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.
726 Garbage Collector Names
727 -----------------------
729 Each function may specify a garbage collector name, which is simply a
734 define void @f() gc "name" { ... }
736 The compiler declares the supported values of *name*. Specifying a
737 collector which will cause the compiler to alter its output in order to
738 support the named garbage collection algorithm.
745 Attribute groups are groups of attributes that are referenced by objects within
746 the IR. They are important for keeping ``.ll`` files readable, because a lot of
747 functions will use the same set of attributes. In the degenerative case of a
748 ``.ll`` file that corresponds to a single ``.c`` file, the single attribute
749 group will capture the important command line flags used to build that file.
751 An attribute group is a module-level object. To use an attribute group, an
752 object references the attribute group's ID (e.g. ``#37``). An object may refer
753 to more than one attribute group. In that situation, the attributes from the
754 different groups are merged.
756 Here is an example of attribute groups for a function that should always be
757 inlined, has a stack alignment of 4, and which shouldn't use SSE instructions:
761 ; Target-independent attributes:
762 #0 = attributes { alwaysinline alignstack=4 }
764 ; Target-dependent attributes:
765 #1 = attributes { "no-sse" }
767 ; Function @f has attributes: alwaysinline, alignstack=4, and "no-sse".
768 define void @f() #0 #1 { ... }
775 Function attributes are set to communicate additional information about
776 a function. Function attributes are considered to be part of the
777 function, not of the function type, so functions with different function
778 attributes can have the same function type.
780 Function attributes are simple keywords that follow the type specified.
781 If multiple attributes are needed, they are space separated. For
786 define void @f() noinline { ... }
787 define void @f() alwaysinline { ... }
788 define void @f() alwaysinline optsize { ... }
789 define void @f() optsize { ... }
792 This attribute indicates that the address safety analysis is enabled
795 This attribute indicates that, when emitting the prologue and
796 epilogue, the backend should forcibly align the stack pointer.
797 Specify the desired alignment, which must be a power of two, in
800 This attribute indicates that the inliner should attempt to inline
801 this function into callers whenever possible, ignoring any active
802 inlining size threshold for this caller.
804 This attribute suppresses lazy symbol binding for the function. This
805 may make calls to the function faster, at the cost of extra program
806 startup time if the function is not called during program startup.
808 This attribute indicates that the source code contained a hint that
809 inlining this function is desirable (such as the "inline" keyword in
810 C/C++). It is just a hint; it imposes no requirements on the
813 This attribute disables prologue / epilogue emission for the
814 function. This can have very system-specific consequences.
816 This attribute indicates that calls to the function cannot be
817 duplicated. A call to a ``noduplicate`` function may be moved
818 within its parent function, but may not be duplicated within
821 A function containing a ``noduplicate`` call may still
822 be an inlining candidate, provided that the call is not
823 duplicated by inlining. That implies that the function has
824 internal linkage and only has one call site, so the original
825 call is dead after inlining.
827 This attributes disables implicit floating point instructions.
829 This attribute indicates that the inliner should never inline this
830 function in any situation. This attribute may not be used together
831 with the ``alwaysinline`` attribute.
833 This attribute indicates that the code generator should not use a
834 red zone, even if the target-specific ABI normally permits it.
836 This function attribute indicates that the function never returns
837 normally. This produces undefined behavior at runtime if the
838 function ever does dynamically return.
840 This function attribute indicates that the function never returns
841 with an unwind or exceptional control flow. If the function does
842 unwind, its runtime behavior is undefined.
844 This attribute suggests that optimization passes and code generator
845 passes make choices that keep the code size of this function low,
846 and otherwise do optimizations specifically to reduce code size.
848 This attribute indicates that the function computes its result (or
849 decides to unwind an exception) based strictly on its arguments,
850 without dereferencing any pointer arguments or otherwise accessing
851 any mutable state (e.g. memory, control registers, etc) visible to
852 caller functions. It does not write through any pointer arguments
853 (including ``byval`` arguments) and never changes any state visible
854 to callers. This means that it cannot unwind exceptions by calling
855 the ``C++`` exception throwing methods.
857 This attribute indicates that the function does not write through
858 any pointer arguments (including ``byval`` arguments) or otherwise
859 modify any state (e.g. memory, control registers, etc) visible to
860 caller functions. It may dereference pointer arguments and read
861 state that may be set in the caller. A readonly function always
862 returns the same value (or unwinds an exception identically) when
863 called with the same set of arguments and global state. It cannot
864 unwind an exception by calling the ``C++`` exception throwing
867 This attribute indicates that this function can return twice. The C
868 ``setjmp`` is an example of such a function. The compiler disables
869 some optimizations (like tail calls) in the caller of these
872 This attribute indicates that the function should emit a stack
873 smashing protector. It is in the form of a "canary" --- a random value
874 placed on the stack before the local variables that's checked upon
875 return from the function to see if it has been overwritten. A
876 heuristic is used to determine if a function needs stack protectors
877 or not. The heuristic used will enable protectors for functions with:
879 - Character arrays larger than ``ssp-buffer-size`` (default 8).
880 - Aggregates containing character arrays larger than ``ssp-buffer-size``.
881 - Calls to alloca() with variable sizes or constant sizes greater than
884 If a function that has an ``ssp`` attribute is inlined into a
885 function that doesn't have an ``ssp`` attribute, then the resulting
886 function will have an ``ssp`` attribute.
888 This attribute indicates that the function should *always* emit a
889 stack smashing protector. This overrides the ``ssp`` function
892 If a function that has an ``sspreq`` attribute is inlined into a
893 function that doesn't have an ``sspreq`` attribute or which has an
894 ``ssp`` or ``sspstrong`` attribute, then the resulting function will have
895 an ``sspreq`` attribute.
897 This attribute indicates that the function should emit a stack smashing
898 protector. This attribute causes a strong heuristic to be used when
899 determining if a function needs stack protectors. The strong heuristic
900 will enable protectors for functions with:
902 - Arrays of any size and type
903 - Aggregates containing an array of any size and type.
905 - Local variables that have had their address taken.
907 This overrides the ``ssp`` function attribute.
909 If a function that has an ``sspstrong`` attribute is inlined into a
910 function that doesn't have an ``sspstrong`` attribute, then the
911 resulting function will have an ``sspstrong`` attribute.
913 This attribute indicates that the ABI being targeted requires that
914 an unwind table entry be produce for this function even if we can
915 show that no exceptions passes by it. This is normally the case for
916 the ELF x86-64 abi, but it can be disabled for some compilation
921 Module-Level Inline Assembly
922 ----------------------------
924 Modules may contain "module-level inline asm" blocks, which corresponds
925 to the GCC "file scope inline asm" blocks. These blocks are internally
926 concatenated by LLVM and treated as a single unit, but may be separated
927 in the ``.ll`` file if desired. The syntax is very simple:
931 module asm "inline asm code goes here"
932 module asm "more can go here"
934 The strings can contain any character by escaping non-printable
935 characters. The escape sequence used is simply "\\xx" where "xx" is the
936 two digit hex code for the number.
938 The inline asm code is simply printed to the machine code .s file when
939 assembly code is generated.
944 A module may specify a target specific data layout string that specifies
945 how data is to be laid out in memory. The syntax for the data layout is
950 target datalayout = "layout specification"
952 The *layout specification* consists of a list of specifications
953 separated by the minus sign character ('-'). Each specification starts
954 with a letter and may include other information after the letter to
955 define some aspect of the data layout. The specifications accepted are
959 Specifies that the target lays out data in big-endian form. That is,
960 the bits with the most significance have the lowest address
963 Specifies that the target lays out data in little-endian form. That
964 is, the bits with the least significance have the lowest address
967 Specifies the natural alignment of the stack in bits. Alignment
968 promotion of stack variables is limited to the natural stack
969 alignment to avoid dynamic stack realignment. The stack alignment
970 must be a multiple of 8-bits. If omitted, the natural stack
971 alignment defaults to "unspecified", which does not prevent any
972 alignment promotions.
973 ``p[n]:<size>:<abi>:<pref>``
974 This specifies the *size* of a pointer and its ``<abi>`` and
975 ``<pref>``\erred alignments for address space ``n``. All sizes are in
976 bits. Specifying the ``<pref>`` alignment is optional. If omitted, the
977 preceding ``:`` should be omitted too. The address space, ``n`` is
978 optional, and if not specified, denotes the default address space 0.
979 The value of ``n`` must be in the range [1,2^23).
980 ``i<size>:<abi>:<pref>``
981 This specifies the alignment for an integer type of a given bit
982 ``<size>``. The value of ``<size>`` must be in the range [1,2^23).
983 ``v<size>:<abi>:<pref>``
984 This specifies the alignment for a vector type of a given bit
986 ``f<size>:<abi>:<pref>``
987 This specifies the alignment for a floating point type of a given bit
988 ``<size>``. Only values of ``<size>`` that are supported by the target
989 will work. 32 (float) and 64 (double) are supported on all targets; 80
990 or 128 (different flavors of long double) are also supported on some
992 ``a<size>:<abi>:<pref>``
993 This specifies the alignment for an aggregate type of a given bit
995 ``s<size>:<abi>:<pref>``
996 This specifies the alignment for a stack object of a given bit
998 ``n<size1>:<size2>:<size3>...``
999 This specifies a set of native integer widths for the target CPU in
1000 bits. For example, it might contain ``n32`` for 32-bit PowerPC,
1001 ``n32:64`` for PowerPC 64, or ``n8:16:32:64`` for X86-64. Elements of
1002 this set are considered to support most general arithmetic operations
1005 When constructing the data layout for a given target, LLVM starts with a
1006 default set of specifications which are then (possibly) overridden by
1007 the specifications in the ``datalayout`` keyword. The default
1008 specifications are given in this list:
1010 - ``E`` - big endian
1011 - ``p:64:64:64`` - 64-bit pointers with 64-bit alignment
1012 - ``S0`` - natural stack alignment is unspecified
1013 - ``i1:8:8`` - i1 is 8-bit (byte) aligned
1014 - ``i8:8:8`` - i8 is 8-bit (byte) aligned
1015 - ``i16:16:16`` - i16 is 16-bit aligned
1016 - ``i32:32:32`` - i32 is 32-bit aligned
1017 - ``i64:32:64`` - i64 has ABI alignment of 32-bits but preferred
1018 alignment of 64-bits
1019 - ``f16:16:16`` - half is 16-bit aligned
1020 - ``f32:32:32`` - float is 32-bit aligned
1021 - ``f64:64:64`` - double is 64-bit aligned
1022 - ``f128:128:128`` - quad is 128-bit aligned
1023 - ``v64:64:64`` - 64-bit vector is 64-bit aligned
1024 - ``v128:128:128`` - 128-bit vector is 128-bit aligned
1025 - ``a0:0:64`` - aggregates are 64-bit aligned
1027 When LLVM is determining the alignment for a given type, it uses the
1030 #. If the type sought is an exact match for one of the specifications,
1031 that specification is used.
1032 #. If no match is found, and the type sought is an integer type, then
1033 the smallest integer type that is larger than the bitwidth of the
1034 sought type is used. If none of the specifications are larger than
1035 the bitwidth then the largest integer type is used. For example,
1036 given the default specifications above, the i7 type will use the
1037 alignment of i8 (next largest) while both i65 and i256 will use the
1038 alignment of i64 (largest specified).
1039 #. If no match is found, and the type sought is a vector type, then the
1040 largest vector type that is smaller than the sought vector type will
1041 be used as a fall back. This happens because <128 x double> can be
1042 implemented in terms of 64 <2 x double>, for example.
1044 The function of the data layout string may not be what you expect.
1045 Notably, this is not a specification from the frontend of what alignment
1046 the code generator should use.
1048 Instead, if specified, the target data layout is required to match what
1049 the ultimate *code generator* expects. This string is used by the
1050 mid-level optimizers to improve code, and this only works if it matches
1051 what the ultimate code generator uses. If you would like to generate IR
1052 that does not embed this target-specific detail into the IR, then you
1053 don't have to specify the string. This will disable some optimizations
1054 that require precise layout information, but this also prevents those
1055 optimizations from introducing target specificity into the IR.
1057 .. _pointeraliasing:
1059 Pointer Aliasing Rules
1060 ----------------------
1062 Any memory access must be done through a pointer value associated with
1063 an address range of the memory access, otherwise the behavior is
1064 undefined. Pointer values are associated with address ranges according
1065 to the following rules:
1067 - A pointer value is associated with the addresses associated with any
1068 value it is *based* on.
1069 - An address of a global variable is associated with the address range
1070 of the variable's storage.
1071 - The result value of an allocation instruction is associated with the
1072 address range of the allocated storage.
1073 - A null pointer in the default address-space is associated with no
1075 - An integer constant other than zero or a pointer value returned from
1076 a function not defined within LLVM may be associated with address
1077 ranges allocated through mechanisms other than those provided by
1078 LLVM. Such ranges shall not overlap with any ranges of addresses
1079 allocated by mechanisms provided by LLVM.
1081 A pointer value is *based* on another pointer value according to the
1084 - A pointer value formed from a ``getelementptr`` operation is *based*
1085 on the first operand of the ``getelementptr``.
1086 - The result value of a ``bitcast`` is *based* on the operand of the
1088 - A pointer value formed by an ``inttoptr`` is *based* on all pointer
1089 values that contribute (directly or indirectly) to the computation of
1090 the pointer's value.
1091 - The "*based* on" relationship is transitive.
1093 Note that this definition of *"based"* is intentionally similar to the
1094 definition of *"based"* in C99, though it is slightly weaker.
1096 LLVM IR does not associate types with memory. The result type of a
1097 ``load`` merely indicates the size and alignment of the memory from
1098 which to load, as well as the interpretation of the value. The first
1099 operand type of a ``store`` similarly only indicates the size and
1100 alignment of the store.
1102 Consequently, type-based alias analysis, aka TBAA, aka
1103 ``-fstrict-aliasing``, is not applicable to general unadorned LLVM IR.
1104 :ref:`Metadata <metadata>` may be used to encode additional information
1105 which specialized optimization passes may use to implement type-based
1110 Volatile Memory Accesses
1111 ------------------------
1113 Certain memory accesses, such as :ref:`load <i_load>`'s,
1114 :ref:`store <i_store>`'s, and :ref:`llvm.memcpy <int_memcpy>`'s may be
1115 marked ``volatile``. The optimizers must not change the number of
1116 volatile operations or change their order of execution relative to other
1117 volatile operations. The optimizers *may* change the order of volatile
1118 operations relative to non-volatile operations. This is not Java's
1119 "volatile" and has no cross-thread synchronization behavior.
1121 IR-level volatile loads and stores cannot safely be optimized into
1122 llvm.memcpy or llvm.memmove intrinsics even when those intrinsics are
1123 flagged volatile. Likewise, the backend should never split or merge
1124 target-legal volatile load/store instructions.
1126 .. admonition:: Rationale
1128 Platforms may rely on volatile loads and stores of natively supported
1129 data width to be executed as single instruction. For example, in C
1130 this holds for an l-value of volatile primitive type with native
1131 hardware support, but not necessarily for aggregate types. The
1132 frontend upholds these expectations, which are intentionally
1133 unspecified in the IR. The rules above ensure that IR transformation
1134 do not violate the frontend's contract with the language.
1138 Memory Model for Concurrent Operations
1139 --------------------------------------
1141 The LLVM IR does not define any way to start parallel threads of
1142 execution or to register signal handlers. Nonetheless, there are
1143 platform-specific ways to create them, and we define LLVM IR's behavior
1144 in their presence. This model is inspired by the C++0x memory model.
1146 For a more informal introduction to this model, see the :doc:`Atomics`.
1148 We define a *happens-before* partial order as the least partial order
1151 - Is a superset of single-thread program order, and
1152 - When a *synchronizes-with* ``b``, includes an edge from ``a`` to
1153 ``b``. *Synchronizes-with* pairs are introduced by platform-specific
1154 techniques, like pthread locks, thread creation, thread joining,
1155 etc., and by atomic instructions. (See also :ref:`Atomic Memory Ordering
1156 Constraints <ordering>`).
1158 Note that program order does not introduce *happens-before* edges
1159 between a thread and signals executing inside that thread.
1161 Every (defined) read operation (load instructions, memcpy, atomic
1162 loads/read-modify-writes, etc.) R reads a series of bytes written by
1163 (defined) write operations (store instructions, atomic
1164 stores/read-modify-writes, memcpy, etc.). For the purposes of this
1165 section, initialized globals are considered to have a write of the
1166 initializer which is atomic and happens before any other read or write
1167 of the memory in question. For each byte of a read R, R\ :sub:`byte`
1168 may see any write to the same byte, except:
1170 - If write\ :sub:`1` happens before write\ :sub:`2`, and
1171 write\ :sub:`2` happens before R\ :sub:`byte`, then
1172 R\ :sub:`byte` does not see write\ :sub:`1`.
1173 - If R\ :sub:`byte` happens before write\ :sub:`3`, then
1174 R\ :sub:`byte` does not see write\ :sub:`3`.
1176 Given that definition, R\ :sub:`byte` is defined as follows:
1178 - If R is volatile, the result is target-dependent. (Volatile is
1179 supposed to give guarantees which can support ``sig_atomic_t`` in
1180 C/C++, and may be used for accesses to addresses which do not behave
1181 like normal memory. It does not generally provide cross-thread
1183 - Otherwise, if there is no write to the same byte that happens before
1184 R\ :sub:`byte`, R\ :sub:`byte` returns ``undef`` for that byte.
1185 - Otherwise, if R\ :sub:`byte` may see exactly one write,
1186 R\ :sub:`byte` returns the value written by that write.
1187 - Otherwise, if R is atomic, and all the writes R\ :sub:`byte` may
1188 see are atomic, it chooses one of the values written. See the :ref:`Atomic
1189 Memory Ordering Constraints <ordering>` section for additional
1190 constraints on how the choice is made.
1191 - Otherwise R\ :sub:`byte` returns ``undef``.
1193 R returns the value composed of the series of bytes it read. This
1194 implies that some bytes within the value may be ``undef`` **without**
1195 the entire value being ``undef``. Note that this only defines the
1196 semantics of the operation; it doesn't mean that targets will emit more
1197 than one instruction to read the series of bytes.
1199 Note that in cases where none of the atomic intrinsics are used, this
1200 model places only one restriction on IR transformations on top of what
1201 is required for single-threaded execution: introducing a store to a byte
1202 which might not otherwise be stored is not allowed in general.
1203 (Specifically, in the case where another thread might write to and read
1204 from an address, introducing a store can change a load that may see
1205 exactly one write into a load that may see multiple writes.)
1209 Atomic Memory Ordering Constraints
1210 ----------------------------------
1212 Atomic instructions (:ref:`cmpxchg <i_cmpxchg>`,
1213 :ref:`atomicrmw <i_atomicrmw>`, :ref:`fence <i_fence>`,
1214 :ref:`atomic load <i_load>`, and :ref:`atomic store <i_store>`) take
1215 an ordering parameter that determines which other atomic instructions on
1216 the same address they *synchronize with*. These semantics are borrowed
1217 from Java and C++0x, but are somewhat more colloquial. If these
1218 descriptions aren't precise enough, check those specs (see spec
1219 references in the :doc:`atomics guide <Atomics>`).
1220 :ref:`fence <i_fence>` instructions treat these orderings somewhat
1221 differently since they don't take an address. See that instruction's
1222 documentation for details.
1224 For a simpler introduction to the ordering constraints, see the
1228 The set of values that can be read is governed by the happens-before
1229 partial order. A value cannot be read unless some operation wrote
1230 it. This is intended to provide a guarantee strong enough to model
1231 Java's non-volatile shared variables. This ordering cannot be
1232 specified for read-modify-write operations; it is not strong enough
1233 to make them atomic in any interesting way.
1235 In addition to the guarantees of ``unordered``, there is a single
1236 total order for modifications by ``monotonic`` operations on each
1237 address. All modification orders must be compatible with the
1238 happens-before order. There is no guarantee that the modification
1239 orders can be combined to a global total order for the whole program
1240 (and this often will not be possible). The read in an atomic
1241 read-modify-write operation (:ref:`cmpxchg <i_cmpxchg>` and
1242 :ref:`atomicrmw <i_atomicrmw>`) reads the value in the modification
1243 order immediately before the value it writes. If one atomic read
1244 happens before another atomic read of the same address, the later
1245 read must see the same value or a later value in the address's
1246 modification order. This disallows reordering of ``monotonic`` (or
1247 stronger) operations on the same address. If an address is written
1248 ``monotonic``-ally by one thread, and other threads ``monotonic``-ally
1249 read that address repeatedly, the other threads must eventually see
1250 the write. This corresponds to the C++0x/C1x
1251 ``memory_order_relaxed``.
1253 In addition to the guarantees of ``monotonic``, a
1254 *synchronizes-with* edge may be formed with a ``release`` operation.
1255 This is intended to model C++'s ``memory_order_acquire``.
1257 In addition to the guarantees of ``monotonic``, if this operation
1258 writes a value which is subsequently read by an ``acquire``
1259 operation, it *synchronizes-with* that operation. (This isn't a
1260 complete description; see the C++0x definition of a release
1261 sequence.) This corresponds to the C++0x/C1x
1262 ``memory_order_release``.
1263 ``acq_rel`` (acquire+release)
1264 Acts as both an ``acquire`` and ``release`` operation on its
1265 address. This corresponds to the C++0x/C1x ``memory_order_acq_rel``.
1266 ``seq_cst`` (sequentially consistent)
1267 In addition to the guarantees of ``acq_rel`` (``acquire`` for an
1268 operation which only reads, ``release`` for an operation which only
1269 writes), there is a global total order on all
1270 sequentially-consistent operations on all addresses, which is
1271 consistent with the *happens-before* partial order and with the
1272 modification orders of all the affected addresses. Each
1273 sequentially-consistent read sees the last preceding write to the
1274 same address in this global order. This corresponds to the C++0x/C1x
1275 ``memory_order_seq_cst`` and Java volatile.
1279 If an atomic operation is marked ``singlethread``, it only *synchronizes
1280 with* or participates in modification and seq\_cst total orderings with
1281 other operations running in the same thread (for example, in signal
1289 LLVM IR floating-point binary ops (:ref:`fadd <i_fadd>`,
1290 :ref:`fsub <i_fsub>`, :ref:`fmul <i_fmul>`, :ref:`fdiv <i_fdiv>`,
1291 :ref:`frem <i_frem>`) have the following flags that can set to enable
1292 otherwise unsafe floating point operations
1295 No NaNs - Allow optimizations to assume the arguments and result are not
1296 NaN. Such optimizations are required to retain defined behavior over
1297 NaNs, but the value of the result is undefined.
1300 No Infs - Allow optimizations to assume the arguments and result are not
1301 +/-Inf. Such optimizations are required to retain defined behavior over
1302 +/-Inf, but the value of the result is undefined.
1305 No Signed Zeros - Allow optimizations to treat the sign of a zero
1306 argument or result as insignificant.
1309 Allow Reciprocal - Allow optimizations to use the reciprocal of an
1310 argument rather than perform division.
1313 Fast - Allow algebraically equivalent transformations that may
1314 dramatically change results in floating point (e.g. reassociate). This
1315 flag implies all the others.
1322 The LLVM type system is one of the most important features of the
1323 intermediate representation. Being typed enables a number of
1324 optimizations to be performed on the intermediate representation
1325 directly, without having to do extra analyses on the side before the
1326 transformation. A strong type system makes it easier to read the
1327 generated code and enables novel analyses and transformations that are
1328 not feasible to perform on normal three address code representations.
1330 Type Classifications
1331 --------------------
1333 The types fall into a few useful classifications:
1342 * - :ref:`integer <t_integer>`
1343 - ``i1``, ``i2``, ``i3``, ... ``i8``, ... ``i16``, ... ``i32``, ...
1346 * - :ref:`floating point <t_floating>`
1347 - ``half``, ``float``, ``double``, ``x86_fp80``, ``fp128``,
1355 - :ref:`integer <t_integer>`, :ref:`floating point <t_floating>`,
1356 :ref:`pointer <t_pointer>`, :ref:`vector <t_vector>`,
1357 :ref:`structure <t_struct>`, :ref:`array <t_array>`,
1358 :ref:`label <t_label>`, :ref:`metadata <t_metadata>`.
1360 * - :ref:`primitive <t_primitive>`
1361 - :ref:`label <t_label>`,
1362 :ref:`void <t_void>`,
1363 :ref:`integer <t_integer>`,
1364 :ref:`floating point <t_floating>`,
1365 :ref:`x86mmx <t_x86mmx>`,
1366 :ref:`metadata <t_metadata>`.
1368 * - :ref:`derived <t_derived>`
1369 - :ref:`array <t_array>`,
1370 :ref:`function <t_function>`,
1371 :ref:`pointer <t_pointer>`,
1372 :ref:`structure <t_struct>`,
1373 :ref:`vector <t_vector>`,
1374 :ref:`opaque <t_opaque>`.
1376 The :ref:`first class <t_firstclass>` types are perhaps the most important.
1377 Values of these types are the only ones which can be produced by
1385 The primitive types are the fundamental building blocks of the LLVM
1396 The integer type is a very simple type that simply specifies an
1397 arbitrary bit width for the integer type desired. Any bit width from 1
1398 bit to 2\ :sup:`23`\ -1 (about 8 million) can be specified.
