8* %ap2)
6504 declare void @llvm.va_start(i8*)
6505 declare void @llvm.va_copy(i8*, i8*)
6506 declare void @llvm.va_end(i8*)
6510 '``llvm.va_start``' Intrinsic
6511 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6518 declare void @llvm.va_start(i8* <arglist>)
6523 The '``llvm.va_start``' intrinsic initializes ``*<arglist>`` for
6524 subsequent use by ``va_arg``.
6529 The argument is a pointer to a ``va_list`` element to initialize.
6534 The '``llvm.va_start``' intrinsic works just like the ``va_start`` macro
6535 available in C. In a target-dependent way, it initializes the
6536 ``va_list`` element to which the argument points, so that the next call
6537 to ``va_arg`` will produce the first variable argument passed to the
6538 function. Unlike the C ``va_start`` macro, this intrinsic does not need
6539 to know the last argument of the function as the compiler can figure
6542 '``llvm.va_end``' Intrinsic
6543 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6550 declare void @llvm.va_end(i8* <arglist>)
6555 The '``llvm.va_end``' intrinsic destroys ``*<arglist>``, which has been
6556 initialized previously with ``llvm.va_start`` or ``llvm.va_copy``.
6561 The argument is a pointer to a ``va_list`` to destroy.
6566 The '``llvm.va_end``' intrinsic works just like the ``va_end`` macro
6567 available in C. In a target-dependent way, it destroys the ``va_list``
6568 element to which the argument points. Calls to
6569 :ref:`llvm.va_start <int_va_start>` and
6570 :ref:`llvm.va_copy <int_va_copy>` must be matched exactly with calls to
6575 '``llvm.va_copy``' Intrinsic
6576 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6583 declare void @llvm.va_copy(i8* <destarglist>, i8* <srcarglist>)
6588 The '``llvm.va_copy``' intrinsic copies the current argument position
6589 from the source argument list to the destination argument list.
6594 The first argument is a pointer to a ``va_list`` element to initialize.
6595 The second argument is a pointer to a ``va_list`` element to copy from.
6600 The '``llvm.va_copy``' intrinsic works just like the ``va_copy`` macro
6601 available in C. In a target-dependent way, it copies the source
6602 ``va_list`` element into the destination ``va_list`` element. This
6603 intrinsic is necessary because the `` llvm.va_start`` intrinsic may be
6604 arbitrarily complex and require, for example, memory allocation.
6606 Accurate Garbage Collection Intrinsics
6607 --------------------------------------
6609 LLVM support for `Accurate Garbage Collection <GarbageCollection.html>`_
6610 (GC) requires the implementation and generation of these intrinsics.
6611 These intrinsics allow identification of :ref:`GC roots on the
6612 stack <int_gcroot>`, as well as garbage collector implementations that
6613 require :ref:`read <int_gcread>` and :ref:`write <int_gcwrite>` barriers.
6614 Front-ends for type-safe garbage collected languages should generate
6615 these intrinsics to make use of the LLVM garbage collectors. For more
6616 details, see `Accurate Garbage Collection with
6617 LLVM <GarbageCollection.html>`_.
6619 The garbage collection intrinsics only operate on objects in the generic
6620 address space (address space zero).
6624 '``llvm.gcroot``' Intrinsic
6625 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6632 declare void @llvm.gcroot(i8** %ptrloc, i8* %metadata)
6637 The '``llvm.gcroot``' intrinsic declares the existence of a GC root to
6638 the code generator, and allows some metadata to be associated with it.
6643 The first argument specifies the address of a stack object that contains
6644 the root pointer. The second pointer (which must be either a constant or
6645 a global value address) contains the meta-data to be associated with the
6651 At runtime, a call to this intrinsic stores a null pointer into the
6652 "ptrloc" location. At compile-time, the code generator generates
6653 information to allow the runtime to find the pointer at GC safe points.
6654 The '``llvm.gcroot``' intrinsic may only be used in a function which
6655 :ref:`specifies a GC algorithm <gc>`.
6659 '``llvm.gcread``' Intrinsic
6660 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6667 declare i8* @llvm.gcread(i8* %ObjPtr, i8** %Ptr)
6672 The '``llvm.gcread``' intrinsic identifies reads of references from heap
6673 locations, allowing garbage collector implementations that require read
6679 The second argument is the address to read from, which should be an
6680 address allocated from the garbage collector. The first object is a
6681 pointer to the start of the referenced object, if needed by the language
6682 runtime (otherwise null).
6687 The '``llvm.gcread``' intrinsic has the same semantics as a load
6688 instruction, but may be replaced with substantially more complex code by
6689 the garbage collector runtime, as needed. The '``llvm.gcread``'
6690 intrinsic may only be used in a function which :ref:`specifies a GC
6695 '``llvm.gcwrite``' Intrinsic
6696 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6703 declare void @llvm.gcwrite(i8* %P1, i8* %Obj, i8** %P2)
6708 The '``llvm.gcwrite``' intrinsic identifies writes of references to heap
6709 locations, allowing garbage collector implementations that require write
6710 barriers (such as generational or reference counting collectors).
6715 The first argument is the reference to store, the second is the start of
6716 the object to store it to, and the third is the address of the field of
6717 Obj to store to. If the runtime does not require a pointer to the
6718 object, Obj may be null.
6723 The '``llvm.gcwrite``' intrinsic has the same semantics as a store
6724 instruction, but may be replaced with substantially more complex code by
6725 the garbage collector runtime, as needed. The '``llvm.gcwrite``'
6726 intrinsic may only be used in a function which :ref:`specifies a GC
6729 Code Generator Intrinsics
6730 -------------------------
6732 These intrinsics are provided by LLVM to expose special features that
6733 may only be implemented with code generator support.
6735 '``llvm.returnaddress``' Intrinsic
6736 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6743 declare i8 *@llvm.returnaddress(i32 <level>)
6748 The '``llvm.returnaddress``' intrinsic attempts to compute a
6749 target-specific value indicating the return address of the current
6750 function or one of its callers.
6755 The argument to this intrinsic indicates which function to return the
6756 address for. Zero indicates the calling function, one indicates its
6757 caller, etc. The argument is **required** to be a constant integer
6763 The '``llvm.returnaddress``' intrinsic either returns a pointer
6764 indicating the return address of the specified call frame, or zero if it
6765 cannot be identified. The value returned by this intrinsic is likely to
6766 be incorrect or 0 for arguments other than zero, so it should only be
6767 used for debugging purposes.
6769 Note that calling this intrinsic does not prevent function inlining or
6770 other aggressive transformations, so the value returned may not be that
6771 of the obvious source-language caller.
6773 '``llvm.frameaddress``' Intrinsic
6774 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6781 declare i8* @llvm.frameaddress(i32 <level>)
6786 The '``llvm.frameaddress``' intrinsic attempts to return the
6787 target-specific frame pointer value for the specified stack frame.
6792 The argument to this intrinsic indicates which function to return the
6793 frame pointer for. Zero indicates the calling function, one indicates
6794 its caller, etc. The argument is **required** to be a constant integer
6800 The '``llvm.frameaddress``' intrinsic either returns a pointer
6801 indicating the frame address of the specified call frame, or zero if it
6802 cannot be identified. The value returned by this intrinsic is likely to
6803 be incorrect or 0 for arguments other than zero, so it should only be
6804 used for debugging purposes.
6806 Note that calling this intrinsic does not prevent function inlining or
6807 other aggressive transformations, so the value returned may not be that
6808 of the obvious source-language caller.
6810 .. _int_read_register:
6811 .. _int_write_register:
6813 '``llvm.read_register``' and '``llvm.write_register``' Intrinsics
6814 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6821 declare i32 @llvm.read_register.i32(metadata)
6822 declare i64 @llvm.read_register.i64(metadata)
6823 declare void @llvm.write_register.i32(metadata, i32 @value)
6824 declare void @llvm.write_register.i64(metadata, i64 @value)
6825 !0 = metadata !{metadata !"sp\00"}
6830 The '``llvm.read_register``' and '``llvm.write_register``' intrinsics
6831 provides access to the named register. The register must be valid on
6832 the architecture being compiled to. The type needs to be compatible
6833 with the register being read.
6838 The '``llvm.read_register``' intrinsic returns the current value of the
6839 register, where possible. The '``llvm.write_register``' intrinsic sets
6840 the current value of the register, where possible.
6842 This is useful to implement named register global variables that need
6843 to always be mapped to a specific register, as is common practice on
6844 bare-metal programs including OS kernels.
