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
132 It also shows a convention that we follow in this document. When
133 demonstrating instructions, we will follow an instruction with a comment
134 that defines the type and name of value produced.
142 LLVM programs are composed of ``Module``'s, each of which is a
143 translation unit of the input programs. Each module consists of
144 functions, global variables, and symbol table entries. Modules may be
145 combined together with the LLVM linker, which merges function (and
146 global variable) definitions, resolves forward declarations, and merges
147 symbol table entries. Here is an example of the "hello world" module:
151 ; Declare the string constant as a global constant.
152 @.str = private unnamed_addr constant [13 x i8] c"hello world\0A\00"
154 ; External declaration of the puts function
155 declare i32 @puts(i8* nocapture) nounwind
157 ; Definition of main function
158 define i32 @main() { ; i32()*
159 ; Convert [13 x i8]* to i8 *...
160 %cast210 = getelementptr [13 x i8]* @.str, i64 0, i64 0
162 ; Call puts function to write out the string to stdout.
163 call i32 @puts(i8* %cast210)
168 !1 = metadata !{i32 42}
171 This example is made up of a :ref:`global variable <globalvars>` named
172 "``.str``", an external declaration of the "``puts``" function, a
173 :ref:`function definition <functionstructure>` for "``main``" and
174 :ref:`named metadata <namedmetadatastructure>` "``foo``".
176 In general, a module is made up of a list of global values (where both
177 functions and global variables are global values). Global values are
178 represented by a pointer to a memory location (in this case, a pointer
179 to an array of char, and a pointer to a function), and have one of the
180 following :ref:`linkage types <linkage>`.
187 All Global Variables and Functions have one of the following types of
191 Global values with "``private``" linkage are only directly
192 accessible by objects in the current module. In particular, linking
193 code into a module with an private global value may cause the
194 private to be renamed as necessary to avoid collisions. Because the
195 symbol is private to the module, all references can be updated. This
196 doesn't show up in any symbol table in the object file.
198 Similar to ``private``, but the symbol is passed through the
199 assembler and evaluated by the linker. Unlike normal strong symbols,
200 they are removed by the linker from the final linked image
201 (executable or dynamic library).
202 ``linker_private_weak``
203 Similar to "``linker_private``", but the symbol is weak. Note that
204 ``linker_private_weak`` symbols are subject to coalescing by the
205 linker. The symbols are removed by the linker from the final linked
206 image (executable or dynamic library).
208 Similar to private, but the value shows as a local symbol
209 (``STB_LOCAL`` in the case of ELF) in the object file. This
210 corresponds to the notion of the '``static``' keyword in C.
211 ``available_externally``
212 Globals with "``available_externally``" linkage are never emitted
213 into the object file corresponding to the LLVM module. They exist to
214 allow inlining and other optimizations to take place given knowledge
215 of the definition of the global, which is known to be somewhere
216 outside the module. Globals with ``available_externally`` linkage
217 are allowed to be discarded at will, and are otherwise the same as
218 ``linkonce_odr``. This linkage type is only allowed on definitions,
221 Globals with "``linkonce``" linkage are merged with other globals of
222 the same name when linkage occurs. This can be used to implement
223 some forms of inline functions, templates, or other code which must
224 be generated in each translation unit that uses it, but where the
225 body may be overridden with a more definitive definition later.
226 Unreferenced ``linkonce`` globals are allowed to be discarded. Note
227 that ``linkonce`` linkage does not actually allow the optimizer to
228 inline the body of this function into callers because it doesn't
229 know if this definition of the function is the definitive definition
230 within the program or whether it will be overridden by a stronger
231 definition. To enable inlining and other optimizations, use
232 "``linkonce_odr``" linkage.
234 "``weak``" linkage has the same merging semantics as ``linkonce``
235 linkage, except that unreferenced globals with ``weak`` linkage may
236 not be discarded. This is used for globals that are declared "weak"
239 "``common``" linkage is most similar to "``weak``" linkage, but they
240 are used for tentative definitions in C, such as "``int X;``" at
241 global scope. Symbols with "``common``" linkage are merged in the
242 same way as ``weak symbols``, and they may not be deleted if
243 unreferenced. ``common`` symbols may not have an explicit section,
244 must have a zero initializer, and may not be marked
245 ':ref:`constant <globalvars>`'. Functions and aliases may not have
248 .. _linkage_appending:
251 "``appending``" linkage may only be applied to global variables of
252 pointer to array type. When two global variables with appending
253 linkage are linked together, the two global arrays are appended
254 together. This is the LLVM, typesafe, equivalent of having the
255 system linker append together "sections" with identical names when
258 The semantics of this linkage follow the ELF object file model: the
259 symbol is weak until linked, if not linked, the symbol becomes null
260 instead of being an undefined reference.
261 ``linkonce_odr``, ``weak_odr``
262 Some languages allow differing globals to be merged, such as two
263 functions with different semantics. Other languages, such as
264 ``C++``, ensure that only equivalent globals are ever merged (the
265 "one definition rule" --- "ODR"). Such languages can use the
266 ``linkonce_odr`` and ``weak_odr`` linkage types to indicate that the
267 global will only be merged with equivalent globals. These linkage
268 types are otherwise the same as their non-``odr`` versions.
269 ``linkonce_odr_auto_hide``
270 Similar to "``linkonce_odr``", but nothing in the translation unit
271 takes the address of this definition. For instance, functions that
272 had an inline definition, but the compiler decided not to inline it.
273 ``linkonce_odr_auto_hide`` may have only ``default`` visibility. The
274 symbols are removed by the linker from the final linked image
275 (executable or dynamic library).
277 If none of the above identifiers are used, the global is externally
278 visible, meaning that it participates in linkage and can be used to
279 resolve external symbol references.
281 The next two types of linkage are targeted for Microsoft Windows
282 platform only. They are designed to support importing (exporting)
283 symbols from (to) DLLs (Dynamic Link Libraries).
286 "``dllimport``" linkage causes the compiler to reference a function
287 or variable via a global pointer to a pointer that is set up by the
288 DLL exporting the symbol. On Microsoft Windows targets, the pointer
289 name is formed by combining ``__imp_`` and the function or variable
292 "``dllexport``" linkage causes the compiler to provide a global
293 pointer to a pointer in a DLL, so that it can be referenced with the
294 ``dllimport`` attribute. On Microsoft Windows targets, the pointer
295 name is formed by combining ``__imp_`` and the function or variable
298 For example, since the "``.LC0``" variable is defined to be internal, if
299 another module defined a "``.LC0``" variable and was linked with this
300 one, one of the two would be renamed, preventing a collision. Since
301 "``main``" and "``puts``" are external (i.e., lacking any linkage
302 declarations), they are accessible outside of the current module.
304 It is illegal for a function *declaration* to have any linkage type
305 other than ``external``, ``dllimport`` or ``extern_weak``.
307 Aliases can have only ``external``, ``internal``, ``weak`` or
308 ``weak_odr`` linkages.
315 LLVM :ref:`functions <functionstructure>`, :ref:`calls <i_call>` and
316 :ref:`invokes <i_invoke>` can all have an optional calling convention
317 specified for the call. The calling convention of any pair of dynamic
318 caller/callee must match, or the behavior of the program is undefined.
319 The following calling conventions are supported by LLVM, and more may be
322 "``ccc``" - The C calling convention
323 This calling convention (the default if no other calling convention
324 is specified) matches the target C calling conventions. This calling
325 convention supports varargs function calls and tolerates some
326 mismatch in the declared prototype and implemented declaration of
327 the function (as does normal C).
328 "``fastcc``" - The fast calling convention
329 This calling convention attempts to make calls as fast as possible
330 (e.g. by passing things in registers). This calling convention
331 allows the target to use whatever tricks it wants to produce fast
332 code for the target, without having to conform to an externally
333 specified ABI (Application Binary Interface). `Tail calls can only
334 be optimized when this, the GHC or the HiPE convention is
335 used. <CodeGenerator.html#id80>`_ This calling convention does not
336 support varargs and requires the prototype of all callees to exactly
337 match the prototype of the function definition.
338 "``coldcc``" - The cold calling convention
339 This calling convention attempts to make code in the caller as
340 efficient as possible under the assumption that the call is not
341 commonly executed. As such, these calls often preserve all registers
342 so that the call does not break any live ranges in the caller side.
343 This calling convention does not support varargs and requires the
344 prototype of all callees to exactly match the prototype of the
346 "``cc 10``" - GHC convention
347 This calling convention has been implemented specifically for use by
348 the `Glasgow Haskell Compiler (GHC) <http://www.haskell.org/ghc>`_.
349 It passes everything in registers, going to extremes to achieve this
350 by disabling callee save registers. This calling convention should
351 not be used lightly but only for specific situations such as an
352 alternative to the *register pinning* performance technique often
353 used when implementing functional programming languages. At the
354 moment only X86 supports this convention and it has the following
357 - On *X86-32* only supports up to 4 bit type parameters. No
358 floating point types are supported.
359 - On *X86-64* only supports up to 10 bit type parameters and 6
360 floating point parameters.
362 This calling convention supports `tail call
363 optimization <CodeGenerator.html#id80>`_ but requires both the
364 caller and callee are using it.
365 "``cc 11``" - The HiPE calling convention
366 This calling convention has been implemented specifically for use by
367 the `High-Performance Erlang
368 (HiPE) <http://www.it.uu.se/research/group/hipe/>`_ compiler, *the*
369 native code compiler of the `Ericsson's Open Source Erlang/OTP
370 system <http://www.erlang.org/download.shtml>`_. It uses more
371 registers for argument passing than the ordinary C calling
372 convention and defines no callee-saved registers. The calling
373 convention properly supports `tail call
374 optimization <CodeGenerator.html#id80>`_ but requires that both the
375 caller and the callee use it. It uses a *register pinning*
376 mechanism, similar to GHC's convention, for keeping frequently
377 accessed runtime components pinned to specific hardware registers.
378 At the moment only X86 supports this convention (both 32 and 64
380 "``cc <n>``" - Numbered convention
381 Any calling convention may be specified by number, allowing
382 target-specific calling conventions to be used. Target specific
383 calling conventions start at 64.
385 More calling conventions can be added/defined on an as-needed basis, to
386 support Pascal conventions or any other well-known target-independent
392 All Global Variables and Functions have one of the following visibility
395 "``default``" - Default style
396 On targets that use the ELF object file format, default visibility
397 means that the declaration is visible to other modules and, in
398 shared libraries, means that the declared entity may be overridden.
399 On Darwin, default visibility means that the declaration is visible
400 to other modules. Default visibility corresponds to "external
401 linkage" in the language.
402 "``hidden``" - Hidden style
403 Two declarations of an object with hidden visibility refer to the
404 same object if they are in the same shared object. Usually, hidden
405 visibility indicates that the symbol will not be placed into the
406 dynamic symbol table, so no other module (executable or shared
407 library) can reference it directly.
408 "``protected``" - Protected style
409 On ELF, protected visibility indicates that the symbol will be
410 placed in the dynamic symbol table, but that references within the
411 defining module will bind to the local symbol. That is, the symbol
412 cannot be overridden by another module.
417 LLVM IR allows you to specify name aliases for certain types. This can
418 make it easier to read the IR and make the IR more condensed
419 (particularly when recursive types are involved). An example of a name
424 %mytype = type { %mytype*, i32 }
426 You may give a name to any :ref:`type <typesystem>` except
427 ":ref:`void <t_void>`". Type name aliases may be used anywhere a type is
428 expected with the syntax "%mytype".
430 Note that type names are aliases for the structural type that they
431 indicate, and that you can therefore specify multiple names for the same
432 type. This often leads to confusing behavior when dumping out a .ll
433 file. Since LLVM IR uses structural typing, the name is not part of the
434 type. When printing out LLVM IR, the printer will pick *one name* to
435 render all types of a particular shape. This means that if you have code
436 where two different source types end up having the same LLVM type, that
437 the dumper will sometimes print the "wrong" or unexpected type. This is
438 an important design point and isn't going to change.
445 Global variables define regions of memory allocated at compilation time
446 instead of run-time. Global variables may optionally be initialized, may
447 have an explicit section to be placed in, and may have an optional
448 explicit alignment specified.
450 A variable may be defined as ``thread_local``, which means that it will
451 not be shared by threads (each thread will have a separated copy of the
452 variable). Not all targets support thread-local variables. Optionally, a
453 TLS model may be specified:
456 For variables that are only used within the current shared library.
458 For variables in modules that will not be loaded dynamically.
460 For variables defined in the executable and only used within it.
462 The models correspond to the ELF TLS models; see `ELF Handling For
463 Thread-Local Storage <http://people.redhat.com/drepper/tls.pdf>`_ for
464 more information on under which circumstances the different models may
465 be used. The target may choose a different TLS model if the specified
466 model is not supported, or if a better choice of model can be made.
468 A variable may be defined as a global ``constant``, which indicates that
469 the contents of the variable will **never** be modified (enabling better
470 optimization, allowing the global data to be placed in the read-only
471 section of an executable, etc). Note that variables that need runtime
472 initialization cannot be marked ``constant`` as there is a store to the
475 LLVM explicitly allows *declarations* of global variables to be marked
476 constant, even if the final definition of the global is not. This
477 capability can be used to enable slightly better optimization of the
478 program, but requires the language definition to guarantee that
479 optimizations based on the 'constantness' are valid for the translation
480 units that do not include the definition.
482 As SSA values, global variables define pointer values that are in scope
483 (i.e. they dominate) all basic blocks in the program. Global variables
484 always define a pointer to their "content" type because they describe a
485 region of memory, and all memory objects in LLVM are accessed through
488 Global variables can be marked with ``unnamed_addr`` which indicates
489 that the address is not significant, only the content. Constants marked
490 like this can be merged with other constants if they have the same
491 initializer. Note that a constant with significant address *can* be
492 merged with a ``unnamed_addr`` constant, the result being a constant
493 whose address is significant.
495 A global variable may be declared to reside in a target-specific
496 numbered address space. For targets that support them, address spaces
497 may affect how optimizations are performed and/or what target
498 instructions are used to access the variable. The default address space
499 is zero. The address space qualifier must precede any other attributes.
501 LLVM allows an explicit section to be specified for globals. If the
502 target supports it, it will emit globals to the section specified.
504 By default, global initializers are optimized by assuming that global
505 variables defined within the module are not modified from their
506 initial values before the start of the global initializer. This is
507 true even for variables potentially accessible from outside the
508 module, including those with external linkage or appearing in
509 ``@llvm.used``. This assumption may be suppressed by marking the
510 variable with ``externally_initialized``.
512 An explicit alignment may be specified for a global, which must be a
513 power of 2. If not present, or if the alignment is set to zero, the
514 alignment of the global is set by the target to whatever it feels
515 convenient. If an explicit alignment is specified, the global is forced
516 to have exactly that alignment. Targets and optimizers are not allowed
517 to over-align the global if the global has an assigned section. In this
518 case, the extra alignment could be observable: for example, code could
519 assume that the globals are densely packed in their section and try to
520 iterate over them as an array, alignment padding would break this
523 For example, the following defines a global in a numbered address space
524 with an initializer, section, and alignment:
528 @G = addrspace(5) constant float 1.0, section "foo", align 4
530 The following example defines a thread-local global with the
531 ``initialexec`` TLS model:
535 @G = thread_local(initialexec) global i32 0, align 4
537 .. _functionstructure:
542 LLVM function definitions consist of the "``define``" keyword, an
543 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
544 style <visibility>`, an optional :ref:`calling convention <callingconv>`,
545 an optional ``unnamed_addr`` attribute, a return type, an optional
546 :ref:`parameter attribute <paramattrs>` for the return type, a function
547 name, a (possibly empty) argument list (each with optional :ref:`parameter
548 attributes <paramattrs>`), optional :ref:`function attributes <fnattrs>`,
549 an optional section, an optional alignment, an optional :ref:`garbage
550 collector name <gc>`, an opening curly brace, a list of basic blocks,
551 and a closing curly brace.
553 LLVM function declarations consist of the "``declare``" keyword, an
554 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
555 style <visibility>`, an optional :ref:`calling convention <callingconv>`,
556 an optional ``unnamed_addr`` attribute, a return type, an optional
557 :ref:`parameter attribute <paramattrs>` for the return type, a function
558 name, a possibly empty list of arguments, an optional alignment, and an
559 optional :ref:`garbage collector name <gc>`.
561 A function definition contains a list of basic blocks, forming the CFG
562 (Control Flow Graph) for the function. Each basic block may optionally
563 start with a label (giving the basic block a symbol table entry),
564 contains a list of instructions, and ends with a
565 :ref:`terminator <terminators>` instruction (such as a branch or function
568 The first basic block in a function is special in two ways: it is
569 immediately executed on entrance to the function, and it is not allowed
570 to have predecessor basic blocks (i.e. there can not be any branches to
571 the entry block of a function). Because the block can have no
572 predecessors, it also cannot have any :ref:`PHI nodes <i_phi>`.
574 LLVM allows an explicit section to be specified for functions. If the
575 target supports it, it will emit functions to the section specified.
577 An explicit alignment may be specified for a function. If not present,
578 or if the alignment is set to zero, the alignment of the function is set
579 by the target to whatever it feels convenient. If an explicit alignment
580 is specified, the function is forced to have at least that much
581 alignment. All alignments must be a power of 2.
583 If the ``unnamed_addr`` attribute is given, the address is know to not
584 be significant and two identical functions can be merged.
588 define [linkage] [visibility]
590 <ResultType> @<FunctionName> ([argument list])
591 [fn Attrs] [section "name"] [align N]
597 Aliases act as "second name" for the aliasee value (which can be either
598 function, global variable, another alias or bitcast of global value).
599 Aliases may have an optional :ref:`linkage type <linkage>`, and an optional
600 :ref:`visibility style <visibility>`.
604 @<Name> = alias [Linkage] [Visibility] <AliaseeTy> @<Aliasee>
606 .. _namedmetadatastructure:
611 Named metadata is a collection of metadata. :ref:`Metadata
612 nodes <metadata>` (but not metadata strings) are the only valid
613 operands for a named metadata.
617 ; Some unnamed metadata nodes, which are referenced by the named metadata.
618 !0 = metadata !{metadata !"zero"}
619 !1 = metadata !{metadata !"one"}
620 !2 = metadata !{metadata !"two"}
622 !name = !{!0, !1, !2}
629 The return type and each parameter of a function type may have a set of
630 *parameter attributes* associated with them. Parameter attributes are
631 used to communicate additional information about the result or
632 parameters of a function. Parameter attributes are considered to be part
633 of the function, not of the function type, so functions with different
634 parameter attributes can have the same function type.
636 Parameter attributes are simple keywords that follow the type specified.
637 If multiple parameter attributes are needed, they are space separated.
642 declare i32 @printf(i8* noalias nocapture, ...)
643 declare i32 @atoi(i8 zeroext)
644 declare signext i8 @returns_signed_char()
646 Note that any attributes for the function result (``nounwind``,
647 ``readonly``) come immediately after the argument list.
649 Currently, only the following parameter attributes are defined:
652 This indicates to the code generator that the parameter or return
653 value should be zero-extended to the extent required by the target's
654 ABI (which is usually 32-bits, but is 8-bits for a i1 on x86-64) by
655 the caller (for a parameter) or the callee (for a return value).
657 This indicates to the code generator that the parameter or return
658 value should be sign-extended to the extent required by the target's
659 ABI (which is usually 32-bits) by the caller (for a parameter) or
660 the callee (for a return value).
662 This indicates that this parameter or return value should be treated
663 in a special target-dependent fashion during while emitting code for
664 a function call or return (usually, by putting it in a register as
665 opposed to memory, though some targets use it to distinguish between
666 two different kinds of registers). Use of this attribute is
669 This indicates that the pointer parameter should really be passed by
670 value to the function. The attribute implies that a hidden copy of
671 the pointee is made between the caller and the callee, so the callee
672 is unable to modify the value in the caller. This attribute is only
673 valid on LLVM pointer arguments. It is generally used to pass
674 structs and arrays by value, but is also valid on pointers to
675 scalars. The copy is considered to belong to the caller not the
676 callee (for example, ``readonly`` functions should not write to
677 ``byval`` parameters). This is not a valid attribute for return
680 The byval attribute also supports specifying an alignment with the
681 align attribute. It indicates the alignment of the stack slot to
682 form and the known alignment of the pointer specified to the call
683 site. If the alignment is not specified, then the code generator
684 makes a target-specific assumption.
687 This indicates that the pointer parameter specifies the address of a
688 structure that is the return value of the function in the source
689 program. This pointer must be guaranteed by the caller to be valid:
690 loads and stores to the structure may be assumed by the callee
691 not to trap and to be properly aligned. This may only be applied to
692 the first parameter. This is not a valid attribute for return
695 This indicates that pointer values `*based* <pointeraliasing>` on
696 the argument or return value do not alias pointer values which are
697 not *based* on it, ignoring certain "irrelevant" dependencies. For a
698 call to the parent function, dependencies between memory references
699 from before or after the call and from those during the call are
700 "irrelevant" to the ``noalias`` keyword for the arguments and return
701 value used in that call. The caller shares the responsibility with
702 the callee for ensuring that these requirements are met. For further
703 details, please see the discussion of the NoAlias response in `alias
704 analysis <AliasAnalysis.html#MustMayNo>`_.
706 Note that this definition of ``noalias`` is intentionally similar
707 to the definition of ``restrict`` in C99 for function arguments,
708 though it is slightly weaker.
710 For function return values, C99's ``restrict`` is not meaningful,
711 while LLVM's ``noalias`` is.
713 This indicates that the callee does not make any copies of the
714 pointer that outlive the callee itself. This is not a valid
715 attribute for return values.
720 This indicates that the pointer parameter can be excised using the
721 :ref:`trampoline intrinsics <int_trampoline>`. This is not a valid
722 attribute for return values and can only be applied to one parameter.
725 This indicates that the value of the function always returns the value
726 of the parameter as its return value. This is an optimization hint to
727 the code generator when generating the caller, allowing tail call
728 optimization and omission of register saves and restores in some cases;
729 it is not checked or enforced when generating the callee. The parameter
730 and the function return type must be valid operands for the
731 :ref:`bitcast instruction <i_bitcast>`. This is not a valid attribute for
732 return values and can only be applied to one parameter.
736 Garbage Collector Names
737 -----------------------
739 Each function may specify a garbage collector name, which is simply a
744 define void @f() gc "name" { ... }
746 The compiler declares the supported values of *name*. Specifying a
747 collector which will cause the compiler to alter its output in order to
748 support the named garbage collection algorithm.
755 Attribute groups are groups of attributes that are referenced by objects within
756 the IR. They are important for keeping ``.ll`` files readable, because a lot of
757 functions will use the same set of attributes. In the degenerative case of a
758 ``.ll`` file that corresponds to a single ``.c`` file, the single attribute
759 group will capture the important command line flags used to build that file.
761 An attribute group is a module-level object. To use an attribute group, an
762 object references the attribute group's ID (e.g. ``#37``). An object may refer
763 to more than one attribute group. In that situation, the attributes from the
764 different groups are merged.
766 Here is an example of attribute groups for a function that should always be
767 inlined, has a stack alignment of 4, and which shouldn't use SSE instructions:
771 ; Target-independent attributes:
772 attributes #0 = { alwaysinline alignstack=4 }
774 ; Target-dependent attributes:
775 attributes #1 = { "no-sse" }
777 ; Function @f has attributes: alwaysinline, alignstack=4, and "no-sse".
778 define void @f() #0 #1 { ... }
785 Function attributes are set to communicate additional information about
786 a function. Function attributes are considered to be part of the
787 function, not of the function type, so functions with different function
788 attributes can have the same function type.
790 Function attributes are simple keywords that follow the type specified.
791 If multiple attributes are needed, they are space separated. For
796 define void @f() noinline { ... }
797 define void @f() alwaysinline { ... }
798 define void @f() alwaysinline optsize { ... }
799 define void @f() optsize { ... }
802 This attribute indicates that, when emitting the prologue and
803 epilogue, the backend should forcibly align the stack pointer.
804 Specify the desired alignment, which must be a power of two, in
807 This attribute indicates that the inliner should attempt to inline
808 this function into callers whenever possible, ignoring any active
809 inlining size threshold for this caller.
811 This attribute suppresses lazy symbol binding for the function. This
812 may make calls to the function faster, at the cost of extra program
813 startup time if the function is not called during program startup.
815 This attribute indicates that the source code contained a hint that
816 inlining this function is desirable (such as the "inline" keyword in
817 C/C++). It is just a hint; it imposes no requirements on the
820 This attribute disables prologue / epilogue emission for the
821 function. This can have very system-specific consequences.
823 This indicates that the callee function at a call site is not
824 recognized as a built-in function. LLVM will retain the original call
825 and not replace it with equivalent code based on the semantics of the
826 built-in function. This is only valid at call sites, not on function
827 declarations or definitions.
829 This attribute indicates that calls to the function cannot be
830 duplicated. A call to a ``noduplicate`` function may be moved
831 within its parent function, but may not be duplicated within
834 A function containing a ``noduplicate`` call may still
835 be an inlining candidate, provided that the call is not
836 duplicated by inlining. That implies that the function has
837 internal linkage and only has one call site, so the original
838 call is dead after inlining.
840 This attributes disables implicit floating point instructions.
842 This attribute indicates that the inliner should never inline this
843 function in any situation. This attribute may not be used together
844 with the ``alwaysinline`` attribute.
846 This attribute indicates that the code generator should not use a
847 red zone, even if the target-specific ABI normally permits it.
849 This function attribute indicates that the function never returns
850 normally. This produces undefined behavior at runtime if the
851 function ever does dynamically return.
853 This function attribute indicates that the function never returns
854 with an unwind or exceptional control flow. If the function does
855 unwind, its runtime behavior is undefined.
857 This attribute suggests that optimization passes and code generator
858 passes make choices that keep the code size of this function low,
859 and otherwise do optimizations specifically to reduce code size.
861 This attribute indicates that the function computes its result (or
862 decides to unwind an exception) based strictly on its arguments,
863 without dereferencing any pointer arguments or otherwise accessing
864 any mutable state (e.g. memory, control registers, etc) visible to
865 caller functions. It does not write through any pointer arguments
866 (including ``byval`` arguments) and never changes any state visible
867 to callers. This means that it cannot unwind exceptions by calling
868 the ``C++`` exception throwing methods.
870 This attribute indicates that the function does not write through
871 any pointer arguments (including ``byval`` arguments) or otherwise
872 modify any state (e.g. memory, control registers, etc) visible to
873 caller functions. It may dereference pointer arguments and read
874 state that may be set in the caller. A readonly function always
875 returns the same value (or unwinds an exception identically) when
876 called with the same set of arguments and global state. It cannot
877 unwind an exception by calling the ``C++`` exception throwing
880 This attribute indicates that this function can return twice. The C
881 ``setjmp`` is an example of such a function. The compiler disables
882 some optimizations (like tail calls) in the caller of these
885 This attribute indicates that AddressSanitizer checks
886 (dynamic address safety analysis) are enabled for this function.
888 This attribute indicates that MemorySanitizer checks (dynamic detection
889 of accesses to uninitialized memory) are enabled for this function.
891 This attribute indicates that ThreadSanitizer checks
892 (dynamic thread safety analysis) are enabled for this function.
894 This attribute indicates that the function should emit a stack
895 smashing protector. It is in the form of a "canary" --- a random value
896 placed on the stack before the local variables that's checked upon
897 return from the function to see if it has been overwritten. A
898 heuristic is used to determine if a function needs stack protectors
899 or not. The heuristic used will enable protectors for functions with:
901 - Character arrays larger than ``ssp-buffer-size`` (default 8).
902 - Aggregates containing character arrays larger than ``ssp-buffer-size``.
903 - Calls to alloca() with variable sizes or constant sizes greater than
906 If a function that has an ``ssp`` attribute is inlined into a
907 function that doesn't have an ``ssp`` attribute, then the resulting
908 function will have an ``ssp`` attribute.
910 This attribute indicates that the function should *always* emit a
911 stack smashing protector. This overrides the ``ssp`` function
914 If a function that has an ``sspreq`` attribute is inlined into a
915 function that doesn't have an ``sspreq`` attribute or which has an
916 ``ssp`` or ``sspstrong`` attribute, then the resulting function will have
917 an ``sspreq`` attribute.
919 This attribute indicates that the function should emit a stack smashing
920 protector. This attribute causes a strong heuristic to be used when
921 determining if a function needs stack protectors. The strong heuristic
922 will enable protectors for functions with:
924 - Arrays of any size and type
925 - Aggregates containing an array of any size and type.
927 - Local variables that have had their address taken.
929 This overrides the ``ssp`` function attribute.
931 If a function that has an ``sspstrong`` attribute is inlined into a
932 function that doesn't have an ``sspstrong`` attribute, then the
933 resulting function will have an ``sspstrong`` attribute.
935 This attribute indicates that the ABI being targeted requires that
936 an unwind table entry be produce for this function even if we can
937 show that no exceptions passes by it. This is normally the case for
938 the ELF x86-64 abi, but it can be disabled for some compilation
943 Module-Level Inline Assembly
944 ----------------------------
946 Modules may contain "module-level inline asm" blocks, which corresponds
947 to the GCC "file scope inline asm" blocks. These blocks are internally
948 concatenated by LLVM and treated as a single unit, but may be separated
949 in the ``.ll`` file if desired. The syntax is very simple:
953 module asm "inline asm code goes here"
954 module asm "more can go here"
956 The strings can contain any character by escaping non-printable
957 characters. The escape sequence used is simply "\\xx" where "xx" is the
958 two digit hex code for the number.
960 The inline asm code is simply printed to the machine code .s file when
961 assembly code is generated.
966 A module may specify a target specific data layout string that specifies
967 how data is to be laid out in memory. The syntax for the data layout is
972 target datalayout = "layout specification"
974 The *layout specification* consists of a list of specifications
975 separated by the minus sign character ('-'). Each specification starts
976 with a letter and may include other information after the letter to
977 define some aspect of the data layout. The specifications accepted are
981 Specifies that the target lays out data in big-endian form. That is,
982 the bits with the most significance have the lowest address
985 Specifies that the target lays out data in little-endian form. That
986 is, the bits with the least significance have the lowest address
989 Specifies the natural alignment of the stack in bits. Alignment
990 promotion of stack variables is limited to the natural stack
991 alignment to avoid dynamic stack realignment. The stack alignment
992 must be a multiple of 8-bits. If omitted, the natural stack
993 alignment defaults to "unspecified", which does not prevent any
994 alignment promotions.
995 ``p[n]:<size>:<abi>:<pref>``
996 This specifies the *size* of a pointer and its ``<abi>`` and
997 ``<pref>``\erred alignments for address space ``n``. All sizes are in
998 bits. Specifying the ``<pref>`` alignment is optional. If omitted, the
999 preceding ``:`` should be omitted too. The address space, ``n`` is
1000 optional, and if not specified, denotes the default address space 0.
1001 The value of ``n`` must be in the range [1,2^23).
1002 ``i<size>:<abi>:<pref>``
1003 This specifies the alignment for an integer type of a given bit
1004 ``<size>``. The value of ``<size>`` must be in the range [1,2^23).