1407 The number of bits the integer will occupy is specified by the ``N``
1413 +----------------+------------------------------------------------+
1414 | ``i1`` | a single-bit integer. |
1415 +----------------+------------------------------------------------+
1416 | ``i32`` | a 32-bit integer. |
1417 +----------------+------------------------------------------------+
1418 | ``i1942652`` | a really big integer of over 1 million bits. |
1419 +----------------+------------------------------------------------+
1423 Floating Point Types
1424 ^^^^^^^^^^^^^^^^^^^^
1433 - 16-bit floating point value
1436 - 32-bit floating point value
1439 - 64-bit floating point value
1442 - 128-bit floating point value (112-bit mantissa)
1445 - 80-bit floating point value (X87)
1448 - 128-bit floating point value (two 64-bits)
1458 The x86mmx type represents a value held in an MMX register on an x86
1459 machine. The operations allowed on it are quite limited: parameters and
1460 return values, load and store, and bitcast. User-specified MMX
1461 instructions are represented as intrinsic or asm calls with arguments
1462 and/or results of this type. There are no arrays, vectors or constants
1480 The void type does not represent any value and has no size.
1497 The label type represents code labels.
1514 The metadata type represents embedded metadata. No derived types may be
1515 created from metadata except for :ref:`function <t_function>` arguments.
1529 The real power in LLVM comes from the derived types in the system. This
1530 is what allows a programmer to represent arrays, functions, pointers,
1531 and other useful types. Each of these types contain one or more element
1532 types which may be a primitive type, or another derived type. For
1533 example, it is possible to have a two dimensional array, using an array
1534 as the element type of another array.
1541 Aggregate Types are a subset of derived types that can contain multiple
1542 member types. :ref:`Arrays <t_array>` and :ref:`structs <t_struct>` are
1543 aggregate types. :ref:`Vectors <t_vector>` are not considered to be
1554 The array type is a very simple derived type that arranges elements
1555 sequentially in memory. The array type requires a size (number of
1556 elements) and an underlying data type.
1563 [<# elements> x <elementtype>]
1565 The number of elements is a constant integer value; ``elementtype`` may
1566 be any type with a size.
1571 +------------------+--------------------------------------+
1572 | ``[40 x i32]`` | Array of 40 32-bit integer values. |
1573 +------------------+--------------------------------------+
1574 | ``[41 x i32]`` | Array of 41 32-bit integer values. |
1575 +------------------+--------------------------------------+
1576 | ``[4 x i8]`` | Array of 4 8-bit integer values. |
1577 +------------------+--------------------------------------+
1579 Here are some examples of multidimensional arrays:
1581 +-----------------------------+----------------------------------------------------------+
1582 | ``[3 x [4 x i32]]`` | 3x4 array of 32-bit integer values. |
1583 +-----------------------------+----------------------------------------------------------+
1584 | ``[12 x [10 x float]]`` | 12x10 array of single precision floating point values. |
1585 +-----------------------------+----------------------------------------------------------+
1586 | ``[2 x [3 x [4 x i16]]]`` | 2x3x4 array of 16-bit integer values. |
1587 +-----------------------------+----------------------------------------------------------+
1589 There is no restriction on indexing beyond the end of the array implied
1590 by a static type (though there are restrictions on indexing beyond the
1591 bounds of an allocated object in some cases). This means that
1592 single-dimension 'variable sized array' addressing can be implemented in
1593 LLVM with a zero length array type. An implementation of 'pascal style
1594 arrays' in LLVM could use the type "``{ i32, [0 x float]}``", for
1605 The function type can be thought of as a function signature. It consists
1606 of a return type and a list of formal parameter types. The return type
1607 of a function type is a first class type or a void type.
1614 <returntype> (<parameter list>)
1616 ...where '``<parameter list>``' is a comma-separated list of type
1617 specifiers. Optionally, the parameter list may include a type ``...``,
1618 which indicates that the function takes a variable number of arguments.
1619 Variable argument functions can access their arguments with the
1620 :ref:`variable argument handling intrinsic <int_varargs>` functions.
1621 '``<returntype>``' is any type except :ref:`label <t_label>`.
1626 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1627 | ``i32 (i32)`` | function taking an ``i32``, returning an ``i32`` |
1628 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1629 | ``float (i16, i32 *) *`` | :ref:`Pointer <t_pointer>` to a function that takes an ``i16`` and a :ref:`pointer <t_pointer>` to ``i32``, returning ``float``. |
1630 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1631 | ``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. |
1632 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1633 | ``{i32, i32} (i32)`` | A function taking an ``i32``, returning a :ref:`structure <t_struct>` containing two ``i32`` values |
1634 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1644 The structure type is used to represent a collection of data members
1645 together in memory. The elements of a structure may be any type that has
1648 Structures in memory are accessed using '``load``' and '``store``' by
1649 getting a pointer to a field with the '``getelementptr``' instruction.
1650 Structures in registers are accessed using the '``extractvalue``' and
1651 '``insertvalue``' instructions.
1653 Structures may optionally be "packed" structures, which indicate that
1654 the alignment of the struct is one byte, and that there is no padding
1655 between the elements. In non-packed structs, padding between field types
1656 is inserted as defined by the DataLayout string in the module, which is
1657 required to match what the underlying code generator expects.
1659 Structures can either be "literal" or "identified". A literal structure
1660 is defined inline with other types (e.g. ``{i32, i32}*``) whereas
1661 identified types are always defined at the top level with a name.
1662 Literal types are uniqued by their contents and can never be recursive
1663 or opaque since there is no way to write one. Identified types can be
1664 recursive, can be opaqued, and are never uniqued.
1671 %T1 = type { <type list> } ; Identified normal struct type
1672 %T2 = type <{ <type list> }> ; Identified packed struct type
1677 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1678 | ``{ i32, i32, i32 }`` | A triple of three ``i32`` values |
1679 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1680 | ``{ 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``. |
1681 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1682 | ``<{ i8, i32 }>`` | A packed struct known to be 5 bytes in size. |
1683 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1687 Opaque Structure Types
1688 ^^^^^^^^^^^^^^^^^^^^^^
1693 Opaque structure types are used to represent named structure types that
1694 do not have a body specified. This corresponds (for example) to the C
1695 notion of a forward declared structure.
1708 +--------------+-------------------+
1709 | ``opaque`` | An opaque type. |
1710 +--------------+-------------------+
1720 The pointer type is used to specify memory locations. Pointers are
1721 commonly used to reference objects in memory.
1723 Pointer types may have an optional address space attribute defining the
1724 numbered address space where the pointed-to object resides. The default
1725 address space is number zero. The semantics of non-zero address spaces
1726 are target-specific.
1728 Note that LLVM does not permit pointers to void (``void*``) nor does it
1729 permit pointers to labels (``label*``). Use ``i8*`` instead.
1741 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1742 | ``[4 x i32]*`` | A :ref:`pointer <t_pointer>` to :ref:`array <t_array>` of four ``i32`` values. |
1743 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1744 | ``i32 (i32*) *`` | A :ref:`pointer <t_pointer>` to a :ref:`function <t_function>` that takes an ``i32*``, returning an ``i32``. |
1745 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1746 | ``i32 addrspace(5)*`` | A :ref:`pointer <t_pointer>` to an ``i32`` value that resides in address space #5. |
1747 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1757 A vector type is a simple derived type that represents a vector of
1758 elements. Vector types are used when multiple primitive data are
1759 operated in parallel using a single instruction (SIMD). A vector type
1760 requires a size (number of elements) and an underlying primitive data
1761 type. Vector types are considered :ref:`first class <t_firstclass>`.
1768 < <# elements> x <elementtype> >
1770 The number of elements is a constant integer value larger than 0;
1771 elementtype may be any integer or floating point type, or a pointer to
1772 these types. Vectors of size zero are not allowed.
1777 +-------------------+--------------------------------------------------+
1778 | ``<4 x i32>`` | Vector of 4 32-bit integer values. |
1779 +-------------------+--------------------------------------------------+
1780 | ``<8 x float>`` | Vector of 8 32-bit floating-point values. |
1781 +-------------------+--------------------------------------------------+
1782 | ``<2 x i64>`` | Vector of 2 64-bit integer values. |
1783 +-------------------+--------------------------------------------------+
1784 | ``<4 x i64*>`` | Vector of 4 pointers to 64-bit integer values. |
1785 +-------------------+--------------------------------------------------+
1790 LLVM has several different basic types of constants. This section
1791 describes them all and their syntax.
1796 **Boolean constants**
1797 The two strings '``true``' and '``false``' are both valid constants
1799 **Integer constants**
1800 Standard integers (such as '4') are constants of the
1801 :ref:`integer <t_integer>` type. Negative numbers may be used with
1803 **Floating point constants**
1804 Floating point constants use standard decimal notation (e.g.
1805 123.421), exponential notation (e.g. 1.23421e+2), or a more precise
1806 hexadecimal notation (see below). The assembler requires the exact
1807 decimal value of a floating-point constant. For example, the
1808 assembler accepts 1.25 but rejects 1.3 because 1.3 is a repeating
1809 decimal in binary. Floating point constants must have a :ref:`floating
1810 point <t_floating>` type.
1811 **Null pointer constants**
1812 The identifier '``null``' is recognized as a null pointer constant
1813 and must be of :ref:`pointer type <t_pointer>`.
1815 The one non-intuitive notation for constants is the hexadecimal form of
1816 floating point constants. For example, the form
1817 '``double 0x432ff973cafa8000``' is equivalent to (but harder to read
1818 than) '``double 4.5e+15``'. The only time hexadecimal floating point
1819 constants are required (and the only time that they are generated by the
1820 disassembler) is when a floating point constant must be emitted but it
1821 cannot be represented as a decimal floating point number in a reasonable
1822 number of digits. For example, NaN's, infinities, and other special
1823 values are represented in their IEEE hexadecimal format so that assembly
1824 and disassembly do not cause any bits to change in the constants.
1826 When using the hexadecimal form, constants of types half, float, and
1827 double are represented using the 16-digit form shown above (which
1828 matches the IEEE754 representation for double); half and float values
1829 must, however, be exactly representable as IEEE 754 half and single
1830 precision, respectively. Hexadecimal format is always used for long
1831 double, and there are three forms of long double. The 80-bit format used
1832 by x86 is represented as ``0xK`` followed by 20 hexadecimal digits. The
1833 128-bit format used by PowerPC (two adjacent doubles) is represented by
1834 ``0xM`` followed by 32 hexadecimal digits. The IEEE 128-bit format is
1835 represented by ``0xL`` followed by 32 hexadecimal digits; no currently
1836 supported target uses this format. Long doubles will only work if they
1837 match the long double format on your target. The IEEE 16-bit format
1838 (half precision) is represented by ``0xH`` followed by 4 hexadecimal
1839 digits. All hexadecimal formats are big-endian (sign bit at the left).
1841 There are no constants of type x86mmx.
1846 Complex constants are a (potentially recursive) combination of simple
1847 constants and smaller complex constants.
1849 **Structure constants**
1850 Structure constants are represented with notation similar to
1851 structure type definitions (a comma separated list of elements,
1852 surrounded by braces (``{}``)). For example:
1853 "``{ i32 4, float 17.0, i32* @G }``", where "``@G``" is declared as
1854 "``@G = external global i32``". Structure constants must have
1855 :ref:`structure type <t_struct>`, and the number and types of elements
1856 must match those specified by the type.
1858 Array constants are represented with notation similar to array type
1859 definitions (a comma separated list of elements, surrounded by
1860 square brackets (``[]``)). For example:
1861 "``[ i32 42, i32 11, i32 74 ]``". Array constants must have
1862 :ref:`array type <t_array>`, and the number and types of elements must
1863 match those specified by the type.
1864 **Vector constants**
1865 Vector constants are represented with notation similar to vector
1866 type definitions (a comma separated list of elements, surrounded by
1867 less-than/greater-than's (``<>``)). For example:
1868 "``< i32 42, i32 11, i32 74, i32 100 >``". Vector constants
1869 must have :ref:`vector type <t_vector>`, and the number and types of
1870 elements must match those specified by the type.
1871 **Zero initialization**
1872 The string '``zeroinitializer``' can be used to zero initialize a
1873 value to zero of *any* type, including scalar and
1874 :ref:`aggregate <t_aggregate>` types. This is often used to avoid
1875 having to print large zero initializers (e.g. for large arrays) and
1876 is always exactly equivalent to using explicit zero initializers.
1878 A metadata node is a structure-like constant with :ref:`metadata
1879 type <t_metadata>`. For example:
1880 "``metadata !{ i32 0, metadata !"test" }``". Unlike other
1881 constants that are meant to be interpreted as part of the
1882 instruction stream, metadata is a place to attach additional
1883 information such as debug info.
1885 Global Variable and Function Addresses
1886 --------------------------------------
1888 The addresses of :ref:`global variables <globalvars>` and
1889 :ref:`functions <functionstructure>` are always implicitly valid
1890 (link-time) constants. These constants are explicitly referenced when
1891 the :ref:`identifier for the global <identifiers>` is used and always have
1892 :ref:`pointer <t_pointer>` type. For example, the following is a legal LLVM
1895 .. code-block:: llvm
1899 @Z = global [2 x i32*] [ i32* @X, i32* @Y ]
1906 The string '``undef``' can be used anywhere a constant is expected, and
1907 indicates that the user of the value may receive an unspecified
1908 bit-pattern. Undefined values may be of any type (other than '``label``'
1909 or '``void``') and be used anywhere a constant is permitted.
1911 Undefined values are useful because they indicate to the compiler that
1912 the program is well defined no matter what value is used. This gives the
1913 compiler more freedom to optimize. Here are some examples of
1914 (potentially surprising) transformations that are valid (in pseudo IR):
1916 .. code-block:: llvm
1926 This is safe because all of the output bits are affected by the undef
1927 bits. Any output bit can have a zero or one depending on the input bits.
1929 .. code-block:: llvm
1940 These logical operations have bits that are not always affected by the
1941 input. For example, if ``%X`` has a zero bit, then the output of the
1942 '``and``' operation will always be a zero for that bit, no matter what
1943 the corresponding bit from the '``undef``' is. As such, it is unsafe to
1944 optimize or assume that the result of the '``and``' is '``undef``'.
1945 However, it is safe to assume that all bits of the '``undef``' could be
1946 0, and optimize the '``and``' to 0. Likewise, it is safe to assume that
1947 all the bits of the '``undef``' operand to the '``or``' could be set,
1948 allowing the '``or``' to be folded to -1.
1950 .. code-block:: llvm
1952 %A = select undef, %X, %Y
1953 %B = select undef, 42, %Y
1954 %C = select %X, %Y, undef
1964 This set of examples shows that undefined '``select``' (and conditional
1965 branch) conditions can go *either way*, but they have to come from one
1966 of the two operands. In the ``%A`` example, if ``%X`` and ``%Y`` were
1967 both known to have a clear low bit, then ``%A`` would have to have a
1968 cleared low bit. However, in the ``%C`` example, the optimizer is
1969 allowed to assume that the '``undef``' operand could be the same as
1970 ``%Y``, allowing the whole '``select``' to be eliminated.
1972 .. code-block:: llvm
1974 %A = xor undef, undef
1991 This example points out that two '``undef``' operands are not
1992 necessarily the same. This can be surprising to people (and also matches
1993 C semantics) where they assume that "``X^X``" is always zero, even if
1994 ``X`` is undefined. This isn't true for a number of reasons, but the
1995 short answer is that an '``undef``' "variable" can arbitrarily change
1996 its value over its "live range". This is true because the variable
1997 doesn't actually *have a live range*. Instead, the value is logically
1998 read from arbitrary registers that happen to be around when needed, so
1999 the value is not necessarily consistent over time. In fact, ``%A`` and
2000 ``%C`` need to have the same semantics or the core LLVM "replace all
2001 uses with" concept would not hold.
2003 .. code-block:: llvm
2011 These examples show the crucial difference between an *undefined value*
2012 and *undefined behavior*. An undefined value (like '``undef``') is
2013 allowed to have an arbitrary bit-pattern. This means that the ``%A``
2014 operation can be constant folded to '``undef``', because the '``undef``'
2015 could be an SNaN, and ``fdiv`` is not (currently) defined on SNaN's.
2016 However, in the second example, we can make a more aggressive
2017 assumption: because the ``undef`` is allowed to be an arbitrary value,
2018 we are allowed to assume that it could be zero. Since a divide by zero
2019 has *undefined behavior*, we are allowed to assume that the operation
2020 does not execute at all. This allows us to delete the divide and all
2021 code after it. Because the undefined operation "can't happen", the
2022 optimizer can assume that it occurs in dead code.
2024 .. code-block:: llvm
2026 a: store undef -> %X
2027 b: store %X -> undef
2032 These examples reiterate the ``fdiv`` example: a store *of* an undefined
2033 value can be assumed to not have any effect; we can assume that the
2034 value is overwritten with bits that happen to match what was already
2035 there. However, a store *to* an undefined location could clobber
2036 arbitrary memory, therefore, it has undefined behavior.
2043 Poison values are similar to :ref:`undef values <undefvalues>`, however
2044 they also represent the fact that an instruction or constant expression
2045 which cannot evoke side effects has nevertheless detected a condition
2046 which results in undefined behavior.
2048 There is currently no way of representing a poison value in the IR; they
2049 only exist when produced by operations such as :ref:`add <i_add>` with
2052 Poison value behavior is defined in terms of value *dependence*:
2054 - Values other than :ref:`phi <i_phi>` nodes depend on their operands.
2055 - :ref:`Phi <i_phi>` nodes depend on the operand corresponding to
2056 their dynamic predecessor basic block.
2057 - Function arguments depend on the corresponding actual argument values
2058 in the dynamic callers of their functions.
2059 - :ref:`Call <i_call>` instructions depend on the :ref:`ret <i_ret>`
2060 instructions that dynamically transfer control back to them.
2061 - :ref:`Invoke <i_invoke>` instructions depend on the
2062 :ref:`ret <i_ret>`, :ref:`resume <i_resume>`, or exception-throwing
2063 call instructions that dynamically transfer control back to them.
2064 - Non-volatile loads and stores depend on the most recent stores to all
2065 of the referenced memory addresses, following the order in the IR
2066 (including loads and stores implied by intrinsics such as
2067 :ref:`@llvm.memcpy <int_memcpy>`.)
2068 - An instruction with externally visible side effects depends on the
2069 most recent preceding instruction with externally visible side
2070 effects, following the order in the IR. (This includes :ref:`volatile
2071 operations <volatile>`.)
2072 - An instruction *control-depends* on a :ref:`terminator
2073 instruction <terminators>` if the terminator instruction has
2074 multiple successors and the instruction is always executed when
2075 control transfers to one of the successors, and may not be executed
2076 when control is transferred to another.
2077 - Additionally, an instruction also *control-depends* on a terminator
2078 instruction if the set of instructions it otherwise depends on would
2079 be different if the terminator had transferred control to a different
2081 - Dependence is transitive.
2083 Poison Values have the same behavior as :ref:`undef values <undefvalues>`,
2084 with the additional affect that any instruction which has a *dependence*
2085 on a poison value has undefined behavior.
2087 Here are some examples:
2089 .. code-block:: llvm
2092 %poison = sub nuw i32 0, 1 ; Results in a poison value.
2093 %still_poison = and i32 %poison, 0 ; 0, but also poison.
2094 %poison_yet_again = getelementptr i32* @h, i32 %still_poison
2095 store i32 0, i32* %poison_yet_again ; memory at @h[0] is poisoned
2097 store i32 %poison, i32* @g ; Poison value stored to memory.
2098 %poison2 = load i32* @g ; Poison value loaded back from memory.
2100 store volatile i32 %poison, i32* @g ; External observation; undefined behavior.
2102 %narrowaddr = bitcast i32* @g to i16*
2103 %wideaddr = bitcast i32* @g to i64*
2104 %poison3 = load i16* %narrowaddr ; Returns a poison value.
2105 %poison4 = load i64* %wideaddr ; Returns a poison value.
2107 %cmp = icmp slt i32 %poison, 0 ; Returns a poison value.
2108 br i1 %cmp, label %true, label %end ; Branch to either destination.
2111 store volatile i32 0, i32* @g ; This is control-dependent on %cmp, so
2112 ; it has undefined behavior.
2116 %p = phi i32 [ 0, %entry ], [ 1, %true ]
2117 ; Both edges into this PHI are
2118 ; control-dependent on %cmp, so this
2119 ; always results in a poison value.
2121 store volatile i32 0, i32* @g ; This would depend on the store in %true
2122 ; if %cmp is true, or the store in %entry
2123 ; otherwise, so this is undefined behavior.
2125 br i1 %cmp, label %second_true, label %second_end
2126 ; The same branch again, but this time the
2127 ; true block doesn't have side effects.
2134 store volatile i32 0, i32* @g ; This time, the instruction always depends
2135 ; on the store in %end. Also, it is
2136 ; control-equivalent to %end, so this is
2137 ; well-defined (ignoring earlier undefined
2138 ; behavior in this example).
2142 Addresses of Basic Blocks
2143 -------------------------
2145 ``blockaddress(@function, %block)``
2147 The '``blockaddress``' constant computes the address of the specified
2148 basic block in the specified function, and always has an ``i8*`` type.
2149 Taking the address of the entry block is illegal.
2151 This value only has defined behavior when used as an operand to the
2152 ':ref:`indirectbr <i_indirectbr>`' instruction, or for comparisons
2153 against null. Pointer equality tests between labels addresses results in
2154 undefined behavior --- though, again, comparison against null is ok, and
2155 no label is equal to the null pointer. This may be passed around as an
2156 opaque pointer sized value as long as the bits are not inspected. This
2157 allows ``ptrtoint`` and arithmetic to be performed on these values so
2158 long as the original value is reconstituted before the ``indirectbr``
2161 Finally, some targets may provide defined semantics when using the value
2162 as the operand to an inline assembly, but that is target specific.
2164 Constant Expressions
2165 --------------------
2167 Constant expressions are used to allow expressions involving other
2168 constants to be used as constants. Constant expressions may be of any
2169 :ref:`first class <t_firstclass>` type and may involve any LLVM operation
2170 that does not have side effects (e.g. load and call are not supported).
2171 The following is the syntax for constant expressions:
2173 ``trunc (CST to TYPE)``
2174 Truncate a constant to another type. The bit size of CST must be
2175 larger than the bit size of TYPE. Both types must be integers.
2176 ``zext (CST to TYPE)``
2177 Zero extend a constant to another type. The bit size of CST must be
2178 smaller than the bit size of TYPE. Both types must be integers.
2179 ``sext (CST to TYPE)``
2180 Sign extend a constant to another type. The bit size of CST must be
2181 smaller than the bit size of TYPE. Both types must be integers.
2182 ``fptrunc (CST to TYPE)``
2183 Truncate a floating point constant to another floating point type.
2184 The size of CST must be larger than the size of TYPE. Both types
2185 must be floating point.
2186 ``fpext (CST to TYPE)``
2187 Floating point extend a constant to another type. The size of CST
2188 must be smaller or equal to the size of TYPE. Both types must be
2190 ``fptoui (CST to TYPE)``
2191 Convert a floating point constant to the corresponding unsigned
2192 integer constant. TYPE must be a scalar or vector integer type. CST
2193 must be of scalar or vector floating point type. Both CST and TYPE
2194 must be scalars, or vectors of the same number of elements. If the
2195 value won't fit in the integer type, the results are undefined.
2196 ``fptosi (CST to TYPE)``
2197 Convert a floating point constant to the corresponding signed
2198 integer constant. TYPE must be a scalar or vector integer type. CST
2199 must be of scalar or vector floating point type. Both CST and TYPE
2200 must be scalars, or vectors of the same number of elements. If the
2201 value won't fit in the integer type, the results are undefined.
2202 ``uitofp (CST to TYPE)``
2203 Convert an unsigned integer constant to the corresponding floating
2204 point constant. TYPE must be a scalar or vector floating point type.
2205 CST must be of scalar or vector integer type. Both CST and TYPE must
2206 be scalars, or vectors of the same number of elements. If the value
2207 won't fit in the floating point type, the results are undefined.
2208 ``sitofp (CST to TYPE)``
2209 Convert a signed integer constant to the corresponding floating
2210 point constant. TYPE must be a scalar or vector floating point type.
2211 CST must be of scalar or vector integer type. Both CST and TYPE must
2212 be scalars, or vectors of the same number of elements. If the value
2213 won't fit in the floating point type, the results are undefined.
2214 ``ptrtoint (CST to TYPE)``
2215 Convert a pointer typed constant to the corresponding integer
2216 constant ``TYPE`` must be an integer type. ``CST`` must be of
2217 pointer type. The ``CST`` value is zero extended, truncated, or
2218 unchanged to make it fit in ``TYPE``.
2219 ``inttoptr (CST to TYPE)``
2220 Convert an integer constant to a pointer constant. TYPE must be a
2221 pointer type. CST must be of integer type. The CST value is zero
2222 extended, truncated, or unchanged to make it fit in a pointer size.
2223 This one is *really* dangerous!
2224 ``bitcast (CST to TYPE)``
2225 Convert a constant, CST, to another TYPE. The constraints of the
2226 operands are the same as those for the :ref:`bitcast
2227 instruction <i_bitcast>`.
2228 ``getelementptr (CSTPTR, IDX0, IDX1, ...)``, ``getelementptr inbounds (CSTPTR, IDX0, IDX1, ...)``
2229 Perform the :ref:`getelementptr operation <i_getelementptr>` on
2230 constants. As with the :ref:`getelementptr <i_getelementptr>`
2231 instruction, the index list may have zero or more indexes, which are
2232 required to make sense for the type of "CSTPTR".