6846 The compiler doesn't check for register availability or use of the used
6847 register in surrounding code, including inline assembly. Because of that,
6848 allocatable registers are not supported.
6850 Warning: So far it only works with the stack pointer on selected
6851 architectures (ARM, ARM64, AArch64, PowerPC and x86_64). Significant amount of
6852 work is needed to support other registers and even more so, allocatable
6857 '``llvm.stacksave``' Intrinsic
6858 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6865 declare i8* @llvm.stacksave()
6870 The '``llvm.stacksave``' intrinsic is used to remember the current state
6871 of the function stack, for use with
6872 :ref:`llvm.stackrestore <int_stackrestore>`. This is useful for
6873 implementing language features like scoped automatic variable sized
6879 This intrinsic returns a opaque pointer value that can be passed to
6880 :ref:`llvm.stackrestore <int_stackrestore>`. When an
6881 ``llvm.stackrestore`` intrinsic is executed with a value saved from
6882 ``llvm.stacksave``, it effectively restores the state of the stack to
6883 the state it was in when the ``llvm.stacksave`` intrinsic executed. In
6884 practice, this pops any :ref:`alloca <i_alloca>` blocks from the stack that
6885 were allocated after the ``llvm.stacksave`` was executed.
6887 .. _int_stackrestore:
6889 '``llvm.stackrestore``' Intrinsic
6890 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6897 declare void @llvm.stackrestore(i8* %ptr)
6902 The '``llvm.stackrestore``' intrinsic is used to restore the state of
6903 the function stack to the state it was in when the corresponding
6904 :ref:`llvm.stacksave <int_stacksave>` intrinsic executed. This is
6905 useful for implementing language features like scoped automatic variable
6906 sized arrays in C99.
6911 See the description for :ref:`llvm.stacksave <int_stacksave>`.
6913 '``llvm.prefetch``' Intrinsic
6914 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6921 declare void @llvm.prefetch(i8* <address>, i32 <rw>, i32 <locality>, i32 <cache type>)
6926 The '``llvm.prefetch``' intrinsic is a hint to the code generator to
6927 insert a prefetch instruction if supported; otherwise, it is a noop.
6928 Prefetches have no effect on the behavior of the program but can change
6929 its performance characteristics.
6934 ``address`` is the address to be prefetched, ``rw`` is the specifier
6935 determining if the fetch should be for a read (0) or write (1), and
6936 ``locality`` is a temporal locality specifier ranging from (0) - no
6937 locality, to (3) - extremely local keep in cache. The ``cache type``
6938 specifies whether the prefetch is performed on the data (1) or
6939 instruction (0) cache. The ``rw``, ``locality`` and ``cache type``
6940 arguments must be constant integers.
6945 This intrinsic does not modify the behavior of the program. In
6946 particular, prefetches cannot trap and do not produce a value. On
6947 targets that support this intrinsic, the prefetch can provide hints to
6948 the processor cache for better performance.
6950 '``llvm.pcmarker``' Intrinsic
6951 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6958 declare void @llvm.pcmarker(i32 <id>)
6963 The '``llvm.pcmarker``' intrinsic is a method to export a Program
6964 Counter (PC) in a region of code to simulators and other tools. The
6965 method is target specific, but it is expected that the marker will use
6966 exported symbols to transmit the PC of the marker. The marker makes no
6967 guarantees that it will remain with any specific instruction after
6968 optimizations. It is possible that the presence of a marker will inhibit
6969 optimizations. The intended use is to be inserted after optimizations to
6970 allow correlations of simulation runs.
6975 ``id`` is a numerical id identifying the marker.
6980 This intrinsic does not modify the behavior of the program. Backends
6981 that do not support this intrinsic may ignore it.
6983 '``llvm.readcyclecounter``' Intrinsic
6984 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6991 declare i64 @llvm.readcyclecounter()
6996 The '``llvm.readcyclecounter``' intrinsic provides access to the cycle
6997 counter register (or similar low latency, high accuracy clocks) on those
6998 targets that support it. On X86, it should map to RDTSC. On Alpha, it
6999 should map to RPCC. As the backing counters overflow quickly (on the
7000 order of 9 seconds on alpha), this should only be used for small
7006 When directly supported, reading the cycle counter should not modify any
7007 memory. Implementations are allowed to either return a application
7008 specific value or a system wide value. On backends without support, this
7009 is lowered to a constant 0.
7011 Note that runtime support may be conditional on the privilege-level code is
7012 running at and the host platform.
7014 '``llvm.clear_cache``' Intrinsic
7015 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7022 declare void @llvm.clear_cache(i8*, i8*)
7027 The '``llvm.clear_cache``' intrinsic ensures visibility of modifications
7028 in the specified range to the execution unit of the processor. On
7029 targets with non-unified instruction and data cache, the implementation
7030 flushes the instruction cache.
7035 On platforms with coherent instruction and data caches (e.g. x86), this
7036 intrinsic is a nop. On platforms with non-coherent instruction and data
7037 cache (e.g. ARM, MIPS), the intrinsic is lowered either to appropriate
7038 instructions or a system call, if cache flushing requires special
7041 The default behavior is to emit a call to ``__clear_cache`` from the run
7044 This instrinsic does *not* empty the instruction pipeline. Modifications
7045 of the current function are outside the scope of the intrinsic.
7047 Standard C Library Intrinsics
7048 -----------------------------
7050 LLVM provides intrinsics for a few important standard C library
7051 functions. These intrinsics allow source-language front-ends to pass
7052 information about the alignment of the pointer arguments to the code
7053 generator, providing opportunity for more efficient code generation.
7057 '``llvm.memcpy``' Intrinsic
7058 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7063 This is an overloaded intrinsic. You can use ``llvm.memcpy`` on any
7064 integer bit width and for different address spaces. Not all targets
7065 support all bit widths however.
7069 declare void @llvm.memcpy.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
7070 i32 <len>, i32 <align>, i1 <isvolatile>)
7071 declare void @llvm.memcpy.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
7072 i64 <len>, i32 <align>, i1 <isvolatile>)
7077 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
7078 source location to the destination location.
7080 Note that, unlike the standard libc function, the ``llvm.memcpy.*``
7081 intrinsics do not return a value, takes extra alignment/isvolatile
7082 arguments and the pointers can be in specified address spaces.
7087 The first argument is a pointer to the destination, the second is a
7088 pointer to the source. The third argument is an integer argument
7089 specifying the number of bytes to copy, the fourth argument is the
7090 alignment of the source and destination locations, and the fifth is a
7091 boolean indicating a volatile access.
7093 If the call to this intrinsic has an alignment value that is not 0 or 1,
7094 then the caller guarantees that both the source and destination pointers
7095 are aligned to that boundary.
7097 If the ``isvolatile`` parameter is ``true``, the ``llvm.memcpy`` call is
7098 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
7099 very cleanly specified and it is unwise to depend on it.
7104 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
7105 source location to the destination location, which are not allowed to
7106 overlap. It copies "len" bytes of memory over. If the argument is known
7107 to be aligned to some boundary, this can be specified as the fourth
7108 argument, otherwise it should be set to 0 or 1 (both meaning no alignment).
7110 '``llvm.memmove``' Intrinsic
7111 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7116 This is an overloaded intrinsic. You can use llvm.memmove on any integer
7117 bit width and for different address space. Not all targets support all
7122 declare void @llvm.memmove.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
7123 i32 <len>, i32 <align>, i1 <isvolatile>)
7124 declare void @llvm.memmove.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
7125 i64 <len>, i32 <align>, i1 <isvolatile>)
7130 The '``llvm.memmove.*``' intrinsics move a block of memory from the
7131 source location to the destination location. It is similar to the
7132 '``llvm.memcpy``' intrinsic but allows the two memory locations to
7135 Note that, unlike the standard libc function, the ``llvm.memmove.*``
7136 intrinsics do not return a value, takes extra alignment/isvolatile
7137 arguments and the pointers can be in specified address spaces.
7142 The first argument is a pointer to the destination, the second is a
7143 pointer to the source. The third argument is an integer argument
7144 specifying the number of bytes to copy, the fourth argument is the
7145 alignment of the source and destination locations, and the fifth is a
7146 boolean indicating a volatile access.
7148 If the call to this intrinsic has an alignment value that is not 0 or 1,
7149 then the caller guarantees that the source and destination pointers are
7150 aligned to that boundary.