1005 ``v<size>:<abi>:<pref>``
1006 This specifies the alignment for a vector type of a given bit
1008 ``f<size>:<abi>:<pref>``
1009 This specifies the alignment for a floating point type of a given bit
1010 ``<size>``. Only values of ``<size>`` that are supported by the target
1011 will work. 32 (float) and 64 (double) are supported on all targets; 80
1012 or 128 (different flavors of long double) are also supported on some
1014 ``a<size>:<abi>:<pref>``
1015 This specifies the alignment for an aggregate type of a given bit
1017 ``s<size>:<abi>:<pref>``
1018 This specifies the alignment for a stack object of a given bit
1020 ``n<size1>:<size2>:<size3>...``
1021 This specifies a set of native integer widths for the target CPU in
1022 bits. For example, it might contain ``n32`` for 32-bit PowerPC,
1023 ``n32:64`` for PowerPC 64, or ``n8:16:32:64`` for X86-64. Elements of
1024 this set are considered to support most general arithmetic operations
1027 When constructing the data layout for a given target, LLVM starts with a
1028 default set of specifications which are then (possibly) overridden by
1029 the specifications in the ``datalayout`` keyword. The default
1030 specifications are given in this list:
1032 - ``E`` - big endian
1033 - ``p:64:64:64`` - 64-bit pointers with 64-bit alignment
1034 - ``S0`` - natural stack alignment is unspecified
1035 - ``i1:8:8`` - i1 is 8-bit (byte) aligned
1036 - ``i8:8:8`` - i8 is 8-bit (byte) aligned
1037 - ``i16:16:16`` - i16 is 16-bit aligned
1038 - ``i32:32:32`` - i32 is 32-bit aligned
1039 - ``i64:32:64`` - i64 has ABI alignment of 32-bits but preferred
1040 alignment of 64-bits
1041 - ``f16:16:16`` - half is 16-bit aligned
1042 - ``f32:32:32`` - float is 32-bit aligned
1043 - ``f64:64:64`` - double is 64-bit aligned
1044 - ``f128:128:128`` - quad is 128-bit aligned
1045 - ``v64:64:64`` - 64-bit vector is 64-bit aligned
1046 - ``v128:128:128`` - 128-bit vector is 128-bit aligned
1047 - ``a0:0:64`` - aggregates are 64-bit aligned
1049 When LLVM is determining the alignment for a given type, it uses the
1052 #. If the type sought is an exact match for one of the specifications,
1053 that specification is used.
1054 #. If no match is found, and the type sought is an integer type, then
1055 the smallest integer type that is larger than the bitwidth of the
1056 sought type is used. If none of the specifications are larger than
1057 the bitwidth then the largest integer type is used. For example,
1058 given the default specifications above, the i7 type will use the
1059 alignment of i8 (next largest) while both i65 and i256 will use the
1060 alignment of i64 (largest specified).
1061 #. If no match is found, and the type sought is a vector type, then the
1062 largest vector type that is smaller than the sought vector type will
1063 be used as a fall back. This happens because <128 x double> can be
1064 implemented in terms of 64 <2 x double>, for example.
1066 The function of the data layout string may not be what you expect.
1067 Notably, this is not a specification from the frontend of what alignment
1068 the code generator should use.
1070 Instead, if specified, the target data layout is required to match what
1071 the ultimate *code generator* expects. This string is used by the
1072 mid-level optimizers to improve code, and this only works if it matches
1073 what the ultimate code generator uses. If you would like to generate IR
1074 that does not embed this target-specific detail into the IR, then you
1075 don't have to specify the string. This will disable some optimizations
1076 that require precise layout information, but this also prevents those
1077 optimizations from introducing target specificity into the IR.
1079 .. _pointeraliasing:
1081 Pointer Aliasing Rules
1082 ----------------------
1084 Any memory access must be done through a pointer value associated with
1085 an address range of the memory access, otherwise the behavior is
1086 undefined. Pointer values are associated with address ranges according
1087 to the following rules:
1089 - A pointer value is associated with the addresses associated with any
1090 value it is *based* on.
1091 - An address of a global variable is associated with the address range
1092 of the variable's storage.
1093 - The result value of an allocation instruction is associated with the
1094 address range of the allocated storage.
1095 - A null pointer in the default address-space is associated with no
1097 - An integer constant other than zero or a pointer value returned from
1098 a function not defined within LLVM may be associated with address
1099 ranges allocated through mechanisms other than those provided by
1100 LLVM. Such ranges shall not overlap with any ranges of addresses
1101 allocated by mechanisms provided by LLVM.
1103 A pointer value is *based* on another pointer value according to the
1106 - A pointer value formed from a ``getelementptr`` operation is *based*
1107 on the first operand of the ``getelementptr``.
1108 - The result value of a ``bitcast`` is *based* on the operand of the
1110 - A pointer value formed by an ``inttoptr`` is *based* on all pointer
1111 values that contribute (directly or indirectly) to the computation of
1112 the pointer's value.
1113 - The "*based* on" relationship is transitive.
1115 Note that this definition of *"based"* is intentionally similar to the
1116 definition of *"based"* in C99, though it is slightly weaker.
1118 LLVM IR does not associate types with memory. The result type of a
1119 ``load`` merely indicates the size and alignment of the memory from
1120 which to load, as well as the interpretation of the value. The first
1121 operand type of a ``store`` similarly only indicates the size and
1122 alignment of the store.
1124 Consequently, type-based alias analysis, aka TBAA, aka
1125 ``-fstrict-aliasing``, is not applicable to general unadorned LLVM IR.
1126 :ref:`Metadata <metadata>` may be used to encode additional information
1127 which specialized optimization passes may use to implement type-based
1132 Volatile Memory Accesses
1133 ------------------------
1135 Certain memory accesses, such as :ref:`load <i_load>`'s,
1136 :ref:`store <i_store>`'s, and :ref:`llvm.memcpy <int_memcpy>`'s may be
1137 marked ``volatile``. The optimizers must not change the number of
1138 volatile operations or change their order of execution relative to other
1139 volatile operations. The optimizers *may* change the order of volatile
1140 operations relative to non-volatile operations. This is not Java's
1141 "volatile" and has no cross-thread synchronization behavior.
1143 IR-level volatile loads and stores cannot safely be optimized into
1144 llvm.memcpy or llvm.memmove intrinsics even when those intrinsics are
1145 flagged volatile. Likewise, the backend should never split or merge
1146 target-legal volatile load/store instructions.
1148 .. admonition:: Rationale
1150 Platforms may rely on volatile loads and stores of natively supported
1151 data width to be executed as single instruction. For example, in C
1152 this holds for an l-value of volatile primitive type with native
1153 hardware support, but not necessarily for aggregate types. The
1154 frontend upholds these expectations, which are intentionally
1155 unspecified in the IR. The rules above ensure that IR transformation
1156 do not violate the frontend's contract with the language.
1160 Memory Model for Concurrent Operations
1161 --------------------------------------
1163 The LLVM IR does not define any way to start parallel threads of
1164 execution or to register signal handlers. Nonetheless, there are
1165 platform-specific ways to create them, and we define LLVM IR's behavior
1166 in their presence. This model is inspired by the C++0x memory model.
1168 For a more informal introduction to this model, see the :doc:`Atomics`.
1170 We define a *happens-before* partial order as the least partial order
1173 - Is a superset of single-thread program order, and
1174 - When a *synchronizes-with* ``b``, includes an edge from ``a`` to
1175 ``b``. *Synchronizes-with* pairs are introduced by platform-specific
1176 techniques, like pthread locks, thread creation, thread joining,
1177 etc., and by atomic instructions. (See also :ref:`Atomic Memory Ordering
1178 Constraints <ordering>`).
1180 Note that program order does not introduce *happens-before* edges
1181 between a thread and signals executing inside that thread.
1183 Every (defined) read operation (load instructions, memcpy, atomic
1184 loads/read-modify-writes, etc.) R reads a series of bytes written by
1185 (defined) write operations (store instructions, atomic
1186 stores/read-modify-writes, memcpy, etc.). For the purposes of this
1187 section, initialized globals are considered to have a write of the
1188 initializer which is atomic and happens before any other read or write
1189 of the memory in question. For each byte of a read R, R\ :sub:`byte`
1190 may see any write to the same byte, except:
1192 - If write\ :sub:`1` happens before write\ :sub:`2`, and
1193 write\ :sub:`2` happens before R\ :sub:`byte`, then
1194 R\ :sub:`byte` does not see write\ :sub:`1`.
1195 - If R\ :sub:`byte` happens before write\ :sub:`3`, then
1196 R\ :sub:`byte` does not see write\ :sub:`3`.
1198 Given that definition, R\ :sub:`byte` is defined as follows:
1200 - If R is volatile, the result is target-dependent. (Volatile is
1201 supposed to give guarantees which can support ``sig_atomic_t`` in
1202 C/C++, and may be used for accesses to addresses which do not behave
1203 like normal memory. It does not generally provide cross-thread
1205 - Otherwise, if there is no write to the same byte that happens before
1206 R\ :sub:`byte`, R\ :sub:`byte` returns ``undef`` for that byte.
1207 - Otherwise, if R\ :sub:`byte` may see exactly one write,
1208 R\ :sub:`byte` returns the value written by that write.
1209 - Otherwise, if R is atomic, and all the writes R\ :sub:`byte` may
1210 see are atomic, it chooses one of the values written. See the :ref:`Atomic
1211 Memory Ordering Constraints <ordering>` section for additional
1212 constraints on how the choice is made.
1213 - Otherwise R\ :sub:`byte` returns ``undef``.
1215 R returns the value composed of the series of bytes it read. This
1216 implies that some bytes within the value may be ``undef`` **without**
1217 the entire value being ``undef``. Note that this only defines the
1218 semantics of the operation; it doesn't mean that targets will emit more
1219 than one instruction to read the series of bytes.
1221 Note that in cases where none of the atomic intrinsics are used, this
1222 model places only one restriction on IR transformations on top of what
1223 is required for single-threaded execution: introducing a store to a byte
1224 which might not otherwise be stored is not allowed in general.
1225 (Specifically, in the case where another thread might write to and read
1226 from an address, introducing a store can change a load that may see
1227 exactly one write into a load that may see multiple writes.)
1231 Atomic Memory Ordering Constraints
1232 ----------------------------------
1234 Atomic instructions (:ref:`cmpxchg <i_cmpxchg>`,
1235 :ref:`atomicrmw <i_atomicrmw>`, :ref:`fence <i_fence>`,
1236 :ref:`atomic load <i_load>`, and :ref:`atomic store <i_store>`) take
1237 an ordering parameter that determines which other atomic instructions on
1238 the same address they *synchronize with*. These semantics are borrowed
1239 from Java and C++0x, but are somewhat more colloquial. If these
1240 descriptions aren't precise enough, check those specs (see spec
1241 references in the :doc:`atomics guide <Atomics>`).
1242 :ref:`fence <i_fence>` instructions treat these orderings somewhat
1243 differently since they don't take an address. See that instruction's
1244 documentation for details.
1246 For a simpler introduction to the ordering constraints, see the
1250 The set of values that can be read is governed by the happens-before
1251 partial order. A value cannot be read unless some operation wrote
1252 it. This is intended to provide a guarantee strong enough to model
1253 Java's non-volatile shared variables. This ordering cannot be
1254 specified for read-modify-write operations; it is not strong enough
1255 to make them atomic in any interesting way.
1257 In addition to the guarantees of ``unordered``, there is a single
1258 total order for modifications by ``monotonic`` operations on each
1259 address. All modification orders must be compatible with the
1260 happens-before order. There is no guarantee that the modification
1261 orders can be combined to a global total order for the whole program
1262 (and this often will not be possible). The read in an atomic
1263 read-modify-write operation (:ref:`cmpxchg <i_cmpxchg>` and
1264 :ref:`atomicrmw <i_atomicrmw>`) reads the value in the modification
1265 order immediately before the value it writes. If one atomic read
1266 happens before another atomic read of the same address, the later
1267 read must see the same value or a later value in the address's
1268 modification order. This disallows reordering of ``monotonic`` (or
1269 stronger) operations on the same address. If an address is written
1270 ``monotonic``-ally by one thread, and other threads ``monotonic``-ally
1271 read that address repeatedly, the other threads must eventually see
1272 the write. This corresponds to the C++0x/C1x
1273 ``memory_order_relaxed``.
1275 In addition to the guarantees of ``monotonic``, a
1276 *synchronizes-with* edge may be formed with a ``release`` operation.
1277 This is intended to model C++'s ``memory_order_acquire``.
1279 In addition to the guarantees of ``monotonic``, if this operation
1280 writes a value which is subsequently read by an ``acquire``
1281 operation, it *synchronizes-with* that operation. (This isn't a
1282 complete description; see the C++0x definition of a release
1283 sequence.) This corresponds to the C++0x/C1x
1284 ``memory_order_release``.
1285 ``acq_rel`` (acquire+release)
1286 Acts as both an ``acquire`` and ``release`` operation on its
1287 address. This corresponds to the C++0x/C1x ``memory_order_acq_rel``.
1288 ``seq_cst`` (sequentially consistent)
1289 In addition to the guarantees of ``acq_rel`` (``acquire`` for an
1290 operation which only reads, ``release`` for an operation which only
1291 writes), there is a global total order on all
1292 sequentially-consistent operations on all addresses, which is
1293 consistent with the *happens-before* partial order and with the
1294 modification orders of all the affected addresses. Each
1295 sequentially-consistent read sees the last preceding write to the
1296 same address in this global order. This corresponds to the C++0x/C1x
1297 ``memory_order_seq_cst`` and Java volatile.
1301 If an atomic operation is marked ``singlethread``, it only *synchronizes
1302 with* or participates in modification and seq\_cst total orderings with
1303 other operations running in the same thread (for example, in signal
1311 LLVM IR floating-point binary ops (:ref:`fadd <i_fadd>`,
1312 :ref:`fsub <i_fsub>`, :ref:`fmul <i_fmul>`, :ref:`fdiv <i_fdiv>`,
1313 :ref:`frem <i_frem>`) have the following flags that can set to enable
1314 otherwise unsafe floating point operations
1317 No NaNs - Allow optimizations to assume the arguments and result are not
1318 NaN. Such optimizations are required to retain defined behavior over
1319 NaNs, but the value of the result is undefined.
1322 No Infs - Allow optimizations to assume the arguments and result are not
1323 +/-Inf. Such optimizations are required to retain defined behavior over
1324 +/-Inf, but the value of the result is undefined.
1327 No Signed Zeros - Allow optimizations to treat the sign of a zero
1328 argument or result as insignificant.
1331 Allow Reciprocal - Allow optimizations to use the reciprocal of an
1332 argument rather than perform division.
1335 Fast - Allow algebraically equivalent transformations that may
1336 dramatically change results in floating point (e.g. reassociate). This
1337 flag implies all the others.
1344 The LLVM type system is one of the most important features of the
1345 intermediate representation. Being typed enables a number of
1346 optimizations to be performed on the intermediate representation
1347 directly, without having to do extra analyses on the side before the
1348 transformation. A strong type system makes it easier to read the
1349 generated code and enables novel analyses and transformations that are
1350 not feasible to perform on normal three address code representations.
1352 Type Classifications
1353 --------------------
1355 The types fall into a few useful classifications:
1364 * - :ref:`integer <t_integer>`
1365 - ``i1``, ``i2``, ``i3``, ... ``i8``, ... ``i16``, ... ``i32``, ...
1368 * - :ref:`floating point <t_floating>`
1369 - ``half``, ``float``, ``double``, ``x86_fp80``, ``fp128``,
1377 - :ref:`integer <t_integer>`, :ref:`floating point <t_floating>`,
1378 :ref:`pointer <t_pointer>`, :ref:`vector <t_vector>`,
1379 :ref:`structure <t_struct>`, :ref:`array <t_array>`,
1380 :ref:`label <t_label>`, :ref:`metadata <t_metadata>`.
1382 * - :ref:`primitive <t_primitive>`
1383 - :ref:`label <t_label>`,
1384 :ref:`void <t_void>`,
1385 :ref:`integer <t_integer>`,
1386 :ref:`floating point <t_floating>`,
1387 :ref:`x86mmx <t_x86mmx>`,
1388 :ref:`metadata <t_metadata>`.
1390 * - :ref:`derived <t_derived>`
1391 - :ref:`array <t_array>`,
1392 :ref:`function <t_function>`,
1393 :ref:`pointer <t_pointer>`,
1394 :ref:`structure <t_struct>`,
1395 :ref:`vector <t_vector>`,
1396 :ref:`opaque <t_opaque>`.
1398 The :ref:`first class <t_firstclass>` types are perhaps the most important.
1399 Values of these types are the only ones which can be produced by
1407 The primitive types are the fundamental building blocks of the LLVM
1418 The integer type is a very simple type that simply specifies an
1419 arbitrary bit width for the integer type desired. Any bit width from 1
1420 bit to 2\ :sup:`23`\ -1 (about 8 million) can be specified.
1429 The number of bits the integer will occupy is specified by the ``N``
1435 +----------------+------------------------------------------------+
1436 | ``i1`` | a single-bit integer. |
1437 +----------------+------------------------------------------------+
1438 | ``i32`` | a 32-bit integer. |
1439 +----------------+------------------------------------------------+
1440 | ``i1942652`` | a really big integer of over 1 million bits. |
1441 +----------------+------------------------------------------------+
1445 Floating Point Types
1446 ^^^^^^^^^^^^^^^^^^^^
1455 - 16-bit floating point value
1458 - 32-bit floating point value
1461 - 64-bit floating point value
1464 - 128-bit floating point value (112-bit mantissa)
1467 - 80-bit floating point value (X87)
1470 - 128-bit floating point value (two 64-bits)
1480 The x86mmx type represents a value held in an MMX register on an x86
1481 machine. The operations allowed on it are quite limited: parameters and
1482 return values, load and store, and bitcast. User-specified MMX
1483 instructions are represented as intrinsic or asm calls with arguments
1484 and/or results of this type. There are no arrays, vectors or constants
1502 The void type does not represent any value and has no size.
1519 The label type represents code labels.
1536 The metadata type represents embedded metadata. No derived types may be
1537 created from metadata except for :ref:`function <t_function>` arguments.
1551 The real power in LLVM comes from the derived types in the system. This
1552 is what allows a programmer to represent arrays, functions, pointers,
1553 and other useful types. Each of these types contain one or more element
1554 types which may be a primitive type, or another derived type. For
1555 example, it is possible to have a two dimensional array, using an array
1556 as the element type of another array.
1563 Aggregate Types are a subset of derived types that can contain multiple
1564 member types. :ref:`Arrays <t_array>` and :ref:`structs <t_struct>` are
1565 aggregate types. :ref:`Vectors <t_vector>` are not considered to be
1576 The array type is a very simple derived type that arranges elements
1577 sequentially in memory. The array type requires a size (number of
1578 elements) and an underlying data type.
1585 [<# elements> x <elementtype>]
1587 The number of elements is a constant integer value; ``elementtype`` may
1588 be any type with a size.
1593 +------------------+--------------------------------------+
1594 | ``[40 x i32]`` | Array of 40 32-bit integer values. |
1595 +------------------+--------------------------------------+
1596 | ``[41 x i32]`` | Array of 41 32-bit integer values. |
1597 +------------------+--------------------------------------+
1598 | ``[4 x i8]`` | Array of 4 8-bit integer values. |
1599 +------------------+--------------------------------------+
1601 Here are some examples of multidimensional arrays:
1603 +-----------------------------+----------------------------------------------------------+
1604 | ``[3 x [4 x i32]]`` | 3x4 array of 32-bit integer values. |
1605 +-----------------------------+----------------------------------------------------------+
1606 | ``[12 x [10 x float]]`` | 12x10 array of single precision floating point values. |
1607 +-----------------------------+----------------------------------------------------------+
1608 | ``[2 x [3 x [4 x i16]]]`` | 2x3x4 array of 16-bit integer values. |
1609 +-----------------------------+----------------------------------------------------------+
1611 There is no restriction on indexing beyond the end of the array implied
1612 by a static type (though there are restrictions on indexing beyond the
1613 bounds of an allocated object in some cases). This means that
1614 single-dimension 'variable sized array' addressing can be implemented in
1615 LLVM with a zero length array type. An implementation of 'pascal style
1616 arrays' in LLVM could use the type "``{ i32, [0 x float]}``", for
1627 The function type can be thought of as a function signature. It consists
1628 of a return type and a list of formal parameter types. The return type
1629 of a function type is a first class type or a void type.
1636 <returntype> (<parameter list>)
1638 ...where '``<parameter list>``' is a comma-separated list of type
1639 specifiers. Optionally, the parameter list may include a type ``...``,
1640 which indicates that the function takes a variable number of arguments.
1641 Variable argument functions can access their arguments with the
1642 :ref:`variable argument handling intrinsic <int_varargs>` functions.
1643 '``<returntype>``' is any type except :ref:`label <t_label>`.
1648 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1649 | ``i32 (i32)`` | function taking an ``i32``, returning an ``i32`` |
1650 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1651 | ``float (i16, i32 *) *`` | :ref:`Pointer <t_pointer>` to a function that takes an ``i16`` and a :ref:`pointer <t_pointer>` to ``i32``, returning ``float``. |
1652 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1653 | ``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. |
1654 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1655 | ``{i32, i32} (i32)`` | A function taking an ``i32``, returning a :ref:`structure <t_struct>` containing two ``i32`` values |
1656 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1666 The structure type is used to represent a collection of data members
1667 together in memory. The elements of a structure may be any type that has
1670 Structures in memory are accessed using '``load``' and '``store``' by
1671 getting a pointer to a field with the '``getelementptr``' instruction.
1672 Structures in registers are accessed using the '``extractvalue``' and
1673 '``insertvalue``' instructions.
1675 Structures may optionally be "packed" structures, which indicate that
1676 the alignment of the struct is one byte, and that there is no padding
1677 between the elements. In non-packed structs, padding between field types
1678 is inserted as defined by the DataLayout string in the module, which is
1679 required to match what the underlying code generator expects.
1681 Structures can either be "literal" or "identified". A literal structure
1682 is defined inline with other types (e.g. ``{i32, i32}*``) whereas
1683 identified types are always defined at the top level with a name.
1684 Literal types are uniqued by their contents and can never be recursive
1685 or opaque since there is no way to write one. Identified types can be
1686 recursive, can be opaqued, and are never uniqued.
1693 %T1 = type { <type list> } ; Identified normal struct type
1694 %T2 = type <{ <type list> }> ; Identified packed struct type
1699 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1700 | ``{ i32, i32, i32 }`` | A triple of three ``i32`` values |
1701 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1702 | ``{ 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``. |
1703 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1704 | ``<{ i8, i32 }>`` | A packed struct known to be 5 bytes in size. |
1705 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1709 Opaque Structure Types
1710 ^^^^^^^^^^^^^^^^^^^^^^
1715 Opaque structure types are used to represent named structure types that
1716 do not have a body specified. This corresponds (for example) to the C
1717 notion of a forward declared structure.
1730 +--------------+-------------------+
1731 | ``opaque`` | An opaque type. |
1732 +--------------+-------------------+
1742 The pointer type is used to specify memory locations. Pointers are
1743 commonly used to reference objects in memory.
1745 Pointer types may have an optional address space attribute defining the
1746 numbered address space where the pointed-to object resides. The default
1747 address space is number zero. The semantics of non-zero address spaces
1748 are target-specific.
1750 Note that LLVM does not permit pointers to void (``void*``) nor does it
1751 permit pointers to labels (``label*``). Use ``i8*`` instead.
1763 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1764 | ``[4 x i32]*`` | A :ref:`pointer <t_pointer>` to :ref:`array <t_array>` of four ``i32`` values. |
1765 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1766 | ``i32 (i32*) *`` | A :ref:`pointer <t_pointer>` to a :ref:`function <t_function>` that takes an ``i32*``, returning an ``i32``. |
1767 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1768 | ``i32 addrspace(5)*`` | A :ref:`pointer <t_pointer>` to an ``i32`` value that resides in address space #5. |
1769 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1779 A vector type is a simple derived type that represents a vector of
1780 elements. Vector types are used when multiple primitive data are
1781 operated in parallel using a single instruction (SIMD). A vector type
1782 requires a size (number of elements) and an underlying primitive data
1783 type. Vector types are considered :ref:`first class <t_firstclass>`.
1790 < <# elements> x <elementtype> >
1792 The number of elements is a constant integer value larger than 0;
1793 elementtype may be any integer or floating point type, or a pointer to
1794 these types. Vectors of size zero are not allowed.
1799 +-------------------+--------------------------------------------------+
1800 | ``<4 x i32>`` | Vector of 4 32-bit integer values. |
1801 +-------------------+--------------------------------------------------+
1802 | ``<8 x float>`` | Vector of 8 32-bit floating-point values. |
1803 +-------------------+--------------------------------------------------+
1804 | ``<2 x i64>`` | Vector of 2 64-bit integer values. |
1805 +-------------------+--------------------------------------------------+
1806 | ``<4 x i64*>`` | Vector of 4 pointers to 64-bit integer values. |
1807 +-------------------+--------------------------------------------------+
1812 LLVM has several different basic types of constants. This section
1813 describes them all and their syntax.
1818 **Boolean constants**
1819 The two strings '``true``' and '``false``' are both valid constants
1821 **Integer constants**
1822 Standard integers (such as '4') are constants of the
1823 :ref:`integer <t_integer>` type. Negative numbers may be used with
1825 **Floating point constants**
1826 Floating point constants use standard decimal notation (e.g.
1827 123.421), exponential notation (e.g. 1.23421e+2), or a more precise
1828 hexadecimal notation (see below). The assembler requires the exact
1829 decimal value of a floating-point constant. For example, the
1830 assembler accepts 1.25 but rejects 1.3 because 1.3 is a repeating
1831 decimal in binary. Floating point constants must have a :ref:`floating
1832 point <t_floating>` type.
1833 **Null pointer constants**
1834 The identifier '``null``' is recognized as a null pointer constant
1835 and must be of :ref:`pointer type <t_pointer>`.
1837 The one non-intuitive notation for constants is the hexadecimal form of
1838 floating point constants. For example, the form
1839 '``double 0x432ff973cafa8000``' is equivalent to (but harder to read
1840 than) '``double 4.5e+15``'. The only time hexadecimal floating point
1841 constants are required (and the only time that they are generated by the
1842 disassembler) is when a floating point constant must be emitted but it
1843 cannot be represented as a decimal floating point number in a reasonable
1844 number of digits. For example, NaN's, infinities, and other special
1845 values are represented in their IEEE hexadecimal format so that assembly
1846 and disassembly do not cause any bits to change in the constants.
1848 When using the hexadecimal form, constants of types half, float, and
1849 double are represented using the 16-digit form shown above (which
1850 matches the IEEE754 representation for double); half and float values
1851 must, however, be exactly representable as IEEE 754 half and single
1852 precision, respectively. Hexadecimal format is always used for long
1853 double, and there are three forms of long double. The 80-bit format used
1854 by x86 is represented as ``0xK`` followed by 20 hexadecimal digits. The
1855 128-bit format used by PowerPC (two adjacent doubles) is represented by
1856 ``0xM`` followed by 32 hexadecimal digits. The IEEE 128-bit format is
1857 represented by ``0xL`` followed by 32 hexadecimal digits. Long doubles
1858 will only work if they match the long double format on your target.
1859 The IEEE 16-bit format (half precision) is represented by ``0xH``
1860 followed by 4 hexadecimal digits. All hexadecimal formats are big-endian
1861 (sign bit at the left).
1863 There are no constants of type x86mmx.
1868 Complex constants are a (potentially recursive) combination of simple
1869 constants and smaller complex constants.
1871 **Structure constants**
1872 Structure constants are represented with notation similar to
1873 structure type definitions (a comma separated list of elements,
1874 surrounded by braces (``{}``)). For example:
1875 "``{ i32 4, float 17.0, i32* @G }``", where "``@G``" is declared as
1876 "``@G = external global i32``". Structure constants must have
1877 :ref:`structure type <t_struct>`, and the number and types of elements
1878 must match those specified by the type.
1880 Array constants are represented with notation similar to array type
1881 definitions (a comma separated list of elements, surrounded by
1882 square brackets (``[]``)). For example:
1883 "``[ i32 42, i32 11, i32 74 ]``". Array constants must have
1884 :ref:`array type <t_array>`, and the number and types of elements must
1885 match those specified by the type.
1886 **Vector constants**
1887 Vector constants are represented with notation similar to vector
1888 type definitions (a comma separated list of elements, surrounded by
1889 less-than/greater-than's (``<>``)). For example:
1890 "``< i32 42, i32 11, i32 74, i32 100 >``". Vector constants
1891 must have :ref:`vector type <t_vector>`, and the number and types of
1892 elements must match those specified by the type.
1893 **Zero initialization**
1894 The string '``zeroinitializer``' can be used to zero initialize a
1895 value to zero of *any* type, including scalar and
1896 :ref:`aggregate <t_aggregate>` types. This is often used to avoid
1897 having to print large zero initializers (e.g. for large arrays) and
1898 is always exactly equivalent to using explicit zero initializers.
1900 A metadata node is a structure-like constant with :ref:`metadata
1901 type <t_metadata>`. For example:
1902 "``metadata !{ i32 0, metadata !"test" }``". Unlike other
1903 constants that are meant to be interpreted as part of the
1904 instruction stream, metadata is a place to attach additional
1905 information such as debug info.
1907 Global Variable and Function Addresses
1908 --------------------------------------
1910 The addresses of :ref:`global variables <globalvars>` and
1911 :ref:`functions <functionstructure>` are always implicitly valid
1912 (link-time) constants. These constants are explicitly referenced when
1913 the :ref:`identifier for the global <identifiers>` is used and always have
1914 :ref:`pointer <t_pointer>` type. For example, the following is a legal LLVM
1917 .. code-block:: llvm
1921 @Z = global [2 x i32*] [ i32* @X, i32* @Y ]
1928 The string '``undef``' can be used anywhere a constant is expected, and
1929 indicates that the user of the value may receive an unspecified
1930 bit-pattern. Undefined values may be of any type (other than '``label``'
1931 or '``void``') and be used anywhere a constant is permitted.
1933 Undefined values are useful because they indicate to the compiler that
1934 the program is well defined no matter what value is used. This gives the
1935 compiler more freedom to optimize. Here are some examples of
1936 (potentially surprising) transformations that are valid (in pseudo IR):
1938 .. code-block:: llvm
1948 This is safe because all of the output bits are affected by the undef
1949 bits. Any output bit can have a zero or one depending on the input bits.
1951 .. code-block:: llvm
1962 These logical operations have bits that are not always affected by the
1963 input. For example, if ``%X`` has a zero bit, then the output of the
1964 '``and``' operation will always be a zero for that bit, no matter what
1965 the corresponding bit from the '``undef``' is. As such, it is unsafe to
1966 optimize or assume that the result of the '``and``' is '``undef``'.
1967 However, it is safe to assume that all bits of the '``undef``' could be
1968 0, and optimize the '``and``' to 0. Likewise, it is safe to assume that
1969 all the bits of the '``undef``' operand to the '``or``' could be set,
1970 allowing the '``or``' to be folded to -1.