2233 ``select (COND, VAL1, VAL2)``
2234 Perform the :ref:`select operation <i_select>` on constants.
2235 ``icmp COND (VAL1, VAL2)``
2236 Performs the :ref:`icmp operation <i_icmp>` on constants.
2237 ``fcmp COND (VAL1, VAL2)``
2238 Performs the :ref:`fcmp operation <i_fcmp>` on constants.
2239 ``extractelement (VAL, IDX)``
2240 Perform the :ref:`extractelement operation <i_extractelement>` on
2242 ``insertelement (VAL, ELT, IDX)``
2243 Perform the :ref:`insertelement operation <i_insertelement>` on
2245 ``shufflevector (VEC1, VEC2, IDXMASK)``
2246 Perform the :ref:`shufflevector operation <i_shufflevector>` on
2248 ``extractvalue (VAL, IDX0, IDX1, ...)``
2249 Perform the :ref:`extractvalue operation <i_extractvalue>` on
2250 constants. The index list is interpreted in a similar manner as
2251 indices in a ':ref:`getelementptr <i_getelementptr>`' operation. At
2252 least one index value must be specified.
2253 ``insertvalue (VAL, ELT, IDX0, IDX1, ...)``
2254 Perform the :ref:`insertvalue operation <i_insertvalue>` on constants.
2255 The index list is interpreted in a similar manner as indices in a
2256 ':ref:`getelementptr <i_getelementptr>`' operation. At least one index
2257 value must be specified.
2258 ``OPCODE (LHS, RHS)``
2259 Perform the specified operation of the LHS and RHS constants. OPCODE
2260 may be any of the :ref:`binary <binaryops>` or :ref:`bitwise
2261 binary <bitwiseops>` operations. The constraints on operands are
2262 the same as those for the corresponding instruction (e.g. no bitwise
2263 operations on floating point values are allowed).
2268 Inline Assembler Expressions
2269 ----------------------------
2271 LLVM supports inline assembler expressions (as opposed to :ref:`Module-Level
2272 Inline Assembly <moduleasm>`) through the use of a special value. This
2273 value represents the inline assembler as a string (containing the
2274 instructions to emit), a list of operand constraints (stored as a
2275 string), a flag that indicates whether or not the inline asm expression
2276 has side effects, and a flag indicating whether the function containing
2277 the asm needs to align its stack conservatively. An example inline
2278 assembler expression is:
2280 .. code-block:: llvm
2282 i32 (i32) asm "bswap $0", "=r,r"
2284 Inline assembler expressions may **only** be used as the callee operand
2285 of a :ref:`call <i_call>` or an :ref:`invoke <i_invoke>` instruction.
2286 Thus, typically we have:
2288 .. code-block:: llvm
2290 %X = call i32 asm "bswap $0", "=r,r"(i32 %Y)
2292 Inline asms with side effects not visible in the constraint list must be
2293 marked as having side effects. This is done through the use of the
2294 '``sideeffect``' keyword, like so:
2296 .. code-block:: llvm
2298 call void asm sideeffect "eieio", ""()
2300 In some cases inline asms will contain code that will not work unless
2301 the stack is aligned in some way, such as calls or SSE instructions on
2302 x86, yet will not contain code that does that alignment within the asm.
2303 The compiler should make conservative assumptions about what the asm
2304 might contain and should generate its usual stack alignment code in the
2305 prologue if the '``alignstack``' keyword is present:
2307 .. code-block:: llvm
2309 call void asm alignstack "eieio", ""()
2311 Inline asms also support using non-standard assembly dialects. The
2312 assumed dialect is ATT. When the '``inteldialect``' keyword is present,
2313 the inline asm is using the Intel dialect. Currently, ATT and Intel are
2314 the only supported dialects. An example is:
2316 .. code-block:: llvm
2318 call void asm inteldialect "eieio", ""()
2320 If multiple keywords appear the '``sideeffect``' keyword must come
2321 first, the '``alignstack``' keyword second and the '``inteldialect``'
2327 The call instructions that wrap inline asm nodes may have a
2328 "``!srcloc``" MDNode attached to it that contains a list of constant
2329 integers. If present, the code generator will use the integer as the
2330 location cookie value when report errors through the ``LLVMContext``
2331 error reporting mechanisms. This allows a front-end to correlate backend
2332 errors that occur with inline asm back to the source code that produced
2335 .. code-block:: llvm
2337 call void asm sideeffect "something bad", ""(), !srcloc !42
2339 !42 = !{ i32 1234567 }
2341 It is up to the front-end to make sense of the magic numbers it places
2342 in the IR. If the MDNode contains multiple constants, the code generator
2343 will use the one that corresponds to the line of the asm that the error
2348 Metadata Nodes and Metadata Strings
2349 -----------------------------------
2351 LLVM IR allows metadata to be attached to instructions in the program
2352 that can convey extra information about the code to the optimizers and
2353 code generator. One example application of metadata is source-level
2354 debug information. There are two metadata primitives: strings and nodes.
2355 All metadata has the ``metadata`` type and is identified in syntax by a
2356 preceding exclamation point ('``!``').
2358 A metadata string is a string surrounded by double quotes. It can
2359 contain any character by escaping non-printable characters with
2360 "``\xx``" where "``xx``" is the two digit hex code. For example:
2363 Metadata nodes are represented with notation similar to structure
2364 constants (a comma separated list of elements, surrounded by braces and
2365 preceded by an exclamation point). Metadata nodes can have any values as
2366 their operand. For example:
2368 .. code-block:: llvm
2370 !{ metadata !"test\00", i32 10}
2372 A :ref:`named metadata <namedmetadatastructure>` is a collection of
2373 metadata nodes, which can be looked up in the module symbol table. For
2376 .. code-block:: llvm
2378 !foo = metadata !{!4, !3}
2380 Metadata can be used as function arguments. Here ``llvm.dbg.value``
2381 function is using two metadata arguments:
2383 .. code-block:: llvm
2385 call void @llvm.dbg.value(metadata !24, i64 0, metadata !25)
2387 Metadata can be attached with an instruction. Here metadata ``!21`` is
2388 attached to the ``add`` instruction using the ``!dbg`` identifier:
2390 .. code-block:: llvm
2392 %indvar.next = add i64 %indvar, 1, !dbg !21
2394 More information about specific metadata nodes recognized by the
2395 optimizers and code generator is found below.
2400 In LLVM IR, memory does not have types, so LLVM's own type system is not
2401 suitable for doing TBAA. Instead, metadata is added to the IR to
2402 describe a type system of a higher level language. This can be used to
2403 implement typical C/C++ TBAA, but it can also be used to implement
2404 custom alias analysis behavior for other languages.
2406 The current metadata format is very simple. TBAA metadata nodes have up
2407 to three fields, e.g.:
2409 .. code-block:: llvm
2411 !0 = metadata !{ metadata !"an example type tree" }
2412 !1 = metadata !{ metadata !"int", metadata !0 }
2413 !2 = metadata !{ metadata !"float", metadata !0 }
2414 !3 = metadata !{ metadata !"const float", metadata !2, i64 1 }
2416 The first field is an identity field. It can be any value, usually a
2417 metadata string, which uniquely identifies the type. The most important
2418 name in the tree is the name of the root node. Two trees with different
2419 root node names are entirely disjoint, even if they have leaves with
2422 The second field identifies the type's parent node in the tree, or is
2423 null or omitted for a root node. A type is considered to alias all of
2424 its descendants and all of its ancestors in the tree. Also, a type is
2425 considered to alias all types in other trees, so that bitcode produced
2426 from multiple front-ends is handled conservatively.
2428 If the third field is present, it's an integer which if equal to 1
2429 indicates that the type is "constant" (meaning
2430 ``pointsToConstantMemory`` should return true; see `other useful
2431 AliasAnalysis methods <AliasAnalysis.html#OtherItfs>`_).
2433 '``tbaa.struct``' Metadata
2434 ^^^^^^^^^^^^^^^^^^^^^^^^^^
2436 The :ref:`llvm.memcpy <int_memcpy>` is often used to implement
2437 aggregate assignment operations in C and similar languages, however it
2438 is defined to copy a contiguous region of memory, which is more than
2439 strictly necessary for aggregate types which contain holes due to
2440 padding. Also, it doesn't contain any TBAA information about the fields
2443 ``!tbaa.struct`` metadata can describe which memory subregions in a
2444 memcpy are padding and what the TBAA tags of the struct are.
2446 The current metadata format is very simple. ``!tbaa.struct`` metadata
2447 nodes are a list of operands which are in conceptual groups of three.
2448 For each group of three, the first operand gives the byte offset of a
2449 field in bytes, the second gives its size in bytes, and the third gives
2452 .. code-block:: llvm
2454 !4 = metadata !{ i64 0, i64 4, metadata !1, i64 8, i64 4, metadata !2 }
2456 This describes a struct with two fields. The first is at offset 0 bytes
2457 with size 4 bytes, and has tbaa tag !1. The second is at offset 8 bytes
2458 and has size 4 bytes and has tbaa tag !2.
2460 Note that the fields need not be contiguous. In this example, there is a
2461 4 byte gap between the two fields. This gap represents padding which
2462 does not carry useful data and need not be preserved.
2464 '``fpmath``' Metadata
2465 ^^^^^^^^^^^^^^^^^^^^^
2467 ``fpmath`` metadata may be attached to any instruction of floating point
2468 type. It can be used to express the maximum acceptable error in the
2469 result of that instruction, in ULPs, thus potentially allowing the
2470 compiler to use a more efficient but less accurate method of computing
2471 it. ULP is defined as follows:
2473 If ``x`` is a real number that lies between two finite consecutive
2474 floating-point numbers ``a`` and ``b``, without being equal to one
2475 of them, then ``ulp(x) = |b - a|``, otherwise ``ulp(x)`` is the
2476 distance between the two non-equal finite floating-point numbers
2477 nearest ``x``. Moreover, ``ulp(NaN)`` is ``NaN``.
2479 The metadata node shall consist of a single positive floating point
2480 number representing the maximum relative error, for example:
2482 .. code-block:: llvm
2484 !0 = metadata !{ float 2.5 } ; maximum acceptable inaccuracy is 2.5 ULPs
2486 '``range``' Metadata
2487 ^^^^^^^^^^^^^^^^^^^^
2489 ``range`` metadata may be attached only to loads of integer types. It
2490 expresses the possible ranges the loaded value is in. The ranges are
2491 represented with a flattened list of integers. The loaded value is known
2492 to be in the union of the ranges defined by each consecutive pair. Each
2493 pair has the following properties:
2495 - The type must match the type loaded by the instruction.
2496 - The pair ``a,b`` represents the range ``[a,b)``.
2497 - Both ``a`` and ``b`` are constants.
2498 - The range is allowed to wrap.
2499 - The range should not represent the full or empty set. That is,
2502 In addition, the pairs must be in signed order of the lower bound and
2503 they must be non-contiguous.
2507 .. code-block:: llvm
2509 %a = load i8* %x, align 1, !range !0 ; Can only be 0 or 1
2510 %b = load i8* %y, align 1, !range !1 ; Can only be 255 (-1), 0 or 1
2511 %c = load i8* %z, align 1, !range !2 ; Can only be 0, 1, 3, 4 or 5
2512 %d = load i8* %z, align 1, !range !3 ; Can only be -2, -1, 3, 4 or 5
2514 !0 = metadata !{ i8 0, i8 2 }
2515 !1 = metadata !{ i8 255, i8 2 }
2516 !2 = metadata !{ i8 0, i8 2, i8 3, i8 6 }
2517 !3 = metadata !{ i8 -2, i8 0, i8 3, i8 6 }
2519 Module Flags Metadata
2520 =====================
2522 Information about the module as a whole is difficult to convey to LLVM's
2523 subsystems. The LLVM IR isn't sufficient to transmit this information.
2524 The ``llvm.module.flags`` named metadata exists in order to facilitate
2525 this. These flags are in the form of key / value pairs --- much like a
2526 dictionary --- making it easy for any subsystem who cares about a flag to
2529 The ``llvm.module.flags`` metadata contains a list of metadata triplets.
2530 Each triplet has the following form:
2532 - The first element is a *behavior* flag, which specifies the behavior
2533 when two (or more) modules are merged together, and it encounters two
2534 (or more) metadata with the same ID. The supported behaviors are
2536 - The second element is a metadata string that is a unique ID for the
2537 metadata. Each module may only have one flag entry for each unique ID (not
2538 including entries with the **Require** behavior).
2539 - The third element is the value of the flag.
2541 When two (or more) modules are merged together, the resulting
2542 ``llvm.module.flags`` metadata is the union of the modules' flags. That is, for
2543 each unique metadata ID string, there will be exactly one entry in the merged
2544 modules ``llvm.module.flags`` metadata table, and the value for that entry will
2545 be determined by the merge behavior flag, as described below. The only exception
2546 is that entries with the *Require* behavior are always preserved.
2548 The following behaviors are supported:
2559 Emits an error if two values disagree, otherwise the resulting value
2560 is that of the operands.
2564 Emits a warning if two values disagree. The result value will be the
2565 operand for the flag from the first module being linked.
2569 Adds a requirement that another module flag be present and have a
2570 specified value after linking is performed. The value must be a
2571 metadata pair, where the first element of the pair is the ID of the
2572 module flag to be restricted, and the second element of the pair is
2573 the value the module flag should be restricted to. This behavior can
2574 be used to restrict the allowable results (via triggering of an
2575 error) of linking IDs with the **Override** behavior.
2579 Uses the specified value, regardless of the behavior or value of the
2580 other module. If both modules specify **Override**, but the values
2581 differ, an error will be emitted.
2585 Appends the two values, which are required to be metadata nodes.
2589 Appends the two values, which are required to be metadata
2590 nodes. However, duplicate entries in the second list are dropped
2591 during the append operation.
2593 It is an error for a particular unique flag ID to have multiple behaviors,
2594 except in the case of **Require** (which adds restrictions on another metadata
2595 value) or **Override**.
2597 An example of module flags:
2599 .. code-block:: llvm
2601 !0 = metadata !{ i32 1, metadata !"foo", i32 1 }
2602 !1 = metadata !{ i32 4, metadata !"bar", i32 37 }
2603 !2 = metadata !{ i32 2, metadata !"qux", i32 42 }
2604 !3 = metadata !{ i32 3, metadata !"qux",
2606 metadata !"foo", i32 1
2609 !llvm.module.flags = !{ !0, !1, !2, !3 }
2611 - Metadata ``!0`` has the ID ``!"foo"`` and the value '1'. The behavior
2612 if two or more ``!"foo"`` flags are seen is to emit an error if their
2613 values are not equal.
2615 - Metadata ``!1`` has the ID ``!"bar"`` and the value '37'. The
2616 behavior if two or more ``!"bar"`` flags are seen is to use the value
2619 - Metadata ``!2`` has the ID ``!"qux"`` and the value '42'. The
2620 behavior if two or more ``!"qux"`` flags are seen is to emit a
2621 warning if their values are not equal.
2623 - Metadata ``!3`` has the ID ``!"qux"`` and the value:
2627 metadata !{ metadata !"foo", i32 1 }
2629 The behavior is to emit an error if the ``llvm.module.flags`` does not
2630 contain a flag with the ID ``!"foo"`` that has the value '1' after linking is
2633 Objective-C Garbage Collection Module Flags Metadata
2634 ----------------------------------------------------
2636 On the Mach-O platform, Objective-C stores metadata about garbage
2637 collection in a special section called "image info". The metadata
2638 consists of a version number and a bitmask specifying what types of
2639 garbage collection are supported (if any) by the file. If two or more
2640 modules are linked together their garbage collection metadata needs to
2641 be merged rather than appended together.
2643 The Objective-C garbage collection module flags metadata consists of the
2644 following key-value pairs:
2653 * - ``Objective-C Version``
2654 - **[Required]** --- The Objective-C ABI version. Valid values are 1 and 2.
2656 * - ``Objective-C Image Info Version``
2657 - **[Required]** --- The version of the image info section. Currently
2660 * - ``Objective-C Image Info Section``
2661 - **[Required]** --- The section to place the metadata. Valid values are
2662 ``"__OBJC, __image_info, regular"`` for Objective-C ABI version 1, and
2663 ``"__DATA,__objc_imageinfo, regular, no_dead_strip"`` for
2664 Objective-C ABI version 2.
2666 * - ``Objective-C Garbage Collection``
2667 - **[Required]** --- Specifies whether garbage collection is supported or
2668 not. Valid values are 0, for no garbage collection, and 2, for garbage
2669 collection supported.
2671 * - ``Objective-C GC Only``
2672 - **[Optional]** --- Specifies that only garbage collection is supported.
2673 If present, its value must be 6. This flag requires that the
2674 ``Objective-C Garbage Collection`` flag have the value 2.
2676 Some important flag interactions:
2678 - If a module with ``Objective-C Garbage Collection`` set to 0 is
2679 merged with a module with ``Objective-C Garbage Collection`` set to
2680 2, then the resulting module has the
2681 ``Objective-C Garbage Collection`` flag set to 0.
2682 - A module with ``Objective-C Garbage Collection`` set to 0 cannot be
2683 merged with a module with ``Objective-C GC Only`` set to 6.
2685 Automatic Linker Flags Module Flags Metadata
2686 --------------------------------------------
2688 Some targets support embedding flags to the linker inside individual object
2689 files. Typically this is used in conjunction with language extensions which
2690 allow source files to explicitly declare the libraries they depend on, and have
2691 these automatically be transmitted to the linker via object files.
2693 These flags are encoded in the IR using metadata in the module flags section,
2694 using the ``Linker Options`` key. The merge behavior for this flag is required
2695 to be ``AppendUnique``, and the value for the key is expected to be a metadata
2696 node which should be a list of other metadata nodes, each of which should be a
2697 list of metadata strings defining linker options.
2699 For example, the following metadata section specifies two separate sets of
2700 linker options, presumably to link against ``libz`` and the ``Cocoa``
2703 !0 = metadata !{ i32 6, metadata !"Linker Options",
2705 metadata !{ metadata !"-lz" },
2706 metadata !{ metadata !"-framework", metadata !"Cocoa" } } }
2707 !llvm.module.flags = !{ !0 }
2709 The metadata encoding as lists of lists of options, as opposed to a collapsed
2710 list of options, is chosen so that the IR encoding can use multiple option
2711 strings to specify e.g., a single library, while still having that specifier be
2712 preserved as an atomic element that can be recognized by a target specific
2713 assembly writer or object file emitter.
2715 Each individual option is required to be either a valid option for the target's
2716 linker, or an option that is reserved by the target specific assembly writer or
2717 object file emitter. No other aspect of these options is defined by the IR.
2719 Intrinsic Global Variables
2720 ==========================
2722 LLVM has a number of "magic" global variables that contain data that
2723 affect code generation or other IR semantics. These are documented here.
2724 All globals of this sort should have a section specified as
2725 "``llvm.metadata``". This section and all globals that start with
2726 "``llvm.``" are reserved for use by LLVM.
2728 The '``llvm.used``' Global Variable
2729 -----------------------------------
2731 The ``@llvm.used`` global is an array with i8\* element type which has
2732 :ref:`appending linkage <linkage_appending>`. This array contains a list of
2733 pointers to global variables and functions which may optionally have a
2734 pointer cast formed of bitcast or getelementptr. For example, a legal
2737 .. code-block:: llvm
2742 @llvm.used = appending global [2 x i8*] [
2744 i8* bitcast (i32* @Y to i8*)
2745 ], section "llvm.metadata"
2747 If a global variable appears in the ``@llvm.used`` list, then the
2748 compiler, assembler, and linker are required to treat the symbol as if
2749 there is a reference to the global that it cannot see. For example, if a
2750 variable has internal linkage and no references other than that from the
2751 ``@llvm.used`` list, it cannot be deleted. This is commonly used to
2752 represent references from inline asms and other things the compiler
2753 cannot "see", and corresponds to "``attribute((used))``" in GNU C.
2755 On some targets, the code generator must emit a directive to the
2756 assembler or object file to prevent the assembler and linker from
2757 molesting the symbol.
2759 The '``llvm.compiler.used``' Global Variable
2760 --------------------------------------------
2762 The ``@llvm.compiler.used`` directive is the same as the ``@llvm.used``
2763 directive, except that it only prevents the compiler from touching the
2764 symbol. On targets that support it, this allows an intelligent linker to
2765 optimize references to the symbol without being impeded as it would be
2768 This is a rare construct that should only be used in rare circumstances,
2769 and should not be exposed to source languages.
2771 The '``llvm.global_ctors``' Global Variable
2772 -------------------------------------------
2774 .. code-block:: llvm
2776 %0 = type { i32, void ()* }
2777 @llvm.global_ctors = appending global [1 x %0] [%0 { i32 65535, void ()* @ctor }]
2779 The ``@llvm.global_ctors`` array contains a list of constructor
2780 functions and associated priorities. The functions referenced by this
2781 array will be called in ascending order of priority (i.e. lowest first)
2782 when the module is loaded. The order of functions with the same priority
2785 The '``llvm.global_dtors``' Global Variable
2786 -------------------------------------------
2788 .. code-block:: llvm
2790 %0 = type { i32, void ()* }
2791 @llvm.global_dtors = appending global [1 x %0] [%0 { i32 65535, void ()* @dtor }]
2793 The ``@llvm.global_dtors`` array contains a list of destructor functions
2794 and associated priorities. The functions referenced by this array will
2795 be called in descending order of priority (i.e. highest first) when the
2796 module is loaded. The order of functions with the same priority is not
2799 Instruction Reference
2800 =====================
2802 The LLVM instruction set consists of several different classifications
2803 of instructions: :ref:`terminator instructions <terminators>`, :ref:`binary
2804 instructions <binaryops>`, :ref:`bitwise binary
2805 instructions <bitwiseops>`, :ref:`memory instructions <memoryops>`, and
2806 :ref:`other instructions <otherops>`.
2810 Terminator Instructions
2811 -----------------------
2813 As mentioned :ref:`previously <functionstructure>`, every basic block in a
2814 program ends with a "Terminator" instruction, which indicates which
2815 block should be executed after the current block is finished. These
2816 terminator instructions typically yield a '``void``' value: they produce
2817 control flow, not values (the one exception being the
2818 ':ref:`invoke <i_invoke>`' instruction).
2820 The terminator instructions are: ':ref:`ret <i_ret>`',
2821 ':ref:`br <i_br>`', ':ref:`switch <i_switch>`',
2822 ':ref:`indirectbr <i_indirectbr>`', ':ref:`invoke <i_invoke>`',
2823 ':ref:`resume <i_resume>`', and ':ref:`unreachable <i_unreachable>`'.
2827 '``ret``' Instruction
2828 ^^^^^^^^^^^^^^^^^^^^^
2835 ret <type> <value> ; Return a value from a non-void function
2836 ret void ; Return from void function
2841 The '``ret``' instruction is used to return control flow (and optionally
2842 a value) from a function back to the caller.
2844 There are two forms of the '``ret``' instruction: one that returns a
2845 value and then causes control flow, and one that just causes control
2851 The '``ret``' instruction optionally accepts a single argument, the
2852 return value. The type of the return value must be a ':ref:`first
2853 class <t_firstclass>`' type.
2855 A function is not :ref:`well formed <wellformed>` if it it has a non-void
2856 return type and contains a '``ret``' instruction with no return value or
2857 a return value with a type that does not match its type, or if it has a
2858 void return type and contains a '``ret``' instruction with a return
2864 When the '``ret``' instruction is executed, control flow returns back to
2865 the calling function's context. If the caller is a
2866 ":ref:`call <i_call>`" instruction, execution continues at the
2867 instruction after the call. If the caller was an
2868 ":ref:`invoke <i_invoke>`" instruction, execution continues at the
2869 beginning of the "normal" destination block. If the instruction returns
2870 a value, that value shall set the call or invoke instruction's return
2876 .. code-block:: llvm
2878 ret i32 5 ; Return an integer value of 5
2879 ret void ; Return from a void function
2880 ret { i32, i8 } { i32 4, i8 2 } ; Return a struct of values 4 and 2
2884 '``br``' Instruction
2885 ^^^^^^^^^^^^^^^^^^^^
2892 br i1 <cond>, label <iftrue>, label <iffalse>
2893 br label <dest> ; Unconditional branch
2898 The '``br``' instruction is used to cause control flow to transfer to a
2899 different basic block in the current function. There are two forms of
2900 this instruction, corresponding to a conditional branch and an
2901 unconditional branch.
2906 The conditional branch form of the '``br``' instruction takes a single
2907 '``i1``' value and two '``label``' values. The unconditional form of the
2908 '``br``' instruction takes a single '``label``' value as a target.
2913 Upon execution of a conditional '``br``' instruction, the '``i1``'
2914 argument is evaluated. If the value is ``true``, control flows to the
2915 '``iftrue``' ``label`` argument. If "cond" is ``false``, control flows
2916 to the '``iffalse``' ``label`` argument.
2921 .. code-block:: llvm
2924 %cond = icmp eq i32 %a, %b
2925 br i1 %cond, label %IfEqual, label %IfUnequal
2933 '``switch``' Instruction
2934 ^^^^^^^^^^^^^^^^^^^^^^^^
2941 switch <intty> <value>, label <defaultdest> [ <intty> <val>, label <dest> ... ]
2946 The '``switch``' instruction is used to transfer control flow to one of
2947 several different places. It is a generalization of the '``br``'
2948 instruction, allowing a branch to occur to one of many possible
2954 The '``switch``' instruction uses three parameters: an integer
2955 comparison value '``value``', a default '``label``' destination, and an
2956 array of pairs of comparison value constants and '``label``'s. The table
2957 is not allowed to contain duplicate constant entries.