7152 If the ``isvolatile`` parameter is ``true``, the ``llvm.memmove`` call
7153 is a :ref:`volatile operation <volatile>`. The detailed access behavior is
7154 not very cleanly specified and it is unwise to depend on it.
7159 The '``llvm.memmove.*``' intrinsics copy a block of memory from the
7160 source location to the destination location, which may overlap. It
7161 copies "len" bytes of memory over. If the argument is known to be
7162 aligned to some boundary, this can be specified as the fourth argument,
7163 otherwise it should be set to 0 or 1 (both meaning no alignment).
7165 '``llvm.memset.*``' Intrinsics
7166 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7171 This is an overloaded intrinsic. You can use llvm.memset on any integer
7172 bit width and for different address spaces. However, not all targets
7173 support all bit widths.
7177 declare void @llvm.memset.p0i8.i32(i8* <dest>, i8 <val>,
7178 i32 <len>, i32 <align>, i1 <isvolatile>)
7179 declare void @llvm.memset.p0i8.i64(i8* <dest>, i8 <val>,
7180 i64 <len>, i32 <align>, i1 <isvolatile>)
7185 The '``llvm.memset.*``' intrinsics fill a block of memory with a
7186 particular byte value.
7188 Note that, unlike the standard libc function, the ``llvm.memset``
7189 intrinsic does not return a value and takes extra alignment/volatile
7190 arguments. Also, the destination can be in an arbitrary address space.
7195 The first argument is a pointer to the destination to fill, the second
7196 is the byte value with which to fill it, the third argument is an
7197 integer argument specifying the number of bytes to fill, and the fourth
7198 argument is the known alignment of the destination location.
7200 If the call to this intrinsic has an alignment value that is not 0 or 1,
7201 then the caller guarantees that the destination pointer is aligned to
7204 If the ``isvolatile`` parameter is ``true``, the ``llvm.memset`` call is
7205 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
7206 very cleanly specified and it is unwise to depend on it.
7211 The '``llvm.memset.*``' intrinsics fill "len" bytes of memory starting
7212 at the destination location. If the argument is known to be aligned to
7213 some boundary, this can be specified as the fourth argument, otherwise
7214 it should be set to 0 or 1 (both meaning no alignment).
7216 '``llvm.sqrt.*``' Intrinsic
7217 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7222 This is an overloaded intrinsic. You can use ``llvm.sqrt`` on any
7223 floating point or vector of floating point type. Not all targets support
7228 declare float @llvm.sqrt.f32(float %Val)
7229 declare double @llvm.sqrt.f64(double %Val)
7230 declare x86_fp80 @llvm.sqrt.f80(x86_fp80 %Val)
7231 declare fp128 @llvm.sqrt.f128(fp128 %Val)
7232 declare ppc_fp128 @llvm.sqrt.ppcf128(ppc_fp128 %Val)
7237 The '``llvm.sqrt``' intrinsics return the sqrt of the specified operand,
7238 returning the same value as the libm '``sqrt``' functions would. Unlike
7239 ``sqrt`` in libm, however, ``llvm.sqrt`` has undefined behavior for
7240 negative numbers other than -0.0 (which allows for better optimization,
7241 because there is no need to worry about errno being set).
7242 ``llvm.sqrt(-0.0)`` is defined to return -0.0 like IEEE sqrt.
7247 The argument and return value are floating point numbers of the same
7253 This function returns the sqrt of the specified operand if it is a
7254 nonnegative floating point number.
7256 '``llvm.powi.*``' Intrinsic
7257 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7262 This is an overloaded intrinsic. You can use ``llvm.powi`` on any
7263 floating point or vector of floating point type. Not all targets support
7268 declare float @llvm.powi.f32(float %Val, i32 %power)
7269 declare double @llvm.powi.f64(double %Val, i32 %power)
7270 declare x86_fp80 @llvm.powi.f80(x86_fp80 %Val, i32 %power)
7271 declare fp128 @llvm.powi.f128(fp128 %Val, i32 %power)
7272 declare ppc_fp128 @llvm.powi.ppcf128(ppc_fp128 %Val, i32 %power)
7277 The '``llvm.powi.*``' intrinsics return the first operand raised to the
7278 specified (positive or negative) power. The order of evaluation of
7279 multiplications is not defined. When a vector of floating point type is
7280 used, the second argument remains a scalar integer value.
7285 The second argument is an integer power, and the first is a value to
7286 raise to that power.
7291 This function returns the first value raised to the second power with an
7292 unspecified sequence of rounding operations.
7294 '``llvm.sin.*``' Intrinsic
7295 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7300 This is an overloaded intrinsic. You can use ``llvm.sin`` on any
7301 floating point or vector of floating point type. Not all targets support
7306 declare float @llvm.sin.f32(float %Val)
7307 declare double @llvm.sin.f64(double %Val)
7308 declare x86_fp80 @llvm.sin.f80(x86_fp80 %Val)
7309 declare fp128 @llvm.sin.f128(fp128 %Val)
7310 declare ppc_fp128 @llvm.sin.ppcf128(ppc_fp128 %Val)
7315 The '``llvm.sin.*``' intrinsics return the sine of the operand.
7320 The argument and return value are floating point numbers of the same
7326 This function returns the sine of the specified operand, returning the
7327 same values as the libm ``sin`` functions would, and handles error
7328 conditions in the same way.
7330 '``llvm.cos.*``' Intrinsic
7331 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7336 This is an overloaded intrinsic. You can use ``llvm.cos`` on any
7337 floating point or vector of floating point type. Not all targets support
7342 declare float @llvm.cos.f32(float %Val)
7343 declare double @llvm.cos.f64(double %Val)
7344 declare x86_fp80 @llvm.cos.f80(x86_fp80 %Val)
7345 declare fp128 @llvm.cos.f128(fp128 %Val)
7346 declare ppc_fp128 @llvm.cos.ppcf128(ppc_fp128 %Val)
7351 The '``llvm.cos.*``' intrinsics return the cosine of the operand.
7356 The argument and return value are floating point numbers of the same
7362 This function returns the cosine of the specified operand, returning the
7363 same values as the libm ``cos`` functions would, and handles error
7364 conditions in the same way.
7366 '``llvm.pow.*``' Intrinsic
7367 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7372 This is an overloaded intrinsic. You can use ``llvm.pow`` on any
7373 floating point or vector of floating point type. Not all targets support
7378 declare float @llvm.pow.f32(float %Val, float %Power)
7379 declare double @llvm.pow.f64(double %Val, double %Power)
7380 declare x86_fp80 @llvm.pow.f80(x86_fp80 %Val, x86_fp80 %Power)
7381 declare fp128 @llvm.pow.f128(fp128 %Val, fp128 %Power)
7382 declare ppc_fp128 @llvm.pow.ppcf128(ppc_fp128 %Val, ppc_fp128 Power)
7387 The '``llvm.pow.*``' intrinsics return the first operand raised to the
7388 specified (positive or negative) power.
7393 The second argument is a floating point power, and the first is a value
7394 to raise to that power.
7399 This function returns the first value raised to the second power,
7400 returning the same values as the libm ``pow`` functions would, and
7401 handles error conditions in the same way.
7403 '``llvm.exp.*``' Intrinsic
7404 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7409 This is an overloaded intrinsic. You can use ``llvm.exp`` on any
7410 floating point or vector of floating point type. Not all targets support
7415 declare float @llvm.exp.f32(float %Val)
7416 declare double @llvm.exp.f64(double %Val)
7417 declare x86_fp80 @llvm.exp.f80(x86_fp80 %Val)
7418 declare fp128 @llvm.exp.f128(fp128 %Val)
7419 declare ppc_fp128 @llvm.exp.ppcf128(ppc_fp128 %Val)
7424 The '``llvm.exp.*``' intrinsics perform the exp function.
7429 The argument and return value are floating point numbers of the same
7435 This function returns the same values as the libm ``exp`` functions
7436 would, and handles error conditions in the same way.
7438 '``llvm.exp2.*``' Intrinsic
7439 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7444 This is an overloaded intrinsic. You can use ``llvm.exp2`` on any
7445 floating point or vector of floating point type. Not all targets support
7450 declare float @llvm.exp2.f32(float %Val)
7451 declare double @llvm.exp2.f64(double %Val)
7452 declare x86_fp80 @llvm.exp2.f80(x86_fp80 %Val)
7453 declare fp128 @llvm.exp2.f128(fp128 %Val)
7454 declare ppc_fp128 @llvm.exp2.ppcf128(ppc_fp128 %Val)
7459 The '``llvm.exp2.*``' intrinsics perform the exp2 function.
7464 The argument and return value are floating point numbers of the same
7470 This function returns the same values as the libm ``exp2`` functions
7471 would, and handles error conditions in the same way.