1972 .. code-block:: llvm
1974 %A = select undef, %X, %Y
1975 %B = select undef, 42, %Y
1976 %C = select %X, %Y, undef
1986 This set of examples shows that undefined '``select``' (and conditional
1987 branch) conditions can go *either way*, but they have to come from one
1988 of the two operands. In the ``%A`` example, if ``%X`` and ``%Y`` were
1989 both known to have a clear low bit, then ``%A`` would have to have a
1990 cleared low bit. However, in the ``%C`` example, the optimizer is
1991 allowed to assume that the '``undef``' operand could be the same as
1992 ``%Y``, allowing the whole '``select``' to be eliminated.
1994 .. code-block:: llvm
1996 %A = xor undef, undef
2013 This example points out that two '``undef``' operands are not
2014 necessarily the same. This can be surprising to people (and also matches
2015 C semantics) where they assume that "``X^X``" is always zero, even if
2016 ``X`` is undefined. This isn't true for a number of reasons, but the
2017 short answer is that an '``undef``' "variable" can arbitrarily change
2018 its value over its "live range". This is true because the variable
2019 doesn't actually *have a live range*. Instead, the value is logically
2020 read from arbitrary registers that happen to be around when needed, so
2021 the value is not necessarily consistent over time. In fact, ``%A`` and
2022 ``%C`` need to have the same semantics or the core LLVM "replace all
2023 uses with" concept would not hold.
2025 .. code-block:: llvm
2033 These examples show the crucial difference between an *undefined value*
2034 and *undefined behavior*. An undefined value (like '``undef``') is
2035 allowed to have an arbitrary bit-pattern. This means that the ``%A``
2036 operation can be constant folded to '``undef``', because the '``undef``'
2037 could be an SNaN, and ``fdiv`` is not (currently) defined on SNaN's.
2038 However, in the second example, we can make a more aggressive
2039 assumption: because the ``undef`` is allowed to be an arbitrary value,
2040 we are allowed to assume that it could be zero. Since a divide by zero
2041 has *undefined behavior*, we are allowed to assume that the operation
2042 does not execute at all. This allows us to delete the divide and all
2043 code after it. Because the undefined operation "can't happen", the
2044 optimizer can assume that it occurs in dead code.
2046 .. code-block:: llvm
2048 a: store undef -> %X
2049 b: store %X -> undef
2054 These examples reiterate the ``fdiv`` example: a store *of* an undefined
2055 value can be assumed to not have any effect; we can assume that the
2056 value is overwritten with bits that happen to match what was already
2057 there. However, a store *to* an undefined location could clobber
2058 arbitrary memory, therefore, it has undefined behavior.
2065 Poison values are similar to :ref:`undef values <undefvalues>`, however
2066 they also represent the fact that an instruction or constant expression
2067 which cannot evoke side effects has nevertheless detected a condition
2068 which results in undefined behavior.
2070 There is currently no way of representing a poison value in the IR; they
2071 only exist when produced by operations such as :ref:`add <i_add>` with
2074 Poison value behavior is defined in terms of value *dependence*:
2076 - Values other than :ref:`phi <i_phi>` nodes depend on their operands.
2077 - :ref:`Phi <i_phi>` nodes depend on the operand corresponding to
2078 their dynamic predecessor basic block.
2079 - Function arguments depend on the corresponding actual argument values
2080 in the dynamic callers of their functions.
2081 - :ref:`Call <i_call>` instructions depend on the :ref:`ret <i_ret>`
2082 instructions that dynamically transfer control back to them.
2083 - :ref:`Invoke <i_invoke>` instructions depend on the
2084 :ref:`ret <i_ret>`, :ref:`resume <i_resume>`, or exception-throwing
2085 call instructions that dynamically transfer control back to them.
2086 - Non-volatile loads and stores depend on the most recent stores to all
2087 of the referenced memory addresses, following the order in the IR
2088 (including loads and stores implied by intrinsics such as
2089 :ref:`@llvm.memcpy <int_memcpy>`.)
2090 - An instruction with externally visible side effects depends on the
2091 most recent preceding instruction with externally visible side
2092 effects, following the order in the IR. (This includes :ref:`volatile
2093 operations <volatile>`.)
2094 - An instruction *control-depends* on a :ref:`terminator
2095 instruction <terminators>` if the terminator instruction has
2096 multiple successors and the instruction is always executed when
2097 control transfers to one of the successors, and may not be executed
2098 when control is transferred to another.
2099 - Additionally, an instruction also *control-depends* on a terminator
2100 instruction if the set of instructions it otherwise depends on would
2101 be different if the terminator had transferred control to a different
2103 - Dependence is transitive.
2105 Poison Values have the same behavior as :ref:`undef values <undefvalues>`,
2106 with the additional affect that any instruction which has a *dependence*
2107 on a poison value has undefined behavior.
2109 Here are some examples:
2111 .. code-block:: llvm
2114 %poison = sub nuw i32 0, 1 ; Results in a poison value.
2115 %still_poison = and i32 %poison, 0 ; 0, but also poison.
2116 %poison_yet_again = getelementptr i32* @h, i32 %still_poison
2117 store i32 0, i32* %poison_yet_again ; memory at @h[0] is poisoned
2119 store i32 %poison, i32* @g ; Poison value stored to memory.
2120 %poison2 = load i32* @g ; Poison value loaded back from memory.
2122 store volatile i32 %poison, i32* @g ; External observation; undefined behavior.
2124 %narrowaddr = bitcast i32* @g to i16*
2125 %wideaddr = bitcast i32* @g to i64*
2126 %poison3 = load i16* %narrowaddr ; Returns a poison value.
2127 %poison4 = load i64* %wideaddr ; Returns a poison value.
2129 %cmp = icmp slt i32 %poison, 0 ; Returns a poison value.
2130 br i1 %cmp, label %true, label %end ; Branch to either destination.
2133 store volatile i32 0, i32* @g ; This is control-dependent on %cmp, so
2134 ; it has undefined behavior.
2138 %p = phi i32 [ 0, %entry ], [ 1, %true ]
2139 ; Both edges into this PHI are
2140 ; control-dependent on %cmp, so this
2141 ; always results in a poison value.
2143 store volatile i32 0, i32* @g ; This would depend on the store in %true
2144 ; if %cmp is true, or the store in %entry
2145 ; otherwise, so this is undefined behavior.
2147 br i1 %cmp, label %second_true, label %second_end
2148 ; The same branch again, but this time the
2149 ; true block doesn't have side effects.
2156 store volatile i32 0, i32* @g ; This time, the instruction always depends
2157 ; on the store in %end. Also, it is
2158 ; control-equivalent to %end, so this is
2159 ; well-defined (ignoring earlier undefined
2160 ; behavior in this example).
2164 Addresses of Basic Blocks
2165 -------------------------
2167 ``blockaddress(@function, %block)``
2169 The '``blockaddress``' constant computes the address of the specified
2170 basic block in the specified function, and always has an ``i8*`` type.
2171 Taking the address of the entry block is illegal.
2173 This value only has defined behavior when used as an operand to the
2174 ':ref:`indirectbr <i_indirectbr>`' instruction, or for comparisons
2175 against null. Pointer equality tests between labels addresses results in
2176 undefined behavior --- though, again, comparison against null is ok, and
2177 no label is equal to the null pointer. This may be passed around as an
2178 opaque pointer sized value as long as the bits are not inspected. This
2179 allows ``ptrtoint`` and arithmetic to be performed on these values so
2180 long as the original value is reconstituted before the ``indirectbr``
2183 Finally, some targets may provide defined semantics when using the value
2184 as the operand to an inline assembly, but that is target specific.
2186 Constant Expressions
2187 --------------------
2189 Constant expressions are used to allow expressions involving other
2190 constants to be used as constants. Constant expressions may be of any
2191 :ref:`first class <t_firstclass>` type and may involve any LLVM operation
2192 that does not have side effects (e.g. load and call are not supported).
2193 The following is the syntax for constant expressions:
2195 ``trunc (CST to TYPE)``
2196 Truncate a constant to another type. The bit size of CST must be
2197 larger than the bit size of TYPE. Both types must be integers.
2198 ``zext (CST to TYPE)``
2199 Zero extend a constant to another type. The bit size of CST must be
2200 smaller than the bit size of TYPE. Both types must be integers.
2201 ``sext (CST to TYPE)``
2202 Sign extend a constant to another type. The bit size of CST must be
2203 smaller than the bit size of TYPE. Both types must be integers.
2204 ``fptrunc (CST to TYPE)``
2205 Truncate a floating point constant to another floating point type.
2206 The size of CST must be larger than the size of TYPE. Both types
2207 must be floating point.
2208 ``fpext (CST to TYPE)``
2209 Floating point extend a constant to another type. The size of CST
2210 must be smaller or equal to the size of TYPE. Both types must be
2212 ``fptoui (CST to TYPE)``
2213 Convert a floating point constant to the corresponding unsigned
2214 integer constant. TYPE must be a scalar or vector integer type. CST
2215 must be of scalar or vector floating point type. Both CST and TYPE
2216 must be scalars, or vectors of the same number of elements. If the
2217 value won't fit in the integer type, the results are undefined.
2218 ``fptosi (CST to TYPE)``
2219 Convert a floating point constant to the corresponding signed
2220 integer constant. TYPE must be a scalar or vector integer type. CST
2221 must be of scalar or vector floating point type. Both CST and TYPE
2222 must be scalars, or vectors of the same number of elements. If the
2223 value won't fit in the integer type, the results are undefined.
2224 ``uitofp (CST to TYPE)``
2225 Convert an unsigned integer constant to the corresponding floating
2226 point constant. TYPE must be a scalar or vector floating point type.
2227 CST must be of scalar or vector integer type. Both CST and TYPE must
2228 be scalars, or vectors of the same number of elements. If the value
2229 won't fit in the floating point type, the results are undefined.
2230 ``sitofp (CST to TYPE)``
2231 Convert a signed integer constant to the corresponding floating
2232 point constant. TYPE must be a scalar or vector floating point type.
2233 CST must be of scalar or vector integer type. Both CST and TYPE must
2234 be scalars, or vectors of the same number of elements. If the value
2235 won't fit in the floating point type, the results are undefined.
2236 ``ptrtoint (CST to TYPE)``
2237 Convert a pointer typed constant to the corresponding integer
2238 constant. ``TYPE`` must be an integer type. ``CST`` must be of
2239 pointer type. The ``CST`` value is zero extended, truncated, or
2240 unchanged to make it fit in ``TYPE``.
2241 ``inttoptr (CST to TYPE)``
2242 Convert an integer constant to a pointer constant. TYPE must be a
2243 pointer type. CST must be of integer type. The CST value is zero
2244 extended, truncated, or unchanged to make it fit in a pointer size.
2245 This one is *really* dangerous!
2246 ``bitcast (CST to TYPE)``
2247 Convert a constant, CST, to another TYPE. The constraints of the
2248 operands are the same as those for the :ref:`bitcast
2249 instruction <i_bitcast>`.
2250 ``getelementptr (CSTPTR, IDX0, IDX1, ...)``, ``getelementptr inbounds (CSTPTR, IDX0, IDX1, ...)``
2251 Perform the :ref:`getelementptr operation <i_getelementptr>` on
2252 constants. As with the :ref:`getelementptr <i_getelementptr>`
2253 instruction, the index list may have zero or more indexes, which are
2254 required to make sense for the type of "CSTPTR".
2255 ``select (COND, VAL1, VAL2)``
2256 Perform the :ref:`select operation <i_select>` on constants.
2257 ``icmp COND (VAL1, VAL2)``
2258 Performs the :ref:`icmp operation <i_icmp>` on constants.
2259 ``fcmp COND (VAL1, VAL2)``
2260 Performs the :ref:`fcmp operation <i_fcmp>` on constants.
2261 ``extractelement (VAL, IDX)``
2262 Perform the :ref:`extractelement operation <i_extractelement>` on
2264 ``insertelement (VAL, ELT, IDX)``
2265 Perform the :ref:`insertelement operation <i_insertelement>` on
2267 ``shufflevector (VEC1, VEC2, IDXMASK)``
2268 Perform the :ref:`shufflevector operation <i_shufflevector>` on
2270 ``extractvalue (VAL, IDX0, IDX1, ...)``
2271 Perform the :ref:`extractvalue operation <i_extractvalue>` on
2272 constants. The index list is interpreted in a similar manner as
2273 indices in a ':ref:`getelementptr <i_getelementptr>`' operation. At
2274 least one index value must be specified.
2275 ``insertvalue (VAL, ELT, IDX0, IDX1, ...)``
2276 Perform the :ref:`insertvalue operation <i_insertvalue>` on constants.
2277 The index list is interpreted in a similar manner as indices in a
2278 ':ref:`getelementptr <i_getelementptr>`' operation. At least one index
2279 value must be specified.
2280 ``OPCODE (LHS, RHS)``
2281 Perform the specified operation of the LHS and RHS constants. OPCODE
2282 may be any of the :ref:`binary <binaryops>` or :ref:`bitwise
2283 binary <bitwiseops>` operations. The constraints on operands are
2284 the same as those for the corresponding instruction (e.g. no bitwise
2285 operations on floating point values are allowed).
2290 Inline Assembler Expressions
2291 ----------------------------
2293 LLVM supports inline assembler expressions (as opposed to :ref:`Module-Level
2294 Inline Assembly <moduleasm>`) through the use of a special value. This
2295 value represents the inline assembler as a string (containing the
2296 instructions to emit), a list of operand constraints (stored as a
2297 string), a flag that indicates whether or not the inline asm expression
2298 has side effects, and a flag indicating whether the function containing
2299 the asm needs to align its stack conservatively. An example inline
2300 assembler expression is:
2302 .. code-block:: llvm
2304 i32 (i32) asm "bswap $0", "=r,r"
2306 Inline assembler expressions may **only** be used as the callee operand
2307 of a :ref:`call <i_call>` or an :ref:`invoke <i_invoke>` instruction.
2308 Thus, typically we have:
2310 .. code-block:: llvm
2312 %X = call i32 asm "bswap $0", "=r,r"(i32 %Y)
2314 Inline asms with side effects not visible in the constraint list must be
2315 marked as having side effects. This is done through the use of the
2316 '``sideeffect``' keyword, like so:
2318 .. code-block:: llvm
2320 call void asm sideeffect "eieio", ""()
2322 In some cases inline asms will contain code that will not work unless
2323 the stack is aligned in some way, such as calls or SSE instructions on
2324 x86, yet will not contain code that does that alignment within the asm.
2325 The compiler should make conservative assumptions about what the asm
2326 might contain and should generate its usual stack alignment code in the
2327 prologue if the '``alignstack``' keyword is present:
2329 .. code-block:: llvm
2331 call void asm alignstack "eieio", ""()
2333 Inline asms also support using non-standard assembly dialects. The
2334 assumed dialect is ATT. When the '``inteldialect``' keyword is present,
2335 the inline asm is using the Intel dialect. Currently, ATT and Intel are
2336 the only supported dialects. An example is:
2338 .. code-block:: llvm
2340 call void asm inteldialect "eieio", ""()
2342 If multiple keywords appear the '``sideeffect``' keyword must come
2343 first, the '``alignstack``' keyword second and the '``inteldialect``'
2349 The call instructions that wrap inline asm nodes may have a
2350 "``!srcloc``" MDNode attached to it that contains a list of constant
2351 integers. If present, the code generator will use the integer as the
2352 location cookie value when report errors through the ``LLVMContext``
2353 error reporting mechanisms. This allows a front-end to correlate backend
2354 errors that occur with inline asm back to the source code that produced
2357 .. code-block:: llvm
2359 call void asm sideeffect "something bad", ""(), !srcloc !42
2361 !42 = !{ i32 1234567 }
2363 It is up to the front-end to make sense of the magic numbers it places
2364 in the IR. If the MDNode contains multiple constants, the code generator
2365 will use the one that corresponds to the line of the asm that the error
2370 Metadata Nodes and Metadata Strings
2371 -----------------------------------
2373 LLVM IR allows metadata to be attached to instructions in the program
2374 that can convey extra information about the code to the optimizers and
2375 code generator. One example application of metadata is source-level
2376 debug information. There are two metadata primitives: strings and nodes.
2377 All metadata has the ``metadata`` type and is identified in syntax by a
2378 preceding exclamation point ('``!``').
2380 A metadata string is a string surrounded by double quotes. It can
2381 contain any character by escaping non-printable characters with
2382 "``\xx``" where "``xx``" is the two digit hex code. For example:
2385 Metadata nodes are represented with notation similar to structure
2386 constants (a comma separated list of elements, surrounded by braces and
2387 preceded by an exclamation point). Metadata nodes can have any values as
2388 their operand. For example:
2390 .. code-block:: llvm
2392 !{ metadata !"test\00", i32 10}
2394 A :ref:`named metadata <namedmetadatastructure>` is a collection of
2395 metadata nodes, which can be looked up in the module symbol table. For
2398 .. code-block:: llvm
2400 !foo = metadata !{!4, !3}
2402 Metadata can be used as function arguments. Here ``llvm.dbg.value``
2403 function is using two metadata arguments:
2405 .. code-block:: llvm
2407 call void @llvm.dbg.value(metadata !24, i64 0, metadata !25)
2409 Metadata can be attached with an instruction. Here metadata ``!21`` is
2410 attached to the ``add`` instruction using the ``!dbg`` identifier:
2412 .. code-block:: llvm
2414 %indvar.next = add i64 %indvar, 1, !dbg !21
2416 More information about specific metadata nodes recognized by the
2417 optimizers and code generator is found below.
2422 In LLVM IR, memory does not have types, so LLVM's own type system is not
2423 suitable for doing TBAA. Instead, metadata is added to the IR to
2424 describe a type system of a higher level language. This can be used to
2425 implement typical C/C++ TBAA, but it can also be used to implement
2426 custom alias analysis behavior for other languages.
2428 The current metadata format is very simple. TBAA metadata nodes have up
2429 to three fields, e.g.:
2431 .. code-block:: llvm
2433 !0 = metadata !{ metadata !"an example type tree" }
2434 !1 = metadata !{ metadata !"int", metadata !0 }
2435 !2 = metadata !{ metadata !"float", metadata !0 }
2436 !3 = metadata !{ metadata !"const float", metadata !2, i64 1 }
2438 The first field is an identity field. It can be any value, usually a
2439 metadata string, which uniquely identifies the type. The most important
2440 name in the tree is the name of the root node. Two trees with different
2441 root node names are entirely disjoint, even if they have leaves with
2444 The second field identifies the type's parent node in the tree, or is
2445 null or omitted for a root node. A type is considered to alias all of
2446 its descendants and all of its ancestors in the tree. Also, a type is
2447 considered to alias all types in other trees, so that bitcode produced
2448 from multiple front-ends is handled conservatively.
2450 If the third field is present, it's an integer which if equal to 1
2451 indicates that the type is "constant" (meaning
2452 ``pointsToConstantMemory`` should return true; see `other useful
2453 AliasAnalysis methods <AliasAnalysis.html#OtherItfs>`_).
2455 '``tbaa.struct``' Metadata
2456 ^^^^^^^^^^^^^^^^^^^^^^^^^^
2458 The :ref:`llvm.memcpy <int_memcpy>` is often used to implement
2459 aggregate assignment operations in C and similar languages, however it
2460 is defined to copy a contiguous region of memory, which is more than
2461 strictly necessary for aggregate types which contain holes due to
2462 padding. Also, it doesn't contain any TBAA information about the fields
2465 ``!tbaa.struct`` metadata can describe which memory subregions in a
2466 memcpy are padding and what the TBAA tags of the struct are.
2468 The current metadata format is very simple. ``!tbaa.struct`` metadata
2469 nodes are a list of operands which are in conceptual groups of three.
2470 For each group of three, the first operand gives the byte offset of a
2471 field in bytes, the second gives its size in bytes, and the third gives
2474 .. code-block:: llvm
2476 !4 = metadata !{ i64 0, i64 4, metadata !1, i64 8, i64 4, metadata !2 }
2478 This describes a struct with two fields. The first is at offset 0 bytes
2479 with size 4 bytes, and has tbaa tag !1. The second is at offset 8 bytes
2480 and has size 4 bytes and has tbaa tag !2.
2482 Note that the fields need not be contiguous. In this example, there is a
2483 4 byte gap between the two fields. This gap represents padding which
2484 does not carry useful data and need not be preserved.
2486 '``fpmath``' Metadata
2487 ^^^^^^^^^^^^^^^^^^^^^
2489 ``fpmath`` metadata may be attached to any instruction of floating point
2490 type. It can be used to express the maximum acceptable error in the
2491 result of that instruction, in ULPs, thus potentially allowing the
2492 compiler to use a more efficient but less accurate method of computing
2493 it. ULP is defined as follows:
2495 If ``x`` is a real number that lies between two finite consecutive
2496 floating-point numbers ``a`` and ``b``, without being equal to one
2497 of them, then ``ulp(x) = |b - a|``, otherwise ``ulp(x)`` is the
2498 distance between the two non-equal finite floating-point numbers
2499 nearest ``x``. Moreover, ``ulp(NaN)`` is ``NaN``.
2501 The metadata node shall consist of a single positive floating point
2502 number representing the maximum relative error, for example:
2504 .. code-block:: llvm
2506 !0 = metadata !{ float 2.5 } ; maximum acceptable inaccuracy is 2.5 ULPs
2508 '``range``' Metadata
2509 ^^^^^^^^^^^^^^^^^^^^
2511 ``range`` metadata may be attached only to loads of integer types. It
2512 expresses the possible ranges the loaded value is in. The ranges are
2513 represented with a flattened list of integers. The loaded value is known
2514 to be in the union of the ranges defined by each consecutive pair. Each
2515 pair has the following properties:
2517 - The type must match the type loaded by the instruction.
2518 - The pair ``a,b`` represents the range ``[a,b)``.
2519 - Both ``a`` and ``b`` are constants.
2520 - The range is allowed to wrap.
2521 - The range should not represent the full or empty set. That is,
2524 In addition, the pairs must be in signed order of the lower bound and
2525 they must be non-contiguous.
2529 .. code-block:: llvm
2531 %a = load i8* %x, align 1, !range !0 ; Can only be 0 or 1
2532 %b = load i8* %y, align 1, !range !1 ; Can only be 255 (-1), 0 or 1
2533 %c = load i8* %z, align 1, !range !2 ; Can only be 0, 1, 3, 4 or 5
2534 %d = load i8* %z, align 1, !range !3 ; Can only be -2, -1, 3, 4 or 5
2536 !0 = metadata !{ i8 0, i8 2 }
2537 !1 = metadata !{ i8 255, i8 2 }
2538 !2 = metadata !{ i8 0, i8 2, i8 3, i8 6 }
2539 !3 = metadata !{ i8 -2, i8 0, i8 3, i8 6 }
2544 It is sometimes useful to attach information to loop constructs. Currently,
2545 loop metadata is implemented as metadata attached to the branch instruction
2546 in the loop latch block. This type of metadata refer to a metadata node that is
2547 guaranteed to be separate for each loop. The loop-level metadata is prefixed
2550 The loop identifier metadata is implemented using a metadata that refers to
2551 itself to avoid merging it with any other identifier metadata, e.g.,
2552 during module linkage or function inlining. That is, each loop should refer
2553 to their own identification metadata even if they reside in separate functions.
2554 The following example contains loop identifier metadata for two separate loop
2557 .. code-block:: llvm
2559 !0 = metadata !{ metadata !0 }
2560 !1 = metadata !{ metadata !1 }
2563 '``llvm.loop.parallel``' Metadata
2564 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2566 This loop metadata can be used to communicate that a loop should be considered
2567 a parallel loop. The semantics of parallel loops in this case is the one
2568 with the strongest cross-iteration instruction ordering freedom: the
2569 iterations in the loop can be considered completely independent of each
2570 other (also known as embarrassingly parallel loops).
2572 This metadata can originate from a programming language with parallel loop
2573 constructs. In such a case it is completely the programmer's responsibility
2574 to ensure the instructions from the different iterations of the loop can be
2575 executed in an arbitrary order, in parallel, or intertwined. No loop-carried
2576 dependency checking at all must be expected from the compiler.
2578 In order to fulfill the LLVM requirement for metadata to be safely ignored,
2579 it is important to ensure that a parallel loop is converted to
2580 a sequential loop in case an optimization (agnostic of the parallel loop
2581 semantics) converts the loop back to such. This happens when new memory
2582 accesses that do not fulfill the requirement of free ordering across iterations
2583 are added to the loop. Therefore, this metadata is required, but not
2584 sufficient, to consider the loop at hand a parallel loop. For a loop
2585 to be parallel, all its memory accessing instructions need to be
2586 marked with the ``llvm.mem.parallel_loop_access`` metadata that refer
2587 to the same loop identifier metadata that identify the loop at hand.
2592 Metadata types used to annotate memory accesses with information helpful
2593 for optimizations are prefixed with ``llvm.mem``.
2595 '``llvm.mem.parallel_loop_access``' Metadata
2596 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2598 For a loop to be parallel, in addition to using
2599 the ``llvm.loop.parallel`` metadata to mark the loop latch branch instruction,
2600 also all of the memory accessing instructions in the loop body need to be
2601 marked with the ``llvm.mem.parallel_loop_access`` metadata. If there
2602 is at least one memory accessing instruction not marked with the metadata,
2603 the loop, despite it possibly using the ``llvm.loop.parallel`` metadata,
2604 must be considered a sequential loop. This causes parallel loops to be
2605 converted to sequential loops due to optimization passes that are unaware of
2606 the parallel semantics and that insert new memory instructions to the loop
2609 Example of a loop that is considered parallel due to its correct use of
2610 both ``llvm.loop.parallel`` and ``llvm.mem.parallel_loop_access``
2611 metadata types that refer to the same loop identifier metadata.
2613 .. code-block:: llvm
2617 %0 = load i32* %arrayidx, align 4, !llvm.mem.parallel_loop_access !0
2619 store i32 %0, i32* %arrayidx4, align 4, !llvm.mem.parallel_loop_access !0
2621 br i1 %exitcond, label %for.end, label %for.body, !llvm.loop.parallel !0
2625 !0 = metadata !{ metadata !0 }
2627 It is also possible to have nested parallel loops. In that case the
2628 memory accesses refer to a list of loop identifier metadata nodes instead of
2629 the loop identifier metadata node directly:
2631 .. code-block:: llvm
2638 %0 = load i32* %arrayidx, align 4, !llvm.mem.parallel_loop_access !0
2640 store i32 %0, i32* %arrayidx4, align 4, !llvm.mem.parallel_loop_access !0
2642 br i1 %exitcond, label %inner.for.end, label %inner.for.body, !llvm.loop.parallel !1
2646 %0 = load i32* %arrayidx, align 4, !llvm.mem.parallel_loop_access !0
2648 store i32 %0, i32* %arrayidx4, align 4, !llvm.mem.parallel_loop_access !0
2650 br i1 %exitcond, label %outer.for.end, label %outer.for.body, !llvm.loop.parallel !2
2652 outer.for.end: ; preds = %for.body
2654 !0 = metadata !{ metadata !1, metadata !2 } ; a list of parallel loop identifiers
2655 !1 = metadata !{ metadata !1 } ; an identifier for the inner parallel loop
2656 !2 = metadata !{ metadata !2 } ; an identifier for the outer parallel loop
2659 Module Flags Metadata
2660 =====================
2662 Information about the module as a whole is difficult to convey to LLVM's
2663 subsystems. The LLVM IR isn't sufficient to transmit this information.
2664 The ``llvm.module.flags`` named metadata exists in order to facilitate
2665 this. These flags are in the form of key / value pairs --- much like a
2666 dictionary --- making it easy for any subsystem who cares about a flag to
2669 The ``llvm.module.flags`` metadata contains a list of metadata triplets.
2670 Each triplet has the following form:
2672 - The first element is a *behavior* flag, which specifies the behavior
2673 when two (or more) modules are merged together, and it encounters two
2674 (or more) metadata with the same ID. The supported behaviors are
2676 - The second element is a metadata string that is a unique ID for the
2677 metadata. Each module may only have one flag entry for each unique ID (not
2678 including entries with the **Require** behavior).
2679 - The third element is the value of the flag.
2681 When two (or more) modules are merged together, the resulting
2682 ``llvm.module.flags`` metadata is the union of the modules' flags. That is, for
2683 each unique metadata ID string, there will be exactly one entry in the merged
2684 modules ``llvm.module.flags`` metadata table, and the value for that entry will
2685 be determined by the merge behavior flag, as described below. The only exception
2686 is that entries with the *Require* behavior are always preserved.
2688 The following behaviors are supported:
2699 Emits an error if two values disagree, otherwise the resulting value
2700 is that of the operands.
2704 Emits a warning if two values disagree. The result value will be the
2705 operand for the flag from the first module being linked.
2709 Adds a requirement that another module flag be present and have a
2710 specified value after linking is performed. The value must be a
2711 metadata pair, where the first element of the pair is the ID of the
2712 module flag to be restricted, and the second element of the pair is
2713 the value the module flag should be restricted to. This behavior can
2714 be used to restrict the allowable results (via triggering of an
2715 error) of linking IDs with the **Override** behavior.
2719 Uses the specified value, regardless of the behavior or value of the
2720 other module. If both modules specify **Override**, but the values
2721 differ, an error will be emitted.
2725 Appends the two values, which are required to be metadata nodes.
2729 Appends the two values, which are required to be metadata
2730 nodes. However, duplicate entries in the second list are dropped
2731 during the append operation.
2733 It is an error for a particular unique flag ID to have multiple behaviors,
2734 except in the case of **Require** (which adds restrictions on another metadata
2735 value) or **Override**.
2737 An example of module flags:
2739 .. code-block:: llvm
2741 !0 = metadata !{ i32 1, metadata !"foo", i32 1 }
2742 !1 = metadata !{ i32 4, metadata !"bar", i32 37 }
2743 !2 = metadata !{ i32 2, metadata !"qux", i32 42 }
2744 !3 = metadata !{ i32 3, metadata !"qux",
2746 metadata !"foo", i32 1
2749 !llvm.module.flags = !{ !0, !1, !2, !3 }
2751 - Metadata ``!0`` has the ID ``!"foo"`` and the value '1'. The behavior
2752 if two or more ``!"foo"`` flags are seen is to emit an error if their
2753 values are not equal.
2755 - Metadata ``!1`` has the ID ``!"bar"`` and the value '37'. The
2756 behavior if two or more ``!"bar"`` flags are seen is to use the value
2759 - Metadata ``!2`` has the ID ``!"qux"`` and the value '42'. The
2760 behavior if two or more ``!"qux"`` flags are seen is to emit a
2761 warning if their values are not equal.
2763 - Metadata ``!3`` has the ID ``!"qux"`` and the value:
2767 metadata !{ metadata !"foo", i32 1 }
2769 The behavior is to emit an error if the ``llvm.module.flags`` does not
2770 contain a flag with the ID ``!"foo"`` that has the value '1' after linking is
2773 Objective-C Garbage Collection Module Flags Metadata
2774 ----------------------------------------------------
2776 On the Mach-O platform, Objective-C stores metadata about garbage
2777 collection in a special section called "image info". The metadata
2778 consists of a version number and a bitmask specifying what types of
2779 garbage collection are supported (if any) by the file. If two or more
2780 modules are linked together their garbage collection metadata needs to
2781 be merged rather than appended together.