2962 The ``switch`` instruction specifies a table of values and destinations.
2963 When the '``switch``' instruction is executed, this table is searched
2964 for the given value. If the value is found, control flow is transferred
2965 to the corresponding destination; otherwise, control flow is transferred
2966 to the default destination.
2971 Depending on properties of the target machine and the particular
2972 ``switch`` instruction, this instruction may be code generated in
2973 different ways. For example, it could be generated as a series of
2974 chained conditional branches or with a lookup table.
2979 .. code-block:: llvm
2981 ; Emulate a conditional br instruction
2982 %Val = zext i1 %value to i32
2983 switch i32 %Val, label %truedest [ i32 0, label %falsedest ]
2985 ; Emulate an unconditional br instruction
2986 switch i32 0, label %dest [ ]
2988 ; Implement a jump table:
2989 switch i32 %val, label %otherwise [ i32 0, label %onzero
2991 i32 2, label %ontwo ]
2995 '``indirectbr``' Instruction
2996 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3003 indirectbr <somety>* <address>, [ label <dest1>, label <dest2>, ... ]
3008 The '``indirectbr``' instruction implements an indirect branch to a
3009 label within the current function, whose address is specified by
3010 "``address``". Address must be derived from a
3011 :ref:`blockaddress <blockaddress>` constant.
3016 The '``address``' argument is the address of the label to jump to. The
3017 rest of the arguments indicate the full set of possible destinations
3018 that the address may point to. Blocks are allowed to occur multiple
3019 times in the destination list, though this isn't particularly useful.
3021 This destination list is required so that dataflow analysis has an
3022 accurate understanding of the CFG.
3027 Control transfers to the block specified in the address argument. All
3028 possible destination blocks must be listed in the label list, otherwise
3029 this instruction has undefined behavior. This implies that jumps to
3030 labels defined in other functions have undefined behavior as well.
3035 This is typically implemented with a jump through a register.
3040 .. code-block:: llvm
3042 indirectbr i8* %Addr, [ label %bb1, label %bb2, label %bb3 ]
3046 '``invoke``' Instruction
3047 ^^^^^^^^^^^^^^^^^^^^^^^^
3054 <result> = invoke [cconv] [ret attrs] <ptr to function ty> <function ptr val>(<function args>) [fn attrs]
3055 to label <normal label> unwind label <exception label>
3060 The '``invoke``' instruction causes control to transfer to a specified
3061 function, with the possibility of control flow transfer to either the
3062 '``normal``' label or the '``exception``' label. If the callee function
3063 returns with the "``ret``" instruction, control flow will return to the
3064 "normal" label. If the callee (or any indirect callees) returns via the
3065 ":ref:`resume <i_resume>`" instruction or other exception handling
3066 mechanism, control is interrupted and continued at the dynamically
3067 nearest "exception" label.
3069 The '``exception``' label is a `landing
3070 pad <ExceptionHandling.html#overview>`_ for the exception. As such,
3071 '``exception``' label is required to have the
3072 ":ref:`landingpad <i_landingpad>`" instruction, which contains the
3073 information about the behavior of the program after unwinding happens,
3074 as its first non-PHI instruction. The restrictions on the
3075 "``landingpad``" instruction's tightly couples it to the "``invoke``"
3076 instruction, so that the important information contained within the
3077 "``landingpad``" instruction can't be lost through normal code motion.
3082 This instruction requires several arguments:
3084 #. The optional "cconv" marker indicates which :ref:`calling
3085 convention <callingconv>` the call should use. If none is
3086 specified, the call defaults to using C calling conventions.
3087 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
3088 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
3090 #. '``ptr to function ty``': shall be the signature of the pointer to
3091 function value being invoked. In most cases, this is a direct
3092 function invocation, but indirect ``invoke``'s are just as possible,
3093 branching off an arbitrary pointer to function value.
3094 #. '``function ptr val``': An LLVM value containing a pointer to a
3095 function to be invoked.
3096 #. '``function args``': argument list whose types match the function
3097 signature argument types and parameter attributes. All arguments must
3098 be of :ref:`first class <t_firstclass>` type. If the function signature
3099 indicates the function accepts a variable number of arguments, the
3100 extra arguments can be specified.
3101 #. '``normal label``': the label reached when the called function
3102 executes a '``ret``' instruction.
3103 #. '``exception label``': the label reached when a callee returns via
3104 the :ref:`resume <i_resume>` instruction or other exception handling
3106 #. The optional :ref:`function attributes <fnattrs>` list. Only
3107 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
3108 attributes are valid here.
3113 This instruction is designed to operate as a standard '``call``'
3114 instruction in most regards. The primary difference is that it
3115 establishes an association with a label, which is used by the runtime
3116 library to unwind the stack.
3118 This instruction is used in languages with destructors to ensure that
3119 proper cleanup is performed in the case of either a ``longjmp`` or a
3120 thrown exception. Additionally, this is important for implementation of
3121 '``catch``' clauses in high-level languages that support them.
3123 For the purposes of the SSA form, the definition of the value returned
3124 by the '``invoke``' instruction is deemed to occur on the edge from the
3125 current block to the "normal" label. If the callee unwinds then no
3126 return value is available.
3131 .. code-block:: llvm
3133 %retval = invoke i32 @Test(i32 15) to label %Continue
3134 unwind label %TestCleanup ; {i32}:retval set
3135 %retval = invoke coldcc i32 %Testfnptr(i32 15) to label %Continue
3136 unwind label %TestCleanup ; {i32}:retval set
3140 '``resume``' Instruction
3141 ^^^^^^^^^^^^^^^^^^^^^^^^
3148 resume <type> <value>
3153 The '``resume``' instruction is a terminator instruction that has no
3159 The '``resume``' instruction requires one argument, which must have the
3160 same type as the result of any '``landingpad``' instruction in the same
3166 The '``resume``' instruction resumes propagation of an existing
3167 (in-flight) exception whose unwinding was interrupted with a
3168 :ref:`landingpad <i_landingpad>` instruction.
3173 .. code-block:: llvm
3175 resume { i8*, i32 } %exn
3179 '``unreachable``' Instruction
3180 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3192 The '``unreachable``' instruction has no defined semantics. This
3193 instruction is used to inform the optimizer that a particular portion of
3194 the code is not reachable. This can be used to indicate that the code
3195 after a no-return function cannot be reached, and other facts.
3200 The '``unreachable``' instruction has no defined semantics.
3207 Binary operators are used to do most of the computation in a program.
3208 They require two operands of the same type, execute an operation on
3209 them, and produce a single value. The operands might represent multiple
3210 data, as is the case with the :ref:`vector <t_vector>` data type. The
3211 result value has the same type as its operands.
3213 There are several different binary operators:
3217 '``add``' Instruction
3218 ^^^^^^^^^^^^^^^^^^^^^
3225 <result> = add <ty> <op1>, <op2> ; yields {ty}:result
3226 <result> = add nuw <ty> <op1>, <op2> ; yields {ty}:result
3227 <result> = add nsw <ty> <op1>, <op2> ; yields {ty}:result
3228 <result> = add nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3233 The '``add``' instruction returns the sum of its two operands.
3238 The two arguments to the '``add``' instruction must be
3239 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3240 arguments must have identical types.
3245 The value produced is the integer sum of the two operands.
3247 If the sum has unsigned overflow, the result returned is the
3248 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
3251 Because LLVM integers use a two's complement representation, this
3252 instruction is appropriate for both signed and unsigned integers.
3254 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3255 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3256 result value of the ``add`` is a :ref:`poison value <poisonvalues>` if
3257 unsigned and/or signed overflow, respectively, occurs.
3262 .. code-block:: llvm
3264 <result> = add i32 4, %var ; yields {i32}:result = 4 + %var
3268 '``fadd``' Instruction
3269 ^^^^^^^^^^^^^^^^^^^^^^
3276 <result> = fadd [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3281 The '``fadd``' instruction returns the sum of its two operands.
3286 The two arguments to the '``fadd``' instruction must be :ref:`floating
3287 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3288 Both arguments must have identical types.
3293 The value produced is the floating point sum of the two operands. This
3294 instruction can also take any number of :ref:`fast-math flags <fastmath>`,
3295 which are optimization hints to enable otherwise unsafe floating point
3301 .. code-block:: llvm
3303 <result> = fadd float 4.0, %var ; yields {float}:result = 4.0 + %var
3305 '``sub``' Instruction
3306 ^^^^^^^^^^^^^^^^^^^^^
3313 <result> = sub <ty> <op1>, <op2> ; yields {ty}:result
3314 <result> = sub nuw <ty> <op1>, <op2> ; yields {ty}:result
3315 <result> = sub nsw <ty> <op1>, <op2> ; yields {ty}:result
3316 <result> = sub nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3321 The '``sub``' instruction returns the difference of its two operands.
3323 Note that the '``sub``' instruction is used to represent the '``neg``'
3324 instruction present in most other intermediate representations.
3329 The two arguments to the '``sub``' instruction must be
3330 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3331 arguments must have identical types.
3336 The value produced is the integer difference of the two operands.
3338 If the difference has unsigned overflow, the result returned is the
3339 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
3342 Because LLVM integers use a two's complement representation, this
3343 instruction is appropriate for both signed and unsigned integers.
3345 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3346 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3347 result value of the ``sub`` is a :ref:`poison value <poisonvalues>` if
3348 unsigned and/or signed overflow, respectively, occurs.
3353 .. code-block:: llvm
3355 <result> = sub i32 4, %var ; yields {i32}:result = 4 - %var
3356 <result> = sub i32 0, %val ; yields {i32}:result = -%var
3360 '``fsub``' Instruction
3361 ^^^^^^^^^^^^^^^^^^^^^^
3368 <result> = fsub [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3373 The '``fsub``' instruction returns the difference of its two operands.
3375 Note that the '``fsub``' instruction is used to represent the '``fneg``'
3376 instruction present in most other intermediate representations.
3381 The two arguments to the '``fsub``' instruction must be :ref:`floating
3382 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3383 Both arguments must have identical types.
3388 The value produced is the floating point difference of the two operands.
3389 This instruction can also take any number of :ref:`fast-math
3390 flags <fastmath>`, which are optimization hints to enable otherwise
3391 unsafe floating point optimizations:
3396 .. code-block:: llvm
3398 <result> = fsub float 4.0, %var ; yields {float}:result = 4.0 - %var
3399 <result> = fsub float -0.0, %val ; yields {float}:result = -%var
3401 '``mul``' Instruction
3402 ^^^^^^^^^^^^^^^^^^^^^
3409 <result> = mul <ty> <op1>, <op2> ; yields {ty}:result
3410 <result> = mul nuw <ty> <op1>, <op2> ; yields {ty}:result
3411 <result> = mul nsw <ty> <op1>, <op2> ; yields {ty}:result
3412 <result> = mul nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3417 The '``mul``' instruction returns the product of its two operands.
3422 The two arguments to the '``mul``' instruction must be
3423 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3424 arguments must have identical types.
3429 The value produced is the integer product of the two operands.
3431 If the result of the multiplication has unsigned overflow, the result
3432 returned is the mathematical result modulo 2\ :sup:`n`\ , where n is the
3433 bit width of the result.
3435 Because LLVM integers use a two's complement representation, and the
3436 result is the same width as the operands, this instruction returns the
3437 correct result for both signed and unsigned integers. If a full product
3438 (e.g. ``i32`` * ``i32`` -> ``i64``) is needed, the operands should be
3439 sign-extended or zero-extended as appropriate to the width of the full
3442 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3443 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3444 result value of the ``mul`` is a :ref:`poison value <poisonvalues>` if
3445 unsigned and/or signed overflow, respectively, occurs.
3450 .. code-block:: llvm
3452 <result> = mul i32 4, %var ; yields {i32}:result = 4 * %var
3456 '``fmul``' Instruction
3457 ^^^^^^^^^^^^^^^^^^^^^^
3464 <result> = fmul [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3469 The '``fmul``' instruction returns the product of its two operands.
3474 The two arguments to the '``fmul``' instruction must be :ref:`floating
3475 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3476 Both arguments must have identical types.
3481 The value produced is the floating point product of the two operands.
3482 This instruction can also take any number of :ref:`fast-math
3483 flags <fastmath>`, which are optimization hints to enable otherwise
3484 unsafe floating point optimizations:
3489 .. code-block:: llvm
3491 <result> = fmul float 4.0, %var ; yields {float}:result = 4.0 * %var
3493 '``udiv``' Instruction
3494 ^^^^^^^^^^^^^^^^^^^^^^
3501 <result> = udiv <ty> <op1>, <op2> ; yields {ty}:result
3502 <result> = udiv exact <ty> <op1>, <op2> ; yields {ty}:result
3507 The '``udiv``' instruction returns the quotient of its two operands.
3512 The two arguments to the '``udiv``' instruction must be
3513 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3514 arguments must have identical types.
3519 The value produced is the unsigned integer quotient of the two operands.
3521 Note that unsigned integer division and signed integer division are
3522 distinct operations; for signed integer division, use '``sdiv``'.
3524 Division by zero leads to undefined behavior.
3526 If the ``exact`` keyword is present, the result value of the ``udiv`` is
3527 a :ref:`poison value <poisonvalues>` if %op1 is not a multiple of %op2 (as
3528 such, "((a udiv exact b) mul b) == a").
3533 .. code-block:: llvm
3535 <result> = udiv i32 4, %var ; yields {i32}:result = 4 / %var
3537 '``sdiv``' Instruction
3538 ^^^^^^^^^^^^^^^^^^^^^^
3545 <result> = sdiv <ty> <op1>, <op2> ; yields {ty}:result
3546 <result> = sdiv exact <ty> <op1>, <op2> ; yields {ty}:result
3551 The '``sdiv``' instruction returns the quotient of its two operands.
3556 The two arguments to the '``sdiv``' instruction must be
3557 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3558 arguments must have identical types.
3563 The value produced is the signed integer quotient of the two operands
3564 rounded towards zero.
3566 Note that signed integer division and unsigned integer division are
3567 distinct operations; for unsigned integer division, use '``udiv``'.
3569 Division by zero leads to undefined behavior. Overflow also leads to
3570 undefined behavior; this is a rare case, but can occur, for example, by
3571 doing a 32-bit division of -2147483648 by -1.
3573 If the ``exact`` keyword is present, the result value of the ``sdiv`` is
3574 a :ref:`poison value <poisonvalues>` if the result would be rounded.
3579 .. code-block:: llvm
3581 <result> = sdiv i32 4, %var ; yields {i32}:result = 4 / %var
3585 '``fdiv``' Instruction
3586 ^^^^^^^^^^^^^^^^^^^^^^
3593 <result> = fdiv [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3598 The '``fdiv``' instruction returns the quotient of its two operands.
3603 The two arguments to the '``fdiv``' instruction must be :ref:`floating
3604 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3605 Both arguments must have identical types.
3610 The value produced is the floating point quotient of the two operands.
3611 This instruction can also take any number of :ref:`fast-math
3612 flags <fastmath>`, which are optimization hints to enable otherwise
3613 unsafe floating point optimizations:
3618 .. code-block:: llvm
3620 <result> = fdiv float 4.0, %var ; yields {float}:result = 4.0 / %var
3622 '``urem``' Instruction
3623 ^^^^^^^^^^^^^^^^^^^^^^
3630 <result> = urem <ty> <op1>, <op2> ; yields {ty}:result
3635 The '``urem``' instruction returns the remainder from the unsigned
3636 division of its two arguments.
3641 The two arguments to the '``urem``' instruction must be
3642 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3643 arguments must have identical types.
3648 This instruction returns the unsigned integer *remainder* of a division.
3649 This instruction always performs an unsigned division to get the
3652 Note that unsigned integer remainder and signed integer remainder are
3653 distinct operations; for signed integer remainder, use '``srem``'.
3655 Taking the remainder of a division by zero leads to undefined behavior.
3660 .. code-block:: llvm
3662 <result> = urem i32 4, %var ; yields {i32}:result = 4 % %var
3664 '``srem``' Instruction
3665 ^^^^^^^^^^^^^^^^^^^^^^
3672 <result> = srem <ty> <op1>, <op2> ; yields {ty}:result
3677 The '``srem``' instruction returns the remainder from the signed
3678 division of its two operands. This instruction can also take
3679 :ref:`vector <t_vector>` versions of the values in which case the elements
3685 The two arguments to the '``srem``' instruction must be
3686 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3687 arguments must have identical types.
3692 This instruction returns the *remainder* of a division (where the result
3693 is either zero or has the same sign as the dividend, ``op1``), not the
3694 *modulo* operator (where the result is either zero or has the same sign
3695 as the divisor, ``op2``) of a value. For more information about the
3696 difference, see `The Math
3697 Forum <http://mathforum.org/dr.math/problems/anne.4.28.99.html>`_. For a
3698 table of how this is implemented in various languages, please see
3700 operation <http://en.wikipedia.org/wiki/Modulo_operation>`_.
3702 Note that signed integer remainder and unsigned integer remainder are
3703 distinct operations; for unsigned integer remainder, use '``urem``'.
3705 Taking the remainder of a division by zero leads to undefined behavior.
3706 Overflow also leads to undefined behavior; this is a rare case, but can
3707 occur, for example, by taking the remainder of a 32-bit division of
3708 -2147483648 by -1. (The remainder doesn't actually overflow, but this
3709 rule lets srem be implemented using instructions that return both the
3710 result of the division and the remainder.)
3715 .. code-block:: llvm
3717 <result> = srem i32 4, %var ; yields {i32}:result = 4 % %var
3721 '``frem``' Instruction
3722 ^^^^^^^^^^^^^^^^^^^^^^
3729 <result> = frem [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3734 The '``frem``' instruction returns the remainder from the division of
3740 The two arguments to the '``frem``' instruction must be :ref:`floating
3741 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3742 Both arguments must have identical types.
3747 This instruction returns the *remainder* of a division. The remainder
3748 has the same sign as the dividend. This instruction can also take any
3749 number of :ref:`fast-math flags <fastmath>`, which are optimization hints
3750 to enable otherwise unsafe floating point optimizations:
3755 .. code-block:: llvm
3757 <result> = frem float 4.0, %var ; yields {float}:result = 4.0 % %var
3761 Bitwise Binary Operations
3762 -------------------------
3764 Bitwise binary operators are used to do various forms of bit-twiddling
3765 in a program. They are generally very efficient instructions and can
3766 commonly be strength reduced from other instructions. They require two
3767 operands of the same type, execute an operation on them, and produce a
3768 single value. The resulting value is the same type as its operands.
3770 '``shl``' Instruction
3771 ^^^^^^^^^^^^^^^^^^^^^
3778 <result> = shl <ty> <op1>, <op2> ; yields {ty}:result
3779 <result> = shl nuw <ty> <op1>, <op2> ; yields {ty}:result
3780 <result> = shl nsw <ty> <op1>, <op2> ; yields {ty}:result
3781 <result> = shl nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3786 The '``shl``' instruction returns the first operand shifted to the left
3787 a specified number of bits.
3792 Both arguments to the '``shl``' instruction must be the same
3793 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
3794 '``op2``' is treated as an unsigned value.
3799 The value produced is ``op1`` \* 2\ :sup:`op2` mod 2\ :sup:`n`,
3800 where ``n`` is the width of the result. If ``op2`` is (statically or
3801 dynamically) negative or equal to or larger than the number of bits in
3802 ``op1``, the result is undefined. If the arguments are vectors, each
3803 vector element of ``op1`` is shifted by the corresponding shift amount
3806 If the ``nuw`` keyword is present, then the shift produces a :ref:`poison
3807 value <poisonvalues>` if it shifts out any non-zero bits. If the
3808 ``nsw`` keyword is present, then the shift produces a :ref:`poison
3809 value <poisonvalues>` if it shifts out any bits that disagree with the
3810 resultant sign bit. As such, NUW/NSW have the same semantics as they
3811 would if the shift were expressed as a mul instruction with the same
3812 nsw/nuw bits in (mul %op1, (shl 1, %op2)).
3817 .. code-block:: llvm
3819 <result> = shl i32 4, %var ; yields {i32}: 4 << %var
3820 <result> = shl i32 4, 2 ; yields {i32}: 16
3821 <result> = shl i32 1, 10 ; yields {i32}: 1024
3822 <result> = shl i32 1, 32 ; undefined
3823 <result> = shl <2 x i32> < i32 1, i32 1>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 2, i32 4>
3825 '``lshr``' Instruction
3826 ^^^^^^^^^^^^^^^^^^^^^^
3833 <result> = lshr <ty> <op1>, <op2> ; yields {ty}:result
3834 <result> = lshr exact <ty> <op1>, <op2> ; yields {ty}:result
3839 The '``lshr``' instruction (logical shift right) returns the first
3840 operand shifted to the right a specified number of bits with zero fill.
3845 Both arguments to the '``lshr``' instruction must be the same
3846 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
3847 '``op2``' is treated as an unsigned value.
3852 This instruction always performs a logical shift right operation. The
3853 most significant bits of the result will be filled with zero bits after
3854 the shift. If ``op2`` is (statically or dynamically) equal to or larger
3855 than the number of bits in ``op1``, the result is undefined. If the
3856 arguments are vectors, each vector element of ``op1`` is shifted by the
3857 corresponding shift amount in ``op2``.
3859 If the ``exact`` keyword is present, the result value of the ``lshr`` is
3860 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
3866 .. code-block:: llvm
3868 <result> = lshr i32 4, 1 ; yields {i32}:result = 2
3869 <result> = lshr i32 4, 2 ; yields {i32}:result = 1
3870 <result> = lshr i8 4, 3 ; yields {i8}:result = 0
3871 <result> = lshr i8 -2, 1 ; yields {i8}:result = 0x7FFFFFFF
3872 <result> = lshr i32 1, 32 ; undefined
3873 <result> = lshr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 0x7FFFFFFF, i32 1>
3875 '``ashr``' Instruction
3876 ^^^^^^^^^^^^^^^^^^^^^^
3883 <result> = ashr <ty> <op1>, <op2> ; yields {ty}:result
3884 <result> = ashr exact <ty> <op1>, <op2> ; yields {ty}:result
3889 The '``ashr``' instruction (arithmetic shift right) returns the first
3890 operand shifted to the right a specified number of bits with sign
3896 Both arguments to the '``ashr``' instruction must be the same
3897 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
3898 '``op2``' is treated as an unsigned value.
3903 This instruction always performs an arithmetic shift right operation,
3904 The most significant bits of the result will be filled with the sign bit
3905 of ``op1``. If ``op2`` is (statically or dynamically) equal to or larger
3906 than the number of bits in ``op1``, the result is undefined. If the
3907 arguments are vectors, each vector element of ``op1`` is shifted by the
3908 corresponding shift amount in ``op2``.
3910 If the ``exact`` keyword is present, the result value of the ``ashr`` is
3911 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
3917 .. code-block:: llvm
3919 <result> = ashr i32 4, 1 ; yields {i32}:result = 2
3920 <result> = ashr i32 4, 2 ; yields {i32}:result = 1
3921 <result> = ashr i8 4, 3 ; yields {i8}:result = 0
3922 <result> = ashr i8 -2, 1 ; yields {i8}:result = -1
3923 <result> = ashr i32 1, 32 ; undefined
3924 <result> = ashr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 3> ; yields: result=<2 x i32> < i32 -1, i32 0>
3926 '``and``' Instruction
3927 ^^^^^^^^^^^^^^^^^^^^^
3934 <result> = and <ty> <op1>, <op2> ; yields {ty}:result
3939 The '``and``' instruction returns the bitwise logical and of its two
3945 The two arguments to the '``and``' instruction must be
3946 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3947 arguments must have identical types.
3952 The truth table used for the '``and``' instruction is:
3969 .. code-block:: llvm
3971 <result> = and i32 4, %var ; yields {i32}:result = 4 & %var
3972 <result> = and i32 15, 40 ; yields {i32}:result = 8
3973 <result> = and i32 4, 8 ; yields {i32}:result = 0
3975 '``or``' Instruction
3976 ^^^^^^^^^^^^^^^^^^^^
3983 <result> = or <ty> <op1>, <op2> ; yields {ty}:result
3988 The '``or``' instruction returns the bitwise logical inclusive or of its
3994 The two arguments to the '``or``' instruction must be
3995 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3996 arguments must have identical types.
4001 The truth table used for the '``or``' instruction is:
4020 <result> = or i32 4, %var ; yields {i32}:result = 4 | %var
4021 <result> = or i32 15, 40 ; yields {i32}:result = 47
4022 <result> = or i32 4, 8 ; yields {i32}:result = 12
4024 '``xor``' Instruction
4025 ^^^^^^^^^^^^^^^^^^^^^
4032 <result> = xor <ty> <op1>, <op2> ; yields {ty}:result
4037 The '``xor``' instruction returns the bitwise logical exclusive or of
4038 its two operands. The ``xor`` is used to implement the "one's
4039 complement" operation, which is the "~" operator in C.
4044 The two arguments to the '``xor``' instruction must be
4045 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4046 arguments must have identical types.
4051 The truth table used for the '``xor``' instruction is:
4068 .. code-block:: llvm
4070 <result> = xor i32 4, %var ; yields {i32}:result = 4 ^ %var
4071 <result> = xor i32 15, 40 ; yields {i32}:result = 39
4072 <result> = xor i32 4, 8 ; yields {i32}:result = 12
4073 <result> = xor i32 %V, -1 ; yields {i32}:result = ~%V
4078 LLVM supports several instructions to represent vector operations in a
4079 target-independent manner. These instructions cover the element-access
4080 and vector-specific operations needed to process vectors effectively.