7473 '``llvm.log.*``' Intrinsic
7474 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7479 This is an overloaded intrinsic. You can use ``llvm.log`` on any
7480 floating point or vector of floating point type. Not all targets support
7485 declare float @llvm.log.f32(float %Val)
7486 declare double @llvm.log.f64(double %Val)
7487 declare x86_fp80 @llvm.log.f80(x86_fp80 %Val)
7488 declare fp128 @llvm.log.f128(fp128 %Val)
7489 declare ppc_fp128 @llvm.log.ppcf128(ppc_fp128 %Val)
7494 The '``llvm.log.*``' intrinsics perform the log function.
7499 The argument and return value are floating point numbers of the same
7505 This function returns the same values as the libm ``log`` functions
7506 would, and handles error conditions in the same way.
7508 '``llvm.log10.*``' Intrinsic
7509 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7514 This is an overloaded intrinsic. You can use ``llvm.log10`` on any
7515 floating point or vector of floating point type. Not all targets support
7520 declare float @llvm.log10.f32(float %Val)
7521 declare double @llvm.log10.f64(double %Val)
7522 declare x86_fp80 @llvm.log10.f80(x86_fp80 %Val)
7523 declare fp128 @llvm.log10.f128(fp128 %Val)
7524 declare ppc_fp128 @llvm.log10.ppcf128(ppc_fp128 %Val)
7529 The '``llvm.log10.*``' intrinsics perform the log10 function.
7534 The argument and return value are floating point numbers of the same
7540 This function returns the same values as the libm ``log10`` functions
7541 would, and handles error conditions in the same way.
7543 '``llvm.log2.*``' Intrinsic
7544 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7549 This is an overloaded intrinsic. You can use ``llvm.log2`` on any
7550 floating point or vector of floating point type. Not all targets support
7555 declare float @llvm.log2.f32(float %Val)
7556 declare double @llvm.log2.f64(double %Val)
7557 declare x86_fp80 @llvm.log2.f80(x86_fp80 %Val)
7558 declare fp128 @llvm.log2.f128(fp128 %Val)
7559 declare ppc_fp128 @llvm.log2.ppcf128(ppc_fp128 %Val)
7564 The '``llvm.log2.*``' intrinsics perform the log2 function.
7569 The argument and return value are floating point numbers of the same
7575 This function returns the same values as the libm ``log2`` functions
7576 would, and handles error conditions in the same way.
7578 '``llvm.fma.*``' Intrinsic
7579 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7584 This is an overloaded intrinsic. You can use ``llvm.fma`` on any
7585 floating point or vector of floating point type. Not all targets support
7590 declare float @llvm.fma.f32(float %a, float %b, float %c)
7591 declare double @llvm.fma.f64(double %a, double %b, double %c)
7592 declare x86_fp80 @llvm.fma.f80(x86_fp80 %a, x86_fp80 %b, x86_fp80 %c)
7593 declare fp128 @llvm.fma.f128(fp128 %a, fp128 %b, fp128 %c)
7594 declare ppc_fp128 @llvm.fma.ppcf128(ppc_fp128 %a, ppc_fp128 %b, ppc_fp128 %c)
7599 The '``llvm.fma.*``' intrinsics perform the fused multiply-add
7605 The argument and return value are floating point numbers of the same
7611 This function returns the same values as the libm ``fma`` functions
7612 would, and does not set errno.
7614 '``llvm.fabs.*``' Intrinsic
7615 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7620 This is an overloaded intrinsic. You can use ``llvm.fabs`` on any
7621 floating point or vector of floating point type. Not all targets support
7626 declare float @llvm.fabs.f32(float %Val)
7627 declare double @llvm.fabs.f64(double %Val)
7628 declare x86_fp80 @llvm.fabs.f80(x86_fp80 %Val)
7629 declare fp128 @llvm.fabs.f128(fp128 %Val)
7630 declare ppc_fp128 @llvm.fabs.ppcf128(ppc_fp128 %Val)
7635 The '``llvm.fabs.*``' intrinsics return the absolute value of the
7641 The argument and return value are floating point numbers of the same
7647 This function returns the same values as the libm ``fabs`` functions
7648 would, and handles error conditions in the same way.
7650 '``llvm.copysign.*``' Intrinsic
7651 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7656 This is an overloaded intrinsic. You can use ``llvm.copysign`` on any
7657 floating point or vector of floating point type. Not all targets support
7662 declare float @llvm.copysign.f32(float %Mag, float %Sgn)
7663 declare double @llvm.copysign.f64(double %Mag, double %Sgn)
7664 declare x86_fp80 @llvm.copysign.f80(x86_fp80 %Mag, x86_fp80 %Sgn)
7665 declare fp128 @llvm.copysign.f128(fp128 %Mag, fp128 %Sgn)
7666 declare ppc_fp128 @llvm.copysign.ppcf128(ppc_fp128 %Mag, ppc_fp128 %Sgn)
7671 The '``llvm.copysign.*``' intrinsics return a value with the magnitude of the
7672 first operand and the sign of the second operand.
7677 The arguments and return value are floating point numbers of the same
7683 This function returns the same values as the libm ``copysign``
7684 functions would, and handles error conditions in the same way.
7686 '``llvm.floor.*``' Intrinsic
7687 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7692 This is an overloaded intrinsic. You can use ``llvm.floor`` on any
7693 floating point or vector of floating point type. Not all targets support
7698 declare float @llvm.floor.f32(float %Val)
7699 declare double @llvm.floor.f64(double %Val)
7700 declare x86_fp80 @llvm.floor.f80(x86_fp80 %Val)
7701 declare fp128 @llvm.floor.f128(fp128 %Val)
7702 declare ppc_fp128 @llvm.floor.ppcf128(ppc_fp128 %Val)
7707 The '``llvm.floor.*``' intrinsics return the floor of the operand.
7712 The argument and return value are floating point numbers of the same
7718 This function returns the same values as the libm ``floor`` functions
7719 would, and handles error conditions in the same way.
7721 '``llvm.ceil.*``' Intrinsic
7722 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7727 This is an overloaded intrinsic. You can use ``llvm.ceil`` on any
7728 floating point or vector of floating point type. Not all targets support
7733 declare float @llvm.ceil.f32(float %Val)
7734 declare double @llvm.ceil.f64(double %Val)
7735 declare x86_fp80 @llvm.ceil.f80(x86_fp80 %Val)
7736 declare fp128 @llvm.ceil.f128(fp128 %Val)
7737 declare ppc_fp128 @llvm.ceil.ppcf128(ppc_fp128 %Val)
7742 The '``llvm.ceil.*``' intrinsics return the ceiling of the operand.
7747 The argument and return value are floating point numbers of the same
7753 This function returns the same values as the libm ``ceil`` functions
7754 would, and handles error conditions in the same way.
7756 '``llvm.trunc.*``' Intrinsic
7757 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7762 This is an overloaded intrinsic. You can use ``llvm.trunc`` on any
7763 floating point or vector of floating point type. Not all targets support
7768 declare float @llvm.trunc.f32(float %Val)
7769 declare double @llvm.trunc.f64(double %Val)
7770 declare x86_fp80 @llvm.trunc.f80(x86_fp80 %Val)
7771 declare fp128 @llvm.trunc.f128(fp128 %Val)
7772 declare ppc_fp128 @llvm.trunc.ppcf128(ppc_fp128 %Val)
7777 The '``llvm.trunc.*``' intrinsics returns the operand rounded to the
7778 nearest integer not larger in magnitude than the operand.
7783 The argument and return value are floating point numbers of the same
7789 This function returns the same values as the libm ``trunc`` functions
7790 would, and handles error conditions in the same way.
7792 '``llvm.rint.*``' Intrinsic
7793 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7798 This is an overloaded intrinsic. You can use ``llvm.rint`` on any
7799 floating point or vector of floating point type. Not all targets support
7804 declare float @llvm.rint.f32(float %Val)
7805 declare double @llvm.rint.f64(double %Val)
7806 declare x86_fp80 @llvm.rint.f80(x86_fp80 %Val)
7807 declare fp128 @llvm.rint.f128(fp128 %Val)
7808 declare ppc_fp128 @llvm.rint.ppcf128(ppc_fp128 %Val)
7813 The '``llvm.rint.*``' intrinsics returns the operand rounded to the
7814 nearest integer. It may raise an inexact floating-point exception if the
7815 operand isn't an integer.