2783 The Objective-C garbage collection module flags metadata consists of the
2784 following key-value pairs:
2793 * - ``Objective-C Version``
2794 - **[Required]** --- The Objective-C ABI version. Valid values are 1 and 2.
2796 * - ``Objective-C Image Info Version``
2797 - **[Required]** --- The version of the image info section. Currently
2800 * - ``Objective-C Image Info Section``
2801 - **[Required]** --- The section to place the metadata. Valid values are
2802 ``"__OBJC, __image_info, regular"`` for Objective-C ABI version 1, and
2803 ``"__DATA,__objc_imageinfo, regular, no_dead_strip"`` for
2804 Objective-C ABI version 2.
2806 * - ``Objective-C Garbage Collection``
2807 - **[Required]** --- Specifies whether garbage collection is supported or
2808 not. Valid values are 0, for no garbage collection, and 2, for garbage
2809 collection supported.
2811 * - ``Objective-C GC Only``
2812 - **[Optional]** --- Specifies that only garbage collection is supported.
2813 If present, its value must be 6. This flag requires that the
2814 ``Objective-C Garbage Collection`` flag have the value 2.
2816 Some important flag interactions:
2818 - If a module with ``Objective-C Garbage Collection`` set to 0 is
2819 merged with a module with ``Objective-C Garbage Collection`` set to
2820 2, then the resulting module has the
2821 ``Objective-C Garbage Collection`` flag set to 0.
2822 - A module with ``Objective-C Garbage Collection`` set to 0 cannot be
2823 merged with a module with ``Objective-C GC Only`` set to 6.
2825 Automatic Linker Flags Module Flags Metadata
2826 --------------------------------------------
2828 Some targets support embedding flags to the linker inside individual object
2829 files. Typically this is used in conjunction with language extensions which
2830 allow source files to explicitly declare the libraries they depend on, and have
2831 these automatically be transmitted to the linker via object files.
2833 These flags are encoded in the IR using metadata in the module flags section,
2834 using the ``Linker Options`` key. The merge behavior for this flag is required
2835 to be ``AppendUnique``, and the value for the key is expected to be a metadata
2836 node which should be a list of other metadata nodes, each of which should be a
2837 list of metadata strings defining linker options.
2839 For example, the following metadata section specifies two separate sets of
2840 linker options, presumably to link against ``libz`` and the ``Cocoa``
2843 !0 = metadata !{ i32 6, metadata !"Linker Options",
2845 metadata !{ metadata !"-lz" },
2846 metadata !{ metadata !"-framework", metadata !"Cocoa" } } }
2847 !llvm.module.flags = !{ !0 }
2849 The metadata encoding as lists of lists of options, as opposed to a collapsed
2850 list of options, is chosen so that the IR encoding can use multiple option
2851 strings to specify e.g., a single library, while still having that specifier be
2852 preserved as an atomic element that can be recognized by a target specific
2853 assembly writer or object file emitter.
2855 Each individual option is required to be either a valid option for the target's
2856 linker, or an option that is reserved by the target specific assembly writer or
2857 object file emitter. No other aspect of these options is defined by the IR.
2859 Intrinsic Global Variables
2860 ==========================
2862 LLVM has a number of "magic" global variables that contain data that
2863 affect code generation or other IR semantics. These are documented here.
2864 All globals of this sort should have a section specified as
2865 "``llvm.metadata``". This section and all globals that start with
2866 "``llvm.``" are reserved for use by LLVM.
2868 The '``llvm.used``' Global Variable
2869 -----------------------------------
2871 The ``@llvm.used`` global is an array which has
2872 :ref:`appending linkage <linkage_appending>`. This array contains a list of
2873 pointers to global variables, functions and aliases which may optionally have a
2874 pointer cast formed of bitcast or getelementptr. For example, a legal
2877 .. code-block:: llvm
2882 @llvm.used = appending global [2 x i8*] [
2884 i8* bitcast (i32* @Y to i8*)
2885 ], section "llvm.metadata"
2887 If a symbol appears in the ``@llvm.used`` list, then the compiler, assembler,
2888 and linker are required to treat the symbol as if there is a reference to the
2889 symbol that it cannot see. For example, if a variable has internal linkage and
2890 no references other than that from the ``@llvm.used`` list, it cannot be
2891 deleted. This is commonly used to represent references from inline asms and
2892 other things the compiler cannot "see", and corresponds to
2893 "``attribute((used))``" in GNU C.
2895 On some targets, the code generator must emit a directive to the
2896 assembler or object file to prevent the assembler and linker from
2897 molesting the symbol.
2899 The '``llvm.compiler.used``' Global Variable
2900 --------------------------------------------
2902 The ``@llvm.compiler.used`` directive is the same as the ``@llvm.used``
2903 directive, except that it only prevents the compiler from touching the
2904 symbol. On targets that support it, this allows an intelligent linker to
2905 optimize references to the symbol without being impeded as it would be
2908 This is a rare construct that should only be used in rare circumstances,
2909 and should not be exposed to source languages.
2911 The '``llvm.global_ctors``' Global Variable
2912 -------------------------------------------
2914 .. code-block:: llvm
2916 %0 = type { i32, void ()* }
2917 @llvm.global_ctors = appending global [1 x %0] [%0 { i32 65535, void ()* @ctor }]
2919 The ``@llvm.global_ctors`` array contains a list of constructor
2920 functions and associated priorities. The functions referenced by this
2921 array will be called in ascending order of priority (i.e. lowest first)
2922 when the module is loaded. The order of functions with the same priority
2925 The '``llvm.global_dtors``' Global Variable
2926 -------------------------------------------
2928 .. code-block:: llvm
2930 %0 = type { i32, void ()* }
2931 @llvm.global_dtors = appending global [1 x %0] [%0 { i32 65535, void ()* @dtor }]
2933 The ``@llvm.global_dtors`` array contains a list of destructor functions
2934 and associated priorities. The functions referenced by this array will
2935 be called in descending order of priority (i.e. highest first) when the
2936 module is loaded. The order of functions with the same priority is not
2939 Instruction Reference
2940 =====================
2942 The LLVM instruction set consists of several different classifications
2943 of instructions: :ref:`terminator instructions <terminators>`, :ref:`binary
2944 instructions <binaryops>`, :ref:`bitwise binary
2945 instructions <bitwiseops>`, :ref:`memory instructions <memoryops>`, and
2946 :ref:`other instructions <otherops>`.
2950 Terminator Instructions
2951 -----------------------
2953 As mentioned :ref:`previously <functionstructure>`, every basic block in a
2954 program ends with a "Terminator" instruction, which indicates which
2955 block should be executed after the current block is finished. These
2956 terminator instructions typically yield a '``void``' value: they produce
2957 control flow, not values (the one exception being the
2958 ':ref:`invoke <i_invoke>`' instruction).
2960 The terminator instructions are: ':ref:`ret <i_ret>`',
2961 ':ref:`br <i_br>`', ':ref:`switch <i_switch>`',
2962 ':ref:`indirectbr <i_indirectbr>`', ':ref:`invoke <i_invoke>`',
2963 ':ref:`resume <i_resume>`', and ':ref:`unreachable <i_unreachable>`'.
2967 '``ret``' Instruction
2968 ^^^^^^^^^^^^^^^^^^^^^
2975 ret <type> <value> ; Return a value from a non-void function
2976 ret void ; Return from void function
2981 The '``ret``' instruction is used to return control flow (and optionally
2982 a value) from a function back to the caller.
2984 There are two forms of the '``ret``' instruction: one that returns a
2985 value and then causes control flow, and one that just causes control
2991 The '``ret``' instruction optionally accepts a single argument, the
2992 return value. The type of the return value must be a ':ref:`first
2993 class <t_firstclass>`' type.
2995 A function is not :ref:`well formed <wellformed>` if it it has a non-void
2996 return type and contains a '``ret``' instruction with no return value or
2997 a return value with a type that does not match its type, or if it has a
2998 void return type and contains a '``ret``' instruction with a return
3004 When the '``ret``' instruction is executed, control flow returns back to
3005 the calling function's context. If the caller is a
3006 ":ref:`call <i_call>`" instruction, execution continues at the
3007 instruction after the call. If the caller was an
3008 ":ref:`invoke <i_invoke>`" instruction, execution continues at the
3009 beginning of the "normal" destination block. If the instruction returns
3010 a value, that value shall set the call or invoke instruction's return
3016 .. code-block:: llvm
3018 ret i32 5 ; Return an integer value of 5
3019 ret void ; Return from a void function
3020 ret { i32, i8 } { i32 4, i8 2 } ; Return a struct of values 4 and 2
3024 '``br``' Instruction
3025 ^^^^^^^^^^^^^^^^^^^^
3032 br i1 <cond>, label <iftrue>, label <iffalse>
3033 br label <dest> ; Unconditional branch
3038 The '``br``' instruction is used to cause control flow to transfer to a
3039 different basic block in the current function. There are two forms of
3040 this instruction, corresponding to a conditional branch and an
3041 unconditional branch.
3046 The conditional branch form of the '``br``' instruction takes a single
3047 '``i1``' value and two '``label``' values. The unconditional form of the
3048 '``br``' instruction takes a single '``label``' value as a target.
3053 Upon execution of a conditional '``br``' instruction, the '``i1``'
3054 argument is evaluated. If the value is ``true``, control flows to the
3055 '``iftrue``' ``label`` argument. If "cond" is ``false``, control flows
3056 to the '``iffalse``' ``label`` argument.
3061 .. code-block:: llvm
3064 %cond = icmp eq i32 %a, %b
3065 br i1 %cond, label %IfEqual, label %IfUnequal
3073 '``switch``' Instruction
3074 ^^^^^^^^^^^^^^^^^^^^^^^^
3081 switch <intty> <value>, label <defaultdest> [ <intty> <val>, label <dest> ... ]
3086 The '``switch``' instruction is used to transfer control flow to one of
3087 several different places. It is a generalization of the '``br``'
3088 instruction, allowing a branch to occur to one of many possible
3094 The '``switch``' instruction uses three parameters: an integer
3095 comparison value '``value``', a default '``label``' destination, and an
3096 array of pairs of comparison value constants and '``label``'s. The table
3097 is not allowed to contain duplicate constant entries.
3102 The ``switch`` instruction specifies a table of values and destinations.
3103 When the '``switch``' instruction is executed, this table is searched
3104 for the given value. If the value is found, control flow is transferred
3105 to the corresponding destination; otherwise, control flow is transferred
3106 to the default destination.
3111 Depending on properties of the target machine and the particular
3112 ``switch`` instruction, this instruction may be code generated in
3113 different ways. For example, it could be generated as a series of
3114 chained conditional branches or with a lookup table.
3119 .. code-block:: llvm
3121 ; Emulate a conditional br instruction
3122 %Val = zext i1 %value to i32
3123 switch i32 %Val, label %truedest [ i32 0, label %falsedest ]
3125 ; Emulate an unconditional br instruction
3126 switch i32 0, label %dest [ ]
3128 ; Implement a jump table:
3129 switch i32 %val, label %otherwise [ i32 0, label %onzero
3131 i32 2, label %ontwo ]
3135 '``indirectbr``' Instruction
3136 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3143 indirectbr <somety>* <address>, [ label <dest1>, label <dest2>, ... ]
3148 The '``indirectbr``' instruction implements an indirect branch to a
3149 label within the current function, whose address is specified by
3150 "``address``". Address must be derived from a
3151 :ref:`blockaddress <blockaddress>` constant.
3156 The '``address``' argument is the address of the label to jump to. The
3157 rest of the arguments indicate the full set of possible destinations
3158 that the address may point to. Blocks are allowed to occur multiple
3159 times in the destination list, though this isn't particularly useful.
3161 This destination list is required so that dataflow analysis has an
3162 accurate understanding of the CFG.
3167 Control transfers to the block specified in the address argument. All
3168 possible destination blocks must be listed in the label list, otherwise
3169 this instruction has undefined behavior. This implies that jumps to
3170 labels defined in other functions have undefined behavior as well.
3175 This is typically implemented with a jump through a register.
3180 .. code-block:: llvm
3182 indirectbr i8* %Addr, [ label %bb1, label %bb2, label %bb3 ]
3186 '``invoke``' Instruction
3187 ^^^^^^^^^^^^^^^^^^^^^^^^
3194 <result> = invoke [cconv] [ret attrs] <ptr to function ty> <function ptr val>(<function args>) [fn attrs]
3195 to label <normal label> unwind label <exception label>
3200 The '``invoke``' instruction causes control to transfer to a specified
3201 function, with the possibility of control flow transfer to either the
3202 '``normal``' label or the '``exception``' label. If the callee function
3203 returns with the "``ret``" instruction, control flow will return to the
3204 "normal" label. If the callee (or any indirect callees) returns via the
3205 ":ref:`resume <i_resume>`" instruction or other exception handling
3206 mechanism, control is interrupted and continued at the dynamically
3207 nearest "exception" label.
3209 The '``exception``' label is a `landing
3210 pad <ExceptionHandling.html#overview>`_ for the exception. As such,
3211 '``exception``' label is required to have the
3212 ":ref:`landingpad <i_landingpad>`" instruction, which contains the
3213 information about the behavior of the program after unwinding happens,
3214 as its first non-PHI instruction. The restrictions on the
3215 "``landingpad``" instruction's tightly couples it to the "``invoke``"
3216 instruction, so that the important information contained within the
3217 "``landingpad``" instruction can't be lost through normal code motion.
3222 This instruction requires several arguments:
3224 #. The optional "cconv" marker indicates which :ref:`calling
3225 convention <callingconv>` the call should use. If none is
3226 specified, the call defaults to using C calling conventions.
3227 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
3228 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
3230 #. '``ptr to function ty``': shall be the signature of the pointer to
3231 function value being invoked. In most cases, this is a direct
3232 function invocation, but indirect ``invoke``'s are just as possible,
3233 branching off an arbitrary pointer to function value.
3234 #. '``function ptr val``': An LLVM value containing a pointer to a
3235 function to be invoked.
3236 #. '``function args``': argument list whose types match the function
3237 signature argument types and parameter attributes. All arguments must
3238 be of :ref:`first class <t_firstclass>` type. If the function signature
3239 indicates the function accepts a variable number of arguments, the
3240 extra arguments can be specified.
3241 #. '``normal label``': the label reached when the called function
3242 executes a '``ret``' instruction.
3243 #. '``exception label``': the label reached when a callee returns via
3244 the :ref:`resume <i_resume>` instruction or other exception handling
3246 #. The optional :ref:`function attributes <fnattrs>` list. Only
3247 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
3248 attributes are valid here.
3253 This instruction is designed to operate as a standard '``call``'
3254 instruction in most regards. The primary difference is that it
3255 establishes an association with a label, which is used by the runtime
3256 library to unwind the stack.
3258 This instruction is used in languages with destructors to ensure that
3259 proper cleanup is performed in the case of either a ``longjmp`` or a
3260 thrown exception. Additionally, this is important for implementation of
3261 '``catch``' clauses in high-level languages that support them.
3263 For the purposes of the SSA form, the definition of the value returned
3264 by the '``invoke``' instruction is deemed to occur on the edge from the
3265 current block to the "normal" label. If the callee unwinds then no
3266 return value is available.
3271 .. code-block:: llvm
3273 %retval = invoke i32 @Test(i32 15) to label %Continue
3274 unwind label %TestCleanup ; {i32}:retval set
3275 %retval = invoke coldcc i32 %Testfnptr(i32 15) to label %Continue
3276 unwind label %TestCleanup ; {i32}:retval set
3280 '``resume``' Instruction
3281 ^^^^^^^^^^^^^^^^^^^^^^^^
3288 resume <type> <value>
3293 The '``resume``' instruction is a terminator instruction that has no
3299 The '``resume``' instruction requires one argument, which must have the
3300 same type as the result of any '``landingpad``' instruction in the same
3306 The '``resume``' instruction resumes propagation of an existing
3307 (in-flight) exception whose unwinding was interrupted with a
3308 :ref:`landingpad <i_landingpad>` instruction.
3313 .. code-block:: llvm
3315 resume { i8*, i32 } %exn
3319 '``unreachable``' Instruction
3320 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3332 The '``unreachable``' instruction has no defined semantics. This
3333 instruction is used to inform the optimizer that a particular portion of
3334 the code is not reachable. This can be used to indicate that the code
3335 after a no-return function cannot be reached, and other facts.
3340 The '``unreachable``' instruction has no defined semantics.
3347 Binary operators are used to do most of the computation in a program.
3348 They require two operands of the same type, execute an operation on
3349 them, and produce a single value. The operands might represent multiple
3350 data, as is the case with the :ref:`vector <t_vector>` data type. The
3351 result value has the same type as its operands.
3353 There are several different binary operators:
3357 '``add``' Instruction
3358 ^^^^^^^^^^^^^^^^^^^^^
3365 <result> = add <ty> <op1>, <op2> ; yields {ty}:result
3366 <result> = add nuw <ty> <op1>, <op2> ; yields {ty}:result
3367 <result> = add nsw <ty> <op1>, <op2> ; yields {ty}:result
3368 <result> = add nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3373 The '``add``' instruction returns the sum of its two operands.
3378 The two arguments to the '``add``' instruction must be
3379 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3380 arguments must have identical types.
3385 The value produced is the integer sum of the two operands.
3387 If the sum has unsigned overflow, the result returned is the
3388 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
3391 Because LLVM integers use a two's complement representation, this
3392 instruction is appropriate for both signed and unsigned integers.
3394 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3395 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3396 result value of the ``add`` is a :ref:`poison value <poisonvalues>` if
3397 unsigned and/or signed overflow, respectively, occurs.
3402 .. code-block:: llvm
3404 <result> = add i32 4, %var ; yields {i32}:result = 4 + %var
3408 '``fadd``' Instruction
3409 ^^^^^^^^^^^^^^^^^^^^^^
3416 <result> = fadd [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3421 The '``fadd``' instruction returns the sum of its two operands.
3426 The two arguments to the '``fadd``' instruction must be :ref:`floating
3427 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3428 Both arguments must have identical types.
3433 The value produced is the floating point sum of the two operands. This
3434 instruction can also take any number of :ref:`fast-math flags <fastmath>`,
3435 which are optimization hints to enable otherwise unsafe floating point
3441 .. code-block:: llvm
3443 <result> = fadd float 4.0, %var ; yields {float}:result = 4.0 + %var
3445 '``sub``' Instruction
3446 ^^^^^^^^^^^^^^^^^^^^^
3453 <result> = sub <ty> <op1>, <op2> ; yields {ty}:result
3454 <result> = sub nuw <ty> <op1>, <op2> ; yields {ty}:result
3455 <result> = sub nsw <ty> <op1>, <op2> ; yields {ty}:result
3456 <result> = sub nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3461 The '``sub``' instruction returns the difference of its two operands.
3463 Note that the '``sub``' instruction is used to represent the '``neg``'
3464 instruction present in most other intermediate representations.
3469 The two arguments to the '``sub``' instruction must be
3470 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3471 arguments must have identical types.
3476 The value produced is the integer difference of the two operands.
3478 If the difference has unsigned overflow, the result returned is the
3479 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
3482 Because LLVM integers use a two's complement representation, this
3483 instruction is appropriate for both signed and unsigned integers.
3485 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3486 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3487 result value of the ``sub`` is a :ref:`poison value <poisonvalues>` if
3488 unsigned and/or signed overflow, respectively, occurs.
3493 .. code-block:: llvm
3495 <result> = sub i32 4, %var ; yields {i32}:result = 4 - %var
3496 <result> = sub i32 0, %val ; yields {i32}:result = -%var
3500 '``fsub``' Instruction
3501 ^^^^^^^^^^^^^^^^^^^^^^
3508 <result> = fsub [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3513 The '``fsub``' instruction returns the difference of its two operands.
3515 Note that the '``fsub``' instruction is used to represent the '``fneg``'
3516 instruction present in most other intermediate representations.
3521 The two arguments to the '``fsub``' instruction must be :ref:`floating
3522 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3523 Both arguments must have identical types.
3528 The value produced is the floating point difference of the two operands.
3529 This instruction can also take any number of :ref:`fast-math
3530 flags <fastmath>`, which are optimization hints to enable otherwise
3531 unsafe floating point optimizations:
3536 .. code-block:: llvm
3538 <result> = fsub float 4.0, %var ; yields {float}:result = 4.0 - %var
3539 <result> = fsub float -0.0, %val ; yields {float}:result = -%var
3541 '``mul``' Instruction
3542 ^^^^^^^^^^^^^^^^^^^^^
3549 <result> = mul <ty> <op1>, <op2> ; yields {ty}:result
3550 <result> = mul nuw <ty> <op1>, <op2> ; yields {ty}:result
3551 <result> = mul nsw <ty> <op1>, <op2> ; yields {ty}:result
3552 <result> = mul nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3557 The '``mul``' instruction returns the product of its two operands.
3562 The two arguments to the '``mul``' instruction must be
3563 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3564 arguments must have identical types.
3569 The value produced is the integer product of the two operands.
3571 If the result of the multiplication has unsigned overflow, the result
3572 returned is the mathematical result modulo 2\ :sup:`n`\ , where n is the
3573 bit width of the result.
3575 Because LLVM integers use a two's complement representation, and the
3576 result is the same width as the operands, this instruction returns the
3577 correct result for both signed and unsigned integers. If a full product
3578 (e.g. ``i32`` * ``i32`` -> ``i64``) is needed, the operands should be
3579 sign-extended or zero-extended as appropriate to the width of the full
3582 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3583 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3584 result value of the ``mul`` is a :ref:`poison value <poisonvalues>` if
3585 unsigned and/or signed overflow, respectively, occurs.
3590 .. code-block:: llvm
3592 <result> = mul i32 4, %var ; yields {i32}:result = 4 * %var
3596 '``fmul``' Instruction
3597 ^^^^^^^^^^^^^^^^^^^^^^
3604 <result> = fmul [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3609 The '``fmul``' instruction returns the product of its two operands.
3614 The two arguments to the '``fmul``' instruction must be :ref:`floating
3615 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3616 Both arguments must have identical types.
3621 The value produced is the floating point product of the two operands.
3622 This instruction can also take any number of :ref:`fast-math
3623 flags <fastmath>`, which are optimization hints to enable otherwise
3624 unsafe floating point optimizations:
3629 .. code-block:: llvm
3631 <result> = fmul float 4.0, %var ; yields {float}:result = 4.0 * %var
3633 '``udiv``' Instruction
3634 ^^^^^^^^^^^^^^^^^^^^^^
3641 <result> = udiv <ty> <op1>, <op2> ; yields {ty}:result
3642 <result> = udiv exact <ty> <op1>, <op2> ; yields {ty}:result
3647 The '``udiv``' instruction returns the quotient of its two operands.
3652 The two arguments to the '``udiv``' instruction must be
3653 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3654 arguments must have identical types.
3659 The value produced is the unsigned integer quotient of the two operands.
3661 Note that unsigned integer division and signed integer division are
3662 distinct operations; for signed integer division, use '``sdiv``'.
3664 Division by zero leads to undefined behavior.
3666 If the ``exact`` keyword is present, the result value of the ``udiv`` is
3667 a :ref:`poison value <poisonvalues>` if %op1 is not a multiple of %op2 (as
3668 such, "((a udiv exact b) mul b) == a").
3673 .. code-block:: llvm
3675 <result> = udiv i32 4, %var ; yields {i32}:result = 4 / %var
3677 '``sdiv``' Instruction
3678 ^^^^^^^^^^^^^^^^^^^^^^
3685 <result> = sdiv <ty> <op1>, <op2> ; yields {ty}:result
3686 <result> = sdiv exact <ty> <op1>, <op2> ; yields {ty}:result
3691 The '``sdiv``' instruction returns the quotient of its two operands.
3696 The two arguments to the '``sdiv``' instruction must be
3697 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3698 arguments must have identical types.
3703 The value produced is the signed integer quotient of the two operands
3704 rounded towards zero.
3706 Note that signed integer division and unsigned integer division are
3707 distinct operations; for unsigned integer division, use '``udiv``'.
3709 Division by zero leads to undefined behavior. Overflow also leads to
3710 undefined behavior; this is a rare case, but can occur, for example, by
3711 doing a 32-bit division of -2147483648 by -1.
3713 If the ``exact`` keyword is present, the result value of the ``sdiv`` is
3714 a :ref:`poison value <poisonvalues>` if the result would be rounded.
3719 .. code-block:: llvm
3721 <result> = sdiv i32 4, %var ; yields {i32}:result = 4 / %var
3725 '``fdiv``' Instruction
3726 ^^^^^^^^^^^^^^^^^^^^^^
3733 <result> = fdiv [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3738 The '``fdiv``' instruction returns the quotient of its two operands.
3743 The two arguments to the '``fdiv``' instruction must be :ref:`floating
3744 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3745 Both arguments must have identical types.
3750 The value produced is the floating point quotient of the two operands.
3751 This instruction can also take any number of :ref:`fast-math
3752 flags <fastmath>`, which are optimization hints to enable otherwise
3753 unsafe floating point optimizations:
3758 .. code-block:: llvm
3760 <result> = fdiv float 4.0, %var ; yields {float}:result = 4.0 / %var
3762 '``urem``' Instruction
3763 ^^^^^^^^^^^^^^^^^^^^^^
3770 <result> = urem <ty> <op1>, <op2> ; yields {ty}:result
3775 The '``urem``' instruction returns the remainder from the unsigned
3776 division of its two arguments.
3781 The two arguments to the '``urem``' instruction must be
3782 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3783 arguments must have identical types.
3788 This instruction returns the unsigned integer *remainder* of a division.
3789 This instruction always performs an unsigned division to get the
3792 Note that unsigned integer remainder and signed integer remainder are
3793 distinct operations; for signed integer remainder, use '``srem``'.
3795 Taking the remainder of a division by zero leads to undefined behavior.
3800 .. code-block:: llvm
3802 <result> = urem i32 4, %var ; yields {i32}:result = 4 % %var
3804 '``srem``' Instruction
3805 ^^^^^^^^^^^^^^^^^^^^^^
3812 <result> = srem <ty> <op1>, <op2> ; yields {ty}:result
3817 The '``srem``' instruction returns the remainder from the signed
3818 division of its two operands. This instruction can also take
3819 :ref:`vector <t_vector>` versions of the values in which case the elements
3825 The two arguments to the '``srem``' instruction must be
3826 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3827 arguments must have identical types.
3832 This instruction returns the *remainder* of a division (where the result
3833 is either zero or has the same sign as the dividend, ``op1``), not the
3834 *modulo* operator (where the result is either zero or has the same sign
3835 as the divisor, ``op2``) of a value. For more information about the
3836 difference, see `The Math
3837 Forum <http://mathforum.org/dr.math/problems/anne.4.28.99.html>`_. For a
3838 table of how this is implemented in various languages, please see
3840 operation <http://en.wikipedia.org/wiki/Modulo_operation>`_.
3842 Note that signed integer remainder and unsigned integer remainder are
3843 distinct operations; for unsigned integer remainder, use '``urem``'.
3845 Taking the remainder of a division by zero leads to undefined behavior.
3846 Overflow also leads to undefined behavior; this is a rare case, but can
3847 occur, for example, by taking the remainder of a 32-bit division of
3848 -2147483648 by -1. (The remainder doesn't actually overflow, but this
3849 rule lets srem be implemented using instructions that return both the
3850 result of the division and the remainder.)
3855 .. code-block:: llvm
3857 <result> = srem i32 4, %var ; yields {i32}:result = 4 % %var
3861 '``frem``' Instruction
3862 ^^^^^^^^^^^^^^^^^^^^^^
3869 <result> = frem [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3874 The '``frem``' instruction returns the remainder from the division of
3880 The two arguments to the '``frem``' instruction must be :ref:`floating
3881 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3882 Both arguments must have identical types.
3887 This instruction returns the *remainder* of a division. The remainder
3888 has the same sign as the dividend. This instruction can also take any
3889 number of :ref:`fast-math flags <fastmath>`, which are optimization hints
3890 to enable otherwise unsafe floating point optimizations:
3895 .. code-block:: llvm
3897 <result> = frem float 4.0, %var ; yields {float}:result = 4.0 % %var
3901 Bitwise Binary Operations
3902 -------------------------
3904 Bitwise binary operators are used to do various forms of bit-twiddling
3905 in a program. They are generally very efficient instructions and can
3906 commonly be strength reduced from other instructions. They require two
3907 operands of the same type, execute an operation on them, and produce a
3908 single value. The resulting value is the same type as its operands.
3910 '``shl``' Instruction
3911 ^^^^^^^^^^^^^^^^^^^^^
3918 <result> = shl <ty> <op1>, <op2> ; yields {ty}:result
3919 <result> = shl nuw <ty> <op1>, <op2> ; yields {ty}:result
3920 <result> = shl nsw <ty> <op1>, <op2> ; yields {ty}:result
3921 <result> = shl nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3926 The '``shl``' instruction returns the first operand shifted to the left
3927 a specified number of bits.
3932 Both arguments to the '``shl``' instruction must be the same
3933 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
3934 '``op2``' is treated as an unsigned value.
3939 The value produced is ``op1`` \* 2\ :sup:`op2` mod 2\ :sup:`n`,
3940 where ``n`` is the width of the result. If ``op2`` is (statically or
3941 dynamically) negative or equal to or larger than the number of bits in
3942 ``op1``, the result is undefined. If the arguments are vectors, each
3943 vector element of ``op1`` is shifted by the corresponding shift amount
3946 If the ``nuw`` keyword is present, then the shift produces a :ref:`poison
3947 value <poisonvalues>` if it shifts out any non-zero bits. If the
3948 ``nsw`` keyword is present, then the shift produces a :ref:`poison
3949 value <poisonvalues>` if it shifts out any bits that disagree with the
3950 resultant sign bit. As such, NUW/NSW have the same semantics as they
3951 would if the shift were expressed as a mul instruction with the same
3952 nsw/nuw bits in (mul %op1, (shl 1, %op2)).
3957 .. code-block:: llvm
3959 <result> = shl i32 4, %var ; yields {i32}: 4 << %var
3960 <result> = shl i32 4, 2 ; yields {i32}: 16
3961 <result> = shl i32 1, 10 ; yields {i32}: 1024
3962 <result> = shl i32 1, 32 ; undefined
3963 <result> = shl <2 x i32> < i32 1, i32 1>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 2, i32 4>
3965 '``lshr``' Instruction
3966 ^^^^^^^^^^^^^^^^^^^^^^
3973 <result> = lshr <ty> <op1>, <op2> ; yields {ty}:result
3974 <result> = lshr exact <ty> <op1>, <op2> ; yields {ty}:result
3979 The '``lshr``' instruction (logical shift right) returns the first
3980 operand shifted to the right a specified number of bits with zero fill.
3985 Both arguments to the '``lshr``' instruction must be the same
3986 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
3987 '``op2``' is treated as an unsigned value.