4081 While LLVM does directly support these vector operations, many
4082 sophisticated algorithms will want to use target-specific intrinsics to
4083 take full advantage of a specific target.
4085 .. _i_extractelement:
4087 '``extractelement``' Instruction
4088 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4095 <result> = extractelement <n x <ty>> <val>, i32 <idx> ; yields <ty>
4100 The '``extractelement``' instruction extracts a single scalar element
4101 from a vector at a specified index.
4106 The first operand of an '``extractelement``' instruction is a value of
4107 :ref:`vector <t_vector>` type. The second operand is an index indicating
4108 the position from which to extract the element. The index may be a
4114 The result is a scalar of the same type as the element type of ``val``.
4115 Its value is the value at position ``idx`` of ``val``. If ``idx``
4116 exceeds the length of ``val``, the results are undefined.
4121 .. code-block:: llvm
4123 <result> = extractelement <4 x i32> %vec, i32 0 ; yields i32
4125 .. _i_insertelement:
4127 '``insertelement``' Instruction
4128 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4135 <result> = insertelement <n x <ty>> <val>, <ty> <elt>, i32 <idx> ; yields <n x <ty>>
4140 The '``insertelement``' instruction inserts a scalar element into a
4141 vector at a specified index.
4146 The first operand of an '``insertelement``' instruction is a value of
4147 :ref:`vector <t_vector>` type. The second operand is a scalar value whose
4148 type must equal the element type of the first operand. The third operand
4149 is an index indicating the position at which to insert the value. The
4150 index may be a variable.
4155 The result is a vector of the same type as ``val``. Its element values
4156 are those of ``val`` except at position ``idx``, where it gets the value
4157 ``elt``. If ``idx`` exceeds the length of ``val``, the results are
4163 .. code-block:: llvm
4165 <result> = insertelement <4 x i32> %vec, i32 1, i32 0 ; yields <4 x i32>
4167 .. _i_shufflevector:
4169 '``shufflevector``' Instruction
4170 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4177 <result> = shufflevector <n x <ty>> <v1>, <n x <ty>> <v2>, <m x i32> <mask> ; yields <m x <ty>>
4182 The '``shufflevector``' instruction constructs a permutation of elements
4183 from two input vectors, returning a vector with the same element type as
4184 the input and length that is the same as the shuffle mask.
4189 The first two operands of a '``shufflevector``' instruction are vectors
4190 with the same type. The third argument is a shuffle mask whose element
4191 type is always 'i32'. The result of the instruction is a vector whose
4192 length is the same as the shuffle mask and whose element type is the
4193 same as the element type of the first two operands.
4195 The shuffle mask operand is required to be a constant vector with either
4196 constant integer or undef values.
4201 The elements of the two input vectors are numbered from left to right
4202 across both of the vectors. The shuffle mask operand specifies, for each
4203 element of the result vector, which element of the two input vectors the
4204 result element gets. The element selector may be undef (meaning "don't
4205 care") and the second operand may be undef if performing a shuffle from
4211 .. code-block:: llvm
4213 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
4214 <4 x i32> <i32 0, i32 4, i32 1, i32 5> ; yields <4 x i32>
4215 <result> = shufflevector <4 x i32> %v1, <4 x i32> undef,
4216 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32> - Identity shuffle.
4217 <result> = shufflevector <8 x i32> %v1, <8 x i32> undef,
4218 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32>
4219 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
4220 <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7 > ; yields <8 x i32>
4222 Aggregate Operations
4223 --------------------
4225 LLVM supports several instructions for working with
4226 :ref:`aggregate <t_aggregate>` values.
4230 '``extractvalue``' Instruction
4231 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4238 <result> = extractvalue <aggregate type> <val>, <idx>{, <idx>}*
4243 The '``extractvalue``' instruction extracts the value of a member field
4244 from an :ref:`aggregate <t_aggregate>` value.
4249 The first operand of an '``extractvalue``' instruction is a value of
4250 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The operands are
4251 constant indices to specify which value to extract in a similar manner
4252 as indices in a '``getelementptr``' instruction.
4254 The major differences to ``getelementptr`` indexing are:
4256 - Since the value being indexed is not a pointer, the first index is
4257 omitted and assumed to be zero.
4258 - At least one index must be specified.
4259 - Not only struct indices but also array indices must be in bounds.
4264 The result is the value at the position in the aggregate specified by
4270 .. code-block:: llvm
4272 <result> = extractvalue {i32, float} %agg, 0 ; yields i32
4276 '``insertvalue``' Instruction
4277 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4284 <result> = insertvalue <aggregate type> <val>, <ty> <elt>, <idx>{, <idx>}* ; yields <aggregate type>
4289 The '``insertvalue``' instruction inserts a value into a member field in
4290 an :ref:`aggregate <t_aggregate>` value.
4295 The first operand of an '``insertvalue``' instruction is a value of
4296 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The second operand is
4297 a first-class value to insert. The following operands are constant
4298 indices indicating the position at which to insert the value in a
4299 similar manner as indices in a '``extractvalue``' instruction. The value
4300 to insert must have the same type as the value identified by the
4306 The result is an aggregate of the same type as ``val``. Its value is
4307 that of ``val`` except that the value at the position specified by the
4308 indices is that of ``elt``.
4313 .. code-block:: llvm
4315 %agg1 = insertvalue {i32, float} undef, i32 1, 0 ; yields {i32 1, float undef}
4316 %agg2 = insertvalue {i32, float} %agg1, float %val, 1 ; yields {i32 1, float %val}
4317 %agg3 = insertvalue {i32, {float}} %agg1, float %val, 1, 0 ; yields {i32 1, float %val}
4321 Memory Access and Addressing Operations
4322 ---------------------------------------
4324 A key design point of an SSA-based representation is how it represents
4325 memory. In LLVM, no memory locations are in SSA form, which makes things
4326 very simple. This section describes how to read, write, and allocate
4331 '``alloca``' Instruction
4332 ^^^^^^^^^^^^^^^^^^^^^^^^
4339 <result> = alloca <type>[, <ty> <NumElements>][, align <alignment>] ; yields {type*}:result
4344 The '``alloca``' instruction allocates memory on the stack frame of the
4345 currently executing function, to be automatically released when this
4346 function returns to its caller. The object is always allocated in the
4347 generic address space (address space zero).
4352 The '``alloca``' instruction allocates ``sizeof(<type>)*NumElements``
4353 bytes of memory on the runtime stack, returning a pointer of the
4354 appropriate type to the program. If "NumElements" is specified, it is
4355 the number of elements allocated, otherwise "NumElements" is defaulted
4356 to be one. If a constant alignment is specified, the value result of the
4357 allocation is guaranteed to be aligned to at least that boundary. If not
4358 specified, or if zero, the target can choose to align the allocation on
4359 any convenient boundary compatible with the type.
4361 '``type``' may be any sized type.
4366 Memory is allocated; a pointer is returned. The operation is undefined
4367 if there is insufficient stack space for the allocation. '``alloca``'d
4368 memory is automatically released when the function returns. The
4369 '``alloca``' instruction is commonly used to represent automatic
4370 variables that must have an address available. When the function returns
4371 (either with the ``ret`` or ``resume`` instructions), the memory is
4372 reclaimed. Allocating zero bytes is legal, but the result is undefined.
4373 The order in which memory is allocated (ie., which way the stack grows)
4379 .. code-block:: llvm
4381 %ptr = alloca i32 ; yields {i32*}:ptr
4382 %ptr = alloca i32, i32 4 ; yields {i32*}:ptr
4383 %ptr = alloca i32, i32 4, align 1024 ; yields {i32*}:ptr
4384 %ptr = alloca i32, align 1024 ; yields {i32*}:ptr
4388 '``load``' Instruction
4389 ^^^^^^^^^^^^^^^^^^^^^^
4396 <result> = load [volatile] <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>][, !invariant.load !<index>]
4397 <result> = load atomic [volatile] <ty>* <pointer> [singlethread] <ordering>, align <alignment>
4398 !<index> = !{ i32 1 }
4403 The '``load``' instruction is used to read from memory.
4408 The argument to the '``load``' instruction specifies the memory address
4409 from which to load. The pointer must point to a :ref:`first
4410 class <t_firstclass>` type. If the ``load`` is marked as ``volatile``,
4411 then the optimizer is not allowed to modify the number or order of
4412 execution of this ``load`` with other :ref:`volatile
4413 operations <volatile>`.
4415 If the ``load`` is marked as ``atomic``, it takes an extra
4416 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
4417 ``release`` and ``acq_rel`` orderings are not valid on ``load``
4418 instructions. Atomic loads produce :ref:`defined <memmodel>` results
4419 when they may see multiple atomic stores. The type of the pointee must
4420 be an integer type whose bit width is a power of two greater than or
4421 equal to eight and less than or equal to a target-specific size limit.
4422 ``align`` must be explicitly specified on atomic loads, and the load has
4423 undefined behavior if the alignment is not set to a value which is at
4424 least the size in bytes of the pointee. ``!nontemporal`` does not have
4425 any defined semantics for atomic loads.
4427 The optional constant ``align`` argument specifies the alignment of the
4428 operation (that is, the alignment of the memory address). A value of 0
4429 or an omitted ``align`` argument means that the operation has the abi
4430 alignment for the target. It is the responsibility of the code emitter
4431 to ensure that the alignment information is correct. Overestimating the
4432 alignment results in undefined behavior. Underestimating the alignment
4433 may produce less efficient code. An alignment of 1 is always safe.
4435 The optional ``!nontemporal`` metadata must reference a single
4436 metatadata name <index> corresponding to a metadata node with one
4437 ``i32`` entry of value 1. The existence of the ``!nontemporal``
4438 metatadata on the instruction tells the optimizer and code generator
4439 that this load is not expected to be reused in the cache. The code
4440 generator may select special instructions to save cache bandwidth, such
4441 as the ``MOVNT`` instruction on x86.
4443 The optional ``!invariant.load`` metadata must reference a single
4444 metatadata name <index> corresponding to a metadata node with no
4445 entries. The existence of the ``!invariant.load`` metatadata on the
4446 instruction tells the optimizer and code generator that this load
4447 address points to memory which does not change value during program
4448 execution. The optimizer may then move this load around, for example, by
4449 hoisting it out of loops using loop invariant code motion.
4454 The location of memory pointed to is loaded. If the value being loaded
4455 is of scalar type then the number of bytes read does not exceed the
4456 minimum number of bytes needed to hold all bits of the type. For
4457 example, loading an ``i24`` reads at most three bytes. When loading a
4458 value of a type like ``i20`` with a size that is not an integral number
4459 of bytes, the result is undefined if the value was not originally
4460 written using a store of the same type.
4465 .. code-block:: llvm
4467 %ptr = alloca i32 ; yields {i32*}:ptr
4468 store i32 3, i32* %ptr ; yields {void}
4469 %val = load i32* %ptr ; yields {i32}:val = i32 3
4473 '``store``' Instruction
4474 ^^^^^^^^^^^^^^^^^^^^^^^
4481 store [volatile] <ty> <value>, <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>] ; yields {void}
4482 store atomic [volatile] <ty> <value>, <ty>* <pointer> [singlethread] <ordering>, align <alignment> ; yields {void}
4487 The '``store``' instruction is used to write to memory.
4492 There are two arguments to the '``store``' instruction: a value to store
4493 and an address at which to store it. The type of the '``<pointer>``'
4494 operand must be a pointer to the :ref:`first class <t_firstclass>` type of
4495 the '``<value>``' operand. If the ``store`` is marked as ``volatile``,
4496 then the optimizer is not allowed to modify the number or order of
4497 execution of this ``store`` with other :ref:`volatile
4498 operations <volatile>`.
4500 If the ``store`` is marked as ``atomic``, it takes an extra
4501 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
4502 ``acquire`` and ``acq_rel`` orderings aren't valid on ``store``
4503 instructions. Atomic loads produce :ref:`defined <memmodel>` results
4504 when they may see multiple atomic stores. The type of the pointee must
4505 be an integer type whose bit width is a power of two greater than or
4506 equal to eight and less than or equal to a target-specific size limit.
4507 ``align`` must be explicitly specified on atomic stores, and the store
4508 has undefined behavior if the alignment is not set to a value which is
4509 at least the size in bytes of the pointee. ``!nontemporal`` does not
4510 have any defined semantics for atomic stores.
4512 The optional constant "align" argument specifies the alignment of the
4513 operation (that is, the alignment of the memory address). A value of 0
4514 or an omitted "align" argument means that the operation has the abi
4515 alignment for the target. It is the responsibility of the code emitter
4516 to ensure that the alignment information is correct. Overestimating the
4517 alignment results in an undefined behavior. Underestimating the
4518 alignment may produce less efficient code. An alignment of 1 is always
4521 The optional !nontemporal metadata must reference a single metatadata
4522 name <index> corresponding to a metadata node with one i32 entry of
4523 value 1. The existence of the !nontemporal metatadata on the instruction
4524 tells the optimizer and code generator that this load is not expected to
4525 be reused in the cache. The code generator may select special
4526 instructions to save cache bandwidth, such as the MOVNT instruction on
4532 The contents of memory are updated to contain '``<value>``' at the
4533 location specified by the '``<pointer>``' operand. If '``<value>``' is
4534 of scalar type then the number of bytes written does not exceed the
4535 minimum number of bytes needed to hold all bits of the type. For
4536 example, storing an ``i24`` writes at most three bytes. When writing a
4537 value of a type like ``i20`` with a size that is not an integral number
4538 of bytes, it is unspecified what happens to the extra bits that do not
4539 belong to the type, but they will typically be overwritten.
4544 .. code-block:: llvm
4546 %ptr = alloca i32 ; yields {i32*}:ptr
4547 store i32 3, i32* %ptr ; yields {void}
4548 %val = load i32* %ptr ; yields {i32}:val = i32 3
4552 '``fence``' Instruction
4553 ^^^^^^^^^^^^^^^^^^^^^^^
4560 fence [singlethread] <ordering> ; yields {void}
4565 The '``fence``' instruction is used to introduce happens-before edges
4571 '``fence``' instructions take an :ref:`ordering <ordering>` argument which
4572 defines what *synchronizes-with* edges they add. They can only be given
4573 ``acquire``, ``release``, ``acq_rel``, and ``seq_cst`` orderings.
4578 A fence A which has (at least) ``release`` ordering semantics
4579 *synchronizes with* a fence B with (at least) ``acquire`` ordering
4580 semantics if and only if there exist atomic operations X and Y, both
4581 operating on some atomic object M, such that A is sequenced before X, X
4582 modifies M (either directly or through some side effect of a sequence
4583 headed by X), Y is sequenced before B, and Y observes M. This provides a
4584 *happens-before* dependency between A and B. Rather than an explicit
4585 ``fence``, one (but not both) of the atomic operations X or Y might
4586 provide a ``release`` or ``acquire`` (resp.) ordering constraint and
4587 still *synchronize-with* the explicit ``fence`` and establish the
4588 *happens-before* edge.
4590 A ``fence`` which has ``seq_cst`` ordering, in addition to having both
4591 ``acquire`` and ``release`` semantics specified above, participates in
4592 the global program order of other ``seq_cst`` operations and/or fences.
4594 The optional ":ref:`singlethread <singlethread>`" argument specifies
4595 that the fence only synchronizes with other fences in the same thread.
4596 (This is useful for interacting with signal handlers.)
4601 .. code-block:: llvm
4603 fence acquire ; yields {void}
4604 fence singlethread seq_cst ; yields {void}
4608 '``cmpxchg``' Instruction
4609 ^^^^^^^^^^^^^^^^^^^^^^^^^
4616 cmpxchg [volatile] <ty>* <pointer>, <ty> <cmp>, <ty> <new> [singlethread] <ordering> ; yields {ty}
4621 The '``cmpxchg``' instruction is used to atomically modify memory. It
4622 loads a value in memory and compares it to a given value. If they are
4623 equal, it stores a new value into the memory.
4628 There are three arguments to the '``cmpxchg``' instruction: an address
4629 to operate on, a value to compare to the value currently be at that
4630 address, and a new value to place at that address if the compared values
4631 are equal. The type of '<cmp>' must be an integer type whose bit width
4632 is a power of two greater than or equal to eight and less than or equal
4633 to a target-specific size limit. '<cmp>' and '<new>' must have the same
4634 type, and the type of '<pointer>' must be a pointer to that type. If the
4635 ``cmpxchg`` is marked as ``volatile``, then the optimizer is not allowed
4636 to modify the number or order of execution of this ``cmpxchg`` with
4637 other :ref:`volatile operations <volatile>`.
4639 The :ref:`ordering <ordering>` argument specifies how this ``cmpxchg``
4640 synchronizes with other atomic operations.
4642 The optional "``singlethread``" argument declares that the ``cmpxchg``
4643 is only atomic with respect to code (usually signal handlers) running in
4644 the same thread as the ``cmpxchg``. Otherwise the cmpxchg is atomic with
4645 respect to all other code in the system.
4647 The pointer passed into cmpxchg must have alignment greater than or
4648 equal to the size in memory of the operand.
4653 The contents of memory at the location specified by the '``<pointer>``'
4654 operand is read and compared to '``<cmp>``'; if the read value is the
4655 equal, '``<new>``' is written. The original value at the location is
4658 A successful ``cmpxchg`` is a read-modify-write instruction for the purpose
4659 of identifying release sequences. A failed ``cmpxchg`` is equivalent to an
4660 atomic load with an ordering parameter determined by dropping any
4661 ``release`` part of the ``cmpxchg``'s ordering.
4666 .. code-block:: llvm
4669 %orig = atomic load i32* %ptr unordered ; yields {i32}
4673 %cmp = phi i32 [ %orig, %entry ], [%old, %loop]
4674 %squared = mul i32 %cmp, %cmp
4675 %old = cmpxchg i32* %ptr, i32 %cmp, i32 %squared ; yields {i32}
4676 %success = icmp eq i32 %cmp, %old
4677 br i1 %success, label %done, label %loop
4684 '``atomicrmw``' Instruction
4685 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
4692 atomicrmw [volatile] <operation> <ty>* <pointer>, <ty> <value> [singlethread] <ordering> ; yields {ty}
4697 The '``atomicrmw``' instruction is used to atomically modify memory.
4702 There are three arguments to the '``atomicrmw``' instruction: an
4703 operation to apply, an address whose value to modify, an argument to the
4704 operation. The operation must be one of the following keywords:
4718 The type of '<value>' must be an integer type whose bit width is a power
4719 of two greater than or equal to eight and less than or equal to a
4720 target-specific size limit. The type of the '``<pointer>``' operand must
4721 be a pointer to that type. If the ``atomicrmw`` is marked as
4722 ``volatile``, then the optimizer is not allowed to modify the number or
4723 order of execution of this ``atomicrmw`` with other :ref:`volatile
4724 operations <volatile>`.
4729 The contents of memory at the location specified by the '``<pointer>``'
4730 operand are atomically read, modified, and written back. The original
4731 value at the location is returned. The modification is specified by the
4734 - xchg: ``*ptr = val``
4735 - add: ``*ptr = *ptr + val``
4736 - sub: ``*ptr = *ptr - val``
4737 - and: ``*ptr = *ptr & val``
4738 - nand: ``*ptr = ~(*ptr & val)``
4739 - or: ``*ptr = *ptr | val``
4740 - xor: ``*ptr = *ptr ^ val``
4741 - max: ``*ptr = *ptr > val ? *ptr : val`` (using a signed comparison)
4742 - min: ``*ptr = *ptr < val ? *ptr : val`` (using a signed comparison)
4743 - umax: ``*ptr = *ptr > val ? *ptr : val`` (using an unsigned
4745 - umin: ``*ptr = *ptr < val ? *ptr : val`` (using an unsigned
4751 .. code-block:: llvm
4753 %old = atomicrmw add i32* %ptr, i32 1 acquire ; yields {i32}
4755 .. _i_getelementptr:
4757 '``getelementptr``' Instruction
4758 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4765 <result> = getelementptr <pty>* <ptrval>{, <ty> <idx>}*
4766 <result> = getelementptr inbounds <pty>* <ptrval>{, <ty> <idx>}*
4767 <result> = getelementptr <ptr vector> ptrval, <vector index type> idx
4772 The '``getelementptr``' instruction is used to get the address of a
4773 subelement of an :ref:`aggregate <t_aggregate>` data structure. It performs
4774 address calculation only and does not access memory.
4779 The first argument is always a pointer or a vector of pointers, and
4780 forms the basis of the calculation. The remaining arguments are indices
4781 that indicate which of the elements of the aggregate object are indexed.
4782 The interpretation of each index is dependent on the type being indexed
4783 into. The first index always indexes the pointer value given as the
4784 first argument, the second index indexes a value of the type pointed to
4785 (not necessarily the value directly pointed to, since the first index
4786 can be non-zero), etc. The first type indexed into must be a pointer
4787 value, subsequent types can be arrays, vectors, and structs. Note that
4788 subsequent types being indexed into can never be pointers, since that
4789 would require loading the pointer before continuing calculation.
4791 The type of each index argument depends on the type it is indexing into.
4792 When indexing into a (optionally packed) structure, only ``i32`` integer
4793 **constants** are allowed (when using a vector of indices they must all
4794 be the **same** ``i32`` integer constant). When indexing into an array,
4795 pointer or vector, integers of any width are allowed, and they are not
4796 required to be constant. These integers are treated as signed values
4799 For example, let's consider a C code fragment and how it gets compiled
4815 int *foo(struct ST *s) {
4816 return &s[1].Z.B[5][13];
4819 The LLVM code generated by Clang is:
4821 .. code-block:: llvm
4823 %struct.RT = type { i8, [10 x [20 x i32]], i8 }
4824 %struct.ST = type { i32, double, %struct.RT }
4826 define i32* @foo(%struct.ST* %s) nounwind uwtable readnone optsize ssp {
4828 %arrayidx = getelementptr inbounds %struct.ST* %s, i64 1, i32 2, i32 1, i64 5, i64 13
4835 In the example above, the first index is indexing into the
4836 '``%struct.ST*``' type, which is a pointer, yielding a '``%struct.ST``'
4837 = '``{ i32, double, %struct.RT }``' type, a structure. The second index
4838 indexes into the third element of the structure, yielding a
4839 '``%struct.RT``' = '``{ i8 , [10 x [20 x i32]], i8 }``' type, another
4840 structure. The third index indexes into the second element of the
4841 structure, yielding a '``[10 x [20 x i32]]``' type, an array. The two
4842 dimensions of the array are subscripted into, yielding an '``i32``'
4843 type. The '``getelementptr``' instruction returns a pointer to this
4844 element, thus computing a value of '``i32*``' type.
4846 Note that it is perfectly legal to index partially through a structure,
4847 returning a pointer to an inner element. Because of this, the LLVM code
4848 for the given testcase is equivalent to:
4850 .. code-block:: llvm
4852 define i32* @foo(%struct.ST* %s) {
4853 %t1 = getelementptr %struct.ST* %s, i32 1 ; yields %struct.ST*:%t1
4854 %t2 = getelementptr %struct.ST* %t1, i32 0, i32 2 ; yields %struct.RT*:%t2
4855 %t3 = getelementptr %struct.RT* %t2, i32 0, i32 1 ; yields [10 x [20 x i32]]*:%t3
4856 %t4 = getelementptr [10 x [20 x i32]]* %t3, i32 0, i32 5 ; yields [20 x i32]*:%t4
4857 %t5 = getelementptr [20 x i32]* %t4, i32 0, i32 13 ; yields i32*:%t5
4861 If the ``inbounds`` keyword is present, the result value of the
4862 ``getelementptr`` is a :ref:`poison value <poisonvalues>` if the base
4863 pointer is not an *in bounds* address of an allocated object, or if any
4864 of the addresses that would be formed by successive addition of the
4865 offsets implied by the indices to the base address with infinitely
4866 precise signed arithmetic are not an *in bounds* address of that
4867 allocated object. The *in bounds* addresses for an allocated object are
4868 all the addresses that point into the object, plus the address one byte
4869 past the end. In cases where the base is a vector of pointers the
4870 ``inbounds`` keyword applies to each of the computations element-wise.
4872 If the ``inbounds`` keyword is not present, the offsets are added to the
4873 base address with silently-wrapping two's complement arithmetic. If the
4874 offsets have a different width from the pointer, they are sign-extended
4875 or truncated to the width of the pointer. The result value of the
4876 ``getelementptr`` may be outside the object pointed to by the base
4877 pointer. The result value may not necessarily be used to access memory
4878 though, even if it happens to point into allocated storage. See the
4879 :ref:`Pointer Aliasing Rules <pointeraliasing>` section for more
4882 The getelementptr instruction is often confusing. For some more insight
4883 into how it works, see :doc:`the getelementptr FAQ <GetElementPtr>`.
4888 .. code-block:: llvm
4890 ; yields [12 x i8]*:aptr
4891 %aptr = getelementptr {i32, [12 x i8]}* %saptr, i64 0, i32 1
4893 %vptr = getelementptr {i32, <2 x i8>}* %svptr, i64 0, i32 1, i32 1
4895 %eptr = getelementptr [12 x i8]* %aptr, i64 0, i32 1
4897 %iptr = getelementptr [10 x i32]* @arr, i16 0, i16 0
4899 In cases where the pointer argument is a vector of pointers, each index
4900 must be a vector with the same number of elements. For example:
4902 .. code-block:: llvm
4904 %A = getelementptr <4 x i8*> %ptrs, <4 x i64> %offsets,
4906 Conversion Operations
4907 ---------------------
4909 The instructions in this category are the conversion instructions
4910 (casting) which all take a single operand and a type. They perform
4911 various bit conversions on the operand.