7820 The argument and return value are floating point numbers of the same
7826 This function returns the same values as the libm ``rint`` functions
7827 would, and handles error conditions in the same way.
7829 '``llvm.nearbyint.*``' Intrinsic
7830 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7835 This is an overloaded intrinsic. You can use ``llvm.nearbyint`` on any
7836 floating point or vector of floating point type. Not all targets support
7841 declare float @llvm.nearbyint.f32(float %Val)
7842 declare double @llvm.nearbyint.f64(double %Val)
7843 declare x86_fp80 @llvm.nearbyint.f80(x86_fp80 %Val)
7844 declare fp128 @llvm.nearbyint.f128(fp128 %Val)
7845 declare ppc_fp128 @llvm.nearbyint.ppcf128(ppc_fp128 %Val)
7850 The '``llvm.nearbyint.*``' intrinsics returns the operand rounded to the
7856 The argument and return value are floating point numbers of the same
7862 This function returns the same values as the libm ``nearbyint``
7863 functions would, and handles error conditions in the same way.
7865 '``llvm.round.*``' Intrinsic
7866 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7871 This is an overloaded intrinsic. You can use ``llvm.round`` on any
7872 floating point or vector of floating point type. Not all targets support
7877 declare float @llvm.round.f32(float %Val)
7878 declare double @llvm.round.f64(double %Val)
7879 declare x86_fp80 @llvm.round.f80(x86_fp80 %Val)
7880 declare fp128 @llvm.round.f128(fp128 %Val)
7881 declare ppc_fp128 @llvm.round.ppcf128(ppc_fp128 %Val)
7886 The '``llvm.round.*``' intrinsics returns the operand rounded to the
7892 The argument and return value are floating point numbers of the same
7898 This function returns the same values as the libm ``round``
7899 functions would, and handles error conditions in the same way.
7901 Bit Manipulation Intrinsics
7902 ---------------------------
7904 LLVM provides intrinsics for a few important bit manipulation
7905 operations. These allow efficient code generation for some algorithms.
7907 '``llvm.bswap.*``' Intrinsics
7908 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7913 This is an overloaded intrinsic function. You can use bswap on any
7914 integer type that is an even number of bytes (i.e. BitWidth % 16 == 0).
7918 declare i16 @llvm.bswap.i16(i16 <id>)
7919 declare i32 @llvm.bswap.i32(i32 <id>)
7920 declare i64 @llvm.bswap.i64(i64 <id>)
7925 The '``llvm.bswap``' family of intrinsics is used to byte swap integer
7926 values with an even number of bytes (positive multiple of 16 bits).
7927 These are useful for performing operations on data that is not in the
7928 target's native byte order.
7933 The ``llvm.bswap.i16`` intrinsic returns an i16 value that has the high
7934 and low byte of the input i16 swapped. Similarly, the ``llvm.bswap.i32``
7935 intrinsic returns an i32 value that has the four bytes of the input i32
7936 swapped, so that if the input bytes are numbered 0, 1, 2, 3 then the
7937 returned i32 will have its bytes in 3, 2, 1, 0 order. The
7938 ``llvm.bswap.i48``, ``llvm.bswap.i64`` and other intrinsics extend this
7939 concept to additional even-byte lengths (6 bytes, 8 bytes and more,
7942 '``llvm.ctpop.*``' Intrinsic
7943 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7948 This is an overloaded intrinsic. You can use llvm.ctpop on any integer
7949 bit width, or on any vector with integer elements. Not all targets
7950 support all bit widths or vector types, however.
7954 declare i8 @llvm.ctpop.i8(i8 <src>)
7955 declare i16 @llvm.ctpop.i16(i16 <src>)
7956 declare i32 @llvm.ctpop.i32(i32 <src>)
7957 declare i64 @llvm.ctpop.i64(i64 <src>)
7958 declare i256 @llvm.ctpop.i256(i256 <src>)
7959 declare <2 x i32> @llvm.ctpop.v2i32(<2 x i32> <src>)
7964 The '``llvm.ctpop``' family of intrinsics counts the number of bits set
7970 The only argument is the value to be counted. The argument may be of any
7971 integer type, or a vector with integer elements. The return type must
7972 match the argument type.
7977 The '``llvm.ctpop``' intrinsic counts the 1's in a variable, or within
7978 each element of a vector.
7980 '``llvm.ctlz.*``' Intrinsic
7981 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7986 This is an overloaded intrinsic. You can use ``llvm.ctlz`` on any
7987 integer bit width, or any vector whose elements are integers. Not all
7988 targets support all bit widths or vector types, however.
7992 declare i8 @llvm.ctlz.i8 (i8 <src>, i1 <is_zero_undef>)
7993 declare i16 @llvm.ctlz.i16 (i16 <src>, i1 <is_zero_undef>)
7994 declare i32 @llvm.ctlz.i32 (i32 <src>, i1 <is_zero_undef>)
7995 declare i64 @llvm.ctlz.i64 (i64 <src>, i1 <is_zero_undef>)
7996 declare i256 @llvm.ctlz.i256(i256 <src>, i1 <is_zero_undef>)
7997 declase <2 x i32> @llvm.ctlz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
8002 The '``llvm.ctlz``' family of intrinsic functions counts the number of
8003 leading zeros in a variable.
8008 The first argument is the value to be counted. This argument may be of
8009 any integer type, or a vectory with integer element type. The return
8010 type must match the first argument type.
8012 The second argument must be a constant and is a flag to indicate whether
8013 the intrinsic should ensure that a zero as the first argument produces a
8014 defined result. Historically some architectures did not provide a
8015 defined result for zero values as efficiently, and many algorithms are
8016 now predicated on avoiding zero-value inputs.
8021 The '``llvm.ctlz``' intrinsic counts the leading (most significant)
8022 zeros in a variable, or within each element of the vector. If
8023 ``src == 0`` then the result is the size in bits of the type of ``src``
8024 if ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
8025 ``llvm.ctlz(i32 2) = 30``.
8027 '``llvm.cttz.*``' Intrinsic
8028 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8033 This is an overloaded intrinsic. You can use ``llvm.cttz`` on any
8034 integer bit width, or any vector of integer elements. Not all targets
8035 support all bit widths or vector types, however.
8039 declare i8 @llvm.cttz.i8 (i8 <src>, i1 <is_zero_undef>)
8040 declare i16 @llvm.cttz.i16 (i16 <src>, i1 <is_zero_undef>)
8041 declare i32 @llvm.cttz.i32 (i32 <src>, i1 <is_zero_undef>)
8042 declare i64 @llvm.cttz.i64 (i64 <src>, i1 <is_zero_undef>)
8043 declare i256 @llvm.cttz.i256(i256 <src>, i1 <is_zero_undef>)
8044 declase <2 x i32> @llvm.cttz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
8049 The '``llvm.cttz``' family of intrinsic functions counts the number of
8055 The first argument is the value to be counted. This argument may be of
8056 any integer type, or a vectory with integer element type. The return
8057 type must match the first argument type.
8059 The second argument must be a constant and is a flag to indicate whether
8060 the intrinsic should ensure that a zero as the first argument produces a
8061 defined result. Historically some architectures did not provide a
8062 defined result for zero values as efficiently, and many algorithms are
8063 now predicated on avoiding zero-value inputs.
8068 The '``llvm.cttz``' intrinsic counts the trailing (least significant)
8069 zeros in a variable, or within each element of a vector. If ``src == 0``
8070 then the result is the size in bits of the type of ``src`` if
8071 ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
8072 ``llvm.cttz(2) = 1``.
8074 Arithmetic with Overflow Intrinsics
8075 -----------------------------------
8077 LLVM provides intrinsics for some arithmetic with overflow operations.
8079 '``llvm.sadd.with.overflow.*``' Intrinsics
8080 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8085 This is an overloaded intrinsic. You can use ``llvm.sadd.with.overflow``
8086 on any integer bit width.
8090 declare {i16, i1} @llvm.sadd.with.overflow.i16(i16 %a, i16 %b)
8091 declare {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
8092 declare {i64, i1} @llvm.sadd.with.overflow.i64(i64 %a, i64 %b)
8097 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
8098 a signed addition of the two arguments, and indicate whether an overflow
8099 occurred during the signed summation.