3992 This instruction always performs a logical shift right operation. The
3993 most significant bits of the result will be filled with zero bits after
3994 the shift. If ``op2`` is (statically or dynamically) equal to or larger
3995 than the number of bits in ``op1``, the result is undefined. If the
3996 arguments are vectors, each vector element of ``op1`` is shifted by the
3997 corresponding shift amount in ``op2``.
3999 If the ``exact`` keyword is present, the result value of the ``lshr`` is
4000 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
4006 .. code-block:: llvm
4008 <result> = lshr i32 4, 1 ; yields {i32}:result = 2
4009 <result> = lshr i32 4, 2 ; yields {i32}:result = 1
4010 <result> = lshr i8 4, 3 ; yields {i8}:result = 0
4011 <result> = lshr i8 -2, 1 ; yields {i8}:result = 0x7F
4012 <result> = lshr i32 1, 32 ; undefined
4013 <result> = lshr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 0x7FFFFFFF, i32 1>
4015 '``ashr``' Instruction
4016 ^^^^^^^^^^^^^^^^^^^^^^
4023 <result> = ashr <ty> <op1>, <op2> ; yields {ty}:result
4024 <result> = ashr exact <ty> <op1>, <op2> ; yields {ty}:result
4029 The '``ashr``' instruction (arithmetic shift right) returns the first
4030 operand shifted to the right a specified number of bits with sign
4036 Both arguments to the '``ashr``' instruction must be the same
4037 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4038 '``op2``' is treated as an unsigned value.
4043 This instruction always performs an arithmetic shift right operation,
4044 The most significant bits of the result will be filled with the sign bit
4045 of ``op1``. If ``op2`` is (statically or dynamically) equal to or larger
4046 than the number of bits in ``op1``, the result is undefined. If the
4047 arguments are vectors, each vector element of ``op1`` is shifted by the
4048 corresponding shift amount in ``op2``.
4050 If the ``exact`` keyword is present, the result value of the ``ashr`` is
4051 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
4057 .. code-block:: llvm
4059 <result> = ashr i32 4, 1 ; yields {i32}:result = 2
4060 <result> = ashr i32 4, 2 ; yields {i32}:result = 1
4061 <result> = ashr i8 4, 3 ; yields {i8}:result = 0
4062 <result> = ashr i8 -2, 1 ; yields {i8}:result = -1
4063 <result> = ashr i32 1, 32 ; undefined
4064 <result> = ashr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 3> ; yields: result=<2 x i32> < i32 -1, i32 0>
4066 '``and``' Instruction
4067 ^^^^^^^^^^^^^^^^^^^^^
4074 <result> = and <ty> <op1>, <op2> ; yields {ty}:result
4079 The '``and``' instruction returns the bitwise logical and of its two
4085 The two arguments to the '``and``' instruction must be
4086 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4087 arguments must have identical types.
4092 The truth table used for the '``and``' instruction is:
4109 .. code-block:: llvm
4111 <result> = and i32 4, %var ; yields {i32}:result = 4 & %var
4112 <result> = and i32 15, 40 ; yields {i32}:result = 8
4113 <result> = and i32 4, 8 ; yields {i32}:result = 0
4115 '``or``' Instruction
4116 ^^^^^^^^^^^^^^^^^^^^
4123 <result> = or <ty> <op1>, <op2> ; yields {ty}:result
4128 The '``or``' instruction returns the bitwise logical inclusive or of its
4134 The two arguments to the '``or``' instruction must be
4135 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4136 arguments must have identical types.
4141 The truth table used for the '``or``' instruction is:
4160 <result> = or i32 4, %var ; yields {i32}:result = 4 | %var
4161 <result> = or i32 15, 40 ; yields {i32}:result = 47
4162 <result> = or i32 4, 8 ; yields {i32}:result = 12
4164 '``xor``' Instruction
4165 ^^^^^^^^^^^^^^^^^^^^^
4172 <result> = xor <ty> <op1>, <op2> ; yields {ty}:result
4177 The '``xor``' instruction returns the bitwise logical exclusive or of
4178 its two operands. The ``xor`` is used to implement the "one's
4179 complement" operation, which is the "~" operator in C.
4184 The two arguments to the '``xor``' instruction must be
4185 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4186 arguments must have identical types.
4191 The truth table used for the '``xor``' instruction is:
4208 .. code-block:: llvm
4210 <result> = xor i32 4, %var ; yields {i32}:result = 4 ^ %var
4211 <result> = xor i32 15, 40 ; yields {i32}:result = 39
4212 <result> = xor i32 4, 8 ; yields {i32}:result = 12
4213 <result> = xor i32 %V, -1 ; yields {i32}:result = ~%V
4218 LLVM supports several instructions to represent vector operations in a
4219 target-independent manner. These instructions cover the element-access
4220 and vector-specific operations needed to process vectors effectively.
4221 While LLVM does directly support these vector operations, many
4222 sophisticated algorithms will want to use target-specific intrinsics to
4223 take full advantage of a specific target.
4225 .. _i_extractelement:
4227 '``extractelement``' Instruction
4228 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4235 <result> = extractelement <n x <ty>> <val>, i32 <idx> ; yields <ty>
4240 The '``extractelement``' instruction extracts a single scalar element
4241 from a vector at a specified index.
4246 The first operand of an '``extractelement``' instruction is a value of
4247 :ref:`vector <t_vector>` type. The second operand is an index indicating
4248 the position from which to extract the element. The index may be a
4254 The result is a scalar of the same type as the element type of ``val``.
4255 Its value is the value at position ``idx`` of ``val``. If ``idx``
4256 exceeds the length of ``val``, the results are undefined.
4261 .. code-block:: llvm
4263 <result> = extractelement <4 x i32> %vec, i32 0 ; yields i32
4265 .. _i_insertelement:
4267 '``insertelement``' Instruction
4268 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4275 <result> = insertelement <n x <ty>> <val>, <ty> <elt>, i32 <idx> ; yields <n x <ty>>
4280 The '``insertelement``' instruction inserts a scalar element into a
4281 vector at a specified index.
4286 The first operand of an '``insertelement``' instruction is a value of
4287 :ref:`vector <t_vector>` type. The second operand is a scalar value whose
4288 type must equal the element type of the first operand. The third operand
4289 is an index indicating the position at which to insert the value. The
4290 index may be a variable.
4295 The result is a vector of the same type as ``val``. Its element values
4296 are those of ``val`` except at position ``idx``, where it gets the value
4297 ``elt``. If ``idx`` exceeds the length of ``val``, the results are
4303 .. code-block:: llvm
4305 <result> = insertelement <4 x i32> %vec, i32 1, i32 0 ; yields <4 x i32>
4307 .. _i_shufflevector:
4309 '``shufflevector``' Instruction
4310 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4317 <result> = shufflevector <n x <ty>> <v1>, <n x <ty>> <v2>, <m x i32> <mask> ; yields <m x <ty>>
4322 The '``shufflevector``' instruction constructs a permutation of elements
4323 from two input vectors, returning a vector with the same element type as
4324 the input and length that is the same as the shuffle mask.
4329 The first two operands of a '``shufflevector``' instruction are vectors
4330 with the same type. The third argument is a shuffle mask whose element
4331 type is always 'i32'. The result of the instruction is a vector whose
4332 length is the same as the shuffle mask and whose element type is the
4333 same as the element type of the first two operands.
4335 The shuffle mask operand is required to be a constant vector with either
4336 constant integer or undef values.
4341 The elements of the two input vectors are numbered from left to right
4342 across both of the vectors. The shuffle mask operand specifies, for each
4343 element of the result vector, which element of the two input vectors the
4344 result element gets. The element selector may be undef (meaning "don't
4345 care") and the second operand may be undef if performing a shuffle from
4351 .. code-block:: llvm
4353 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
4354 <4 x i32> <i32 0, i32 4, i32 1, i32 5> ; yields <4 x i32>
4355 <result> = shufflevector <4 x i32> %v1, <4 x i32> undef,
4356 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32> - Identity shuffle.
4357 <result> = shufflevector <8 x i32> %v1, <8 x i32> undef,
4358 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32>
4359 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
4360 <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7 > ; yields <8 x i32>
4362 Aggregate Operations
4363 --------------------
4365 LLVM supports several instructions for working with
4366 :ref:`aggregate <t_aggregate>` values.
4370 '``extractvalue``' Instruction
4371 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4378 <result> = extractvalue <aggregate type> <val>, <idx>{, <idx>}*
4383 The '``extractvalue``' instruction extracts the value of a member field
4384 from an :ref:`aggregate <t_aggregate>` value.
4389 The first operand of an '``extractvalue``' instruction is a value of
4390 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The operands are
4391 constant indices to specify which value to extract in a similar manner
4392 as indices in a '``getelementptr``' instruction.
4394 The major differences to ``getelementptr`` indexing are:
4396 - Since the value being indexed is not a pointer, the first index is
4397 omitted and assumed to be zero.
4398 - At least one index must be specified.
4399 - Not only struct indices but also array indices must be in bounds.
4404 The result is the value at the position in the aggregate specified by
4410 .. code-block:: llvm
4412 <result> = extractvalue {i32, float} %agg, 0 ; yields i32
4416 '``insertvalue``' Instruction
4417 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4424 <result> = insertvalue <aggregate type> <val>, <ty> <elt>, <idx>{, <idx>}* ; yields <aggregate type>
4429 The '``insertvalue``' instruction inserts a value into a member field in
4430 an :ref:`aggregate <t_aggregate>` value.
4435 The first operand of an '``insertvalue``' instruction is a value of
4436 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The second operand is
4437 a first-class value to insert. The following operands are constant
4438 indices indicating the position at which to insert the value in a
4439 similar manner as indices in a '``extractvalue``' instruction. The value
4440 to insert must have the same type as the value identified by the
4446 The result is an aggregate of the same type as ``val``. Its value is
4447 that of ``val`` except that the value at the position specified by the
4448 indices is that of ``elt``.
4453 .. code-block:: llvm
4455 %agg1 = insertvalue {i32, float} undef, i32 1, 0 ; yields {i32 1, float undef}
4456 %agg2 = insertvalue {i32, float} %agg1, float %val, 1 ; yields {i32 1, float %val}
4457 %agg3 = insertvalue {i32, {float}} %agg1, float %val, 1, 0 ; yields {i32 1, float %val}
4461 Memory Access and Addressing Operations
4462 ---------------------------------------
4464 A key design point of an SSA-based representation is how it represents
4465 memory. In LLVM, no memory locations are in SSA form, which makes things
4466 very simple. This section describes how to read, write, and allocate
4471 '``alloca``' Instruction
4472 ^^^^^^^^^^^^^^^^^^^^^^^^
4479 <result> = alloca <type>[, <ty> <NumElements>][, align <alignment>] ; yields {type*}:result
4484 The '``alloca``' instruction allocates memory on the stack frame of the
4485 currently executing function, to be automatically released when this
4486 function returns to its caller. The object is always allocated in the
4487 generic address space (address space zero).
4492 The '``alloca``' instruction allocates ``sizeof(<type>)*NumElements``
4493 bytes of memory on the runtime stack, returning a pointer of the
4494 appropriate type to the program. If "NumElements" is specified, it is
4495 the number of elements allocated, otherwise "NumElements" is defaulted
4496 to be one. If a constant alignment is specified, the value result of the
4497 allocation is guaranteed to be aligned to at least that boundary. If not
4498 specified, or if zero, the target can choose to align the allocation on
4499 any convenient boundary compatible with the type.
4501 '``type``' may be any sized type.
4506 Memory is allocated; a pointer is returned. The operation is undefined
4507 if there is insufficient stack space for the allocation. '``alloca``'d
4508 memory is automatically released when the function returns. The
4509 '``alloca``' instruction is commonly used to represent automatic
4510 variables that must have an address available. When the function returns
4511 (either with the ``ret`` or ``resume`` instructions), the memory is
4512 reclaimed. Allocating zero bytes is legal, but the result is undefined.
4513 The order in which memory is allocated (ie., which way the stack grows)
4519 .. code-block:: llvm
4521 %ptr = alloca i32 ; yields {i32*}:ptr
4522 %ptr = alloca i32, i32 4 ; yields {i32*}:ptr
4523 %ptr = alloca i32, i32 4, align 1024 ; yields {i32*}:ptr
4524 %ptr = alloca i32, align 1024 ; yields {i32*}:ptr
4528 '``load``' Instruction
4529 ^^^^^^^^^^^^^^^^^^^^^^
4536 <result> = load [volatile] <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>][, !invariant.load !<index>]
4537 <result> = load atomic [volatile] <ty>* <pointer> [singlethread] <ordering>, align <alignment>
4538 !<index> = !{ i32 1 }
4543 The '``load``' instruction is used to read from memory.
4548 The argument to the ``load`` instruction specifies the memory address
4549 from which to load. The pointer must point to a :ref:`first
4550 class <t_firstclass>` type. If the ``load`` is marked as ``volatile``,
4551 then the optimizer is not allowed to modify the number or order of
4552 execution of this ``load`` with other :ref:`volatile
4553 operations <volatile>`.
4555 If the ``load`` is marked as ``atomic``, it takes an extra
4556 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
4557 ``release`` and ``acq_rel`` orderings are not valid on ``load``
4558 instructions. Atomic loads produce :ref:`defined <memmodel>` results
4559 when they may see multiple atomic stores. The type of the pointee must
4560 be an integer type whose bit width is a power of two greater than or
4561 equal to eight and less than or equal to a target-specific size limit.
4562 ``align`` must be explicitly specified on atomic loads, and the load has
4563 undefined behavior if the alignment is not set to a value which is at
4564 least the size in bytes of the pointee. ``!nontemporal`` does not have
4565 any defined semantics for atomic loads.
4567 The optional constant ``align`` argument specifies the alignment of the
4568 operation (that is, the alignment of the memory address). A value of 0
4569 or an omitted ``align`` argument means that the operation has the ABI
4570 alignment for the target. It is the responsibility of the code emitter
4571 to ensure that the alignment information is correct. Overestimating the
4572 alignment results in undefined behavior. Underestimating the alignment
4573 may produce less efficient code. An alignment of 1 is always safe.
4575 The optional ``!nontemporal`` metadata must reference a single
4576 metatadata name ``<index>`` corresponding to a metadata node with one
4577 ``i32`` entry of value 1. The existence of the ``!nontemporal``
4578 metatadata on the instruction tells the optimizer and code generator
4579 that this load is not expected to be reused in the cache. The code
4580 generator may select special instructions to save cache bandwidth, such
4581 as the ``MOVNT`` instruction on x86.
4583 The optional ``!invariant.load`` metadata must reference a single
4584 metatadata name ``<index>`` corresponding to a metadata node with no
4585 entries. The existence of the ``!invariant.load`` metatadata on the
4586 instruction tells the optimizer and code generator that this load
4587 address points to memory which does not change value during program
4588 execution. The optimizer may then move this load around, for example, by
4589 hoisting it out of loops using loop invariant code motion.
4594 The location of memory pointed to is loaded. If the value being loaded
4595 is of scalar type then the number of bytes read does not exceed the
4596 minimum number of bytes needed to hold all bits of the type. For
4597 example, loading an ``i24`` reads at most three bytes. When loading a
4598 value of a type like ``i20`` with a size that is not an integral number
4599 of bytes, the result is undefined if the value was not originally
4600 written using a store of the same type.
4605 .. code-block:: llvm
4607 %ptr = alloca i32 ; yields {i32*}:ptr
4608 store i32 3, i32* %ptr ; yields {void}
4609 %val = load i32* %ptr ; yields {i32}:val = i32 3
4613 '``store``' Instruction
4614 ^^^^^^^^^^^^^^^^^^^^^^^
4621 store [volatile] <ty> <value>, <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>] ; yields {void}
4622 store atomic [volatile] <ty> <value>, <ty>* <pointer> [singlethread] <ordering>, align <alignment> ; yields {void}
4627 The '``store``' instruction is used to write to memory.
4632 There are two arguments to the ``store`` instruction: a value to store
4633 and an address at which to store it. The type of the ``<pointer>``
4634 operand must be a pointer to the :ref:`first class <t_firstclass>` type of
4635 the ``<value>`` operand. If the ``store`` is marked as ``volatile``,
4636 then the optimizer is not allowed to modify the number or order of
4637 execution of this ``store`` with other :ref:`volatile
4638 operations <volatile>`.
4640 If the ``store`` is marked as ``atomic``, it takes an extra
4641 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
4642 ``acquire`` and ``acq_rel`` orderings aren't valid on ``store``
4643 instructions. Atomic loads produce :ref:`defined <memmodel>` results
4644 when they may see multiple atomic stores. The type of the pointee must
4645 be an integer type whose bit width is a power of two greater than or
4646 equal to eight and less than or equal to a target-specific size limit.
4647 ``align`` must be explicitly specified on atomic stores, and the store
4648 has undefined behavior if the alignment is not set to a value which is
4649 at least the size in bytes of the pointee. ``!nontemporal`` does not
4650 have any defined semantics for atomic stores.
4652 The optional constant ``align`` argument specifies the alignment of the
4653 operation (that is, the alignment of the memory address). A value of 0
4654 or an omitted ``align`` argument means that the operation has the ABI
4655 alignment for the target. It is the responsibility of the code emitter
4656 to ensure that the alignment information is correct. Overestimating the
4657 alignment results in undefined behavior. Underestimating the
4658 alignment may produce less efficient code. An alignment of 1 is always
4661 The optional ``!nontemporal`` metadata must reference a single metatadata
4662 name ``<index>`` corresponding to a metadata node with one ``i32`` entry of
4663 value 1. The existence of the ``!nontemporal`` metatadata on the instruction
4664 tells the optimizer and code generator that this load is not expected to
4665 be reused in the cache. The code generator may select special
4666 instructions to save cache bandwidth, such as the MOVNT instruction on
4672 The contents of memory are updated to contain ``<value>`` at the
4673 location specified by the ``<pointer>`` operand. If ``<value>`` is
4674 of scalar type then the number of bytes written does not exceed the
4675 minimum number of bytes needed to hold all bits of the type. For
4676 example, storing an ``i24`` writes at most three bytes. When writing a
4677 value of a type like ``i20`` with a size that is not an integral number
4678 of bytes, it is unspecified what happens to the extra bits that do not
4679 belong to the type, but they will typically be overwritten.
4684 .. code-block:: llvm
4686 %ptr = alloca i32 ; yields {i32*}:ptr
4687 store i32 3, i32* %ptr ; yields {void}
4688 %val = load i32* %ptr ; yields {i32}:val = i32 3
4692 '``fence``' Instruction
4693 ^^^^^^^^^^^^^^^^^^^^^^^
4700 fence [singlethread] <ordering> ; yields {void}
4705 The '``fence``' instruction is used to introduce happens-before edges
4711 '``fence``' instructions take an :ref:`ordering <ordering>` argument which
4712 defines what *synchronizes-with* edges they add. They can only be given
4713 ``acquire``, ``release``, ``acq_rel``, and ``seq_cst`` orderings.
4718 A fence A which has (at least) ``release`` ordering semantics
4719 *synchronizes with* a fence B with (at least) ``acquire`` ordering
4720 semantics if and only if there exist atomic operations X and Y, both
4721 operating on some atomic object M, such that A is sequenced before X, X
4722 modifies M (either directly or through some side effect of a sequence
4723 headed by X), Y is sequenced before B, and Y observes M. This provides a
4724 *happens-before* dependency between A and B. Rather than an explicit
4725 ``fence``, one (but not both) of the atomic operations X or Y might
4726 provide a ``release`` or ``acquire`` (resp.) ordering constraint and
4727 still *synchronize-with* the explicit ``fence`` and establish the
4728 *happens-before* edge.
4730 A ``fence`` which has ``seq_cst`` ordering, in addition to having both
4731 ``acquire`` and ``release`` semantics specified above, participates in
4732 the global program order of other ``seq_cst`` operations and/or fences.
4734 The optional ":ref:`singlethread <singlethread>`" argument specifies
4735 that the fence only synchronizes with other fences in the same thread.
4736 (This is useful for interacting with signal handlers.)
4741 .. code-block:: llvm
4743 fence acquire ; yields {void}
4744 fence singlethread seq_cst ; yields {void}
4748 '``cmpxchg``' Instruction
4749 ^^^^^^^^^^^^^^^^^^^^^^^^^
4756 cmpxchg [volatile] <ty>* <pointer>, <ty> <cmp>, <ty> <new> [singlethread] <ordering> ; yields {ty}
4761 The '``cmpxchg``' instruction is used to atomically modify memory. It
4762 loads a value in memory and compares it to a given value. If they are
4763 equal, it stores a new value into the memory.
4768 There are three arguments to the '``cmpxchg``' instruction: an address
4769 to operate on, a value to compare to the value currently be at that
4770 address, and a new value to place at that address if the compared values
4771 are equal. The type of '<cmp>' must be an integer type whose bit width
4772 is a power of two greater than or equal to eight and less than or equal
4773 to a target-specific size limit. '<cmp>' and '<new>' must have the same
4774 type, and the type of '<pointer>' must be a pointer to that type. If the
4775 ``cmpxchg`` is marked as ``volatile``, then the optimizer is not allowed
4776 to modify the number or order of execution of this ``cmpxchg`` with
4777 other :ref:`volatile operations <volatile>`.
4779 The :ref:`ordering <ordering>` argument specifies how this ``cmpxchg``
4780 synchronizes with other atomic operations.
4782 The optional "``singlethread``" argument declares that the ``cmpxchg``
4783 is only atomic with respect to code (usually signal handlers) running in
4784 the same thread as the ``cmpxchg``. Otherwise the cmpxchg is atomic with
4785 respect to all other code in the system.
4787 The pointer passed into cmpxchg must have alignment greater than or
4788 equal to the size in memory of the operand.
4793 The contents of memory at the location specified by the '``<pointer>``'
4794 operand is read and compared to '``<cmp>``'; if the read value is the
4795 equal, '``<new>``' is written. The original value at the location is
4798 A successful ``cmpxchg`` is a read-modify-write instruction for the purpose
4799 of identifying release sequences. A failed ``cmpxchg`` is equivalent to an
4800 atomic load with an ordering parameter determined by dropping any
4801 ``release`` part of the ``cmpxchg``'s ordering.
4806 .. code-block:: llvm
4809 %orig = atomic load i32* %ptr unordered ; yields {i32}
4813 %cmp = phi i32 [ %orig, %entry ], [%old, %loop]
4814 %squared = mul i32 %cmp, %cmp
4815 %old = cmpxchg i32* %ptr, i32 %cmp, i32 %squared ; yields {i32}
4816 %success = icmp eq i32 %cmp, %old
4817 br i1 %success, label %done, label %loop
4824 '``atomicrmw``' Instruction
4825 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
4832 atomicrmw [volatile] <operation> <ty>* <pointer>, <ty> <value> [singlethread] <ordering> ; yields {ty}
4837 The '``atomicrmw``' instruction is used to atomically modify memory.
4842 There are three arguments to the '``atomicrmw``' instruction: an
4843 operation to apply, an address whose value to modify, an argument to the
4844 operation. The operation must be one of the following keywords:
4858 The type of '<value>' must be an integer type whose bit width is a power
4859 of two greater than or equal to eight and less than or equal to a
4860 target-specific size limit. The type of the '``<pointer>``' operand must
4861 be a pointer to that type. If the ``atomicrmw`` is marked as
4862 ``volatile``, then the optimizer is not allowed to modify the number or
4863 order of execution of this ``atomicrmw`` with other :ref:`volatile
4864 operations <volatile>`.
4869 The contents of memory at the location specified by the '``<pointer>``'
4870 operand are atomically read, modified, and written back. The original
4871 value at the location is returned. The modification is specified by the
4874 - xchg: ``*ptr = val``
4875 - add: ``*ptr = *ptr + val``
4876 - sub: ``*ptr = *ptr - val``
4877 - and: ``*ptr = *ptr & val``
4878 - nand: ``*ptr = ~(*ptr & val)``
4879 - or: ``*ptr = *ptr | val``
4880 - xor: ``*ptr = *ptr ^ val``
4881 - max: ``*ptr = *ptr > val ? *ptr : val`` (using a signed comparison)
4882 - min: ``*ptr = *ptr < val ? *ptr : val`` (using a signed comparison)
4883 - umax: ``*ptr = *ptr > val ? *ptr : val`` (using an unsigned
4885 - umin: ``*ptr = *ptr < val ? *ptr : val`` (using an unsigned
4891 .. code-block:: llvm
4893 %old = atomicrmw add i32* %ptr, i32 1 acquire ; yields {i32}
4895 .. _i_getelementptr:
4897 '``getelementptr``' Instruction
4898 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4905 <result> = getelementptr <pty>* <ptrval>{, <ty> <idx>}*
4906 <result> = getelementptr inbounds <pty>* <ptrval>{, <ty> <idx>}*
4907 <result> = getelementptr <ptr vector> ptrval, <vector index type> idx
4912 The '``getelementptr``' instruction is used to get the address of a
4913 subelement of an :ref:`aggregate <t_aggregate>` data structure. It performs
4914 address calculation only and does not access memory.
4919 The first argument is always a pointer or a vector of pointers, and
4920 forms the basis of the calculation. The remaining arguments are indices
4921 that indicate which of the elements of the aggregate object are indexed.
4922 The interpretation of each index is dependent on the type being indexed
4923 into. The first index always indexes the pointer value given as the
4924 first argument, the second index indexes a value of the type pointed to
4925 (not necessarily the value directly pointed to, since the first index
4926 can be non-zero), etc. The first type indexed into must be a pointer
4927 value, subsequent types can be arrays, vectors, and structs. Note that
4928 subsequent types being indexed into can never be pointers, since that
4929 would require loading the pointer before continuing calculation.
4931 The type of each index argument depends on the type it is indexing into.
4932 When indexing into a (optionally packed) structure, only ``i32`` integer
4933 **constants** are allowed (when using a vector of indices they must all
4934 be the **same** ``i32`` integer constant). When indexing into an array,
4935 pointer or vector, integers of any width are allowed, and they are not
4936 required to be constant. These integers are treated as signed values
4939 For example, let's consider a C code fragment and how it gets compiled
4955 int *foo(struct ST *s) {
4956 return &s[1].Z.B[5][13];
4959 The LLVM code generated by Clang is:
4961 .. code-block:: llvm
4963 %struct.RT = type { i8, [10 x [20 x i32]], i8 }
4964 %struct.ST = type { i32, double, %struct.RT }
4966 define i32* @foo(%struct.ST* %s) nounwind uwtable readnone optsize ssp {
4968 %arrayidx = getelementptr inbounds %struct.ST* %s, i64 1, i32 2, i32 1, i64 5, i64 13
4975 In the example above, the first index is indexing into the
4976 '``%struct.ST*``' type, which is a pointer, yielding a '``%struct.ST``'
4977 = '``{ i32, double, %struct.RT }``' type, a structure. The second index
4978 indexes into the third element of the structure, yielding a
4979 '``%struct.RT``' = '``{ i8 , [10 x [20 x i32]], i8 }``' type, another
4980 structure. The third index indexes into the second element of the
4981 structure, yielding a '``[10 x [20 x i32]]``' type, an array. The two
4982 dimensions of the array are subscripted into, yielding an '``i32``'
4983 type. The '``getelementptr``' instruction returns a pointer to this
4984 element, thus computing a value of '``i32*``' type.
4986 Note that it is perfectly legal to index partially through a structure,
4987 returning a pointer to an inner element. Because of this, the LLVM code
4988 for the given testcase is equivalent to:
4990 .. code-block:: llvm
4992 define i32* @foo(%struct.ST* %s) {
4993 %t1 = getelementptr %struct.ST* %s, i32 1 ; yields %struct.ST*:%t1
4994 %t2 = getelementptr %struct.ST* %t1, i32 0, i32 2 ; yields %struct.RT*:%t2
4995 %t3 = getelementptr %struct.RT* %t2, i32 0, i32 1 ; yields [10 x [20 x i32]]*:%t3
4996 %t4 = getelementptr [10 x [20 x i32]]* %t3, i32 0, i32 5 ; yields [20 x i32]*:%t4
4997 %t5 = getelementptr [20 x i32]* %t4, i32 0, i32 13 ; yields i32*:%t5
5001 If the ``inbounds`` keyword is present, the result value of the
5002 ``getelementptr`` is a :ref:`poison value <poisonvalues>` if the base
5003 pointer is not an *in bounds* address of an allocated object, or if any
5004 of the addresses that would be formed by successive addition of the
5005 offsets implied by the indices to the base address with infinitely
5006 precise signed arithmetic are not an *in bounds* address of that
5007 allocated object. The *in bounds* addresses for an allocated object are
5008 all the addresses that point into the object, plus the address one byte
5009 past the end. In cases where the base is a vector of pointers the
5010 ``inbounds`` keyword applies to each of the computations element-wise.
5012 If the ``inbounds`` keyword is not present, the offsets are added to the
5013 base address with silently-wrapping two's complement arithmetic. If the
5014 offsets have a different width from the pointer, they are sign-extended
5015 or truncated to the width of the pointer. The result value of the
5016 ``getelementptr`` may be outside the object pointed to by the base
5017 pointer. The result value may not necessarily be used to access memory
5018 though, even if it happens to point into allocated storage. See the
5019 :ref:`Pointer Aliasing Rules <pointeraliasing>` section for more
5022 The getelementptr instruction is often confusing. For some more insight
5023 into how it works, see :doc:`the getelementptr FAQ <GetElementPtr>`.
5028 .. code-block:: llvm
5030 ; yields [12 x i8]*:aptr
5031 %aptr = getelementptr {i32, [12 x i8]}* %saptr, i64 0, i32 1
5033 %vptr = getelementptr {i32, <2 x i8>}* %svptr, i64 0, i32 1, i32 1
5035 %eptr = getelementptr [12 x i8]* %aptr, i64 0, i32 1
5037 %iptr = getelementptr [10 x i32]* @arr, i16 0, i16 0
5039 In cases where the pointer argument is a vector of pointers, each index
5040 must be a vector with the same number of elements. For example:
5042 .. code-block:: llvm
5044 %A = getelementptr <4 x i8*> %ptrs, <4 x i64> %offsets,
5046 Conversion Operations
5047 ---------------------
5049 The instructions in this category are the conversion instructions
5050 (casting) which all take a single operand and a type. They perform
5051 various bit conversions on the operand.
5053 '``trunc .. to``' Instruction
5054 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5061 <result> = trunc <ty> <value> to <ty2> ; yields ty2
5066 The '``trunc``' instruction truncates its operand to the type ``ty2``.
5071 The '``trunc``' instruction takes a value to trunc, and a type to trunc
5072 it to. Both types must be of :ref:`integer <t_integer>` types, or vectors
5073 of the same number of integers. The bit size of the ``value`` must be
5074 larger than the bit size of the destination type, ``ty2``. Equal sized
5075 types are not allowed.
5080 The '``trunc``' instruction truncates the high order bits in ``value``
5081 and converts the remaining bits to ``ty2``. Since the source size must
5082 be larger than the destination size, ``trunc`` cannot be a *no-op cast*.