4913 '``trunc .. to``' Instruction
4914 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4921 <result> = trunc <ty> <value> to <ty2> ; yields ty2
4926 The '``trunc``' instruction truncates its operand to the type ``ty2``.
4931 The '``trunc``' instruction takes a value to trunc, and a type to trunc
4932 it to. Both types must be of :ref:`integer <t_integer>` types, or vectors
4933 of the same number of integers. The bit size of the ``value`` must be
4934 larger than the bit size of the destination type, ``ty2``. Equal sized
4935 types are not allowed.
4940 The '``trunc``' instruction truncates the high order bits in ``value``
4941 and converts the remaining bits to ``ty2``. Since the source size must
4942 be larger than the destination size, ``trunc`` cannot be a *no-op cast*.
4943 It will always truncate bits.
4948 .. code-block:: llvm
4950 %X = trunc i32 257 to i8 ; yields i8:1
4951 %Y = trunc i32 123 to i1 ; yields i1:true
4952 %Z = trunc i32 122 to i1 ; yields i1:false
4953 %W = trunc <2 x i16> <i16 8, i16 7> to <2 x i8> ; yields <i8 8, i8 7>
4955 '``zext .. to``' Instruction
4956 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4963 <result> = zext <ty> <value> to <ty2> ; yields ty2
4968 The '``zext``' instruction zero extends its operand to type ``ty2``.
4973 The '``zext``' instruction takes a value to cast, and a type to cast it
4974 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
4975 the same number of integers. The bit size of the ``value`` must be
4976 smaller than the bit size of the destination type, ``ty2``.
4981 The ``zext`` fills the high order bits of the ``value`` with zero bits
4982 until it reaches the size of the destination type, ``ty2``.
4984 When zero extending from i1, the result will always be either 0 or 1.
4989 .. code-block:: llvm
4991 %X = zext i32 257 to i64 ; yields i64:257
4992 %Y = zext i1 true to i32 ; yields i32:1
4993 %Z = zext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
4995 '``sext .. to``' Instruction
4996 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5003 <result> = sext <ty> <value> to <ty2> ; yields ty2
5008 The '``sext``' sign extends ``value`` to the type ``ty2``.
5013 The '``sext``' instruction takes a value to cast, and a type to cast it
5014 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
5015 the same number of integers. The bit size of the ``value`` must be
5016 smaller than the bit size of the destination type, ``ty2``.
5021 The '``sext``' instruction performs a sign extension by copying the sign
5022 bit (highest order bit) of the ``value`` until it reaches the bit size
5023 of the type ``ty2``.
5025 When sign extending from i1, the extension always results in -1 or 0.
5030 .. code-block:: llvm
5032 %X = sext i8 -1 to i16 ; yields i16 :65535
5033 %Y = sext i1 true to i32 ; yields i32:-1
5034 %Z = sext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
5036 '``fptrunc .. to``' Instruction
5037 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5044 <result> = fptrunc <ty> <value> to <ty2> ; yields ty2
5049 The '``fptrunc``' instruction truncates ``value`` to type ``ty2``.
5054 The '``fptrunc``' instruction takes a :ref:`floating point <t_floating>`
5055 value to cast and a :ref:`floating point <t_floating>` type to cast it to.
5056 The size of ``value`` must be larger than the size of ``ty2``. This
5057 implies that ``fptrunc`` cannot be used to make a *no-op cast*.
5062 The '``fptrunc``' instruction truncates a ``value`` from a larger
5063 :ref:`floating point <t_floating>` type to a smaller :ref:`floating
5064 point <t_floating>` type. If the value cannot fit within the
5065 destination type, ``ty2``, then the results are undefined.
5070 .. code-block:: llvm
5072 %X = fptrunc double 123.0 to float ; yields float:123.0
5073 %Y = fptrunc double 1.0E+300 to float ; yields undefined
5075 '``fpext .. to``' Instruction
5076 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5083 <result> = fpext <ty> <value> to <ty2> ; yields ty2
5088 The '``fpext``' extends a floating point ``value`` to a larger floating
5094 The '``fpext``' instruction takes a :ref:`floating point <t_floating>`
5095 ``value`` to cast, and a :ref:`floating point <t_floating>` type to cast it
5096 to. The source type must be smaller than the destination type.
5101 The '``fpext``' instruction extends the ``value`` from a smaller
5102 :ref:`floating point <t_floating>` type to a larger :ref:`floating
5103 point <t_floating>` type. The ``fpext`` cannot be used to make a
5104 *no-op cast* because it always changes bits. Use ``bitcast`` to make a
5105 *no-op cast* for a floating point cast.
5110 .. code-block:: llvm
5112 %X = fpext float 3.125 to double ; yields double:3.125000e+00
5113 %Y = fpext double %X to fp128 ; yields fp128:0xL00000000000000004000900000000000
5115 '``fptoui .. to``' Instruction
5116 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5123 <result> = fptoui <ty> <value> to <ty2> ; yields ty2
5128 The '``fptoui``' converts a floating point ``value`` to its unsigned
5129 integer equivalent of type ``ty2``.
5134 The '``fptoui``' instruction takes a value to cast, which must be a
5135 scalar or vector :ref:`floating point <t_floating>` value, and a type to
5136 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
5137 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
5138 type with the same number of elements as ``ty``
5143 The '``fptoui``' instruction converts its :ref:`floating
5144 point <t_floating>` operand into the nearest (rounding towards zero)
5145 unsigned integer value. If the value cannot fit in ``ty2``, the results
5151 .. code-block:: llvm
5153 %X = fptoui double 123.0 to i32 ; yields i32:123
5154 %Y = fptoui float 1.0E+300 to i1 ; yields undefined:1
5155 %Z = fptoui float 1.04E+17 to i8 ; yields undefined:1
5157 '``fptosi .. to``' Instruction
5158 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5165 <result> = fptosi <ty> <value> to <ty2> ; yields ty2
5170 The '``fptosi``' instruction converts :ref:`floating point <t_floating>`
5171 ``value`` to type ``ty2``.
5176 The '``fptosi``' instruction takes a value to cast, which must be a
5177 scalar or vector :ref:`floating point <t_floating>` value, and a type to
5178 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
5179 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
5180 type with the same number of elements as ``ty``
5185 The '``fptosi``' instruction converts its :ref:`floating
5186 point <t_floating>` operand into the nearest (rounding towards zero)
5187 signed integer value. If the value cannot fit in ``ty2``, the results
5193 .. code-block:: llvm
5195 %X = fptosi double -123.0 to i32 ; yields i32:-123
5196 %Y = fptosi float 1.0E-247 to i1 ; yields undefined:1
5197 %Z = fptosi float 1.04E+17 to i8 ; yields undefined:1
5199 '``uitofp .. to``' Instruction
5200 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5207 <result> = uitofp <ty> <value> to <ty2> ; yields ty2
5212 The '``uitofp``' instruction regards ``value`` as an unsigned integer
5213 and converts that value to the ``ty2`` type.
5218 The '``uitofp``' instruction takes a value to cast, which must be a
5219 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
5220 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
5221 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
5222 type with the same number of elements as ``ty``
5227 The '``uitofp``' instruction interprets its operand as an unsigned
5228 integer quantity and converts it to the corresponding floating point
5229 value. If the value cannot fit in the floating point value, the results
5235 .. code-block:: llvm
5237 %X = uitofp i32 257 to float ; yields float:257.0
5238 %Y = uitofp i8 -1 to double ; yields double:255.0
5240 '``sitofp .. to``' Instruction
5241 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5248 <result> = sitofp <ty> <value> to <ty2> ; yields ty2
5253 The '``sitofp``' instruction regards ``value`` as a signed integer and
5254 converts that value to the ``ty2`` type.
5259 The '``sitofp``' instruction takes a value to cast, which must be a
5260 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
5261 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
5262 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
5263 type with the same number of elements as ``ty``
5268 The '``sitofp``' instruction interprets its operand as a signed integer
5269 quantity and converts it to the corresponding floating point value. If
5270 the value cannot fit in the floating point value, the results are
5276 .. code-block:: llvm
5278 %X = sitofp i32 257 to float ; yields float:257.0
5279 %Y = sitofp i8 -1 to double ; yields double:-1.0
5283 '``ptrtoint .. to``' Instruction
5284 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5291 <result> = ptrtoint <ty> <value> to <ty2> ; yields ty2
5296 The '``ptrtoint``' instruction converts the pointer or a vector of
5297 pointers ``value`` to the integer (or vector of integers) type ``ty2``.
5302 The '``ptrtoint``' instruction takes a ``value`` to cast, which must be
5303 a a value of type :ref:`pointer <t_pointer>` or a vector of pointers, and a
5304 type to cast it to ``ty2``, which must be an :ref:`integer <t_integer>` or
5305 a vector of integers type.
5310 The '``ptrtoint``' instruction converts ``value`` to integer type
5311 ``ty2`` by interpreting the pointer value as an integer and either
5312 truncating or zero extending that value to the size of the integer type.
5313 If ``value`` is smaller than ``ty2`` then a zero extension is done. If
5314 ``value`` is larger than ``ty2`` then a truncation is done. If they are
5315 the same size, then nothing is done (*no-op cast*) other than a type
5321 .. code-block:: llvm
5323 %X = ptrtoint i32* %P to i8 ; yields truncation on 32-bit architecture
5324 %Y = ptrtoint i32* %P to i64 ; yields zero extension on 32-bit architecture
5325 %Z = ptrtoint <4 x i32*> %P to <4 x i64>; yields vector zero extension for a vector of addresses on 32-bit architecture
5329 '``inttoptr .. to``' Instruction
5330 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5337 <result> = inttoptr <ty> <value> to <ty2> ; yields ty2
5342 The '``inttoptr``' instruction converts an integer ``value`` to a
5343 pointer type, ``ty2``.
5348 The '``inttoptr``' instruction takes an :ref:`integer <t_integer>` value to
5349 cast, and a type to cast it to, which must be a :ref:`pointer <t_pointer>`
5355 The '``inttoptr``' instruction converts ``value`` to type ``ty2`` by
5356 applying either a zero extension or a truncation depending on the size
5357 of the integer ``value``. If ``value`` is larger than the size of a
5358 pointer then a truncation is done. If ``value`` is smaller than the size
5359 of a pointer then a zero extension is done. If they are the same size,
5360 nothing is done (*no-op cast*).
5365 .. code-block:: llvm
5367 %X = inttoptr i32 255 to i32* ; yields zero extension on 64-bit architecture
5368 %Y = inttoptr i32 255 to i32* ; yields no-op on 32-bit architecture
5369 %Z = inttoptr i64 0 to i32* ; yields truncation on 32-bit architecture
5370 %Z = inttoptr <4 x i32> %G to <4 x i8*>; yields truncation of vector G to four pointers
5374 '``bitcast .. to``' Instruction
5375 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5382 <result> = bitcast <ty> <value> to <ty2> ; yields ty2
5387 The '``bitcast``' instruction converts ``value`` to type ``ty2`` without
5393 The '``bitcast``' instruction takes a value to cast, which must be a
5394 non-aggregate first class value, and a type to cast it to, which must
5395 also be a non-aggregate :ref:`first class <t_firstclass>` type. The bit
5396 sizes of ``value`` and the destination type, ``ty2``, must be identical.
5397 If the source type is a pointer, the destination type must also be a
5398 pointer. This instruction supports bitwise conversion of vectors to
5399 integers and to vectors of other types (as long as they have the same
5405 The '``bitcast``' instruction converts ``value`` to type ``ty2``. It is
5406 always a *no-op cast* because no bits change with this conversion. The
5407 conversion is done as if the ``value`` had been stored to memory and
5408 read back as type ``ty2``. Pointer (or vector of pointers) types may
5409 only be converted to other pointer (or vector of pointers) types with
5410 this instruction. To convert pointers to other types, use the
5411 :ref:`inttoptr <i_inttoptr>` or :ref:`ptrtoint <i_ptrtoint>` instructions
5417 .. code-block:: llvm
5419 %X = bitcast i8 255 to i8 ; yields i8 :-1
5420 %Y = bitcast i32* %x to sint* ; yields sint*:%x
5421 %Z = bitcast <2 x int> %V to i64; ; yields i64: %V
5422 %Z = bitcast <2 x i32*> %V to <2 x i64*> ; yields <2 x i64*>
5429 The instructions in this category are the "miscellaneous" instructions,
5430 which defy better classification.
5434 '``icmp``' Instruction
5435 ^^^^^^^^^^^^^^^^^^^^^^
5442 <result> = icmp <cond> <ty> <op1>, <op2> ; yields {i1} or {<N x i1>}:result
5447 The '``icmp``' instruction returns a boolean value or a vector of
5448 boolean values based on comparison of its two integer, integer vector,
5449 pointer, or pointer vector operands.
5454 The '``icmp``' instruction takes three operands. The first operand is
5455 the condition code indicating the kind of comparison to perform. It is
5456 not a value, just a keyword. The possible condition code are:
5459 #. ``ne``: not equal
5460 #. ``ugt``: unsigned greater than
5461 #. ``uge``: unsigned greater or equal
5462 #. ``ult``: unsigned less than
5463 #. ``ule``: unsigned less or equal
5464 #. ``sgt``: signed greater than
5465 #. ``sge``: signed greater or equal
5466 #. ``slt``: signed less than
5467 #. ``sle``: signed less or equal
5469 The remaining two arguments must be :ref:`integer <t_integer>` or
5470 :ref:`pointer <t_pointer>` or integer :ref:`vector <t_vector>` typed. They
5471 must also be identical types.
5476 The '``icmp``' compares ``op1`` and ``op2`` according to the condition
5477 code given as ``cond``. The comparison performed always yields either an
5478 :ref:`i1 <t_integer>` or vector of ``i1`` result, as follows:
5480 #. ``eq``: yields ``true`` if the operands are equal, ``false``
5481 otherwise. No sign interpretation is necessary or performed.
5482 #. ``ne``: yields ``true`` if the operands are unequal, ``false``
5483 otherwise. No sign interpretation is necessary or performed.
5484 #. ``ugt``: interprets the operands as unsigned values and yields
5485 ``true`` if ``op1`` is greater than ``op2``.
5486 #. ``uge``: interprets the operands as unsigned values and yields
5487 ``true`` if ``op1`` is greater than or equal to ``op2``.
5488 #. ``ult``: interprets the operands as unsigned values and yields
5489 ``true`` if ``op1`` is less than ``op2``.
5490 #. ``ule``: interprets the operands as unsigned values and yields
5491 ``true`` if ``op1`` is less than or equal to ``op2``.
5492 #. ``sgt``: interprets the operands as signed values and yields ``true``
5493 if ``op1`` is greater than ``op2``.
5494 #. ``sge``: interprets the operands as signed values and yields ``true``
5495 if ``op1`` is greater than or equal to ``op2``.
5496 #. ``slt``: interprets the operands as signed values and yields ``true``
5497 if ``op1`` is less than ``op2``.
5498 #. ``sle``: interprets the operands as signed values and yields ``true``
5499 if ``op1`` is less than or equal to ``op2``.
5501 If the operands are :ref:`pointer <t_pointer>` typed, the pointer values
5502 are compared as if they were integers.
5504 If the operands are integer vectors, then they are compared element by
5505 element. The result is an ``i1`` vector with the same number of elements
5506 as the values being compared. Otherwise, the result is an ``i1``.
5511 .. code-block:: llvm
5513 <result> = icmp eq i32 4, 5 ; yields: result=false
5514 <result> = icmp ne float* %X, %X ; yields: result=false
5515 <result> = icmp ult i16 4, 5 ; yields: result=true
5516 <result> = icmp sgt i16 4, 5 ; yields: result=false
5517 <result> = icmp ule i16 -4, 5 ; yields: result=false
5518 <result> = icmp sge i16 4, 5 ; yields: result=false
5520 Note that the code generator does not yet support vector types with the
5521 ``icmp`` instruction.
5525 '``fcmp``' Instruction
5526 ^^^^^^^^^^^^^^^^^^^^^^
5533 <result> = fcmp <cond> <ty> <op1>, <op2> ; yields {i1} or {<N x i1>}:result
5538 The '``fcmp``' instruction returns a boolean value or vector of boolean
5539 values based on comparison of its operands.
5541 If the operands are floating point scalars, then the result type is a
5542 boolean (:ref:`i1 <t_integer>`).
5544 If the operands are floating point vectors, then the result type is a
5545 vector of boolean with the same number of elements as the operands being
5551 The '``fcmp``' instruction takes three operands. The first operand is
5552 the condition code indicating the kind of comparison to perform. It is
5553 not a value, just a keyword. The possible condition code are:
5555 #. ``false``: no comparison, always returns false
5556 #. ``oeq``: ordered and equal
5557 #. ``ogt``: ordered and greater than
5558 #. ``oge``: ordered and greater than or equal
5559 #. ``olt``: ordered and less than
5560 #. ``ole``: ordered and less than or equal
5561 #. ``one``: ordered and not equal
5562 #. ``ord``: ordered (no nans)
5563 #. ``ueq``: unordered or equal
5564 #. ``ugt``: unordered or greater than
5565 #. ``uge``: unordered or greater than or equal
5566 #. ``ult``: unordered or less than
5567 #. ``ule``: unordered or less than or equal
5568 #. ``une``: unordered or not equal
5569 #. ``uno``: unordered (either nans)
5570 #. ``true``: no comparison, always returns true
5572 *Ordered* means that neither operand is a QNAN while *unordered* means
5573 that either operand may be a QNAN.
5575 Each of ``val1`` and ``val2`` arguments must be either a :ref:`floating
5576 point <t_floating>` type or a :ref:`vector <t_vector>` of floating point
5577 type. They must have identical types.
5582 The '``fcmp``' instruction compares ``op1`` and ``op2`` according to the
5583 condition code given as ``cond``. If the operands are vectors, then the
5584 vectors are compared element by element. Each comparison performed
5585 always yields an :ref:`i1 <t_integer>` result, as follows:
5587 #. ``false``: always yields ``false``, regardless of operands.
5588 #. ``oeq``: yields ``true`` if both operands are not a QNAN and ``op1``
5589 is equal to ``op2``.
5590 #. ``ogt``: yields ``true`` if both operands are not a QNAN and ``op1``
5591 is greater than ``op2``.
5592 #. ``oge``: yields ``true`` if both operands are not a QNAN and ``op1``
5593 is greater than or equal to ``op2``.
5594 #. ``olt``: yields ``true`` if both operands are not a QNAN and ``op1``
5595 is less than ``op2``.
5596 #. ``ole``: yields ``true`` if both operands are not a QNAN and ``op1``
5597 is less than or equal to ``op2``.
5598 #. ``one``: yields ``true`` if both operands are not a QNAN and ``op1``
5599 is not equal to ``op2``.
5600 #. ``ord``: yields ``true`` if both operands are not a QNAN.
5601 #. ``ueq``: yields ``true`` if either operand is a QNAN or ``op1`` is
5603 #. ``ugt``: yields ``true`` if either operand is a QNAN or ``op1`` is
5604 greater than ``op2``.
5605 #. ``uge``: yields ``true`` if either operand is a QNAN or ``op1`` is
5606 greater than or equal to ``op2``.
5607 #. ``ult``: yields ``true`` if either operand is a QNAN or ``op1`` is
5609 #. ``ule``: yields ``true`` if either operand is a QNAN or ``op1`` is
5610 less than or equal to ``op2``.
5611 #. ``une``: yields ``true`` if either operand is a QNAN or ``op1`` is
5612 not equal to ``op2``.
5613 #. ``uno``: yields ``true`` if either operand is a QNAN.
5614 #. ``true``: always yields ``true``, regardless of operands.
5619 .. code-block:: llvm
5621 <result> = fcmp oeq float 4.0, 5.0 ; yields: result=false
5622 <result> = fcmp one float 4.0, 5.0 ; yields: result=true
5623 <result> = fcmp olt float 4.0, 5.0 ; yields: result=true
5624 <result> = fcmp ueq double 1.0, 2.0 ; yields: result=false
5626 Note that the code generator does not yet support vector types with the
5627 ``fcmp`` instruction.
5631 '``phi``' Instruction
5632 ^^^^^^^^^^^^^^^^^^^^^
5639 <result> = phi <ty> [ <val0>, <label0>], ...
5644 The '``phi``' instruction is used to implement the φ node in the SSA
5645 graph representing the function.
5650 The type of the incoming values is specified with the first type field.
5651 After this, the '``phi``' instruction takes a list of pairs as
5652 arguments, with one pair for each predecessor basic block of the current
5653 block. Only values of :ref:`first class <t_firstclass>` type may be used as
5654 the value arguments to the PHI node. Only labels may be used as the
5657 There must be no non-phi instructions between the start of a basic block
5658 and the PHI instructions: i.e. PHI instructions must be first in a basic
5661 For the purposes of the SSA form, the use of each incoming value is
5662 deemed to occur on the edge from the corresponding predecessor block to
5663 the current block (but after any definition of an '``invoke``'
5664 instruction's return value on the same edge).
5669 At runtime, the '``phi``' instruction logically takes on the value
5670 specified by the pair corresponding to the predecessor basic block that
5671 executed just prior to the current block.
5676 .. code-block:: llvm
5678 Loop: ; Infinite loop that counts from 0 on up...
5679 %indvar = phi i32 [ 0, %LoopHeader ], [ %nextindvar, %Loop ]
5680 %nextindvar = add i32 %indvar, 1
5685 '``select``' Instruction
5686 ^^^^^^^^^^^^^^^^^^^^^^^^
5693 <result> = select selty <cond>, <ty> <val1>, <ty> <val2> ; yields ty
5695 selty is either i1 or {<N x i1>}
5700 The '``select``' instruction is used to choose one value based on a
5701 condition, without branching.
5706 The '``select``' instruction requires an 'i1' value or a vector of 'i1'
5707 values indicating the condition, and two values of the same :ref:`first
5708 class <t_firstclass>` type. If the val1/val2 are vectors and the
5709 condition is a scalar, then entire vectors are selected, not individual
5715 If the condition is an i1 and it evaluates to 1, the instruction returns
5716 the first value argument; otherwise, it returns the second value
5719 If the condition is a vector of i1, then the value arguments must be
5720 vectors of the same size, and the selection is done element by element.
5725 .. code-block:: llvm
5727 %X = select i1 true, i8 17, i8 42 ; yields i8:17
5731 '``call``' Instruction
5732 ^^^^^^^^^^^^^^^^^^^^^^
5739 <result> = [tail] call [cconv] [ret attrs] <ty> [<fnty>*] <fnptrval>(<function args>) [fn attrs]
5744 The '``call``' instruction represents a simple function call.
5749 This instruction requires several arguments:
5751 #. The optional "tail" marker indicates that the callee function does
5752 not access any allocas or varargs in the caller. Note that calls may
5753 be marked "tail" even if they do not occur before a
5754 :ref:`ret <i_ret>` instruction. If the "tail" marker is present, the
5755 function call is eligible for tail call optimization, but `might not
5756 in fact be optimized into a jump <CodeGenerator.html#tailcallopt>`_.
5757 The code generator may optimize calls marked "tail" with either 1)
5758 automatic `sibling call
5759 optimization <CodeGenerator.html#sibcallopt>`_ when the caller and
5760 callee have matching signatures, or 2) forced tail call optimization
5761 when the following extra requirements are met:
5763 - Caller and callee both have the calling convention ``fastcc``.
5764 - The call is in tail position (ret immediately follows call and ret
5765 uses value of call or is void).
5766 - Option ``-tailcallopt`` is enabled, or
5767 ``llvm::GuaranteedTailCallOpt`` is ``true``.
5768 - `Platform specific constraints are
5769 met. <CodeGenerator.html#tailcallopt>`_
5771 #. The optional "cconv" marker indicates which :ref:`calling
5772 convention <callingconv>` the call should use. If none is
5773 specified, the call defaults to using C calling conventions. The
5774 calling convention of the call must match the calling convention of
5775 the target function, or else the behavior is undefined.
5776 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
5777 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
5779 #. '``ty``': the type of the call instruction itself which is also the
5780 type of the return value. Functions that return no value are marked
5782 #. '``fnty``': shall be the signature of the pointer to function value
5783 being invoked. The argument types must match the types implied by
5784 this signature. This type can be omitted if the function is not
5785 varargs and if the function type does not return a pointer to a
5787 #. '``fnptrval``': An LLVM value containing a pointer to a function to
5788 be invoked. In most cases, this is a direct function invocation, but
5789 indirect ``call``'s are just as possible, calling an arbitrary pointer
5791 #. '``function args``': argument list whose types match the function
5792 signature argument types and parameter attributes. All arguments must
5793 be of :ref:`first class <t_firstclass>` type. If the function signature
5794 indicates the function accepts a variable number of arguments, the
5795 extra arguments can be specified.
5796 #. The optional :ref:`function attributes <fnattrs>` list. Only
5797 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
5798 attributes are valid here.
5803 The '``call``' instruction is used to cause control flow to transfer to
5804 a specified function, with its incoming arguments bound to the specified
5805 values. Upon a '``ret``' instruction in the called function, control
5806 flow continues with the instruction after the function call, and the
5807 return value of the function is bound to the result argument.