8104 The arguments (%a and %b) and the first element of the result structure
8105 may be of integer types of any bit width, but they must have the same
8106 bit width. The second element of the result structure must be of type
8107 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
8113 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
8114 a signed addition of the two variables. They return a structure --- the
8115 first element of which is the signed summation, and the second element
8116 of which is a bit specifying if the signed summation resulted in an
8122 .. code-block:: llvm
8124 %res = call {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
8125 %sum = extractvalue {i32, i1} %res, 0
8126 %obit = extractvalue {i32, i1} %res, 1
8127 br i1 %obit, label %overflow, label %normal
8129 '``llvm.uadd.with.overflow.*``' Intrinsics
8130 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8135 This is an overloaded intrinsic. You can use ``llvm.uadd.with.overflow``
8136 on any integer bit width.
8140 declare {i16, i1} @llvm.uadd.with.overflow.i16(i16 %a, i16 %b)
8141 declare {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
8142 declare {i64, i1} @llvm.uadd.with.overflow.i64(i64 %a, i64 %b)
8147 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
8148 an unsigned addition of the two arguments, and indicate whether a carry
8149 occurred during the unsigned summation.
8154 The arguments (%a and %b) and the first element of the result structure
8155 may be of integer types of any bit width, but they must have the same
8156 bit width. The second element of the result structure must be of type
8157 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
8163 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
8164 an unsigned addition of the two arguments. They return a structure --- the
8165 first element of which is the sum, and the second element of which is a
8166 bit specifying if the unsigned summation resulted in a carry.
8171 .. code-block:: llvm
8173 %res = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
8174 %sum = extractvalue {i32, i1} %res, 0
8175 %obit = extractvalue {i32, i1} %res, 1
8176 br i1 %obit, label %carry, label %normal
8178 '``llvm.ssub.with.overflow.*``' Intrinsics
8179 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8184 This is an overloaded intrinsic. You can use ``llvm.ssub.with.overflow``
8185 on any integer bit width.
8189 declare {i16, i1} @llvm.ssub.with.overflow.i16(i16 %a, i16 %b)
8190 declare {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
8191 declare {i64, i1} @llvm.ssub.with.overflow.i64(i64 %a, i64 %b)
8196 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
8197 a signed subtraction of the two arguments, and indicate whether an
8198 overflow occurred during the signed subtraction.
8203 The arguments (%a and %b) and the first element of the result structure
8204 may be of integer types of any bit width, but they must have the same
8205 bit width. The second element of the result structure must be of type
8206 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
8212 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
8213 a signed subtraction of the two arguments. They return a structure --- the
8214 first element of which is the subtraction, and the second element of
8215 which is a bit specifying if the signed subtraction resulted in an
8221 .. code-block:: llvm
8223 %res = call {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
8224 %sum = extractvalue {i32, i1} %res, 0
8225 %obit = extractvalue {i32, i1} %res, 1
8226 br i1 %obit, label %overflow, label %normal
8228 '``llvm.usub.with.overflow.*``' Intrinsics
8229 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8234 This is an overloaded intrinsic. You can use ``llvm.usub.with.overflow``
8235 on any integer bit width.
8239 declare {i16, i1} @llvm.usub.with.overflow.i16(i16 %a, i16 %b)
8240 declare {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
8241 declare {i64, i1} @llvm.usub.with.overflow.i64(i64 %a, i64 %b)
8246 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
8247 an unsigned subtraction of the two arguments, and indicate whether an
8248 overflow occurred during the unsigned subtraction.
8253 The arguments (%a and %b) and the first element of the result structure
8254 may be of integer types of any bit width, but they must have the same
8255 bit width. The second element of the result structure must be of type
8256 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
8262 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
8263 an unsigned subtraction of the two arguments. They return a structure ---
8264 the first element of which is the subtraction, and the second element of
8265 which is a bit specifying if the unsigned subtraction resulted in an
8271 .. code-block:: llvm
8273 %res = call {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
8274 %sum = extractvalue {i32, i1} %res, 0
8275 %obit = extractvalue {i32, i1} %res, 1
8276 br i1 %obit, label %overflow, label %normal
8278 '``llvm.smul.with.overflow.*``' Intrinsics
8279 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8284 This is an overloaded intrinsic. You can use ``llvm.smul.with.overflow``
8285 on any integer bit width.
8289 declare {i16, i1} @llvm.smul.with.overflow.i16(i16 %a, i16 %b)
8290 declare {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
8291 declare {i64, i1} @llvm.smul.with.overflow.i64(i64 %a, i64 %b)
8296 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
8297 a signed multiplication of the two arguments, and indicate whether an
8298 overflow occurred during the signed multiplication.
8303 The arguments (%a and %b) and the first element of the result structure
8304 may be of integer types of any bit width, but they must have the same
8305 bit width. The second element of the result structure must be of type
8306 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
8312 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
8313 a signed multiplication of the two arguments. They return a structure ---
8314 the first element of which is the multiplication, and the second element
8315 of which is a bit specifying if the signed multiplication resulted in an
8321 .. code-block:: llvm
8323 %res = call {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
8324 %sum = extractvalue {i32, i1} %res, 0
8325 %obit = extractvalue {i32, i1} %res, 1
8326 br i1 %obit, label %overflow, label %normal
8328 '``llvm.umul.with.overflow.*``' Intrinsics
8329 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8334 This is an overloaded intrinsic. You can use ``llvm.umul.with.overflow``
8335 on any integer bit width.
8339 declare {i16, i1} @llvm.umul.with.overflow.i16(i16 %a, i16 %b)
8340 declare {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
8341 declare {i64, i1} @llvm.umul.with.overflow.i64(i64 %a, i64 %b)
8346 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
8347 a unsigned multiplication of the two arguments, and indicate whether an
8348 overflow occurred during the unsigned multiplication.
8353 The arguments (%a and %b) and the first element of the result structure
8354 may be of integer types of any bit width, but they must have the same
8355 bit width. The second element of the result structure must be of type
8356 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
8362 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
8363 an unsigned multiplication of the two arguments. They return a structure ---
8364 the first element of which is the multiplication, and the second
8365 element of which is a bit specifying if the unsigned multiplication
8366 resulted in an overflow.
8371 .. code-block:: llvm
8373 %res = call {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
8374 %sum = extractvalue {i32, i1} %res, 0
8375 %obit = extractvalue {i32, i1} %res, 1
8376 br i1 %obit, label %overflow, label %normal
8378 Specialised Arithmetic Intrinsics
8379 ---------------------------------
8381 '``llvm.fmuladd.*``' Intrinsic
8382 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8389 declare float @llvm.fmuladd.f32(float %a, float %b, float %c)
8390 declare double @llvm.fmuladd.f64(double %a, double %b, double %c)
8395 The '``llvm.fmuladd.*``' intrinsic functions represent multiply-add
8396 expressions that can be fused if the code generator determines that (a) the
8397 target instruction set has support for a fused operation, and (b) that the
8398 fused operation is more efficient than the equivalent, separate pair of mul
8399 and add instructions.
8404 The '``llvm.fmuladd.*``' intrinsics each take three arguments: two
8405 multiplicands, a and b, and an addend c.
8414 %0 = call float @llvm.fmuladd.f32(%a, %b, %c)
8416 is equivalent to the expression a \* b + c, except that rounding will
8417 not be performed between the multiplication and addition steps if the
8418 code generator fuses the operations. Fusion is not guaranteed, even if
8419 the target platform supports it. If a fused multiply-add is required the
8420 corresponding llvm.fma.\* intrinsic function should be used
8421 instead. This never sets errno, just as '``llvm.fma.*``'.
8426 .. code-block:: llvm
8428 %r2 = call float @llvm.fmuladd.f32(float %a, float %b, float %c) ; yields {float}:r2 = (a * b) + c
8430 Half Precision Floating Point Intrinsics
8431 ----------------------------------------
8433 For most target platforms, half precision floating point is a
8434 storage-only format. This means that it is a dense encoding (in memory)
8435 but does not support computation in the format.