5083 It will always truncate bits.
5088 .. code-block:: llvm
5090 %X = trunc i32 257 to i8 ; yields i8:1
5091 %Y = trunc i32 123 to i1 ; yields i1:true
5092 %Z = trunc i32 122 to i1 ; yields i1:false
5093 %W = trunc <2 x i16> <i16 8, i16 7> to <2 x i8> ; yields <i8 8, i8 7>
5095 '``zext .. to``' Instruction
5096 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5103 <result> = zext <ty> <value> to <ty2> ; yields ty2
5108 The '``zext``' instruction zero extends its operand to type ``ty2``.
5113 The '``zext``' instruction takes a value to cast, and a type to cast it
5114 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
5115 the same number of integers. The bit size of the ``value`` must be
5116 smaller than the bit size of the destination type, ``ty2``.
5121 The ``zext`` fills the high order bits of the ``value`` with zero bits
5122 until it reaches the size of the destination type, ``ty2``.
5124 When zero extending from i1, the result will always be either 0 or 1.
5129 .. code-block:: llvm
5131 %X = zext i32 257 to i64 ; yields i64:257
5132 %Y = zext i1 true to i32 ; yields i32:1
5133 %Z = zext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
5135 '``sext .. to``' Instruction
5136 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5143 <result> = sext <ty> <value> to <ty2> ; yields ty2
5148 The '``sext``' sign extends ``value`` to the type ``ty2``.
5153 The '``sext``' instruction takes a value to cast, and a type to cast it
5154 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
5155 the same number of integers. The bit size of the ``value`` must be
5156 smaller than the bit size of the destination type, ``ty2``.
5161 The '``sext``' instruction performs a sign extension by copying the sign
5162 bit (highest order bit) of the ``value`` until it reaches the bit size
5163 of the type ``ty2``.
5165 When sign extending from i1, the extension always results in -1 or 0.
5170 .. code-block:: llvm
5172 %X = sext i8 -1 to i16 ; yields i16 :65535
5173 %Y = sext i1 true to i32 ; yields i32:-1
5174 %Z = sext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
5176 '``fptrunc .. to``' Instruction
5177 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5184 <result> = fptrunc <ty> <value> to <ty2> ; yields ty2
5189 The '``fptrunc``' instruction truncates ``value`` to type ``ty2``.
5194 The '``fptrunc``' instruction takes a :ref:`floating point <t_floating>`
5195 value to cast and a :ref:`floating point <t_floating>` type to cast it to.
5196 The size of ``value`` must be larger than the size of ``ty2``. This
5197 implies that ``fptrunc`` cannot be used to make a *no-op cast*.
5202 The '``fptrunc``' instruction truncates a ``value`` from a larger
5203 :ref:`floating point <t_floating>` type to a smaller :ref:`floating
5204 point <t_floating>` type. If the value cannot fit within the
5205 destination type, ``ty2``, then the results are undefined.
5210 .. code-block:: llvm
5212 %X = fptrunc double 123.0 to float ; yields float:123.0
5213 %Y = fptrunc double 1.0E+300 to float ; yields undefined
5215 '``fpext .. to``' Instruction
5216 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5223 <result> = fpext <ty> <value> to <ty2> ; yields ty2
5228 The '``fpext``' extends a floating point ``value`` to a larger floating
5234 The '``fpext``' instruction takes a :ref:`floating point <t_floating>`
5235 ``value`` to cast, and a :ref:`floating point <t_floating>` type to cast it
5236 to. The source type must be smaller than the destination type.
5241 The '``fpext``' instruction extends the ``value`` from a smaller
5242 :ref:`floating point <t_floating>` type to a larger :ref:`floating
5243 point <t_floating>` type. The ``fpext`` cannot be used to make a
5244 *no-op cast* because it always changes bits. Use ``bitcast`` to make a
5245 *no-op cast* for a floating point cast.
5250 .. code-block:: llvm
5252 %X = fpext float 3.125 to double ; yields double:3.125000e+00
5253 %Y = fpext double %X to fp128 ; yields fp128:0xL00000000000000004000900000000000
5255 '``fptoui .. to``' Instruction
5256 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5263 <result> = fptoui <ty> <value> to <ty2> ; yields ty2
5268 The '``fptoui``' converts a floating point ``value`` to its unsigned
5269 integer equivalent of type ``ty2``.
5274 The '``fptoui``' instruction takes a value to cast, which must be a
5275 scalar or vector :ref:`floating point <t_floating>` value, and a type to
5276 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
5277 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
5278 type with the same number of elements as ``ty``
5283 The '``fptoui``' instruction converts its :ref:`floating
5284 point <t_floating>` operand into the nearest (rounding towards zero)
5285 unsigned integer value. If the value cannot fit in ``ty2``, the results
5291 .. code-block:: llvm
5293 %X = fptoui double 123.0 to i32 ; yields i32:123
5294 %Y = fptoui float 1.0E+300 to i1 ; yields undefined:1
5295 %Z = fptoui float 1.04E+17 to i8 ; yields undefined:1
5297 '``fptosi .. to``' Instruction
5298 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5305 <result> = fptosi <ty> <value> to <ty2> ; yields ty2
5310 The '``fptosi``' instruction converts :ref:`floating point <t_floating>`
5311 ``value`` to type ``ty2``.
5316 The '``fptosi``' instruction takes a value to cast, which must be a
5317 scalar or vector :ref:`floating point <t_floating>` value, and a type to
5318 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
5319 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
5320 type with the same number of elements as ``ty``
5325 The '``fptosi``' instruction converts its :ref:`floating
5326 point <t_floating>` operand into the nearest (rounding towards zero)
5327 signed integer value. If the value cannot fit in ``ty2``, the results
5333 .. code-block:: llvm
5335 %X = fptosi double -123.0 to i32 ; yields i32:-123
5336 %Y = fptosi float 1.0E-247 to i1 ; yields undefined:1
5337 %Z = fptosi float 1.04E+17 to i8 ; yields undefined:1
5339 '``uitofp .. to``' Instruction
5340 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5347 <result> = uitofp <ty> <value> to <ty2> ; yields ty2
5352 The '``uitofp``' instruction regards ``value`` as an unsigned integer
5353 and converts that value to the ``ty2`` type.
5358 The '``uitofp``' instruction takes a value to cast, which must be a
5359 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
5360 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
5361 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
5362 type with the same number of elements as ``ty``
5367 The '``uitofp``' instruction interprets its operand as an unsigned
5368 integer quantity and converts it to the corresponding floating point
5369 value. If the value cannot fit in the floating point value, the results
5375 .. code-block:: llvm
5377 %X = uitofp i32 257 to float ; yields float:257.0
5378 %Y = uitofp i8 -1 to double ; yields double:255.0
5380 '``sitofp .. to``' Instruction
5381 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5388 <result> = sitofp <ty> <value> to <ty2> ; yields ty2
5393 The '``sitofp``' instruction regards ``value`` as a signed integer and
5394 converts that value to the ``ty2`` type.
5399 The '``sitofp``' instruction takes a value to cast, which must be a
5400 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
5401 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
5402 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
5403 type with the same number of elements as ``ty``
5408 The '``sitofp``' instruction interprets its operand as a signed integer
5409 quantity and converts it to the corresponding floating point value. If
5410 the value cannot fit in the floating point value, the results are
5416 .. code-block:: llvm
5418 %X = sitofp i32 257 to float ; yields float:257.0
5419 %Y = sitofp i8 -1 to double ; yields double:-1.0
5423 '``ptrtoint .. to``' Instruction
5424 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5431 <result> = ptrtoint <ty> <value> to <ty2> ; yields ty2
5436 The '``ptrtoint``' instruction converts the pointer or a vector of
5437 pointers ``value`` to the integer (or vector of integers) type ``ty2``.
5442 The '``ptrtoint``' instruction takes a ``value`` to cast, which must be
5443 a a value of type :ref:`pointer <t_pointer>` or a vector of pointers, and a
5444 type to cast it to ``ty2``, which must be an :ref:`integer <t_integer>` or
5445 a vector of integers type.
5450 The '``ptrtoint``' instruction converts ``value`` to integer type
5451 ``ty2`` by interpreting the pointer value as an integer and either
5452 truncating or zero extending that value to the size of the integer type.
5453 If ``value`` is smaller than ``ty2`` then a zero extension is done. If
5454 ``value`` is larger than ``ty2`` then a truncation is done. If they are
5455 the same size, then nothing is done (*no-op cast*) other than a type
5461 .. code-block:: llvm
5463 %X = ptrtoint i32* %P to i8 ; yields truncation on 32-bit architecture
5464 %Y = ptrtoint i32* %P to i64 ; yields zero extension on 32-bit architecture
5465 %Z = ptrtoint <4 x i32*> %P to <4 x i64>; yields vector zero extension for a vector of addresses on 32-bit architecture
5469 '``inttoptr .. to``' Instruction
5470 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5477 <result> = inttoptr <ty> <value> to <ty2> ; yields ty2
5482 The '``inttoptr``' instruction converts an integer ``value`` to a
5483 pointer type, ``ty2``.
5488 The '``inttoptr``' instruction takes an :ref:`integer <t_integer>` value to
5489 cast, and a type to cast it to, which must be a :ref:`pointer <t_pointer>`
5495 The '``inttoptr``' instruction converts ``value`` to type ``ty2`` by
5496 applying either a zero extension or a truncation depending on the size
5497 of the integer ``value``. If ``value`` is larger than the size of a
5498 pointer then a truncation is done. If ``value`` is smaller than the size
5499 of a pointer then a zero extension is done. If they are the same size,
5500 nothing is done (*no-op cast*).
5505 .. code-block:: llvm
5507 %X = inttoptr i32 255 to i32* ; yields zero extension on 64-bit architecture
5508 %Y = inttoptr i32 255 to i32* ; yields no-op on 32-bit architecture
5509 %Z = inttoptr i64 0 to i32* ; yields truncation on 32-bit architecture
5510 %Z = inttoptr <4 x i32> %G to <4 x i8*>; yields truncation of vector G to four pointers
5514 '``bitcast .. to``' Instruction
5515 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5522 <result> = bitcast <ty> <value> to <ty2> ; yields ty2
5527 The '``bitcast``' instruction converts ``value`` to type ``ty2`` without
5533 The '``bitcast``' instruction takes a value to cast, which must be a
5534 non-aggregate first class value, and a type to cast it to, which must
5535 also be a non-aggregate :ref:`first class <t_firstclass>` type. The bit
5536 sizes of ``value`` and the destination type, ``ty2``, must be identical.
5537 If the source type is a pointer, the destination type must also be a
5538 pointer. This instruction supports bitwise conversion of vectors to
5539 integers and to vectors of other types (as long as they have the same
5545 The '``bitcast``' instruction converts ``value`` to type ``ty2``. It is
5546 always a *no-op cast* because no bits change with this conversion. The
5547 conversion is done as if the ``value`` had been stored to memory and
5548 read back as type ``ty2``. Pointer (or vector of pointers) types may
5549 only be converted to other pointer (or vector of pointers) types with
5550 this instruction. To convert pointers to other types, use the
5551 :ref:`inttoptr <i_inttoptr>` or :ref:`ptrtoint <i_ptrtoint>` instructions
5557 .. code-block:: llvm
5559 %X = bitcast i8 255 to i8 ; yields i8 :-1
5560 %Y = bitcast i32* %x to sint* ; yields sint*:%x
5561 %Z = bitcast <2 x int> %V to i64; ; yields i64: %V
5562 %Z = bitcast <2 x i32*> %V to <2 x i64*> ; yields <2 x i64*>
5569 The instructions in this category are the "miscellaneous" instructions,
5570 which defy better classification.
5574 '``icmp``' Instruction
5575 ^^^^^^^^^^^^^^^^^^^^^^
5582 <result> = icmp <cond> <ty> <op1>, <op2> ; yields {i1} or {<N x i1>}:result
5587 The '``icmp``' instruction returns a boolean value or a vector of
5588 boolean values based on comparison of its two integer, integer vector,
5589 pointer, or pointer vector operands.
5594 The '``icmp``' instruction takes three operands. The first operand is
5595 the condition code indicating the kind of comparison to perform. It is
5596 not a value, just a keyword. The possible condition code are:
5599 #. ``ne``: not equal
5600 #. ``ugt``: unsigned greater than
5601 #. ``uge``: unsigned greater or equal
5602 #. ``ult``: unsigned less than
5603 #. ``ule``: unsigned less or equal
5604 #. ``sgt``: signed greater than
5605 #. ``sge``: signed greater or equal
5606 #. ``slt``: signed less than
5607 #. ``sle``: signed less or equal
5609 The remaining two arguments must be :ref:`integer <t_integer>` or
5610 :ref:`pointer <t_pointer>` or integer :ref:`vector <t_vector>` typed. They
5611 must also be identical types.
5616 The '``icmp``' compares ``op1`` and ``op2`` according to the condition
5617 code given as ``cond``. The comparison performed always yields either an
5618 :ref:`i1 <t_integer>` or vector of ``i1`` result, as follows:
5620 #. ``eq``: yields ``true`` if the operands are equal, ``false``
5621 otherwise. No sign interpretation is necessary or performed.
5622 #. ``ne``: yields ``true`` if the operands are unequal, ``false``
5623 otherwise. No sign interpretation is necessary or performed.
5624 #. ``ugt``: interprets the operands as unsigned values and yields
5625 ``true`` if ``op1`` is greater than ``op2``.
5626 #. ``uge``: interprets the operands as unsigned values and yields
5627 ``true`` if ``op1`` is greater than or equal to ``op2``.
5628 #. ``ult``: interprets the operands as unsigned values and yields
5629 ``true`` if ``op1`` is less than ``op2``.
5630 #. ``ule``: interprets the operands as unsigned values and yields
5631 ``true`` if ``op1`` is less than or equal to ``op2``.
5632 #. ``sgt``: interprets the operands as signed values and yields ``true``
5633 if ``op1`` is greater than ``op2``.
5634 #. ``sge``: interprets the operands as signed values and yields ``true``
5635 if ``op1`` is greater than or equal to ``op2``.
5636 #. ``slt``: interprets the operands as signed values and yields ``true``
5637 if ``op1`` is less than ``op2``.
5638 #. ``sle``: interprets the operands as signed values and yields ``true``
5639 if ``op1`` is less than or equal to ``op2``.
5641 If the operands are :ref:`pointer <t_pointer>` typed, the pointer values
5642 are compared as if they were integers.
5644 If the operands are integer vectors, then they are compared element by
5645 element. The result is an ``i1`` vector with the same number of elements
5646 as the values being compared. Otherwise, the result is an ``i1``.
5651 .. code-block:: llvm
5653 <result> = icmp eq i32 4, 5 ; yields: result=false
5654 <result> = icmp ne float* %X, %X ; yields: result=false
5655 <result> = icmp ult i16 4, 5 ; yields: result=true
5656 <result> = icmp sgt i16 4, 5 ; yields: result=false
5657 <result> = icmp ule i16 -4, 5 ; yields: result=false
5658 <result> = icmp sge i16 4, 5 ; yields: result=false
5660 Note that the code generator does not yet support vector types with the
5661 ``icmp`` instruction.
5665 '``fcmp``' Instruction
5666 ^^^^^^^^^^^^^^^^^^^^^^
5673 <result> = fcmp <cond> <ty> <op1>, <op2> ; yields {i1} or {<N x i1>}:result
5678 The '``fcmp``' instruction returns a boolean value or vector of boolean
5679 values based on comparison of its operands.
5681 If the operands are floating point scalars, then the result type is a
5682 boolean (:ref:`i1 <t_integer>`).
5684 If the operands are floating point vectors, then the result type is a
5685 vector of boolean with the same number of elements as the operands being
5691 The '``fcmp``' instruction takes three operands. The first operand is
5692 the condition code indicating the kind of comparison to perform. It is
5693 not a value, just a keyword. The possible condition code are:
5695 #. ``false``: no comparison, always returns false
5696 #. ``oeq``: ordered and equal
5697 #. ``ogt``: ordered and greater than
5698 #. ``oge``: ordered and greater than or equal
5699 #. ``olt``: ordered and less than
5700 #. ``ole``: ordered and less than or equal
5701 #. ``one``: ordered and not equal
5702 #. ``ord``: ordered (no nans)
5703 #. ``ueq``: unordered or equal
5704 #. ``ugt``: unordered or greater than
5705 #. ``uge``: unordered or greater than or equal
5706 #. ``ult``: unordered or less than
5707 #. ``ule``: unordered or less than or equal
5708 #. ``une``: unordered or not equal
5709 #. ``uno``: unordered (either nans)
5710 #. ``true``: no comparison, always returns true
5712 *Ordered* means that neither operand is a QNAN while *unordered* means
5713 that either operand may be a QNAN.
5715 Each of ``val1`` and ``val2`` arguments must be either a :ref:`floating
5716 point <t_floating>` type or a :ref:`vector <t_vector>` of floating point
5717 type. They must have identical types.
5722 The '``fcmp``' instruction compares ``op1`` and ``op2`` according to the
5723 condition code given as ``cond``. If the operands are vectors, then the
5724 vectors are compared element by element. Each comparison performed
5725 always yields an :ref:`i1 <t_integer>` result, as follows:
5727 #. ``false``: always yields ``false``, regardless of operands.
5728 #. ``oeq``: yields ``true`` if both operands are not a QNAN and ``op1``
5729 is equal to ``op2``.
5730 #. ``ogt``: yields ``true`` if both operands are not a QNAN and ``op1``
5731 is greater than ``op2``.
5732 #. ``oge``: yields ``true`` if both operands are not a QNAN and ``op1``
5733 is greater than or equal to ``op2``.
5734 #. ``olt``: yields ``true`` if both operands are not a QNAN and ``op1``
5735 is less than ``op2``.
5736 #. ``ole``: yields ``true`` if both operands are not a QNAN and ``op1``
5737 is less than or equal to ``op2``.
5738 #. ``one``: yields ``true`` if both operands are not a QNAN and ``op1``
5739 is not equal to ``op2``.
5740 #. ``ord``: yields ``true`` if both operands are not a QNAN.
5741 #. ``ueq``: yields ``true`` if either operand is a QNAN or ``op1`` is
5743 #. ``ugt``: yields ``true`` if either operand is a QNAN or ``op1`` is
5744 greater than ``op2``.
5745 #. ``uge``: yields ``true`` if either operand is a QNAN or ``op1`` is
5746 greater than or equal to ``op2``.
5747 #. ``ult``: yields ``true`` if either operand is a QNAN or ``op1`` is
5749 #. ``ule``: yields ``true`` if either operand is a QNAN or ``op1`` is
5750 less than or equal to ``op2``.
5751 #. ``une``: yields ``true`` if either operand is a QNAN or ``op1`` is
5752 not equal to ``op2``.
5753 #. ``uno``: yields ``true`` if either operand is a QNAN.
5754 #. ``true``: always yields ``true``, regardless of operands.
5759 .. code-block:: llvm
5761 <result> = fcmp oeq float 4.0, 5.0 ; yields: result=false
5762 <result> = fcmp one float 4.0, 5.0 ; yields: result=true
5763 <result> = fcmp olt float 4.0, 5.0 ; yields: result=true
5764 <result> = fcmp ueq double 1.0, 2.0 ; yields: result=false
5766 Note that the code generator does not yet support vector types with the
5767 ``fcmp`` instruction.
5771 '``phi``' Instruction
5772 ^^^^^^^^^^^^^^^^^^^^^
5779 <result> = phi <ty> [ <val0>, <label0>], ...
5784 The '``phi``' instruction is used to implement the φ node in the SSA
5785 graph representing the function.
5790 The type of the incoming values is specified with the first type field.
5791 After this, the '``phi``' instruction takes a list of pairs as
5792 arguments, with one pair for each predecessor basic block of the current
5793 block. Only values of :ref:`first class <t_firstclass>` type may be used as
5794 the value arguments to the PHI node. Only labels may be used as the
5797 There must be no non-phi instructions between the start of a basic block
5798 and the PHI instructions: i.e. PHI instructions must be first in a basic
5801 For the purposes of the SSA form, the use of each incoming value is
5802 deemed to occur on the edge from the corresponding predecessor block to
5803 the current block (but after any definition of an '``invoke``'
5804 instruction's return value on the same edge).
5809 At runtime, the '``phi``' instruction logically takes on the value
5810 specified by the pair corresponding to the predecessor basic block that
5811 executed just prior to the current block.
5816 .. code-block:: llvm
5818 Loop: ; Infinite loop that counts from 0 on up...
5819 %indvar = phi i32 [ 0, %LoopHeader ], [ %nextindvar, %Loop ]
5820 %nextindvar = add i32 %indvar, 1
5825 '``select``' Instruction
5826 ^^^^^^^^^^^^^^^^^^^^^^^^
5833 <result> = select selty <cond>, <ty> <val1>, <ty> <val2> ; yields ty
5835 selty is either i1 or {<N x i1>}
5840 The '``select``' instruction is used to choose one value based on a
5841 condition, without branching.
5846 The '``select``' instruction requires an 'i1' value or a vector of 'i1'
5847 values indicating the condition, and two values of the same :ref:`first
5848 class <t_firstclass>` type. If the val1/val2 are vectors and the
5849 condition is a scalar, then entire vectors are selected, not individual
5855 If the condition is an i1 and it evaluates to 1, the instruction returns
5856 the first value argument; otherwise, it returns the second value
5859 If the condition is a vector of i1, then the value arguments must be
5860 vectors of the same size, and the selection is done element by element.
5865 .. code-block:: llvm
5867 %X = select i1 true, i8 17, i8 42 ; yields i8:17
5871 '``call``' Instruction
5872 ^^^^^^^^^^^^^^^^^^^^^^
5879 <result> = [tail] call [cconv] [ret attrs] <ty> [<fnty>*] <fnptrval>(<function args>) [fn attrs]
5884 The '``call``' instruction represents a simple function call.
5889 This instruction requires several arguments:
5891 #. The optional "tail" marker indicates that the callee function does
5892 not access any allocas or varargs in the caller. Note that calls may
5893 be marked "tail" even if they do not occur before a
5894 :ref:`ret <i_ret>` instruction. If the "tail" marker is present, the
5895 function call is eligible for tail call optimization, but `might not
5896 in fact be optimized into a jump <CodeGenerator.html#tailcallopt>`_.
5897 The code generator may optimize calls marked "tail" with either 1)
5898 automatic `sibling call
5899 optimization <CodeGenerator.html#sibcallopt>`_ when the caller and
5900 callee have matching signatures, or 2) forced tail call optimization
5901 when the following extra requirements are met:
5903 - Caller and callee both have the calling convention ``fastcc``.
5904 - The call is in tail position (ret immediately follows call and ret
5905 uses value of call or is void).
5906 - Option ``-tailcallopt`` is enabled, or
5907 ``llvm::GuaranteedTailCallOpt`` is ``true``.
5908 - `Platform specific constraints are
5909 met. <CodeGenerator.html#tailcallopt>`_
5911 #. The optional "cconv" marker indicates which :ref:`calling
5912 convention <callingconv>` the call should use. If none is
5913 specified, the call defaults to using C calling conventions. The
5914 calling convention of the call must match the calling convention of
5915 the target function, or else the behavior is undefined.
5916 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
5917 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
5919 #. '``ty``': the type of the call instruction itself which is also the
5920 type of the return value. Functions that return no value are marked
5922 #. '``fnty``': shall be the signature of the pointer to function value
5923 being invoked. The argument types must match the types implied by
5924 this signature. This type can be omitted if the function is not
5925 varargs and if the function type does not return a pointer to a
5927 #. '``fnptrval``': An LLVM value containing a pointer to a function to
5928 be invoked. In most cases, this is a direct function invocation, but
5929 indirect ``call``'s are just as possible, calling an arbitrary pointer
5931 #. '``function args``': argument list whose types match the function
5932 signature argument types and parameter attributes. All arguments must
5933 be of :ref:`first class <t_firstclass>` type. If the function signature
5934 indicates the function accepts a variable number of arguments, the
5935 extra arguments can be specified.
5936 #. The optional :ref:`function attributes <fnattrs>` list. Only
5937 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
5938 attributes are valid here.
5943 The '``call``' instruction is used to cause control flow to transfer to
5944 a specified function, with its incoming arguments bound to the specified
5945 values. Upon a '``ret``' instruction in the called function, control
5946 flow continues with the instruction after the function call, and the
5947 return value of the function is bound to the result argument.
5952 .. code-block:: llvm
5954 %retval = call i32 @test(i32 %argc)
5955 call i32 (i8*, ...)* @printf(i8* %msg, i32 12, i8 42) ; yields i32
5956 %X = tail call i32 @foo() ; yields i32
5957 %Y = tail call fastcc i32 @foo() ; yields i32
5958 call void %foo(i8 97 signext)
5960 %struct.A = type { i32, i8 }
5961 %r = call %struct.A @foo() ; yields { 32, i8 }
5962 %gr = extractvalue %struct.A %r, 0 ; yields i32
5963 %gr1 = extractvalue %struct.A %r, 1 ; yields i8
5964 %Z = call void @foo() noreturn ; indicates that %foo never returns normally
5965 %ZZ = call zeroext i32 @bar() ; Return value is %zero extended
5967 llvm treats calls to some functions with names and arguments that match
5968 the standard C99 library as being the C99 library functions, and may
5969 perform optimizations or generate code for them under that assumption.
5970 This is something we'd like to change in the future to provide better
5971 support for freestanding environments and non-C-based languages.
5975 '``va_arg``' Instruction
5976 ^^^^^^^^^^^^^^^^^^^^^^^^
5983 <resultval> = va_arg <va_list*> <arglist>, <argty>
5988 The '``va_arg``' instruction is used to access arguments passed through
5989 the "variable argument" area of a function call. It is used to implement
5990 the ``va_arg`` macro in C.
5995 This instruction takes a ``va_list*`` value and the type of the
5996 argument. It returns a value of the specified argument type and
5997 increments the ``va_list`` to point to the next argument. The actual
5998 type of ``va_list`` is target specific.
6003 The '``va_arg``' instruction loads an argument of the specified type
6004 from the specified ``va_list`` and causes the ``va_list`` to point to
6005 the next argument. For more information, see the variable argument
6006 handling :ref:`Intrinsic Functions <int_varargs>`.
6008 It is legal for this instruction to be called in a function which does
6009 not take a variable number of arguments, for example, the ``vfprintf``
6012 ``va_arg`` is an LLVM instruction instead of an :ref:`intrinsic
6013 function <intrinsics>` because it takes a type as an argument.
6018 See the :ref:`variable argument processing <int_varargs>` section.
6020 Note that the code generator does not yet fully support va\_arg on many
6021 targets. Also, it does not currently support va\_arg with aggregate
6022 types on any target.
6026 '``landingpad``' Instruction
6027 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6034 <resultval> = landingpad <resultty> personality <type> <pers_fn> <clause>+
6035 <resultval> = landingpad <resultty> personality <type> <pers_fn> cleanup <clause>*
6037 <clause> := catch <type> <value>
6038 <clause> := filter <array constant type> <array constant>
6043 The '``landingpad``' instruction is used by `LLVM's exception handling
6044 system <ExceptionHandling.html#overview>`_ to specify that a basic block
6045 is a landing pad --- one where the exception lands, and corresponds to the
6046 code found in the ``catch`` portion of a ``try``/``catch`` sequence. It
6047 defines values supplied by the personality function (``pers_fn``) upon
6048 re-entry to the function. The ``resultval`` has the type ``resultty``.
6053 This instruction takes a ``pers_fn`` value. This is the personality
6054 function associated with the unwinding mechanism. The optional
6055 ``cleanup`` flag indicates that the landing pad block is a cleanup.
6057 A ``clause`` begins with the clause type --- ``catch`` or ``filter`` --- and
6058 contains the global variable representing the "type" that may be caught
6059 or filtered respectively. Unlike the ``catch`` clause, the ``filter``
6060 clause takes an array constant as its argument. Use
6061 "``[0 x i8**] undef``" for a filter which cannot throw. The
6062 '``landingpad``' instruction must contain *at least* one ``clause`` or
6063 the ``cleanup`` flag.
6068 The '``landingpad``' instruction defines the values which are set by the
6069 personality function (``pers_fn``) upon re-entry to the function, and
6070 therefore the "result type" of the ``landingpad`` instruction. As with
6071 calling conventions, how the personality function results are
6072 represented in LLVM IR is target specific.
6074 The clauses are applied in order from top to bottom. If two
6075 ``landingpad`` instructions are merged together through inlining, the
6076 clauses from the calling function are appended to the list of clauses.
6077 When the call stack is being unwound due to an exception being thrown,
6078 the exception is compared against each ``clause`` in turn. If it doesn't
6079 match any of the clauses, and the ``cleanup`` flag is not set, then
6080 unwinding continues further up the call stack.
6082 The ``landingpad`` instruction has several restrictions:
6084 - A landing pad block is a basic block which is the unwind destination
6085 of an '``invoke``' instruction.
6086 - A landing pad block must have a '``landingpad``' instruction as its
6087 first non-PHI instruction.
6088 - There can be only one '``landingpad``' instruction within the landing
6090 - A basic block that is not a landing pad block may not include a
6091 '``landingpad``' instruction.
6092 - All '``landingpad``' instructions in a function must have the same
6093 personality function.
6098 .. code-block:: llvm
6100 ;; A landing pad which can catch an integer.
6101 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6103 ;; A landing pad that is a cleanup.
6104 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6106 ;; A landing pad which can catch an integer and can only throw a double.
6107 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6109 filter [1 x i8**] [@_ZTId]
6116 LLVM supports the notion of an "intrinsic function". These functions
6117 have well known names and semantics and are required to follow certain
6118 restrictions. Overall, these intrinsics represent an extension mechanism
6119 for the LLVM language that does not require changing all of the
6120 transformations in LLVM when adding to the language (or the bitcode
6121 reader/writer, the parser, etc...).
6123 Intrinsic function names must all start with an "``llvm.``" prefix. This
6124 prefix is reserved in LLVM for intrinsic names; thus, function names may
6125 not begin with this prefix. Intrinsic functions must always be external
6126 functions: you cannot define the body of intrinsic functions. Intrinsic
6127 functions may only be used in call or invoke instructions: it is illegal
6128 to take the address of an intrinsic function. Additionally, because
6129 intrinsic functions are part of the LLVM language, it is required if any
6130 are added that they be documented here.
6132 Some intrinsic functions can be overloaded, i.e., the intrinsic
6133 represents a family of functions that perform the same operation but on
6134 different data types. Because LLVM can represent over 8 million
6135 different integer types, overloading is used commonly to allow an
6136 intrinsic function to operate on any integer type. One or more of the
6137 argument types or the result type can be overloaded to accept any
6138 integer type. Argument types may also be defined as exactly matching a
6139 previous argument's type or the result type. This allows an intrinsic
6140 function which accepts multiple arguments, but needs all of them to be
6141 of the same type, to only be overloaded with respect to a single
6142 argument or the result.