5812 .. code-block:: llvm
5814 %retval = call i32 @test(i32 %argc)
5815 call i32 (i8*, ...)* @printf(i8* %msg, i32 12, i8 42) ; yields i32
5816 %X = tail call i32 @foo() ; yields i32
5817 %Y = tail call fastcc i32 @foo() ; yields i32
5818 call void %foo(i8 97 signext)
5820 %struct.A = type { i32, i8 }
5821 %r = call %struct.A @foo() ; yields { 32, i8 }
5822 %gr = extractvalue %struct.A %r, 0 ; yields i32
5823 %gr1 = extractvalue %struct.A %r, 1 ; yields i8
5824 %Z = call void @foo() noreturn ; indicates that %foo never returns normally
5825 %ZZ = call zeroext i32 @bar() ; Return value is %zero extended
5827 llvm treats calls to some functions with names and arguments that match
5828 the standard C99 library as being the C99 library functions, and may
5829 perform optimizations or generate code for them under that assumption.
5830 This is something we'd like to change in the future to provide better
5831 support for freestanding environments and non-C-based languages.
5835 '``va_arg``' Instruction
5836 ^^^^^^^^^^^^^^^^^^^^^^^^
5843 <resultval> = va_arg <va_list*> <arglist>, <argty>
5848 The '``va_arg``' instruction is used to access arguments passed through
5849 the "variable argument" area of a function call. It is used to implement
5850 the ``va_arg`` macro in C.
5855 This instruction takes a ``va_list*`` value and the type of the
5856 argument. It returns a value of the specified argument type and
5857 increments the ``va_list`` to point to the next argument. The actual
5858 type of ``va_list`` is target specific.
5863 The '``va_arg``' instruction loads an argument of the specified type
5864 from the specified ``va_list`` and causes the ``va_list`` to point to
5865 the next argument. For more information, see the variable argument
5866 handling :ref:`Intrinsic Functions <int_varargs>`.
5868 It is legal for this instruction to be called in a function which does
5869 not take a variable number of arguments, for example, the ``vfprintf``
5872 ``va_arg`` is an LLVM instruction instead of an :ref:`intrinsic
5873 function <intrinsics>` because it takes a type as an argument.
5878 See the :ref:`variable argument processing <int_varargs>` section.
5880 Note that the code generator does not yet fully support va\_arg on many
5881 targets. Also, it does not currently support va\_arg with aggregate
5882 types on any target.
5886 '``landingpad``' Instruction
5887 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5894 <resultval> = landingpad <resultty> personality <type> <pers_fn> <clause>+
5895 <resultval> = landingpad <resultty> personality <type> <pers_fn> cleanup <clause>*
5897 <clause> := catch <type> <value>
5898 <clause> := filter <array constant type> <array constant>
5903 The '``landingpad``' instruction is used by `LLVM's exception handling
5904 system <ExceptionHandling.html#overview>`_ to specify that a basic block
5905 is a landing pad --- one where the exception lands, and corresponds to the
5906 code found in the ``catch`` portion of a ``try``/``catch`` sequence. It
5907 defines values supplied by the personality function (``pers_fn``) upon
5908 re-entry to the function. The ``resultval`` has the type ``resultty``.
5913 This instruction takes a ``pers_fn`` value. This is the personality
5914 function associated with the unwinding mechanism. The optional
5915 ``cleanup`` flag indicates that the landing pad block is a cleanup.
5917 A ``clause`` begins with the clause type --- ``catch`` or ``filter`` --- and
5918 contains the global variable representing the "type" that may be caught
5919 or filtered respectively. Unlike the ``catch`` clause, the ``filter``
5920 clause takes an array constant as its argument. Use
5921 "``[0 x i8**] undef``" for a filter which cannot throw. The
5922 '``landingpad``' instruction must contain *at least* one ``clause`` or
5923 the ``cleanup`` flag.
5928 The '``landingpad``' instruction defines the values which are set by the
5929 personality function (``pers_fn``) upon re-entry to the function, and
5930 therefore the "result type" of the ``landingpad`` instruction. As with
5931 calling conventions, how the personality function results are
5932 represented in LLVM IR is target specific.
5934 The clauses are applied in order from top to bottom. If two
5935 ``landingpad`` instructions are merged together through inlining, the
5936 clauses from the calling function are appended to the list of clauses.
5937 When the call stack is being unwound due to an exception being thrown,
5938 the exception is compared against each ``clause`` in turn. If it doesn't
5939 match any of the clauses, and the ``cleanup`` flag is not set, then
5940 unwinding continues further up the call stack.
5942 The ``landingpad`` instruction has several restrictions:
5944 - A landing pad block is a basic block which is the unwind destination
5945 of an '``invoke``' instruction.
5946 - A landing pad block must have a '``landingpad``' instruction as its
5947 first non-PHI instruction.
5948 - There can be only one '``landingpad``' instruction within the landing
5950 - A basic block that is not a landing pad block may not include a
5951 '``landingpad``' instruction.
5952 - All '``landingpad``' instructions in a function must have the same
5953 personality function.
5958 .. code-block:: llvm
5960 ;; A landing pad which can catch an integer.
5961 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
5963 ;; A landing pad that is a cleanup.
5964 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
5966 ;; A landing pad which can catch an integer and can only throw a double.
5967 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
5969 filter [1 x i8**] [@_ZTId]
5976 LLVM supports the notion of an "intrinsic function". These functions
5977 have well known names and semantics and are required to follow certain
5978 restrictions. Overall, these intrinsics represent an extension mechanism
5979 for the LLVM language that does not require changing all of the
5980 transformations in LLVM when adding to the language (or the bitcode
5981 reader/writer, the parser, etc...).
5983 Intrinsic function names must all start with an "``llvm.``" prefix. This
5984 prefix is reserved in LLVM for intrinsic names; thus, function names may
5985 not begin with this prefix. Intrinsic functions must always be external
5986 functions: you cannot define the body of intrinsic functions. Intrinsic
5987 functions may only be used in call or invoke instructions: it is illegal
5988 to take the address of an intrinsic function. Additionally, because
5989 intrinsic functions are part of the LLVM language, it is required if any
5990 are added that they be documented here.
5992 Some intrinsic functions can be overloaded, i.e., the intrinsic
5993 represents a family of functions that perform the same operation but on
5994 different data types. Because LLVM can represent over 8 million
5995 different integer types, overloading is used commonly to allow an
5996 intrinsic function to operate on any integer type. One or more of the
5997 argument types or the result type can be overloaded to accept any
5998 integer type. Argument types may also be defined as exactly matching a
5999 previous argument's type or the result type. This allows an intrinsic
6000 function which accepts multiple arguments, but needs all of them to be
6001 of the same type, to only be overloaded with respect to a single
6002 argument or the result.
6004 Overloaded intrinsics will have the names of its overloaded argument
6005 types encoded into its function name, each preceded by a period. Only
6006 those types which are overloaded result in a name suffix. Arguments
6007 whose type is matched against another type do not. For example, the
6008 ``llvm.ctpop`` function can take an integer of any width and returns an
6009 integer of exactly the same integer width. This leads to a family of
6010 functions such as ``i8 @llvm.ctpop.i8(i8 %val)`` and
6011 ``i29 @llvm.ctpop.i29(i29 %val)``. Only one type, the return type, is
6012 overloaded, and only one type suffix is required. Because the argument's
6013 type is matched against the return type, it does not require its own
6016 To learn how to add an intrinsic function, please see the `Extending
6017 LLVM Guide <ExtendingLLVM.html>`_.
6021 Variable Argument Handling Intrinsics
6022 -------------------------------------
6024 Variable argument support is defined in LLVM with the
6025 :ref:`va_arg <i_va_arg>` instruction and these three intrinsic
6026 functions. These functions are related to the similarly named macros
6027 defined in the ``<stdarg.h>`` header file.
6029 All of these functions operate on arguments that use a target-specific
6030 value type "``va_list``". The LLVM assembly language reference manual
6031 does not define what this type is, so all transformations should be
6032 prepared to handle these functions regardless of the type used.
6034 This example shows how the :ref:`va_arg <i_va_arg>` instruction and the
6035 variable argument handling intrinsic functions are used.
6037 .. code-block:: llvm
6039 define i32 @test(i32 %X, ...) {
6040 ; Initialize variable argument processing
6042 %ap2 = bitcast i8** %ap to i8*
6043 call void @llvm.va_start(i8* %ap2)
6045 ; Read a single integer argument
6046 %tmp = va_arg i8** %ap, i32
6048 ; Demonstrate usage of llvm.va_copy and llvm.va_end
6050 %aq2 = bitcast i8** %aq to i8*
6051 call void @llvm.va_copy(i8* %aq2, i8* %ap2)
6052 call void @llvm.va_end(i8* %aq2)
6054 ; Stop processing of arguments.
6055 call void @llvm.va_end(i8* %ap2)
6059 declare void @llvm.va_start(i8*)
6060 declare void @llvm.va_copy(i8*, i8*)
6061 declare void @llvm.va_end(i8*)
6065 '``llvm.va_start``' Intrinsic
6066 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6073 declare void %llvm.va_start(i8* <arglist>)
6078 The '``llvm.va_start``' intrinsic initializes ``*<arglist>`` for
6079 subsequent use by ``va_arg``.
6084 The argument is a pointer to a ``va_list`` element to initialize.
6089 The '``llvm.va_start``' intrinsic works just like the ``va_start`` macro
6090 available in C. In a target-dependent way, it initializes the
6091 ``va_list`` element to which the argument points, so that the next call
6092 to ``va_arg`` will produce the first variable argument passed to the
6093 function. Unlike the C ``va_start`` macro, this intrinsic does not need
6094 to know the last argument of the function as the compiler can figure
6097 '``llvm.va_end``' Intrinsic
6098 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6105 declare void @llvm.va_end(i8* <arglist>)
6110 The '``llvm.va_end``' intrinsic destroys ``*<arglist>``, which has been
6111 initialized previously with ``llvm.va_start`` or ``llvm.va_copy``.
6116 The argument is a pointer to a ``va_list`` to destroy.
6121 The '``llvm.va_end``' intrinsic works just like the ``va_end`` macro
6122 available in C. In a target-dependent way, it destroys the ``va_list``
6123 element to which the argument points. Calls to
6124 :ref:`llvm.va_start <int_va_start>` and
6125 :ref:`llvm.va_copy <int_va_copy>` must be matched exactly with calls to
6130 '``llvm.va_copy``' Intrinsic
6131 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6138 declare void @llvm.va_copy(i8* <destarglist>, i8* <srcarglist>)
6143 The '``llvm.va_copy``' intrinsic copies the current argument position
6144 from the source argument list to the destination argument list.
6149 The first argument is a pointer to a ``va_list`` element to initialize.
6150 The second argument is a pointer to a ``va_list`` element to copy from.
6155 The '``llvm.va_copy``' intrinsic works just like the ``va_copy`` macro
6156 available in C. In a target-dependent way, it copies the source
6157 ``va_list`` element into the destination ``va_list`` element. This
6158 intrinsic is necessary because the `` llvm.va_start`` intrinsic may be
6159 arbitrarily complex and require, for example, memory allocation.
6161 Accurate Garbage Collection Intrinsics
6162 --------------------------------------
6164 LLVM support for `Accurate Garbage Collection <GarbageCollection.html>`_
6165 (GC) requires the implementation and generation of these intrinsics.
6166 These intrinsics allow identification of :ref:`GC roots on the
6167 stack <int_gcroot>`, as well as garbage collector implementations that
6168 require :ref:`read <int_gcread>` and :ref:`write <int_gcwrite>` barriers.
6169 Front-ends for type-safe garbage collected languages should generate
6170 these intrinsics to make use of the LLVM garbage collectors. For more
6171 details, see `Accurate Garbage Collection with
6172 LLVM <GarbageCollection.html>`_.
6174 The garbage collection intrinsics only operate on objects in the generic
6175 address space (address space zero).
6179 '``llvm.gcroot``' Intrinsic
6180 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6187 declare void @llvm.gcroot(i8** %ptrloc, i8* %metadata)
6192 The '``llvm.gcroot``' intrinsic declares the existence of a GC root to
6193 the code generator, and allows some metadata to be associated with it.
6198 The first argument specifies the address of a stack object that contains
6199 the root pointer. The second pointer (which must be either a constant or
6200 a global value address) contains the meta-data to be associated with the
6206 At runtime, a call to this intrinsic stores a null pointer into the
6207 "ptrloc" location. At compile-time, the code generator generates
6208 information to allow the runtime to find the pointer at GC safe points.
6209 The '``llvm.gcroot``' intrinsic may only be used in a function which
6210 :ref:`specifies a GC algorithm <gc>`.
6214 '``llvm.gcread``' Intrinsic
6215 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6222 declare i8* @llvm.gcread(i8* %ObjPtr, i8** %Ptr)
6227 The '``llvm.gcread``' intrinsic identifies reads of references from heap
6228 locations, allowing garbage collector implementations that require read
6234 The second argument is the address to read from, which should be an
6235 address allocated from the garbage collector. The first object is a
6236 pointer to the start of the referenced object, if needed by the language
6237 runtime (otherwise null).
6242 The '``llvm.gcread``' intrinsic has the same semantics as a load
6243 instruction, but may be replaced with substantially more complex code by
6244 the garbage collector runtime, as needed. The '``llvm.gcread``'
6245 intrinsic may only be used in a function which :ref:`specifies a GC
6250 '``llvm.gcwrite``' Intrinsic
6251 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6258 declare void @llvm.gcwrite(i8* %P1, i8* %Obj, i8** %P2)
6263 The '``llvm.gcwrite``' intrinsic identifies writes of references to heap
6264 locations, allowing garbage collector implementations that require write
6265 barriers (such as generational or reference counting collectors).
6270 The first argument is the reference to store, the second is the start of
6271 the object to store it to, and the third is the address of the field of
6272 Obj to store to. If the runtime does not require a pointer to the
6273 object, Obj may be null.
6278 The '``llvm.gcwrite``' intrinsic has the same semantics as a store
6279 instruction, but may be replaced with substantially more complex code by
6280 the garbage collector runtime, as needed. The '``llvm.gcwrite``'
6281 intrinsic may only be used in a function which :ref:`specifies a GC
6284 Code Generator Intrinsics
6285 -------------------------
6287 These intrinsics are provided by LLVM to expose special features that
6288 may only be implemented with code generator support.
6290 '``llvm.returnaddress``' Intrinsic
6291 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6298 declare i8 *@llvm.returnaddress(i32 <level>)
6303 The '``llvm.returnaddress``' intrinsic attempts to compute a
6304 target-specific value indicating the return address of the current
6305 function or one of its callers.
6310 The argument to this intrinsic indicates which function to return the
6311 address for. Zero indicates the calling function, one indicates its
6312 caller, etc. The argument is **required** to be a constant integer
6318 The '``llvm.returnaddress``' intrinsic either returns a pointer
6319 indicating the return address of the specified call frame, or zero if it
6320 cannot be identified. The value returned by this intrinsic is likely to
6321 be incorrect or 0 for arguments other than zero, so it should only be
6322 used for debugging purposes.
6324 Note that calling this intrinsic does not prevent function inlining or
6325 other aggressive transformations, so the value returned may not be that
6326 of the obvious source-language caller.
6328 '``llvm.frameaddress``' Intrinsic
6329 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6336 declare i8* @llvm.frameaddress(i32 <level>)
6341 The '``llvm.frameaddress``' intrinsic attempts to return the
6342 target-specific frame pointer value for the specified stack frame.
6347 The argument to this intrinsic indicates which function to return the
6348 frame pointer for. Zero indicates the calling function, one indicates
6349 its caller, etc. The argument is **required** to be a constant integer
6355 The '``llvm.frameaddress``' intrinsic either returns a pointer
6356 indicating the frame address of the specified call frame, or zero if it
6357 cannot be identified. The value returned by this intrinsic is likely to
6358 be incorrect or 0 for arguments other than zero, so it should only be
6359 used for debugging purposes.
6361 Note that calling this intrinsic does not prevent function inlining or
6362 other aggressive transformations, so the value returned may not be that
6363 of the obvious source-language caller.
6367 '``llvm.stacksave``' Intrinsic
6368 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6375 declare i8* @llvm.stacksave()
6380 The '``llvm.stacksave``' intrinsic is used to remember the current state
6381 of the function stack, for use with
6382 :ref:`llvm.stackrestore <int_stackrestore>`. This is useful for
6383 implementing language features like scoped automatic variable sized
6389 This intrinsic returns a opaque pointer value that can be passed to
6390 :ref:`llvm.stackrestore <int_stackrestore>`. When an
6391 ``llvm.stackrestore`` intrinsic is executed with a value saved from
6392 ``llvm.stacksave``, it effectively restores the state of the stack to
6393 the state it was in when the ``llvm.stacksave`` intrinsic executed. In
6394 practice, this pops any :ref:`alloca <i_alloca>` blocks from the stack that
6395 were allocated after the ``llvm.stacksave`` was executed.
6397 .. _int_stackrestore:
6399 '``llvm.stackrestore``' Intrinsic
6400 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6407 declare void @llvm.stackrestore(i8* %ptr)
6412 The '``llvm.stackrestore``' intrinsic is used to restore the state of
6413 the function stack to the state it was in when the corresponding
6414 :ref:`llvm.stacksave <int_stacksave>` intrinsic executed. This is
6415 useful for implementing language features like scoped automatic variable
6416 sized arrays in C99.
6421 See the description for :ref:`llvm.stacksave <int_stacksave>`.
6423 '``llvm.prefetch``' Intrinsic
6424 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6431 declare void @llvm.prefetch(i8* <address>, i32 <rw>, i32 <locality>, i32 <cache type>)
6436 The '``llvm.prefetch``' intrinsic is a hint to the code generator to
6437 insert a prefetch instruction if supported; otherwise, it is a noop.
6438 Prefetches have no effect on the behavior of the program but can change
6439 its performance characteristics.
6444 ``address`` is the address to be prefetched, ``rw`` is the specifier
6445 determining if the fetch should be for a read (0) or write (1), and
6446 ``locality`` is a temporal locality specifier ranging from (0) - no
6447 locality, to (3) - extremely local keep in cache. The ``cache type``
6448 specifies whether the prefetch is performed on the data (1) or
6449 instruction (0) cache. The ``rw``, ``locality`` and ``cache type``
6450 arguments must be constant integers.
6455 This intrinsic does not modify the behavior of the program. In
6456 particular, prefetches cannot trap and do not produce a value. On
6457 targets that support this intrinsic, the prefetch can provide hints to
6458 the processor cache for better performance.
6460 '``llvm.pcmarker``' Intrinsic
6461 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6468 declare void @llvm.pcmarker(i32 <id>)
6473 The '``llvm.pcmarker``' intrinsic is a method to export a Program
6474 Counter (PC) in a region of code to simulators and other tools. The
6475 method is target specific, but it is expected that the marker will use
6476 exported symbols to transmit the PC of the marker. The marker makes no
6477 guarantees that it will remain with any specific instruction after
6478 optimizations. It is possible that the presence of a marker will inhibit
6479 optimizations. The intended use is to be inserted after optimizations to
6480 allow correlations of simulation runs.
6485 ``id`` is a numerical id identifying the marker.
6490 This intrinsic does not modify the behavior of the program. Backends
6491 that do not support this intrinsic may ignore it.
6493 '``llvm.readcyclecounter``' Intrinsic
6494 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6501 declare i64 @llvm.readcyclecounter()
6506 The '``llvm.readcyclecounter``' intrinsic provides access to the cycle
6507 counter register (or similar low latency, high accuracy clocks) on those
6508 targets that support it. On X86, it should map to RDTSC. On Alpha, it
6509 should map to RPCC. As the backing counters overflow quickly (on the
6510 order of 9 seconds on alpha), this should only be used for small
6516 When directly supported, reading the cycle counter should not modify any
6517 memory. Implementations are allowed to either return a application
6518 specific value or a system wide value. On backends without support, this
6519 is lowered to a constant 0.
6521 Standard C Library Intrinsics
6522 -----------------------------
6524 LLVM provides intrinsics for a few important standard C library
6525 functions. These intrinsics allow source-language front-ends to pass
6526 information about the alignment of the pointer arguments to the code
6527 generator, providing opportunity for more efficient code generation.
6531 '``llvm.memcpy``' Intrinsic
6532 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6537 This is an overloaded intrinsic. You can use ``llvm.memcpy`` on any
6538 integer bit width and for different address spaces. Not all targets
6539 support all bit widths however.
6543 declare void @llvm.memcpy.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
6544 i32 <len>, i32 <align>, i1 <isvolatile>)
6545 declare void @llvm.memcpy.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
6546 i64 <len>, i32 <align>, i1 <isvolatile>)
6551 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
6552 source location to the destination location.
6554 Note that, unlike the standard libc function, the ``llvm.memcpy.*``
6555 intrinsics do not return a value, takes extra alignment/isvolatile
6556 arguments and the pointers can be in specified address spaces.
6561 The first argument is a pointer to the destination, the second is a
6562 pointer to the source. The third argument is an integer argument
6563 specifying the number of bytes to copy, the fourth argument is the
6564 alignment of the source and destination locations, and the fifth is a
6565 boolean indicating a volatile access.
6567 If the call to this intrinsic has an alignment value that is not 0 or 1,
6568 then the caller guarantees that both the source and destination pointers
6569 are aligned to that boundary.
6571 If the ``isvolatile`` parameter is ``true``, the ``llvm.memcpy`` call is
6572 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
6573 very cleanly specified and it is unwise to depend on it.
6578 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
6579 source location to the destination location, which are not allowed to
6580 overlap. It copies "len" bytes of memory over. If the argument is known
6581 to be aligned to some boundary, this can be specified as the fourth
6582 argument, otherwise it should be set to 0 or 1.
6584 '``llvm.memmove``' Intrinsic
6585 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6590 This is an overloaded intrinsic. You can use llvm.memmove on any integer
6591 bit width and for different address space. Not all targets support all
6596 declare void @llvm.memmove.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
6597 i32 <len>, i32 <align>, i1 <isvolatile>)
6598 declare void @llvm.memmove.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
6599 i64 <len>, i32 <align>, i1 <isvolatile>)
6604 The '``llvm.memmove.*``' intrinsics move a block of memory from the
6605 source location to the destination location. It is similar to the
6606 '``llvm.memcpy``' intrinsic but allows the two memory locations to
6609 Note that, unlike the standard libc function, the ``llvm.memmove.*``
6610 intrinsics do not return a value, takes extra alignment/isvolatile
6611 arguments and the pointers can be in specified address spaces.
6616 The first argument is a pointer to the destination, the second is a
6617 pointer to the source. The third argument is an integer argument
6618 specifying the number of bytes to copy, the fourth argument is the
6619 alignment of the source and destination locations, and the fifth is a
6620 boolean indicating a volatile access.
6622 If the call to this intrinsic has an alignment value that is not 0 or 1,
6623 then the caller guarantees that the source and destination pointers are
6624 aligned to that boundary.
6626 If the ``isvolatile`` parameter is ``true``, the ``llvm.memmove`` call
6627 is a :ref:`volatile operation <volatile>`. The detailed access behavior is
6628 not very cleanly specified and it is unwise to depend on it.
6633 The '``llvm.memmove.*``' intrinsics copy a block of memory from the
6634 source location to the destination location, which may overlap. It
6635 copies "len" bytes of memory over. If the argument is known to be
6636 aligned to some boundary, this can be specified as the fourth argument,
6637 otherwise it should be set to 0 or 1.
6639 '``llvm.memset.*``' Intrinsics
6640 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6645 This is an overloaded intrinsic. You can use llvm.memset on any integer
6646 bit width and for different address spaces. However, not all targets
6647 support all bit widths.
6651 declare void @llvm.memset.p0i8.i32(i8* <dest>, i8 <val>,
6652 i32 <len>, i32 <align>, i1 <isvolatile>)
6653 declare void @llvm.memset.p0i8.i64(i8* <dest>, i8 <val>,
6654 i64 <len>, i32 <align>, i1 <isvolatile>)
6659 The '``llvm.memset.*``' intrinsics fill a block of memory with a
6660 particular byte value.
6662 Note that, unlike the standard libc function, the ``llvm.memset``
6663 intrinsic does not return a value and takes extra alignment/volatile
6664 arguments. Also, the destination can be in an arbitrary address space.
6669 The first argument is a pointer to the destination to fill, the second
6670 is the byte value with which to fill it, the third argument is an
6671 integer argument specifying the number of bytes to fill, and the fourth
6672 argument is the known alignment of the destination location.
6674 If the call to this intrinsic has an alignment value that is not 0 or 1,
6675 then the caller guarantees that the destination pointer is aligned to
6678 If the ``isvolatile`` parameter is ``true``, the ``llvm.memset`` call is
6679 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
6680 very cleanly specified and it is unwise to depend on it.
6685 The '``llvm.memset.*``' intrinsics fill "len" bytes of memory starting
6686 at the destination location. If the argument is known to be aligned to
6687 some boundary, this can be specified as the fourth argument, otherwise
6688 it should be set to 0 or 1.
6690 '``llvm.sqrt.*``' Intrinsic
6691 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6696 This is an overloaded intrinsic. You can use ``llvm.sqrt`` on any
6697 floating point or vector of floating point type. Not all targets support
6702 declare float @llvm.sqrt.f32(float %Val)
6703 declare double @llvm.sqrt.f64(double %Val)
6704 declare x86_fp80 @llvm.sqrt.f80(x86_fp80 %Val)
6705 declare fp128 @llvm.sqrt.f128(fp128 %Val)
6706 declare ppc_fp128 @llvm.sqrt.ppcf128(ppc_fp128 %Val)
6711 The '``llvm.sqrt``' intrinsics return the sqrt of the specified operand,
6712 returning the same value as the libm '``sqrt``' functions would. Unlike
6713 ``sqrt`` in libm, however, ``llvm.sqrt`` has undefined behavior for
6714 negative numbers other than -0.0 (which allows for better optimization,
6715 because there is no need to worry about errno being set).