8437 This means that code must first load the half-precision floating point
8438 value as an i16, then convert it to float with
8439 :ref:`llvm.convert.from.fp16 <int_convert_from_fp16>`. Computation can
8440 then be performed on the float value (including extending to double
8441 etc). To store the value back to memory, it is first converted to float
8442 if needed, then converted to i16 with
8443 :ref:`llvm.convert.to.fp16 <int_convert_to_fp16>`, then storing as an
8446 .. _int_convert_to_fp16:
8448 '``llvm.convert.to.fp16``' Intrinsic
8449 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8456 declare i16 @llvm.convert.to.fp16(f32 %a)
8461 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion
8462 from single precision floating point format to half precision floating
8468 The intrinsic function contains single argument - the value to be
8474 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion
8475 from single precision floating point format to half precision floating
8476 point format. The return value is an ``i16`` which contains the
8482 .. code-block:: llvm
8484 %res = call i16 @llvm.convert.to.fp16(f32 %a)
8485 store i16 %res, i16* @x, align 2
8487 .. _int_convert_from_fp16:
8489 '``llvm.convert.from.fp16``' Intrinsic
8490 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8497 declare f32 @llvm.convert.from.fp16(i16 %a)
8502 The '``llvm.convert.from.fp16``' intrinsic function performs a
8503 conversion from half precision floating point format to single precision
8504 floating point format.
8509 The intrinsic function contains single argument - the value to be
8515 The '``llvm.convert.from.fp16``' intrinsic function performs a
8516 conversion from half single precision floating point format to single
8517 precision floating point format. The input half-float value is
8518 represented by an ``i16`` value.
8523 .. code-block:: llvm
8525 %a = load i16* @x, align 2
8526 %res = call f32 @llvm.convert.from.fp16(i16 %a)
8531 The LLVM debugger intrinsics (which all start with ``llvm.dbg.``
8532 prefix), are described in the `LLVM Source Level
8533 Debugging <SourceLevelDebugging.html#format_common_intrinsics>`_
8536 Exception Handling Intrinsics
8537 -----------------------------
8539 The LLVM exception handling intrinsics (which all start with
8540 ``llvm.eh.`` prefix), are described in the `LLVM Exception
8541 Handling <ExceptionHandling.html#format_common_intrinsics>`_ document.
8545 Trampoline Intrinsics
8546 ---------------------
8548 These intrinsics make it possible to excise one parameter, marked with
8549 the :ref:`nest <nest>` attribute, from a function. The result is a
8550 callable function pointer lacking the nest parameter - the caller does
8551 not need to provide a value for it. Instead, the value to use is stored
8552 in advance in a "trampoline", a block of memory usually allocated on the
8553 stack, which also contains code to splice the nest value into the
8554 argument list. This is used to implement the GCC nested function address
8557 For example, if the function is ``i32 f(i8* nest %c, i32 %x, i32 %y)``
8558 then the resulting function pointer has signature ``i32 (i32, i32)*``.
8559 It can be created as follows:
8561 .. code-block:: llvm
8563 %tramp = alloca [10 x i8], align 4 ; size and alignment only correct for X86
8564 %tramp1 = getelementptr [10 x i8]* %tramp, i32 0, i32 0
8565 call i8* @llvm.init.trampoline(i8* %tramp1, i8* bitcast (i32 (i8*, i32, i32)* @f to i8*), i8* %nval)
8566 %p = call i8* @llvm.adjust.trampoline(i8* %tramp1)
8567 %fp = bitcast i8* %p to i32 (i32, i32)*
8569 The call ``%val = call i32 %fp(i32 %x, i32 %y)`` is then equivalent to
8570 ``%val = call i32 %f(i8* %nval, i32 %x, i32 %y)``.
8574 '``llvm.init.trampoline``' Intrinsic
8575 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8582 declare void @llvm.init.trampoline(i8* <tramp>, i8* <func>, i8* <nval>)
8587 This fills the memory pointed to by ``tramp`` with executable code,
8588 turning it into a trampoline.
8593 The ``llvm.init.trampoline`` intrinsic takes three arguments, all
8594 pointers. The ``tramp`` argument must point to a sufficiently large and
8595 sufficiently aligned block of memory; this memory is written to by the
8596 intrinsic. Note that the size and the alignment are target-specific -
8597 LLVM currently provides no portable way of determining them, so a
8598 front-end that generates this intrinsic needs to have some
8599 target-specific knowledge. The ``func`` argument must hold a function
8600 bitcast to an ``i8*``.
8605 The block of memory pointed to by ``tramp`` is filled with target
8606 dependent code, turning it into a function. Then ``tramp`` needs to be
8607 passed to :ref:`llvm.adjust.trampoline <int_at>` to get a pointer which can
8608 be :ref:`bitcast (to a new function) and called <int_trampoline>`. The new
8609 function's signature is the same as that of ``func`` with any arguments
8610 marked with the ``nest`` attribute removed. At most one such ``nest``
8611 argument is allowed, and it must be of pointer type. Calling the new
8612 function is equivalent to calling ``func`` with the same argument list,
8613 but with ``nval`` used for the missing ``nest`` argument. If, after
8614 calling ``llvm.init.trampoline``, the memory pointed to by ``tramp`` is
8615 modified, then the effect of any later call to the returned function
8616 pointer is undefined.
8620 '``llvm.adjust.trampoline``' Intrinsic
8621 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8628 declare i8* @llvm.adjust.trampoline(i8* <tramp>)
8633 This performs any required machine-specific adjustment to the address of
8634 a trampoline (passed as ``tramp``).
8639 ``tramp`` must point to a block of memory which already has trampoline
8640 code filled in by a previous call to
8641 :ref:`llvm.init.trampoline <int_it>`.
8646 On some architectures the address of the code to be executed needs to be
8647 different to the address where the trampoline is actually stored. This
8648 intrinsic returns the executable address corresponding to ``tramp``
8649 after performing the required machine specific adjustments. The pointer
8650 returned can then be :ref:`bitcast and executed <int_trampoline>`.
8655 This class of intrinsics exists to information about the lifetime of
8656 memory objects and ranges where variables are immutable.
8660 '``llvm.lifetime.start``' Intrinsic
8661 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8668 declare void @llvm.lifetime.start(i64 <size>, i8* nocapture <ptr>)
8673 The '``llvm.lifetime.start``' intrinsic specifies the start of a memory
8679 The first argument is a constant integer representing the size of the
8680 object, or -1 if it is variable sized. The second argument is a pointer
8686 This intrinsic indicates that before this point in the code, the value
8687 of the memory pointed to by ``ptr`` is dead. This means that it is known
8688 to never be used and has an undefined value. A load from the pointer
8689 that precedes this intrinsic can be replaced with ``'undef'``.
8693 '``llvm.lifetime.end``' Intrinsic
8694 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8701 declare void @llvm.lifetime.end(i64 <size>, i8* nocapture <ptr>)
8706 The '``llvm.lifetime.end``' intrinsic specifies the end of a memory
8712 The first argument is a constant integer representing the size of the
8713 object, or -1 if it is variable sized. The second argument is a pointer
8719 This intrinsic indicates that after this point in the code, the value of
8720 the memory pointed to by ``ptr`` is dead. This means that it is known to
8721 never be used and has an undefined value. Any stores into the memory
8722 object following this intrinsic may be removed as dead.
8724 '``llvm.invariant.start``' Intrinsic
8725 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8732 declare {}* @llvm.invariant.start(i64 <size>, i8* nocapture <ptr>)
8737 The '``llvm.invariant.start``' intrinsic specifies that the contents of
8738 a memory object will not change.
8743 The first argument is a constant integer representing the size of the
8744 object, or -1 if it is variable sized. The second argument is a pointer
8750 This intrinsic indicates that until an ``llvm.invariant.end`` that uses
8751 the return value, the referenced memory location is constant and
8754 '``llvm.invariant.end``' Intrinsic
8755 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8762 declare void @llvm.invariant.end({}* <start>, i64 <size>, i8* nocapture <ptr>)
8767 The '``llvm.invariant.end``' intrinsic specifies that the contents of a
8768 memory object are mutable.
8773 The first argument is the matching ``llvm.invariant.start`` intrinsic.
8774 The second argument is a constant integer representing the size of the
8775 object, or -1 if it is variable sized and the third argument is a
8776 pointer to the object.
8781 This intrinsic indicates that the memory is mutable again.
8786 This class of intrinsics is designed to be generic and has no specific
8789 '``llvm.var.annotation``' Intrinsic
8790 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8797 declare void @llvm.var.annotation(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
8802 The '``llvm.var.annotation``' intrinsic.
8807 The first argument is a pointer to a value, the second is a pointer to a
8808 global string, the third is a pointer to a global string which is the
8809 source file name, and the last argument is the line number.