6144 Overloaded intrinsics will have the names of its overloaded argument
6145 types encoded into its function name, each preceded by a period. Only
6146 those types which are overloaded result in a name suffix. Arguments
6147 whose type is matched against another type do not. For example, the
6148 ``llvm.ctpop`` function can take an integer of any width and returns an
6149 integer of exactly the same integer width. This leads to a family of
6150 functions such as ``i8 @llvm.ctpop.i8(i8 %val)`` and
6151 ``i29 @llvm.ctpop.i29(i29 %val)``. Only one type, the return type, is
6152 overloaded, and only one type suffix is required. Because the argument's
6153 type is matched against the return type, it does not require its own
6156 To learn how to add an intrinsic function, please see the `Extending
6157 LLVM Guide <ExtendingLLVM.html>`_.
6161 Variable Argument Handling Intrinsics
6162 -------------------------------------
6164 Variable argument support is defined in LLVM with the
6165 :ref:`va_arg <i_va_arg>` instruction and these three intrinsic
6166 functions. These functions are related to the similarly named macros
6167 defined in the ``<stdarg.h>`` header file.
6169 All of these functions operate on arguments that use a target-specific
6170 value type "``va_list``". The LLVM assembly language reference manual
6171 does not define what this type is, so all transformations should be
6172 prepared to handle these functions regardless of the type used.
6174 This example shows how the :ref:`va_arg <i_va_arg>` instruction and the
6175 variable argument handling intrinsic functions are used.
6177 .. code-block:: llvm
6179 define i32 @test(i32 %X, ...) {
6180 ; Initialize variable argument processing
6182 %ap2 = bitcast i8** %ap to i8*
6183 call void @llvm.va_start(i8* %ap2)
6185 ; Read a single integer argument
6186 %tmp = va_arg i8** %ap, i32
6188 ; Demonstrate usage of llvm.va_copy and llvm.va_end
6190 %aq2 = bitcast i8** %aq to i8*
6191 call void @llvm.va_copy(i8* %aq2, i8* %ap2)
6192 call void @llvm.va_end(i8* %aq2)
6194 ; Stop processing of arguments.
6195 call void @llvm.va_end(i8* %ap2)
6199 declare void @llvm.va_start(i8*)
6200 declare void @llvm.va_copy(i8*, i8*)
6201 declare void @llvm.va_end(i8*)
6205 '``llvm.va_start``' Intrinsic
6206 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6213 declare void %llvm.va_start(i8* <arglist>)
6218 The '``llvm.va_start``' intrinsic initializes ``*<arglist>`` for
6219 subsequent use by ``va_arg``.
6224 The argument is a pointer to a ``va_list`` element to initialize.
6229 The '``llvm.va_start``' intrinsic works just like the ``va_start`` macro
6230 available in C. In a target-dependent way, it initializes the
6231 ``va_list`` element to which the argument points, so that the next call
6232 to ``va_arg`` will produce the first variable argument passed to the
6233 function. Unlike the C ``va_start`` macro, this intrinsic does not need
6234 to know the last argument of the function as the compiler can figure
6237 '``llvm.va_end``' Intrinsic
6238 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6245 declare void @llvm.va_end(i8* <arglist>)
6250 The '``llvm.va_end``' intrinsic destroys ``*<arglist>``, which has been
6251 initialized previously with ``llvm.va_start`` or ``llvm.va_copy``.
6256 The argument is a pointer to a ``va_list`` to destroy.
6261 The '``llvm.va_end``' intrinsic works just like the ``va_end`` macro
6262 available in C. In a target-dependent way, it destroys the ``va_list``
6263 element to which the argument points. Calls to
6264 :ref:`llvm.va_start <int_va_start>` and
6265 :ref:`llvm.va_copy <int_va_copy>` must be matched exactly with calls to
6270 '``llvm.va_copy``' Intrinsic
6271 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6278 declare void @llvm.va_copy(i8* <destarglist>, i8* <srcarglist>)
6283 The '``llvm.va_copy``' intrinsic copies the current argument position
6284 from the source argument list to the destination argument list.
6289 The first argument is a pointer to a ``va_list`` element to initialize.
6290 The second argument is a pointer to a ``va_list`` element to copy from.
6295 The '``llvm.va_copy``' intrinsic works just like the ``va_copy`` macro
6296 available in C. In a target-dependent way, it copies the source
6297 ``va_list`` element into the destination ``va_list`` element. This
6298 intrinsic is necessary because the `` llvm.va_start`` intrinsic may be
6299 arbitrarily complex and require, for example, memory allocation.
6301 Accurate Garbage Collection Intrinsics
6302 --------------------------------------
6304 LLVM support for `Accurate Garbage Collection <GarbageCollection.html>`_
6305 (GC) requires the implementation and generation of these intrinsics.
6306 These intrinsics allow identification of :ref:`GC roots on the
6307 stack <int_gcroot>`, as well as garbage collector implementations that
6308 require :ref:`read <int_gcread>` and :ref:`write <int_gcwrite>` barriers.
6309 Front-ends for type-safe garbage collected languages should generate
6310 these intrinsics to make use of the LLVM garbage collectors. For more
6311 details, see `Accurate Garbage Collection with
6312 LLVM <GarbageCollection.html>`_.
6314 The garbage collection intrinsics only operate on objects in the generic
6315 address space (address space zero).
6319 '``llvm.gcroot``' Intrinsic
6320 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6327 declare void @llvm.gcroot(i8** %ptrloc, i8* %metadata)
6332 The '``llvm.gcroot``' intrinsic declares the existence of a GC root to
6333 the code generator, and allows some metadata to be associated with it.
6338 The first argument specifies the address of a stack object that contains
6339 the root pointer. The second pointer (which must be either a constant or
6340 a global value address) contains the meta-data to be associated with the
6346 At runtime, a call to this intrinsic stores a null pointer into the
6347 "ptrloc" location. At compile-time, the code generator generates
6348 information to allow the runtime to find the pointer at GC safe points.
6349 The '``llvm.gcroot``' intrinsic may only be used in a function which
6350 :ref:`specifies a GC algorithm <gc>`.
6354 '``llvm.gcread``' Intrinsic
6355 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6362 declare i8* @llvm.gcread(i8* %ObjPtr, i8** %Ptr)
6367 The '``llvm.gcread``' intrinsic identifies reads of references from heap
6368 locations, allowing garbage collector implementations that require read
6374 The second argument is the address to read from, which should be an
6375 address allocated from the garbage collector. The first object is a
6376 pointer to the start of the referenced object, if needed by the language
6377 runtime (otherwise null).
6382 The '``llvm.gcread``' intrinsic has the same semantics as a load
6383 instruction, but may be replaced with substantially more complex code by
6384 the garbage collector runtime, as needed. The '``llvm.gcread``'
6385 intrinsic may only be used in a function which :ref:`specifies a GC
6390 '``llvm.gcwrite``' Intrinsic
6391 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6398 declare void @llvm.gcwrite(i8* %P1, i8* %Obj, i8** %P2)
6403 The '``llvm.gcwrite``' intrinsic identifies writes of references to heap
6404 locations, allowing garbage collector implementations that require write
6405 barriers (such as generational or reference counting collectors).
6410 The first argument is the reference to store, the second is the start of
6411 the object to store it to, and the third is the address of the field of
6412 Obj to store to. If the runtime does not require a pointer to the
6413 object, Obj may be null.
6418 The '``llvm.gcwrite``' intrinsic has the same semantics as a store
6419 instruction, but may be replaced with substantially more complex code by
6420 the garbage collector runtime, as needed. The '``llvm.gcwrite``'
6421 intrinsic may only be used in a function which :ref:`specifies a GC
6424 Code Generator Intrinsics
6425 -------------------------
6427 These intrinsics are provided by LLVM to expose special features that
6428 may only be implemented with code generator support.
6430 '``llvm.returnaddress``' Intrinsic
6431 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6438 declare i8 *@llvm.returnaddress(i32 <level>)
6443 The '``llvm.returnaddress``' intrinsic attempts to compute a
6444 target-specific value indicating the return address of the current
6445 function or one of its callers.
6450 The argument to this intrinsic indicates which function to return the
6451 address for. Zero indicates the calling function, one indicates its
6452 caller, etc. The argument is **required** to be a constant integer
6458 The '``llvm.returnaddress``' intrinsic either returns a pointer
6459 indicating the return address of the specified call frame, or zero if it
6460 cannot be identified. The value returned by this intrinsic is likely to
6461 be incorrect or 0 for arguments other than zero, so it should only be
6462 used for debugging purposes.
6464 Note that calling this intrinsic does not prevent function inlining or
6465 other aggressive transformations, so the value returned may not be that
6466 of the obvious source-language caller.
6468 '``llvm.frameaddress``' Intrinsic
6469 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6476 declare i8* @llvm.frameaddress(i32 <level>)
6481 The '``llvm.frameaddress``' intrinsic attempts to return the
6482 target-specific frame pointer value for the specified stack frame.
6487 The argument to this intrinsic indicates which function to return the
6488 frame pointer for. Zero indicates the calling function, one indicates
6489 its caller, etc. The argument is **required** to be a constant integer
6495 The '``llvm.frameaddress``' intrinsic either returns a pointer
6496 indicating the frame address of the specified call frame, or zero if it
6497 cannot be identified. The value returned by this intrinsic is likely to
6498 be incorrect or 0 for arguments other than zero, so it should only be
6499 used for debugging purposes.
6501 Note that calling this intrinsic does not prevent function inlining or
6502 other aggressive transformations, so the value returned may not be that
6503 of the obvious source-language caller.
6507 '``llvm.stacksave``' Intrinsic
6508 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6515 declare i8* @llvm.stacksave()
6520 The '``llvm.stacksave``' intrinsic is used to remember the current state
6521 of the function stack, for use with
6522 :ref:`llvm.stackrestore <int_stackrestore>`. This is useful for
6523 implementing language features like scoped automatic variable sized
6529 This intrinsic returns a opaque pointer value that can be passed to
6530 :ref:`llvm.stackrestore <int_stackrestore>`. When an
6531 ``llvm.stackrestore`` intrinsic is executed with a value saved from
6532 ``llvm.stacksave``, it effectively restores the state of the stack to
6533 the state it was in when the ``llvm.stacksave`` intrinsic executed. In
6534 practice, this pops any :ref:`alloca <i_alloca>` blocks from the stack that
6535 were allocated after the ``llvm.stacksave`` was executed.
6537 .. _int_stackrestore:
6539 '``llvm.stackrestore``' Intrinsic
6540 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6547 declare void @llvm.stackrestore(i8* %ptr)
6552 The '``llvm.stackrestore``' intrinsic is used to restore the state of
6553 the function stack to the state it was in when the corresponding
6554 :ref:`llvm.stacksave <int_stacksave>` intrinsic executed. This is
6555 useful for implementing language features like scoped automatic variable
6556 sized arrays in C99.
6561 See the description for :ref:`llvm.stacksave <int_stacksave>`.
6563 '``llvm.prefetch``' Intrinsic
6564 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6571 declare void @llvm.prefetch(i8* <address>, i32 <rw>, i32 <locality>, i32 <cache type>)
6576 The '``llvm.prefetch``' intrinsic is a hint to the code generator to
6577 insert a prefetch instruction if supported; otherwise, it is a noop.
6578 Prefetches have no effect on the behavior of the program but can change
6579 its performance characteristics.
6584 ``address`` is the address to be prefetched, ``rw`` is the specifier
6585 determining if the fetch should be for a read (0) or write (1), and
6586 ``locality`` is a temporal locality specifier ranging from (0) - no
6587 locality, to (3) - extremely local keep in cache. The ``cache type``
6588 specifies whether the prefetch is performed on the data (1) or
6589 instruction (0) cache. The ``rw``, ``locality`` and ``cache type``
6590 arguments must be constant integers.
6595 This intrinsic does not modify the behavior of the program. In
6596 particular, prefetches cannot trap and do not produce a value. On
6597 targets that support this intrinsic, the prefetch can provide hints to
6598 the processor cache for better performance.
6600 '``llvm.pcmarker``' Intrinsic
6601 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6608 declare void @llvm.pcmarker(i32 <id>)
6613 The '``llvm.pcmarker``' intrinsic is a method to export a Program
6614 Counter (PC) in a region of code to simulators and other tools. The
6615 method is target specific, but it is expected that the marker will use
6616 exported symbols to transmit the PC of the marker. The marker makes no
6617 guarantees that it will remain with any specific instruction after
6618 optimizations. It is possible that the presence of a marker will inhibit
6619 optimizations. The intended use is to be inserted after optimizations to
6620 allow correlations of simulation runs.
6625 ``id`` is a numerical id identifying the marker.
6630 This intrinsic does not modify the behavior of the program. Backends
6631 that do not support this intrinsic may ignore it.
6633 '``llvm.readcyclecounter``' Intrinsic
6634 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6641 declare i64 @llvm.readcyclecounter()
6646 The '``llvm.readcyclecounter``' intrinsic provides access to the cycle
6647 counter register (or similar low latency, high accuracy clocks) on those
6648 targets that support it. On X86, it should map to RDTSC. On Alpha, it
6649 should map to RPCC. As the backing counters overflow quickly (on the
6650 order of 9 seconds on alpha), this should only be used for small
6656 When directly supported, reading the cycle counter should not modify any
6657 memory. Implementations are allowed to either return a application
6658 specific value or a system wide value. On backends without support, this
6659 is lowered to a constant 0.
6661 Standard C Library Intrinsics
6662 -----------------------------
6664 LLVM provides intrinsics for a few important standard C library
6665 functions. These intrinsics allow source-language front-ends to pass
6666 information about the alignment of the pointer arguments to the code
6667 generator, providing opportunity for more efficient code generation.
6671 '``llvm.memcpy``' Intrinsic
6672 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6677 This is an overloaded intrinsic. You can use ``llvm.memcpy`` on any
6678 integer bit width and for different address spaces. Not all targets
6679 support all bit widths however.
6683 declare void @llvm.memcpy.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
6684 i32 <len>, i32 <align>, i1 <isvolatile>)
6685 declare void @llvm.memcpy.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
6686 i64 <len>, i32 <align>, i1 <isvolatile>)
6691 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
6692 source location to the destination location.
6694 Note that, unlike the standard libc function, the ``llvm.memcpy.*``
6695 intrinsics do not return a value, takes extra alignment/isvolatile
6696 arguments and the pointers can be in specified address spaces.
6701 The first argument is a pointer to the destination, the second is a
6702 pointer to the source. The third argument is an integer argument
6703 specifying the number of bytes to copy, the fourth argument is the
6704 alignment of the source and destination locations, and the fifth is a
6705 boolean indicating a volatile access.
6707 If the call to this intrinsic has an alignment value that is not 0 or 1,
6708 then the caller guarantees that both the source and destination pointers
6709 are aligned to that boundary.
6711 If the ``isvolatile`` parameter is ``true``, the ``llvm.memcpy`` call is
6712 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
6713 very cleanly specified and it is unwise to depend on it.
6718 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
6719 source location to the destination location, which are not allowed to
6720 overlap. It copies "len" bytes of memory over. If the argument is known
6721 to be aligned to some boundary, this can be specified as the fourth
6722 argument, otherwise it should be set to 0 or 1.
6724 '``llvm.memmove``' Intrinsic
6725 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6730 This is an overloaded intrinsic. You can use llvm.memmove on any integer
6731 bit width and for different address space. Not all targets support all
6736 declare void @llvm.memmove.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
6737 i32 <len>, i32 <align>, i1 <isvolatile>)
6738 declare void @llvm.memmove.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
6739 i64 <len>, i32 <align>, i1 <isvolatile>)
6744 The '``llvm.memmove.*``' intrinsics move a block of memory from the
6745 source location to the destination location. It is similar to the
6746 '``llvm.memcpy``' intrinsic but allows the two memory locations to
6749 Note that, unlike the standard libc function, the ``llvm.memmove.*``
6750 intrinsics do not return a value, takes extra alignment/isvolatile
6751 arguments and the pointers can be in specified address spaces.
6756 The first argument is a pointer to the destination, the second is a
6757 pointer to the source. The third argument is an integer argument
6758 specifying the number of bytes to copy, the fourth argument is the
6759 alignment of the source and destination locations, and the fifth is a
6760 boolean indicating a volatile access.
6762 If the call to this intrinsic has an alignment value that is not 0 or 1,
6763 then the caller guarantees that the source and destination pointers are
6764 aligned to that boundary.
6766 If the ``isvolatile`` parameter is ``true``, the ``llvm.memmove`` call
6767 is a :ref:`volatile operation <volatile>`. The detailed access behavior is
6768 not very cleanly specified and it is unwise to depend on it.
6773 The '``llvm.memmove.*``' intrinsics copy a block of memory from the
6774 source location to the destination location, which may overlap. It
6775 copies "len" bytes of memory over. If the argument is known to be
6776 aligned to some boundary, this can be specified as the fourth argument,
6777 otherwise it should be set to 0 or 1.
6779 '``llvm.memset.*``' Intrinsics
6780 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6785 This is an overloaded intrinsic. You can use llvm.memset on any integer
6786 bit width and for different address spaces. However, not all targets
6787 support all bit widths.
6791 declare void @llvm.memset.p0i8.i32(i8* <dest>, i8 <val>,
6792 i32 <len>, i32 <align>, i1 <isvolatile>)
6793 declare void @llvm.memset.p0i8.i64(i8* <dest>, i8 <val>,
6794 i64 <len>, i32 <align>, i1 <isvolatile>)
6799 The '``llvm.memset.*``' intrinsics fill a block of memory with a
6800 particular byte value.
6802 Note that, unlike the standard libc function, the ``llvm.memset``
6803 intrinsic does not return a value and takes extra alignment/volatile
6804 arguments. Also, the destination can be in an arbitrary address space.
6809 The first argument is a pointer to the destination to fill, the second
6810 is the byte value with which to fill it, the third argument is an
6811 integer argument specifying the number of bytes to fill, and the fourth
6812 argument is the known alignment of the destination location.
6814 If the call to this intrinsic has an alignment value that is not 0 or 1,
6815 then the caller guarantees that the destination pointer is aligned to
6818 If the ``isvolatile`` parameter is ``true``, the ``llvm.memset`` call is
6819 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
6820 very cleanly specified and it is unwise to depend on it.
6825 The '``llvm.memset.*``' intrinsics fill "len" bytes of memory starting
6826 at the destination location. If the argument is known to be aligned to
6827 some boundary, this can be specified as the fourth argument, otherwise
6828 it should be set to 0 or 1.
6830 '``llvm.sqrt.*``' Intrinsic
6831 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6836 This is an overloaded intrinsic. You can use ``llvm.sqrt`` on any
6837 floating point or vector of floating point type. Not all targets support
6842 declare float @llvm.sqrt.f32(float %Val)
6843 declare double @llvm.sqrt.f64(double %Val)
6844 declare x86_fp80 @llvm.sqrt.f80(x86_fp80 %Val)
6845 declare fp128 @llvm.sqrt.f128(fp128 %Val)
6846 declare ppc_fp128 @llvm.sqrt.ppcf128(ppc_fp128 %Val)
6851 The '``llvm.sqrt``' intrinsics return the sqrt of the specified operand,
6852 returning the same value as the libm '``sqrt``' functions would. Unlike
6853 ``sqrt`` in libm, however, ``llvm.sqrt`` has undefined behavior for
6854 negative numbers other than -0.0 (which allows for better optimization,
6855 because there is no need to worry about errno being set).
6856 ``llvm.sqrt(-0.0)`` is defined to return -0.0 like IEEE sqrt.
6861 The argument and return value are floating point numbers of the same
6867 This function returns the sqrt of the specified operand if it is a
6868 nonnegative floating point number.
6870 '``llvm.powi.*``' Intrinsic
6871 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6876 This is an overloaded intrinsic. You can use ``llvm.powi`` on any
6877 floating point or vector of floating point type. Not all targets support
6882 declare float @llvm.powi.f32(float %Val, i32 %power)
6883 declare double @llvm.powi.f64(double %Val, i32 %power)
6884 declare x86_fp80 @llvm.powi.f80(x86_fp80 %Val, i32 %power)
6885 declare fp128 @llvm.powi.f128(fp128 %Val, i32 %power)
6886 declare ppc_fp128 @llvm.powi.ppcf128(ppc_fp128 %Val, i32 %power)
6891 The '``llvm.powi.*``' intrinsics return the first operand raised to the
6892 specified (positive or negative) power. The order of evaluation of
6893 multiplications is not defined. When a vector of floating point type is
6894 used, the second argument remains a scalar integer value.
6899 The second argument is an integer power, and the first is a value to
6900 raise to that power.
6905 This function returns the first value raised to the second power with an
6906 unspecified sequence of rounding operations.
6908 '``llvm.sin.*``' Intrinsic
6909 ^^^^^^^^^^^^^^^^^^^^^^^^^^
6914 This is an overloaded intrinsic. You can use ``llvm.sin`` on any
6915 floating point or vector of floating point type. Not all targets support
6920 declare float @llvm.sin.f32(float %Val)
6921 declare double @llvm.sin.f64(double %Val)
6922 declare x86_fp80 @llvm.sin.f80(x86_fp80 %Val)
6923 declare fp128 @llvm.sin.f128(fp128 %Val)
6924 declare ppc_fp128 @llvm.sin.ppcf128(ppc_fp128 %Val)
6929 The '``llvm.sin.*``' intrinsics return the sine of the operand.
6934 The argument and return value are floating point numbers of the same
6940 This function returns the sine of the specified operand, returning the
6941 same values as the libm ``sin`` functions would, and handles error
6942 conditions in the same way.
6944 '``llvm.cos.*``' Intrinsic
6945 ^^^^^^^^^^^^^^^^^^^^^^^^^^
6950 This is an overloaded intrinsic. You can use ``llvm.cos`` on any
6951 floating point or vector of floating point type. Not all targets support
6956 declare float @llvm.cos.f32(float %Val)
6957 declare double @llvm.cos.f64(double %Val)
6958 declare x86_fp80 @llvm.cos.f80(x86_fp80 %Val)
6959 declare fp128 @llvm.cos.f128(fp128 %Val)
6960 declare ppc_fp128 @llvm.cos.ppcf128(ppc_fp128 %Val)
6965 The '``llvm.cos.*``' intrinsics return the cosine of the operand.
6970 The argument and return value are floating point numbers of the same
6976 This function returns the cosine of the specified operand, returning the
6977 same values as the libm ``cos`` functions would, and handles error
6978 conditions in the same way.
6980 '``llvm.pow.*``' Intrinsic
6981 ^^^^^^^^^^^^^^^^^^^^^^^^^^
6986 This is an overloaded intrinsic. You can use ``llvm.pow`` on any
6987 floating point or vector of floating point type. Not all targets support
6992 declare float @llvm.pow.f32(float %Val, float %Power)
6993 declare double @llvm.pow.f64(double %Val, double %Power)
6994 declare x86_fp80 @llvm.pow.f80(x86_fp80 %Val, x86_fp80 %Power)
6995 declare fp128 @llvm.pow.f128(fp128 %Val, fp128 %Power)
6996 declare ppc_fp128 @llvm.pow.ppcf128(ppc_fp128 %Val, ppc_fp128 Power)
7001 The '``llvm.pow.*``' intrinsics return the first operand raised to the
7002 specified (positive or negative) power.
7007 The second argument is a floating point power, and the first is a value
7008 to raise to that power.
7013 This function returns the first value raised to the second power,
7014 returning the same values as the libm ``pow`` functions would, and
7015 handles error conditions in the same way.
7017 '``llvm.exp.*``' Intrinsic
7018 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7023 This is an overloaded intrinsic. You can use ``llvm.exp`` on any
7024 floating point or vector of floating point type. Not all targets support
7029 declare float @llvm.exp.f32(float %Val)
7030 declare double @llvm.exp.f64(double %Val)
7031 declare x86_fp80 @llvm.exp.f80(x86_fp80 %Val)
7032 declare fp128 @llvm.exp.f128(fp128 %Val)
7033 declare ppc_fp128 @llvm.exp.ppcf128(ppc_fp128 %Val)
7038 The '``llvm.exp.*``' intrinsics perform the exp function.
7043 The argument and return value are floating point numbers of the same
7049 This function returns the same values as the libm ``exp`` functions
7050 would, and handles error conditions in the same way.
7052 '``llvm.exp2.*``' Intrinsic
7053 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7058 This is an overloaded intrinsic. You can use ``llvm.exp2`` on any
7059 floating point or vector of floating point type. Not all targets support
7064 declare float @llvm.exp2.f32(float %Val)
7065 declare double @llvm.exp2.f64(double %Val)
7066 declare x86_fp80 @llvm.exp2.f80(x86_fp80 %Val)
7067 declare fp128 @llvm.exp2.f128(fp128 %Val)
7068 declare ppc_fp128 @llvm.exp2.ppcf128(ppc_fp128 %Val)
7073 The '``llvm.exp2.*``' intrinsics perform the exp2 function.
7078 The argument and return value are floating point numbers of the same
7084 This function returns the same values as the libm ``exp2`` functions
7085 would, and handles error conditions in the same way.
7087 '``llvm.log.*``' Intrinsic
7088 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7093 This is an overloaded intrinsic. You can use ``llvm.log`` on any
7094 floating point or vector of floating point type. Not all targets support
7099 declare float @llvm.log.f32(float %Val)
7100 declare double @llvm.log.f64(double %Val)
7101 declare x86_fp80 @llvm.log.f80(x86_fp80 %Val)
7102 declare fp128 @llvm.log.f128(fp128 %Val)
7103 declare ppc_fp128 @llvm.log.ppcf128(ppc_fp128 %Val)
7108 The '``llvm.log.*``' intrinsics perform the log function.
7113 The argument and return value are floating point numbers of the same
7119 This function returns the same values as the libm ``log`` functions
7120 would, and handles error conditions in the same way.
7122 '``llvm.log10.*``' Intrinsic
7123 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7128 This is an overloaded intrinsic. You can use ``llvm.log10`` on any
7129 floating point or vector of floating point type. Not all targets support
7134 declare float @llvm.log10.f32(float %Val)
7135 declare double @llvm.log10.f64(double %Val)
7136 declare x86_fp80 @llvm.log10.f80(x86_fp80 %Val)
7137 declare fp128 @llvm.log10.f128(fp128 %Val)
7138 declare ppc_fp128 @llvm.log10.ppcf128(ppc_fp128 %Val)
7143 The '``llvm.log10.*``' intrinsics perform the log10 function.
7148 The argument and return value are floating point numbers of the same
7154 This function returns the same values as the libm ``log10`` functions
7155 would, and handles error conditions in the same way.
7157 '``llvm.log2.*``' Intrinsic
7158 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7163 This is an overloaded intrinsic. You can use ``llvm.log2`` on any
7164 floating point or vector of floating point type. Not all targets support
7169 declare float @llvm.log2.f32(float %Val)
7170 declare double @llvm.log2.f64(double %Val)
7171 declare x86_fp80 @llvm.log2.f80(x86_fp80 %Val)
7172 declare fp128 @llvm.log2.f128(fp128 %Val)
7173 declare ppc_fp128 @llvm.log2.ppcf128(ppc_fp128 %Val)
7178 The '``llvm.log2.*``' intrinsics perform the log2 function.
7183 The argument and return value are floating point numbers of the same
7189 This function returns the same values as the libm ``log2`` functions
7190 would, and handles error conditions in the same way.
7192 '``llvm.fma.*``' Intrinsic
7193 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7198 This is an overloaded intrinsic. You can use ``llvm.fma`` on any
7199 floating point or vector of floating point type. Not all targets support
7204 declare float @llvm.fma.f32(float %a, float %b, float %c)
7205 declare double @llvm.fma.f64(double %a, double %b, double %c)
7206 declare x86_fp80 @llvm.fma.f80(x86_fp80 %a, x86_fp80 %b, x86_fp80 %c)
7207 declare fp128 @llvm.fma.f128(fp128 %a, fp128 %b, fp128 %c)
7208 declare ppc_fp128 @llvm.fma.ppcf128(ppc_fp128 %a, ppc_fp128 %b, ppc_fp128 %c)
7213 The '``llvm.fma.*``' intrinsics perform the fused multiply-add
7219 The argument and return value are floating point numbers of the same
7225 This function returns the same values as the libm ``fma`` functions
7228 '``llvm.fabs.*``' Intrinsic
7229 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7234 This is an overloaded intrinsic. You can use ``llvm.fabs`` on any
7235 floating point or vector of floating point type. Not all targets support
7240 declare float @llvm.fabs.f32(float %Val)
7241 declare double @llvm.fabs.f64(double %Val)
7242 declare x86_fp80 @llvm.fabs.f80(x86_fp80 %Val)
7243 declare fp128 @llvm.fabs.f128(fp128 %Val)
7244 declare ppc_fp128 @llvm.fabs.ppcf128(ppc_fp128 %Val)
7249 The '``llvm.fabs.*``' intrinsics return the absolute value of the
7255 The argument and return value are floating point numbers of the same
7261 This function returns the same values as the libm ``fabs`` functions
7262 would, and handles error conditions in the same way.
7264 '``llvm.floor.*``' Intrinsic
7265 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7270 This is an overloaded intrinsic. You can use ``llvm.floor`` on any
7271 floating point or vector of floating point type. Not all targets support
7276 declare float @llvm.floor.f32(float %Val)
7277 declare double @llvm.floor.f64(double %Val)
7278 declare x86_fp80 @llvm.floor.f80(x86_fp80 %Val)
7279 declare fp128 @llvm.floor.f128(fp128 %Val)
7280 declare ppc_fp128 @llvm.floor.ppcf128(ppc_fp128 %Val)
7285 The '``llvm.floor.*``' intrinsics return the floor of the operand.
7290 The argument and return value are floating point numbers of the same
7296 This function returns the same values as the libm ``floor`` functions
7297 would, and handles error conditions in the same way.
7299 '``llvm.ceil.*``' Intrinsic
7300 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7305 This is an overloaded intrinsic. You can use ``llvm.ceil`` on any
7306 floating point or vector of floating point type. Not all targets support
7311 declare float @llvm.ceil.f32(float %Val)
7312 declare double @llvm.ceil.f64(double %Val)
7313 declare x86_fp80 @llvm.ceil.f80(x86_fp80 %Val)
7314 declare fp128 @llvm.ceil.f128(fp128 %Val)
7315 declare ppc_fp128 @llvm.ceil.ppcf128(ppc_fp128 %Val)
7320 The '``llvm.ceil.*``' intrinsics return the ceiling of the operand.
7325 The argument and return value are floating point numbers of the same
7331 This function returns the same values as the libm ``ceil`` functions
7332 would, and handles error conditions in the same way.
7334 '``llvm.trunc.*``' Intrinsic
7335 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7340 This is an overloaded intrinsic. You can use ``llvm.trunc`` on any
7341 floating point or vector of floating point type. Not all targets support
7346 declare float @llvm.trunc.f32(float %Val)
7347 declare double @llvm.trunc.f64(double %Val)
7348 declare x86_fp80 @llvm.trunc.f80(x86_fp80 %Val)
7349 declare fp128 @llvm.trunc.f128(fp128 %Val)
7350 declare ppc_fp128 @llvm.trunc.ppcf128(ppc_fp128 %Val)
7355 The '``llvm.trunc.*``' intrinsics returns the operand rounded to the
7356 nearest integer not larger in magnitude than the operand.
7361 The argument and return value are floating point numbers of the same
7367 This function returns the same values as the libm ``trunc`` functions
7368 would, and handles error conditions in the same way.
7370 '``llvm.rint.*``' Intrinsic
7371 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7376 This is an overloaded intrinsic. You can use ``llvm.rint`` on any
7377 floating point or vector of floating point type. Not all targets support
7382 declare float @llvm.rint.f32(float %Val)
7383 declare double @llvm.rint.f64(double %Val)
7384 declare x86_fp80 @llvm.rint.f80(x86_fp80 %Val)
7385 declare fp128 @llvm.rint.f128(fp128 %Val)
7386 declare ppc_fp128 @llvm.rint.ppcf128(ppc_fp128 %Val)
7391 The '``llvm.rint.*``' intrinsics returns the operand rounded to the
7392 nearest integer. It may raise an inexact floating-point exception if the
7393 operand isn't an integer.
7398 The argument and return value are floating point numbers of the same
7404 This function returns the same values as the libm ``rint`` functions
7405 would, and handles error conditions in the same way.
7407 '``llvm.nearbyint.*``' Intrinsic
7408 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7413 This is an overloaded intrinsic. You can use ``llvm.nearbyint`` on any
7414 floating point or vector of floating point type. Not all targets support
7419 declare float @llvm.nearbyint.f32(float %Val)
7420 declare double @llvm.nearbyint.f64(double %Val)
7421 declare x86_fp80 @llvm.nearbyint.f80(x86_fp80 %Val)
7422 declare fp128 @llvm.nearbyint.f128(fp128 %Val)
7423 declare ppc_fp128 @llvm.nearbyint.ppcf128(ppc_fp128 %Val)
7428 The '``llvm.nearbyint.*``' intrinsics returns the operand rounded to the
7434 The argument and return value are floating point numbers of the same
7440 This function returns the same values as the libm ``nearbyint``
7441 functions would, and handles error conditions in the same way.
7443 Bit Manipulation Intrinsics
7444 ---------------------------
7446 LLVM provides intrinsics for a few important bit manipulation
7447 operations. These allow efficient code generation for some algorithms.
7449 '``llvm.bswap.*``' Intrinsics
7450 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7455 This is an overloaded intrinsic function. You can use bswap on any
7456 integer type that is an even number of bytes (i.e. BitWidth % 16 == 0).
7460 declare i16 @llvm.bswap.i16(i16 <id>)
7461 declare i32 @llvm.bswap.i32(i32 <id>)
7462 declare i64 @llvm.bswap.i64(i64 <id>)
7467 The '``llvm.bswap``' family of intrinsics is used to byte swap integer
7468 values with an even number of bytes (positive multiple of 16 bits).
7469 These are useful for performing operations on data that is not in the
7470 target's native byte order.
7475 The ``llvm.bswap.i16`` intrinsic returns an i16 value that has the high
7476 and low byte of the input i16 swapped. Similarly, the ``llvm.bswap.i32``
7477 intrinsic returns an i32 value that has the four bytes of the input i32
7478 swapped, so that if the input bytes are numbered 0, 1, 2, 3 then the
7479 returned i32 will have its bytes in 3, 2, 1, 0 order. The
7480 ``llvm.bswap.i48``, ``llvm.bswap.i64`` and other intrinsics extend this
7481 concept to additional even-byte lengths (6 bytes, 8 bytes and more,
7484 '``llvm.ctpop.*``' Intrinsic
7485 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7490 This is an overloaded intrinsic. You can use llvm.ctpop on any integer
7491 bit width, or on any vector with integer elements. Not all targets
7492 support all bit widths or vector types, however.
7496 declare i8 @llvm.ctpop.i8(i8 <src>)
7497 declare i16 @llvm.ctpop.i16(i16 <src>)
7498 declare i32 @llvm.ctpop.i32(i32 <src>)
7499 declare i64 @llvm.ctpop.i64(i64 <src>)
7500 declare i256 @llvm.ctpop.i256(i256 <src>)
7501 declare <2 x i32> @llvm.ctpop.v2i32(<2 x i32> <src>)
7506 The '``llvm.ctpop``' family of intrinsics counts the number of bits set
7512 The only argument is the value to be counted. The argument may be of any
7513 integer type, or a vector with integer elements. The return type must
7514 match the argument type.
7519 The '``llvm.ctpop``' intrinsic counts the 1's in a variable, or within
7520 each element of a vector.
7522 '``llvm.ctlz.*``' Intrinsic
7523 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7528 This is an overloaded intrinsic. You can use ``llvm.ctlz`` on any
7529 integer bit width, or any vector whose elements are integers. Not all
7530 targets support all bit widths or vector types, however.
7534 declare i8 @llvm.ctlz.i8 (i8 <src>, i1 <is_zero_undef>)
7535 declare i16 @llvm.ctlz.i16 (i16 <src>, i1 <is_zero_undef>)
7536 declare i32 @llvm.ctlz.i32 (i32 <src>, i1 <is_zero_undef>)
7537 declare i64 @llvm.ctlz.i64 (i64 <src>, i1 <is_zero_undef>)
7538 declare i256 @llvm.ctlz.i256(i256 <src>, i1 <is_zero_undef>)
7539 declase <2 x i32> @llvm.ctlz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
7544 The '``llvm.ctlz``' family of intrinsic functions counts the number of
7545 leading zeros in a variable.
7550 The first argument is the value to be counted. This argument may be of
7551 any integer type, or a vectory with integer element type. The return
7552 type must match the first argument type.
7554 The second argument must be a constant and is a flag to indicate whether
7555 the intrinsic should ensure that a zero as the first argument produces a
7556 defined result. Historically some architectures did not provide a
7557 defined result for zero values as efficiently, and many algorithms are
7558 now predicated on avoiding zero-value inputs.
7563 The '``llvm.ctlz``' intrinsic counts the leading (most significant)
7564 zeros in a variable, or within each element of the vector. If
7565 ``src == 0`` then the result is the size in bits of the type of ``src``
7566 if ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
7567 ``llvm.ctlz(i32 2) = 30``.
7569 '``llvm.cttz.*``' Intrinsic
7570 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7575 This is an overloaded intrinsic. You can use ``llvm.cttz`` on any
7576 integer bit width, or any vector of integer elements. Not all targets
7577 support all bit widths or vector types, however.
7581 declare i8 @llvm.cttz.i8 (i8 <src>, i1 <is_zero_undef>)
7582 declare i16 @llvm.cttz.i16 (i16 <src>, i1 <is_zero_undef>)
7583 declare i32 @llvm.cttz.i32 (i32 <src>, i1 <is_zero_undef>)
7584 declare i64 @llvm.cttz.i64 (i64 <src>, i1 <is_zero_undef>)
7585 declare i256 @llvm.cttz.i256(i256 <src>, i1 <is_zero_undef>)
7586 declase <2 x i32> @llvm.cttz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
7591 The '``llvm.cttz``' family of intrinsic functions counts the number of
7597 The first argument is the value to be counted. This argument may be of
7598 any integer type, or a vectory with integer element type. The return
7599 type must match the first argument type.
7601 The second argument must be a constant and is a flag to indicate whether
7602 the intrinsic should ensure that a zero as the first argument produces a
7603 defined result. Historically some architectures did not provide a
7604 defined result for zero values as efficiently, and many algorithms are
7605 now predicated on avoiding zero-value inputs.
7610 The '``llvm.cttz``' intrinsic counts the trailing (least significant)
7611 zeros in a variable, or within each element of a vector. If ``src == 0``
7612 then the result is the size in bits of the type of ``src`` if
7613 ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
7614 ``llvm.cttz(2) = 1``.
7616 Arithmetic with Overflow Intrinsics
7617 -----------------------------------
7619 LLVM provides intrinsics for some arithmetic with overflow operations.
7621 '``llvm.sadd.with.overflow.*``' Intrinsics
7622 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7627 This is an overloaded intrinsic. You can use ``llvm.sadd.with.overflow``
7628 on any integer bit width.
7632 declare {i16, i1} @llvm.sadd.with.overflow.i16(i16 %a, i16 %b)
7633 declare {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
7634 declare {i64, i1} @llvm.sadd.with.overflow.i64(i64 %a, i64 %b)
7639 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
7640 a signed addition of the two arguments, and indicate whether an overflow
7641 occurred during the signed summation.
7646 The arguments (%a and %b) and the first element of the result structure
7647 may be of integer types of any bit width, but they must have the same
7648 bit width. The second element of the result structure must be of type
7649 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
7655 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
7656 a signed addition of the two variables. They return a structure --- the
7657 first element of which is the signed summation, and the second element
7658 of which is a bit specifying if the signed summation resulted in an
7664 .. code-block:: llvm
7666 %res = call {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
7667 %sum = extractvalue {i32, i1} %res, 0
7668 %obit = extractvalue {i32, i1} %res, 1
7669 br i1 %obit, label %overflow, label %normal
7671 '``llvm.uadd.with.overflow.*``' Intrinsics
7672 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7677 This is an overloaded intrinsic. You can use ``llvm.uadd.with.overflow``
7678 on any integer bit width.
7682 declare {i16, i1} @llvm.uadd.with.overflow.i16(i16 %a, i16 %b)
7683 declare {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
7684 declare {i64, i1} @llvm.uadd.with.overflow.i64(i64 %a, i64 %b)
7689 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
7690 an unsigned addition of the two arguments, and indicate whether a carry
7691 occurred during the unsigned summation.
7696 The arguments (%a and %b) and the first element of the result structure
7697 may be of integer types of any bit width, but they must have the same
7698 bit width. The second element of the result structure must be of type
7699 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
7705 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
7706 an unsigned addition of the two arguments. They return a structure --- the
7707 first element of which is the sum, and the second element of which is a
7708 bit specifying if the unsigned summation resulted in a carry.
7713 .. code-block:: llvm
7715 %res = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
7716 %sum = extractvalue {i32, i1} %res, 0
7717 %obit = extractvalue {i32, i1} %res, 1
7718 br i1 %obit, label %carry, label %normal
7720 '``llvm.ssub.with.overflow.*``' Intrinsics
7721 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7726 This is an overloaded intrinsic. You can use ``llvm.ssub.with.overflow``
7727 on any integer bit width.
7731 declare {i16, i1} @llvm.ssub.with.overflow.i16(i16 %a, i16 %b)
7732 declare {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
7733 declare {i64, i1} @llvm.ssub.with.overflow.i64(i64 %a, i64 %b)
7738 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
7739 a signed subtraction of the two arguments, and indicate whether an
7740 overflow occurred during the signed subtraction.
7745 The arguments (%a and %b) and the first element of the result structure
7746 may be of integer types of any bit width, but they must have the same
7747 bit width. The second element of the result structure must be of type
7748 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
7754 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
7755 a signed subtraction of the two arguments. They return a structure --- the
7756 first element of which is the subtraction, and the second element of
7757 which is a bit specifying if the signed subtraction resulted in an
7763 .. code-block:: llvm
7765 %res = call {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
7766 %sum = extractvalue {i32, i1} %res, 0
7767 %obit = extractvalue {i32, i1} %res, 1
7768 br i1 %obit, label %overflow, label %normal
7770 '``llvm.usub.with.overflow.*``' Intrinsics
7771 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7776 This is an overloaded intrinsic. You can use ``llvm.usub.with.overflow``
7777 on any integer bit width.
7781 declare {i16, i1} @llvm.usub.with.overflow.i16(i16 %a, i16 %b)
7782 declare {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
7783 declare {i64, i1} @llvm.usub.with.overflow.i64(i64 %a, i64 %b)
7788 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
7789 an unsigned subtraction of the two arguments, and indicate whether an
7790 overflow occurred during the unsigned subtraction.
7795 The arguments (%a and %b) and the first element of the result structure
7796 may be of integer types of any bit width, but they must have the same
7797 bit width. The second element of the result structure must be of type
7798 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
7804 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
7805 an unsigned subtraction of the two arguments. They return a structure ---
7806 the first element of which is the subtraction, and the second element of
7807 which is a bit specifying if the unsigned subtraction resulted in an
7813 .. code-block:: llvm
7815 %res = call {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
7816 %sum = extractvalue {i32, i1} %res, 0
7817 %obit = extractvalue {i32, i1} %res, 1
7818 br i1 %obit, label %overflow, label %normal
7820 '``llvm.smul.with.overflow.*``' Intrinsics
7821 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7826 This is an overloaded intrinsic. You can use ``llvm.smul.with.overflow``
7827 on any integer bit width.
7831 declare {i16, i1} @llvm.smul.with.overflow.i16(i16 %a, i16 %b)
7832 declare {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
7833 declare {i64, i1} @llvm.smul.with.overflow.i64(i64 %a, i64 %b)
7838 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
7839 a signed multiplication of the two arguments, and indicate whether an
7840 overflow occurred during the signed multiplication.
7845 The arguments (%a and %b) and the first element of the result structure
7846 may be of integer types of any bit width, but they must have the same
7847 bit width. The second element of the result structure must be of type
7848 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
7854 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
7855 a signed multiplication of the two arguments. They return a structure ---
7856 the first element of which is the multiplication, and the second element
7857 of which is a bit specifying if the signed multiplication resulted in an
7863 .. code-block:: llvm
7865 %res = call {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
7866 %sum = extractvalue {i32, i1} %res, 0
7867 %obit = extractvalue {i32, i1} %res, 1
7868 br i1 %obit, label %overflow, label %normal
7870 '``llvm.umul.with.overflow.*``' Intrinsics
7871 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7876 This is an overloaded intrinsic. You can use ``llvm.umul.with.overflow``
7877 on any integer bit width.
7881 declare {i16, i1} @llvm.umul.with.overflow.i16(i16 %a, i16 %b)
7882 declare {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
7883 declare {i64, i1} @llvm.umul.with.overflow.i64(i64 %a, i64 %b)
7888 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
7889 a unsigned multiplication of the two arguments, and indicate whether an
7890 overflow occurred during the unsigned multiplication.
7895 The arguments (%a and %b) and the first element of the result structure
7896 may be of integer types of any bit width, but they must have the same
7897 bit width. The second element of the result structure must be of type
7898 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
7904 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
7905 an unsigned multiplication of the two arguments. They return a structure ---
7906 the first element of which is the multiplication, and the second
7907 element of which is a bit specifying if the unsigned multiplication
7908 resulted in an overflow.
7913 .. code-block:: llvm
7915 %res = call {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
7916 %sum = extractvalue {i32, i1} %res, 0
7917 %obit = extractvalue {i32, i1} %res, 1
7918 br i1 %obit, label %overflow, label %normal
7920 Specialised Arithmetic Intrinsics
7921 ---------------------------------
7923 '``llvm.fmuladd.*``' Intrinsic
7924 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7931 declare float @llvm.fmuladd.f32(float %a, float %b, float %c)
7932 declare double @llvm.fmuladd.f64(double %a, double %b, double %c)
7937 The '``llvm.fmuladd.*``' intrinsic functions represent multiply-add
7938 expressions that can be fused if the code generator determines that (a) the
7939 target instruction set has support for a fused operation, and (b) that the
7940 fused operation is more efficient than the equivalent, separate pair of mul
7941 and add instructions.
7946 The '``llvm.fmuladd.*``' intrinsics each take three arguments: two
7947 multiplicands, a and b, and an addend c.
7956 %0 = call float @llvm.fmuladd.f32(%a, %b, %c)
7958 is equivalent to the expression a \* b + c, except that rounding will
7959 not be performed between the multiplication and addition steps if the
7960 code generator fuses the operations. Fusion is not guaranteed, even if
7961 the target platform supports it. If a fused multiply-add is required the
7962 corresponding llvm.fma.\* intrinsic function should be used instead.
7967 .. code-block:: llvm
7969 %r2 = call float @llvm.fmuladd.f32(float %a, float %b, float %c) ; yields {float}:r2 = (a * b) + c
7971 Half Precision Floating Point Intrinsics
7972 ----------------------------------------
7974 For most target platforms, half precision floating point is a
7975 storage-only format. This means that it is a dense encoding (in memory)
7976 but does not support computation in the format.
7978 This means that code must first load the half-precision floating point
7979 value as an i16, then convert it to float with
7980 :ref:`llvm.convert.from.fp16 <int_convert_from_fp16>`. Computation can
7981 then be performed on the float value (including extending to double
7982 etc). To store the value back to memory, it is first converted to float
7983 if needed, then converted to i16 with
7984 :ref:`llvm.convert.to.fp16 <int_convert_to_fp16>`, then storing as an
7987 .. _int_convert_to_fp16:
7989 '``llvm.convert.to.fp16``' Intrinsic
7990 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7997 declare i16 @llvm.convert.to.fp16(f32 %a)
8002 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion
8003 from single precision floating point format to half precision floating
8009 The intrinsic function contains single argument - the value to be
8015 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion
8016 from single precision floating point format to half precision floating
8017 point format. The return value is an ``i16`` which contains the
8023 .. code-block:: llvm
8025 %res = call i16 @llvm.convert.to.fp16(f32 %a)
8026 store i16 %res, i16* @x, align 2
8028 .. _int_convert_from_fp16:
8030 '``llvm.convert.from.fp16``' Intrinsic
8031 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8038 declare f32 @llvm.convert.from.fp16(i16 %a)
8043 The '``llvm.convert.from.fp16``' intrinsic function performs a
8044 conversion from half precision floating point format to single precision
8045 floating point format.
8050 The intrinsic function contains single argument - the value to be
8056 The '``llvm.convert.from.fp16``' intrinsic function performs a
8057 conversion from half single precision floating point format to single
8058 precision floating point format. The input half-float value is
8059 represented by an ``i16`` value.
8064 .. code-block:: llvm
8066 %a = load i16* @x, align 2
8067 %res = call f32 @llvm.convert.from.fp16(i16 %a)
8072 The LLVM debugger intrinsics (which all start with ``llvm.dbg.``
8073 prefix), are described in the `LLVM Source Level
8074 Debugging <SourceLevelDebugging.html#format_common_intrinsics>`_
8077 Exception Handling Intrinsics
8078 -----------------------------
8080 The LLVM exception handling intrinsics (which all start with
8081 ``llvm.eh.`` prefix), are described in the `LLVM Exception
8082 Handling <ExceptionHandling.html#format_common_intrinsics>`_ document.
8086 Trampoline Intrinsics
8087 ---------------------
8089 These intrinsics make it possible to excise one parameter, marked with
8090 the :ref:`nest <nest>` attribute, from a function. The result is a
8091 callable function pointer lacking the nest parameter - the caller does
8092 not need to provide a value for it. Instead, the value to use is stored
8093 in advance in a "trampoline", a block of memory usually allocated on the
8094 stack, which also contains code to splice the nest value into the
8095 argument list. This is used to implement the GCC nested function address
8098 For example, if the function is ``i32 f(i8* nest %c, i32 %x, i32 %y)``
8099 then the resulting function pointer has signature ``i32 (i32, i32)*``.
8100 It can be created as follows:
8102 .. code-block:: llvm
8104 %tramp = alloca [10 x i8], align 4 ; size and alignment only correct for X86
8105 %tramp1 = getelementptr [10 x i8]* %tramp, i32 0, i32 0
8106 call i8* @llvm.init.trampoline(i8* %tramp1, i8* bitcast (i32 (i8*, i32, i32)* @f to i8*), i8* %nval)
8107 %p = call i8* @llvm.adjust.trampoline(i8* %tramp1)
8108 %fp = bitcast i8* %p to i32 (i32, i32)*
8110 The call ``%val = call i32 %fp(i32 %x, i32 %y)`` is then equivalent to
8111 ``%val = call i32 %f(i8* %nval, i32 %x, i32 %y)``.
8115 '``llvm.init.trampoline``' Intrinsic
8116 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8123 declare void @llvm.init.trampoline(i8* <tramp>, i8* <func>, i8* <nval>)
8128 This fills the memory pointed to by ``tramp`` with executable code,
8129 turning it into a trampoline.
8134 The ``llvm.init.trampoline`` intrinsic takes three arguments, all
8135 pointers. The ``tramp`` argument must point to a sufficiently large and
8136 sufficiently aligned block of memory; this memory is written to by the
8137 intrinsic. Note that the size and the alignment are target-specific -
8138 LLVM currently provides no portable way of determining them, so a
8139 front-end that generates this intrinsic needs to have some
8140 target-specific knowledge. The ``func`` argument must hold a function
8141 bitcast to an ``i8*``.
8146 The block of memory pointed to by ``tramp`` is filled with target
8147 dependent code, turning it into a function. Then ``tramp`` needs to be
8148 passed to :ref:`llvm.adjust.trampoline <int_at>` to get a pointer which can
8149 be :ref:`bitcast (to a new function) and called <int_trampoline>`. The new
8150 function's signature is the same as that of ``func`` with any arguments
8151 marked with the ``nest`` attribute removed. At most one such ``nest``
8152 argument is allowed, and it must be of pointer type. Calling the new
8153 function is equivalent to calling ``func`` with the same argument list,
8154 but with ``nval`` used for the missing ``nest`` argument. If, after
8155 calling ``llvm.init.trampoline``, the memory pointed to by ``tramp`` is
8156 modified, then the effect of any later call to the returned function
8157 pointer is undefined.
8161 '``llvm.adjust.trampoline``' Intrinsic
8162 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8169 declare i8* @llvm.adjust.trampoline(i8* <tramp>)
8174 This performs any required machine-specific adjustment to the address of
8175 a trampoline (passed as ``tramp``).
8180 ``tramp`` must point to a block of memory which already has trampoline
8181 code filled in by a previous call to
8182 :ref:`llvm.init.trampoline <int_it>`.
8187 On some architectures the address of the code to be executed needs to be
8188 different to the address where the trampoline is actually stored. This
8189 intrinsic returns the executable address corresponding to ``tramp``
8190 after performing the required machine specific adjustments. The pointer
8191 returned can then be :ref:`bitcast and executed <int_trampoline>`.
8196 This class of intrinsics exists to information about the lifetime of
8197 memory objects and ranges where variables are immutable.
8199 '``llvm.lifetime.start``' Intrinsic
8200 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8207 declare void @llvm.lifetime.start(i64 <size>, i8* nocapture <ptr>)
8212 The '``llvm.lifetime.start``' intrinsic specifies the start of a memory
8218 The first argument is a constant integer representing the size of the
8219 object, or -1 if it is variable sized. The second argument is a pointer
8225 This intrinsic indicates that before this point in the code, the value
8226 of the memory pointed to by ``ptr`` is dead. This means that it is known
8227 to never be used and has an undefined value. A load from the pointer
8228 that precedes this intrinsic can be replaced with ``'undef'``.
8230 '``llvm.lifetime.end``' Intrinsic
8231 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8238 declare void @llvm.lifetime.end(i64 <size>, i8* nocapture <ptr>)
8243 The '``llvm.lifetime.end``' intrinsic specifies the end of a memory
8249 The first argument is a constant integer representing the size of the
8250 object, or -1 if it is variable sized. The second argument is a pointer
8256 This intrinsic indicates that after this point in the code, the value of
8257 the memory pointed to by ``ptr`` is dead. This means that it is known to
8258 never be used and has an undefined value. Any stores into the memory
8259 object following this intrinsic may be removed as dead.
8261 '``llvm.invariant.start``' Intrinsic
8262 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8269 declare {}* @llvm.invariant.start(i64 <size>, i8* nocapture <ptr>)
8274 The '``llvm.invariant.start``' intrinsic specifies that the contents of
8275 a memory object will not change.
8280 The first argument is a constant integer representing the size of the
8281 object, or -1 if it is variable sized. The second argument is a pointer
8287 This intrinsic indicates that until an ``llvm.invariant.end`` that uses
8288 the return value, the referenced memory location is constant and
8291 '``llvm.invariant.end``' Intrinsic
8292 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8299 declare void @llvm.invariant.end({}* <start>, i64 <size>, i8* nocapture <ptr>)
8304 The '``llvm.invariant.end``' intrinsic specifies that the contents of a
8305 memory object are mutable.
8310 The first argument is the matching ``llvm.invariant.start`` intrinsic.
8311 The second argument is a constant integer representing the size of the
8312 object, or -1 if it is variable sized and the third argument is a
8313 pointer to the object.
8318 This intrinsic indicates that the memory is mutable again.
8323 This class of intrinsics is designed to be generic and has no specific
8326 '``llvm.var.annotation``' Intrinsic
8327 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8334 declare void @llvm.var.annotation(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
8339 The '``llvm.var.annotation``' intrinsic.
8344 The first argument is a pointer to a value, the second is a pointer to a
8345 global string, the third is a pointer to a global string which is the
8346 source file name, and the last argument is the line number.
8351 This intrinsic allows annotation of local variables with arbitrary
8352 strings. This can be useful for special purpose optimizations that want
8353 to look for these annotations. These have no other defined use; they are
8354 ignored by code generation and optimization.
8356 '``llvm.ptr.annotation.*``' Intrinsic
8357 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8362 This is an overloaded intrinsic. You can use '``llvm.ptr.annotation``' on a
8363 pointer to an integer of any width. *NOTE* you must specify an address space for
8364 the pointer. The identifier for the default address space is the integer
8369 declare i8* @llvm.ptr.annotation.p<address space>i8(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
8370 declare i16* @llvm.ptr.annotation.p<address space>i16(i16* <val>, i8* <str>, i8* <str>, i32 <int>)
8371 declare i32* @llvm.ptr.annotation.p<address space>i32(i32* <val>, i8* <str>, i8* <str>, i32 <int>)
8372 declare i64* @llvm.ptr.annotation.p<address space>i64(i64* <val>, i8* <str>, i8* <str>, i32 <int>)
8373 declare i256* @llvm.ptr.annotation.p<address space>i256(i256* <val>, i8* <str>, i8* <str>, i32 <int>)
8378 The '``llvm.ptr.annotation``' intrinsic.
8383 The first argument is a pointer to an integer value of arbitrary bitwidth
8384 (result of some expression), the second is a pointer to a global string, the
8385 third is a pointer to a global string which is the source file name, and the
8386 last argument is the line number. It returns the value of the first argument.
8391 This intrinsic allows annotation of a pointer to an integer with arbitrary
8392 strings. This can be useful for special purpose optimizations that want to look
8393 for these annotations. These have no other defined use; they are ignored by code
8394 generation and optimization.
8396 '``llvm.annotation.*``' Intrinsic
8397 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8402 This is an overloaded intrinsic. You can use '``llvm.annotation``' on
8403 any integer bit width.
8407 declare i8 @llvm.annotation.i8(i8 <val>, i8* <str>, i8* <str>, i32 <int>)
8408 declare i16 @llvm.annotation.i16(i16 <val>, i8* <str>, i8* <str>, i32 <int>)
8409 declare i32 @llvm.annotation.i32(i32 <val>, i8* <str>, i8* <str>, i32 <int>)
8410 declare i64 @llvm.annotation.i64(i64 <val>, i8* <str>, i8* <str>, i32 <int>)
8411 declare i256 @llvm.annotation.i256(i256 <val>, i8* <str>, i8* <str>, i32 <int>)
8416 The '``llvm.annotation``' intrinsic.
8421 The first argument is an integer value (result of some expression), the
8422 second is a pointer to a global string, the third is a pointer to a
8423 global string which is the source file name, and the last argument is
8424 the line number. It returns the value of the first argument.
8429 This intrinsic allows annotations to be put on arbitrary expressions
8430 with arbitrary strings. This can be useful for special purpose
8431 optimizations that want to look for these annotations. These have no
8432 other defined use; they are ignored by code generation and optimization.
8434 '``llvm.trap``' Intrinsic
8435 ^^^^^^^^^^^^^^^^^^^^^^^^^
8442 declare void @llvm.trap() noreturn nounwind
8447 The '``llvm.trap``' intrinsic.
8457 This intrinsic is lowered to the target dependent trap instruction. If
8458 the target does not have a trap instruction, this intrinsic will be
8459 lowered to a call of the ``abort()`` function.
8461 '``llvm.debugtrap``' Intrinsic
8462 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8469 declare void @llvm.debugtrap() nounwind
8474 The '``llvm.debugtrap``' intrinsic.
8484 This intrinsic is lowered to code which is intended to cause an
8485 execution trap with the intention of requesting the attention of a
8488 '``llvm.stackprotector``' Intrinsic
8489 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8496 declare void @llvm.stackprotector(i8* <guard>, i8** <slot>)
8501 The ``llvm.stackprotector`` intrinsic takes the ``guard`` and stores it
8502 onto the stack at ``slot``. The stack slot is adjusted to ensure that it
8503 is placed on the stack before local variables.
8508 The ``llvm.stackprotector`` intrinsic requires two pointer arguments.
8509 The first argument is the value loaded from the stack guard
8510 ``@__stack_chk_guard``. The second variable is an ``alloca`` that has
8511 enough space to hold the value of the guard.
8516 This intrinsic causes the prologue/epilogue inserter to force the
8517 position of the ``AllocaInst`` stack slot to be before local variables
8518 on the stack. This is to ensure that if a local variable on the stack is
8519 overwritten, it will destroy the value of the guard. When the function
8520 exits, the guard on the stack is checked against the original guard. If
8521 they are different, then the program aborts by calling the
8522 ``__stack_chk_fail()`` function.
8524 '``llvm.objectsize``' Intrinsic
8525 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8532 declare i32 @llvm.objectsize.i32(i8* <object>, i1 <min>)
8533 declare i64 @llvm.objectsize.i64(i8* <object>, i1 <min>)
8538 The ``llvm.objectsize`` intrinsic is designed to provide information to
8539 the optimizers to determine at compile time whether a) an operation
8540 (like memcpy) will overflow a buffer that corresponds to an object, or
8541 b) that a runtime check for overflow isn't necessary. An object in this
8542 context means an allocation of a specific class, structure, array, or
8548 The ``llvm.objectsize`` intrinsic takes two arguments. The first
8549 argument is a pointer to or into the ``object``. The second argument is
8550 a boolean and determines whether ``llvm.objectsize`` returns 0 (if true)
8551 or -1 (if false) when the object size is unknown. The second argument
8552 only accepts constants.
8557 The ``llvm.objectsize`` intrinsic is lowered to a constant representing
8558 the size of the object concerned. If the size cannot be determined at
8559 compile time, ``llvm.objectsize`` returns ``i32/i64 -1 or 0`` (depending
8560 on the ``min`` argument).
8562 '``llvm.expect``' Intrinsic
8563 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8570 declare i32 @llvm.expect.i32(i32 <val>, i32 <expected_val>)
8571 declare i64 @llvm.expect.i64(i64 <val>, i64 <expected_val>)
8576 The ``llvm.expect`` intrinsic provides information about expected (the
8577 most probable) value of ``val``, which can be used by optimizers.
8582 The ``llvm.expect`` intrinsic takes two arguments. The first argument is
8583 a value. The second argument is an expected value, this needs to be a
8584 constant value, variables are not allowed.
8589 This intrinsic is lowered to the ``val``.
8591 '``llvm.donothing``' Intrinsic
8592 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8599 declare void @llvm.donothing() nounwind readnone
8604 The ``llvm.donothing`` intrinsic doesn't perform any operation. It's the
8605 only intrinsic that can be called with an invoke instruction.
8615 This intrinsic does nothing, and it's removed by optimizers and ignored