8814 This intrinsic allows annotation of local variables with arbitrary
8815 strings. This can be useful for special purpose optimizations that want
8816 to look for these annotations. These have no other defined use; they are
8817 ignored by code generation and optimization.
8819 '``llvm.ptr.annotation.*``' Intrinsic
8820 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8825 This is an overloaded intrinsic. You can use '``llvm.ptr.annotation``' on a
8826 pointer to an integer of any width. *NOTE* you must specify an address space for
8827 the pointer. The identifier for the default address space is the integer
8832 declare i8* @llvm.ptr.annotation.p<address space>i8(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
8833 declare i16* @llvm.ptr.annotation.p<address space>i16(i16* <val>, i8* <str>, i8* <str>, i32 <int>)
8834 declare i32* @llvm.ptr.annotation.p<address space>i32(i32* <val>, i8* <str>, i8* <str>, i32 <int>)
8835 declare i64* @llvm.ptr.annotation.p<address space>i64(i64* <val>, i8* <str>, i8* <str>, i32 <int>)
8836 declare i256* @llvm.ptr.annotation.p<address space>i256(i256* <val>, i8* <str>, i8* <str>, i32 <int>)
8841 The '``llvm.ptr.annotation``' intrinsic.
8846 The first argument is a pointer to an integer value of arbitrary bitwidth
8847 (result of some expression), the second is a pointer to a global string, the
8848 third is a pointer to a global string which is the source file name, and the
8849 last argument is the line number. It returns the value of the first argument.
8854 This intrinsic allows annotation of a pointer to an integer with arbitrary
8855 strings. This can be useful for special purpose optimizations that want to look
8856 for these annotations. These have no other defined use; they are ignored by code
8857 generation and optimization.
8859 '``llvm.annotation.*``' Intrinsic
8860 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8865 This is an overloaded intrinsic. You can use '``llvm.annotation``' on
8866 any integer bit width.
8870 declare i8 @llvm.annotation.i8(i8 <val>, i8* <str>, i8* <str>, i32 <int>)
8871 declare i16 @llvm.annotation.i16(i16 <val>, i8* <str>, i8* <str>, i32 <int>)
8872 declare i32 @llvm.annotation.i32(i32 <val>, i8* <str>, i8* <str>, i32 <int>)
8873 declare i64 @llvm.annotation.i64(i64 <val>, i8* <str>, i8* <str>, i32 <int>)
8874 declare i256 @llvm.annotation.i256(i256 <val>, i8* <str>, i8* <str>, i32 <int>)
8879 The '``llvm.annotation``' intrinsic.
8884 The first argument is an integer value (result of some expression), the
8885 second is a pointer to a global string, the third is a pointer to a
8886 global string which is the source file name, and the last argument is
8887 the line number. It returns the value of the first argument.
8892 This intrinsic allows annotations to be put on arbitrary expressions
8893 with arbitrary strings. This can be useful for special purpose
8894 optimizations that want to look for these annotations. These have no
8895 other defined use; they are ignored by code generation and optimization.
8897 '``llvm.trap``' Intrinsic
8898 ^^^^^^^^^^^^^^^^^^^^^^^^^
8905 declare void @llvm.trap() noreturn nounwind
8910 The '``llvm.trap``' intrinsic.
8920 This intrinsic is lowered to the target dependent trap instruction. If
8921 the target does not have a trap instruction, this intrinsic will be
8922 lowered to a call of the ``abort()`` function.
8924 '``llvm.debugtrap``' Intrinsic
8925 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8932 declare void @llvm.debugtrap() nounwind
8937 The '``llvm.debugtrap``' intrinsic.
8947 This intrinsic is lowered to code which is intended to cause an
8948 execution trap with the intention of requesting the attention of a
8951 '``llvm.stackprotector``' Intrinsic
8952 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8959 declare void @llvm.stackprotector(i8* <guard>, i8** <slot>)
8964 The ``llvm.stackprotector`` intrinsic takes the ``guard`` and stores it
8965 onto the stack at ``slot``. The stack slot is adjusted to ensure that it
8966 is placed on the stack before local variables.
8971 The ``llvm.stackprotector`` intrinsic requires two pointer arguments.
8972 The first argument is the value loaded from the stack guard
8973 ``@__stack_chk_guard``. The second variable is an ``alloca`` that has
8974 enough space to hold the value of the guard.
8979 This intrinsic causes the prologue/epilogue inserter to force the position of
8980 the ``AllocaInst`` stack slot to be before local variables on the stack. This is
8981 to ensure that if a local variable on the stack is overwritten, it will destroy
8982 the value of the guard. When the function exits, the guard on the stack is
8983 checked against the original guard by ``llvm.stackprotectorcheck``. If they are
8984 different, then ``llvm.stackprotectorcheck`` causes the program to abort by
8985 calling the ``__stack_chk_fail()`` function.
8987 '``llvm.stackprotectorcheck``' Intrinsic
8988 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8995 declare void @llvm.stackprotectorcheck(i8** <guard>)
9000 The ``llvm.stackprotectorcheck`` intrinsic compares ``guard`` against an already
9001 created stack protector and if they are not equal calls the
9002 ``__stack_chk_fail()`` function.
9007 The ``llvm.stackprotectorcheck`` intrinsic requires one pointer argument, the
9008 the variable ``@__stack_chk_guard``.
9013 This intrinsic is provided to perform the stack protector check by comparing
9014 ``guard`` with the stack slot created by ``llvm.stackprotector`` and if the
9015 values do not match call the ``__stack_chk_fail()`` function.
9017 The reason to provide this as an IR level intrinsic instead of implementing it
9018 via other IR operations is that in order to perform this operation at the IR
9019 level without an intrinsic, one would need to create additional basic blocks to
9020 handle the success/failure cases. This makes it difficult to stop the stack
9021 protector check from disrupting sibling tail calls in Codegen. With this
9022 intrinsic, we are able to generate the stack protector basic blocks late in
9023 codegen after the tail call decision has occurred.
9025 '``llvm.objectsize``' Intrinsic
9026 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9033 declare i32 @llvm.objectsize.i32(i8* <object>, i1 <min>)
9034 declare i64 @llvm.objectsize.i64(i8* <object>, i1 <min>)
9039 The ``llvm.objectsize`` intrinsic is designed to provide information to
9040 the optimizers to determine at compile time whether a) an operation
9041 (like memcpy) will overflow a buffer that corresponds to an object, or
9042 b) that a runtime check for overflow isn't necessary. An object in this
9043 context means an allocation of a specific class, structure, array, or
9049 The ``llvm.objectsize`` intrinsic takes two arguments. The first
9050 argument is a pointer to or into the ``object``. The second argument is
9051 a boolean and determines whether ``llvm.objectsize`` returns 0 (if true)
9052 or -1 (if false) when the object size is unknown. The second argument
9053 only accepts constants.
9058 The ``llvm.objectsize`` intrinsic is lowered to a constant representing
9059 the size of the object concerned. If the size cannot be determined at
9060 compile time, ``llvm.objectsize`` returns ``i32/i64 -1 or 0`` (depending
9061 on the ``min`` argument).
9063 '``llvm.expect``' Intrinsic
9064 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9069 This is an overloaded intrinsic. You can use ``llvm.expect`` on any
9074 declare i1 @llvm.expect.i1(i1 <val>, i1 <expected_val>)
9075 declare i32 @llvm.expect.i32(i32 <val>, i32 <expected_val>)
9076 declare i64 @llvm.expect.i64(i64 <val>, i64 <expected_val>)
9081 The ``llvm.expect`` intrinsic provides information about expected (the
9082 most probable) value of ``val``, which can be used by optimizers.
9087 The ``llvm.expect`` intrinsic takes two arguments. The first argument is
9088 a value. The second argument is an expected value, this needs to be a
9089 constant value, variables are not allowed.
9094 This intrinsic is lowered to the ``val``.
9096 '``llvm.donothing``' Intrinsic
9097 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9104 declare void @llvm.donothing() nounwind readnone
9109 The ``llvm.donothing`` intrinsic doesn't perform any operation. It's the
9110 only intrinsic that can be called with an invoke instruction.
9120 This intrinsic does nothing, and it's removed by optimizers and ignored
9123 Stack Map Intrinsics
9124 --------------------
9126 LLVM provides experimental intrinsics to support runtime patching
9127 mechanisms commonly desired in dynamic language JITs. These intrinsics
9128 are described in :doc:`StackMaps`.