6716 ``llvm.sqrt(-0.0)`` is defined to return -0.0 like IEEE sqrt.
6721 The argument and return value are floating point numbers of the same
6727 This function returns the sqrt of the specified operand if it is a
6728 nonnegative floating point number.
6730 '``llvm.powi.*``' Intrinsic
6731 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6736 This is an overloaded intrinsic. You can use ``llvm.powi`` on any
6737 floating point or vector of floating point type. Not all targets support
6742 declare float @llvm.powi.f32(float %Val, i32 %power)
6743 declare double @llvm.powi.f64(double %Val, i32 %power)
6744 declare x86_fp80 @llvm.powi.f80(x86_fp80 %Val, i32 %power)
6745 declare fp128 @llvm.powi.f128(fp128 %Val, i32 %power)
6746 declare ppc_fp128 @llvm.powi.ppcf128(ppc_fp128 %Val, i32 %power)
6751 The '``llvm.powi.*``' intrinsics return the first operand raised to the
6752 specified (positive or negative) power. The order of evaluation of
6753 multiplications is not defined. When a vector of floating point type is
6754 used, the second argument remains a scalar integer value.
6759 The second argument is an integer power, and the first is a value to
6760 raise to that power.
6765 This function returns the first value raised to the second power with an
6766 unspecified sequence of rounding operations.
6768 '``llvm.sin.*``' Intrinsic
6769 ^^^^^^^^^^^^^^^^^^^^^^^^^^
6774 This is an overloaded intrinsic. You can use ``llvm.sin`` on any
6775 floating point or vector of floating point type. Not all targets support
6780 declare float @llvm.sin.f32(float %Val)
6781 declare double @llvm.sin.f64(double %Val)
6782 declare x86_fp80 @llvm.sin.f80(x86_fp80 %Val)
6783 declare fp128 @llvm.sin.f128(fp128 %Val)
6784 declare ppc_fp128 @llvm.sin.ppcf128(ppc_fp128 %Val)
6789 The '``llvm.sin.*``' intrinsics return the sine of the operand.
6794 The argument and return value are floating point numbers of the same
6800 This function returns the sine of the specified operand, returning the
6801 same values as the libm ``sin`` functions would, and handles error
6802 conditions in the same way.
6804 '``llvm.cos.*``' Intrinsic
6805 ^^^^^^^^^^^^^^^^^^^^^^^^^^
6810 This is an overloaded intrinsic. You can use ``llvm.cos`` on any
6811 floating point or vector of floating point type. Not all targets support
6816 declare float @llvm.cos.f32(float %Val)
6817 declare double @llvm.cos.f64(double %Val)
6818 declare x86_fp80 @llvm.cos.f80(x86_fp80 %Val)
6819 declare fp128 @llvm.cos.f128(fp128 %Val)
6820 declare ppc_fp128 @llvm.cos.ppcf128(ppc_fp128 %Val)
6825 The '``llvm.cos.*``' intrinsics return the cosine of the operand.
6830 The argument and return value are floating point numbers of the same
6836 This function returns the cosine of the specified operand, returning the
6837 same values as the libm ``cos`` functions would, and handles error
6838 conditions in the same way.
6840 '``llvm.pow.*``' Intrinsic
6841 ^^^^^^^^^^^^^^^^^^^^^^^^^^
6846 This is an overloaded intrinsic. You can use ``llvm.pow`` on any
6847 floating point or vector of floating point type. Not all targets support
6852 declare float @llvm.pow.f32(float %Val, float %Power)
6853 declare double @llvm.pow.f64(double %Val, double %Power)
6854 declare x86_fp80 @llvm.pow.f80(x86_fp80 %Val, x86_fp80 %Power)
6855 declare fp128 @llvm.pow.f128(fp128 %Val, fp128 %Power)
6856 declare ppc_fp128 @llvm.pow.ppcf128(ppc_fp128 %Val, ppc_fp128 Power)
6861 The '``llvm.pow.*``' intrinsics return the first operand raised to the
6862 specified (positive or negative) power.
6867 The second argument is a floating point power, and the first is a value
6868 to raise to that power.
6873 This function returns the first value raised to the second power,
6874 returning the same values as the libm ``pow`` functions would, and
6875 handles error conditions in the same way.
6877 '``llvm.exp.*``' Intrinsic
6878 ^^^^^^^^^^^^^^^^^^^^^^^^^^
6883 This is an overloaded intrinsic. You can use ``llvm.exp`` on any
6884 floating point or vector of floating point type. Not all targets support
6889 declare float @llvm.exp.f32(float %Val)
6890 declare double @llvm.exp.f64(double %Val)
6891 declare x86_fp80 @llvm.exp.f80(x86_fp80 %Val)
6892 declare fp128 @llvm.exp.f128(fp128 %Val)
6893 declare ppc_fp128 @llvm.exp.ppcf128(ppc_fp128 %Val)
6898 The '``llvm.exp.*``' intrinsics perform the exp function.
6903 The argument and return value are floating point numbers of the same
6909 This function returns the same values as the libm ``exp`` functions
6910 would, and handles error conditions in the same way.
6912 '``llvm.exp2.*``' Intrinsic
6913 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6918 This is an overloaded intrinsic. You can use ``llvm.exp2`` on any
6919 floating point or vector of floating point type. Not all targets support
6924 declare float @llvm.exp2.f32(float %Val)
6925 declare double @llvm.exp2.f64(double %Val)
6926 declare x86_fp80 @llvm.exp2.f80(x86_fp80 %Val)
6927 declare fp128 @llvm.exp2.f128(fp128 %Val)
6928 declare ppc_fp128 @llvm.exp2.ppcf128(ppc_fp128 %Val)
6933 The '``llvm.exp2.*``' intrinsics perform the exp2 function.
6938 The argument and return value are floating point numbers of the same
6944 This function returns the same values as the libm ``exp2`` functions
6945 would, and handles error conditions in the same way.
6947 '``llvm.log.*``' Intrinsic
6948 ^^^^^^^^^^^^^^^^^^^^^^^^^^
6953 This is an overloaded intrinsic. You can use ``llvm.log`` on any
6954 floating point or vector of floating point type. Not all targets support
6959 declare float @llvm.log.f32(float %Val)
6960 declare double @llvm.log.f64(double %Val)
6961 declare x86_fp80 @llvm.log.f80(x86_fp80 %Val)
6962 declare fp128 @llvm.log.f128(fp128 %Val)
6963 declare ppc_fp128 @llvm.log.ppcf128(ppc_fp128 %Val)
6968 The '``llvm.log.*``' intrinsics perform the log function.
6973 The argument and return value are floating point numbers of the same
6979 This function returns the same values as the libm ``log`` functions
6980 would, and handles error conditions in the same way.
6982 '``llvm.log10.*``' Intrinsic
6983 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6988 This is an overloaded intrinsic. You can use ``llvm.log10`` on any
6989 floating point or vector of floating point type. Not all targets support
6994 declare float @llvm.log10.f32(float %Val)
6995 declare double @llvm.log10.f64(double %Val)
6996 declare x86_fp80 @llvm.log10.f80(x86_fp80 %Val)
6997 declare fp128 @llvm.log10.f128(fp128 %Val)
6998 declare ppc_fp128 @llvm.log10.ppcf128(ppc_fp128 %Val)
7003 The '``llvm.log10.*``' intrinsics perform the log10 function.
7008 The argument and return value are floating point numbers of the same
7014 This function returns the same values as the libm ``log10`` functions
7015 would, and handles error conditions in the same way.
7017 '``llvm.log2.*``' Intrinsic
7018 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7023 This is an overloaded intrinsic. You can use ``llvm.log2`` on any
7024 floating point or vector of floating point type. Not all targets support
7029 declare float @llvm.log2.f32(float %Val)
7030 declare double @llvm.log2.f64(double %Val)
7031 declare x86_fp80 @llvm.log2.f80(x86_fp80 %Val)
7032 declare fp128 @llvm.log2.f128(fp128 %Val)
7033 declare ppc_fp128 @llvm.log2.ppcf128(ppc_fp128 %Val)
7038 The '``llvm.log2.*``' intrinsics perform the log2 function.
7043 The argument and return value are floating point numbers of the same
7049 This function returns the same values as the libm ``log2`` functions
7050 would, and handles error conditions in the same way.
7052 '``llvm.fma.*``' Intrinsic
7053 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7058 This is an overloaded intrinsic. You can use ``llvm.fma`` on any
7059 floating point or vector of floating point type. Not all targets support
7064 declare float @llvm.fma.f32(float %a, float %b, float %c)
7065 declare double @llvm.fma.f64(double %a, double %b, double %c)
7066 declare x86_fp80 @llvm.fma.f80(x86_fp80 %a, x86_fp80 %b, x86_fp80 %c)
7067 declare fp128 @llvm.fma.f128(fp128 %a, fp128 %b, fp128 %c)
7068 declare ppc_fp128 @llvm.fma.ppcf128(ppc_fp128 %a, ppc_fp128 %b, ppc_fp128 %c)
7073 The '``llvm.fma.*``' intrinsics perform the fused multiply-add
7079 The argument and return value are floating point numbers of the same
7085 This function returns the same values as the libm ``fma`` functions
7088 '``llvm.fabs.*``' Intrinsic
7089 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7094 This is an overloaded intrinsic. You can use ``llvm.fabs`` on any
7095 floating point or vector of floating point type. Not all targets support
7100 declare float @llvm.fabs.f32(float %Val)
7101 declare double @llvm.fabs.f64(double %Val)
7102 declare x86_fp80 @llvm.fabs.f80(x86_fp80 %Val)
7103 declare fp128 @llvm.fabs.f128(fp128 %Val)
7104 declare ppc_fp128 @llvm.fabs.ppcf128(ppc_fp128 %Val)
7109 The '``llvm.fabs.*``' intrinsics return the absolute value of the
7115 The argument and return value are floating point numbers of the same
7121 This function returns the same values as the libm ``fabs`` functions
7122 would, and handles error conditions in the same way.
7124 '``llvm.floor.*``' Intrinsic
7125 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7130 This is an overloaded intrinsic. You can use ``llvm.floor`` on any
7131 floating point or vector of floating point type. Not all targets support
7136 declare float @llvm.floor.f32(float %Val)
7137 declare double @llvm.floor.f64(double %Val)
7138 declare x86_fp80 @llvm.floor.f80(x86_fp80 %Val)
7139 declare fp128 @llvm.floor.f128(fp128 %Val)
7140 declare ppc_fp128 @llvm.floor.ppcf128(ppc_fp128 %Val)
7145 The '``llvm.floor.*``' intrinsics return the floor of the operand.
7150 The argument and return value are floating point numbers of the same
7156 This function returns the same values as the libm ``floor`` functions
7157 would, and handles error conditions in the same way.
7159 '``llvm.ceil.*``' Intrinsic
7160 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7165 This is an overloaded intrinsic. You can use ``llvm.ceil`` on any
7166 floating point or vector of floating point type. Not all targets support
7171 declare float @llvm.ceil.f32(float %Val)
7172 declare double @llvm.ceil.f64(double %Val)
7173 declare x86_fp80 @llvm.ceil.f80(x86_fp80 %Val)
7174 declare fp128 @llvm.ceil.f128(fp128 %Val)
7175 declare ppc_fp128 @llvm.ceil.ppcf128(ppc_fp128 %Val)
7180 The '``llvm.ceil.*``' intrinsics return the ceiling of the operand.
7185 The argument and return value are floating point numbers of the same
7191 This function returns the same values as the libm ``ceil`` functions
7192 would, and handles error conditions in the same way.
7194 '``llvm.trunc.*``' Intrinsic
7195 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7200 This is an overloaded intrinsic. You can use ``llvm.trunc`` on any
7201 floating point or vector of floating point type. Not all targets support
7206 declare float @llvm.trunc.f32(float %Val)
7207 declare double @llvm.trunc.f64(double %Val)
7208 declare x86_fp80 @llvm.trunc.f80(x86_fp80 %Val)
7209 declare fp128 @llvm.trunc.f128(fp128 %Val)
7210 declare ppc_fp128 @llvm.trunc.ppcf128(ppc_fp128 %Val)
7215 The '``llvm.trunc.*``' intrinsics returns the operand rounded to the
7216 nearest integer not larger in magnitude than the operand.
7221 The argument and return value are floating point numbers of the same
7227 This function returns the same values as the libm ``trunc`` functions
7228 would, and handles error conditions in the same way.
7230 '``llvm.rint.*``' Intrinsic
7231 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7236 This is an overloaded intrinsic. You can use ``llvm.rint`` on any
7237 floating point or vector of floating point type. Not all targets support
7242 declare float @llvm.rint.f32(float %Val)
7243 declare double @llvm.rint.f64(double %Val)
7244 declare x86_fp80 @llvm.rint.f80(x86_fp80 %Val)
7245 declare fp128 @llvm.rint.f128(fp128 %Val)
7246 declare ppc_fp128 @llvm.rint.ppcf128(ppc_fp128 %Val)
7251 The '``llvm.rint.*``' intrinsics returns the operand rounded to the
7252 nearest integer. It may raise an inexact floating-point exception if the
7253 operand isn't an integer.
7258 The argument and return value are floating point numbers of the same
7264 This function returns the same values as the libm ``rint`` functions
7265 would, and handles error conditions in the same way.
7267 '``llvm.nearbyint.*``' Intrinsic
7268 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7273 This is an overloaded intrinsic. You can use ``llvm.nearbyint`` on any
7274 floating point or vector of floating point type. Not all targets support
7279 declare float @llvm.nearbyint.f32(float %Val)
7280 declare double @llvm.nearbyint.f64(double %Val)
7281 declare x86_fp80 @llvm.nearbyint.f80(x86_fp80 %Val)
7282 declare fp128 @llvm.nearbyint.f128(fp128 %Val)
7283 declare ppc_fp128 @llvm.nearbyint.ppcf128(ppc_fp128 %Val)
7288 The '``llvm.nearbyint.*``' intrinsics returns the operand rounded to the
7294 The argument and return value are floating point numbers of the same
7300 This function returns the same values as the libm ``nearbyint``
7301 functions would, and handles error conditions in the same way.
7303 Bit Manipulation Intrinsics
7304 ---------------------------
7306 LLVM provides intrinsics for a few important bit manipulation
7307 operations. These allow efficient code generation for some algorithms.
7309 '``llvm.bswap.*``' Intrinsics
7310 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7315 This is an overloaded intrinsic function. You can use bswap on any
7316 integer type that is an even number of bytes (i.e. BitWidth % 16 == 0).
7320 declare i16 @llvm.bswap.i16(i16 <id>)
7321 declare i32 @llvm.bswap.i32(i32 <id>)
7322 declare i64 @llvm.bswap.i64(i64 <id>)
7327 The '``llvm.bswap``' family of intrinsics is used to byte swap integer
7328 values with an even number of bytes (positive multiple of 16 bits).
7329 These are useful for performing operations on data that is not in the
7330 target's native byte order.
7335 The ``llvm.bswap.i16`` intrinsic returns an i16 value that has the high
7336 and low byte of the input i16 swapped. Similarly, the ``llvm.bswap.i32``
7337 intrinsic returns an i32 value that has the four bytes of the input i32
7338 swapped, so that if the input bytes are numbered 0, 1, 2, 3 then the
7339 returned i32 will have its bytes in 3, 2, 1, 0 order. The
7340 ``llvm.bswap.i48``, ``llvm.bswap.i64`` and other intrinsics extend this
7341 concept to additional even-byte lengths (6 bytes, 8 bytes and more,
7344 '``llvm.ctpop.*``' Intrinsic
7345 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7350 This is an overloaded intrinsic. You can use llvm.ctpop on any integer
7351 bit width, or on any vector with integer elements. Not all targets
7352 support all bit widths or vector types, however.
7356 declare i8 @llvm.ctpop.i8(i8 <src>)
7357 declare i16 @llvm.ctpop.i16(i16 <src>)
7358 declare i32 @llvm.ctpop.i32(i32 <src>)
7359 declare i64 @llvm.ctpop.i64(i64 <src>)
7360 declare i256 @llvm.ctpop.i256(i256 <src>)
7361 declare <2 x i32> @llvm.ctpop.v2i32(<2 x i32> <src>)
7366 The '``llvm.ctpop``' family of intrinsics counts the number of bits set
7372 The only argument is the value to be counted. The argument may be of any
7373 integer type, or a vector with integer elements. The return type must
7374 match the argument type.
7379 The '``llvm.ctpop``' intrinsic counts the 1's in a variable, or within
7380 each element of a vector.
7382 '``llvm.ctlz.*``' Intrinsic
7383 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7388 This is an overloaded intrinsic. You can use ``llvm.ctlz`` on any
7389 integer bit width, or any vector whose elements are integers. Not all
7390 targets support all bit widths or vector types, however.
7394 declare i8 @llvm.ctlz.i8 (i8 <src>, i1 <is_zero_undef>)
7395 declare i16 @llvm.ctlz.i16 (i16 <src>, i1 <is_zero_undef>)
7396 declare i32 @llvm.ctlz.i32 (i32 <src>, i1 <is_zero_undef>)
7397 declare i64 @llvm.ctlz.i64 (i64 <src>, i1 <is_zero_undef>)
7398 declare i256 @llvm.ctlz.i256(i256 <src>, i1 <is_zero_undef>)
7399 declase <2 x i32> @llvm.ctlz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
7404 The '``llvm.ctlz``' family of intrinsic functions counts the number of
7405 leading zeros in a variable.
7410 The first argument is the value to be counted. This argument may be of
7411 any integer type, or a vectory with integer element type. The return
7412 type must match the first argument type.
7414 The second argument must be a constant and is a flag to indicate whether
7415 the intrinsic should ensure that a zero as the first argument produces a
7416 defined result. Historically some architectures did not provide a
7417 defined result for zero values as efficiently, and many algorithms are
7418 now predicated on avoiding zero-value inputs.
7423 The '``llvm.ctlz``' intrinsic counts the leading (most significant)
7424 zeros in a variable, or within each element of the vector. If
7425 ``src == 0`` then the result is the size in bits of the type of ``src``
7426 if ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
7427 ``llvm.ctlz(i32 2) = 30``.
7429 '``llvm.cttz.*``' Intrinsic
7430 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7435 This is an overloaded intrinsic. You can use ``llvm.cttz`` on any
7436 integer bit width, or any vector of integer elements. Not all targets
7437 support all bit widths or vector types, however.
7441 declare i8 @llvm.cttz.i8 (i8 <src>, i1 <is_zero_undef>)
7442 declare i16 @llvm.cttz.i16 (i16 <src>, i1 <is_zero_undef>)
7443 declare i32 @llvm.cttz.i32 (i32 <src>, i1 <is_zero_undef>)
7444 declare i64 @llvm.cttz.i64 (i64 <src>, i1 <is_zero_undef>)
7445 declare i256 @llvm.cttz.i256(i256 <src>, i1 <is_zero_undef>)
7446 declase <2 x i32> @llvm.cttz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
7451 The '``llvm.cttz``' family of intrinsic functions counts the number of
7457 The first argument is the value to be counted. This argument may be of
7458 any integer type, or a vectory with integer element type. The return
7459 type must match the first argument type.
7461 The second argument must be a constant and is a flag to indicate whether
7462 the intrinsic should ensure that a zero as the first argument produces a
7463 defined result. Historically some architectures did not provide a
7464 defined result for zero values as efficiently, and many algorithms are
7465 now predicated on avoiding zero-value inputs.
7470 The '``llvm.cttz``' intrinsic counts the trailing (least significant)
7471 zeros in a variable, or within each element of a vector. If ``src == 0``
7472 then the result is the size in bits of the type of ``src`` if
7473 ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
7474 ``llvm.cttz(2) = 1``.
7476 Arithmetic with Overflow Intrinsics
7477 -----------------------------------
7479 LLVM provides intrinsics for some arithmetic with overflow operations.
7481 '``llvm.sadd.with.overflow.*``' Intrinsics
7482 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7487 This is an overloaded intrinsic. You can use ``llvm.sadd.with.overflow``
7488 on any integer bit width.
7492 declare {i16, i1} @llvm.sadd.with.overflow.i16(i16 %a, i16 %b)
7493 declare {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
7494 declare {i64, i1} @llvm.sadd.with.overflow.i64(i64 %a, i64 %b)
7499 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
7500 a signed addition of the two arguments, and indicate whether an overflow
7501 occurred during the signed summation.
7506 The arguments (%a and %b) and the first element of the result structure
7507 may be of integer types of any bit width, but they must have the same
7508 bit width. The second element of the result structure must be of type
7509 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
7515 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
7516 a signed addition of the two variables. They return a structure --- the
7517 first element of which is the signed summation, and the second element
7518 of which is a bit specifying if the signed summation resulted in an
7524 .. code-block:: llvm
7526 %res = call {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
7527 %sum = extractvalue {i32, i1} %res, 0
7528 %obit = extractvalue {i32, i1} %res, 1
7529 br i1 %obit, label %overflow, label %normal
7531 '``llvm.uadd.with.overflow.*``' Intrinsics
7532 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7537 This is an overloaded intrinsic. You can use ``llvm.uadd.with.overflow``
7538 on any integer bit width.
7542 declare {i16, i1} @llvm.uadd.with.overflow.i16(i16 %a, i16 %b)
7543 declare {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
7544 declare {i64, i1} @llvm.uadd.with.overflow.i64(i64 %a, i64 %b)
7549 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
7550 an unsigned addition of the two arguments, and indicate whether a carry
7551 occurred during the unsigned summation.
7556 The arguments (%a and %b) and the first element of the result structure
7557 may be of integer types of any bit width, but they must have the same
7558 bit width. The second element of the result structure must be of type
7559 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
7565 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
7566 an unsigned addition of the two arguments. They return a structure --- the
7567 first element of which is the sum, and the second element of which is a
7568 bit specifying if the unsigned summation resulted in a carry.
7573 .. code-block:: llvm
7575 %res = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
7576 %sum = extractvalue {i32, i1} %res, 0
7577 %obit = extractvalue {i32, i1} %res, 1
7578 br i1 %obit, label %carry, label %normal
7580 '``llvm.ssub.with.overflow.*``' Intrinsics
7581 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7586 This is an overloaded intrinsic. You can use ``llvm.ssub.with.overflow``
7587 on any integer bit width.
7591 declare {i16, i1} @llvm.ssub.with.overflow.i16(i16 %a, i16 %b)
7592 declare {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
7593 declare {i64, i1} @llvm.ssub.with.overflow.i64(i64 %a, i64 %b)
7598 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
7599 a signed subtraction of the two arguments, and indicate whether an
7600 overflow occurred during the signed subtraction.
7605 The arguments (%a and %b) and the first element of the result structure
7606 may be of integer types of any bit width, but they must have the same
7607 bit width. The second element of the result structure must be of type
7608 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
7614 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
7615 a signed subtraction of the two arguments. They return a structure --- the
7616 first element of which is the subtraction, and the second element of
7617 which is a bit specifying if the signed subtraction resulted in an
7623 .. code-block:: llvm
7625 %res = call {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
7626 %sum = extractvalue {i32, i1} %res, 0
7627 %obit = extractvalue {i32, i1} %res, 1
7628 br i1 %obit, label %overflow, label %normal
7630 '``llvm.usub.with.overflow.*``' Intrinsics
7631 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7636 This is an overloaded intrinsic. You can use ``llvm.usub.with.overflow``
7637 on any integer bit width.
7641 declare {i16, i1} @llvm.usub.with.overflow.i16(i16 %a, i16 %b)
7642 declare {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
7643 declare {i64, i1} @llvm.usub.with.overflow.i64(i64 %a, i64 %b)
7648 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
7649 an unsigned subtraction of the two arguments, and indicate whether an
7650 overflow occurred during the unsigned subtraction.
7655 The arguments (%a and %b) and the first element of the result structure
7656 may be of integer types of any bit width, but they must have the same
7657 bit width. The second element of the result structure must be of type
7658 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
7664 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
7665 an unsigned subtraction of the two arguments. They return a structure ---
7666 the first element of which is the subtraction, and the second element of
7667 which is a bit specifying if the unsigned subtraction resulted in an
7673 .. code-block:: llvm
7675 %res = call {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
7676 %sum = extractvalue {i32, i1} %res, 0
7677 %obit = extractvalue {i32, i1} %res, 1
7678 br i1 %obit, label %overflow, label %normal
7680 '``llvm.smul.with.overflow.*``' Intrinsics
7681 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7686 This is an overloaded intrinsic. You can use ``llvm.smul.with.overflow``
7687 on any integer bit width.
7691 declare {i16, i1} @llvm.smul.with.overflow.i16(i16 %a, i16 %b)
7692 declare {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
7693 declare {i64, i1} @llvm.smul.with.overflow.i64(i64 %a, i64 %b)
7698 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
7699 a signed multiplication of the two arguments, and indicate whether an
7700 overflow occurred during the signed multiplication.
7705 The arguments (%a and %b) and the first element of the result structure
7706 may be of integer types of any bit width, but they must have the same
7707 bit width. The second element of the result structure must be of type
7708 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
7714 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
7715 a signed multiplication of the two arguments. They return a structure ---
7716 the first element of which is the multiplication, and the second element
7717 of which is a bit specifying if the signed multiplication resulted in an
7723 .. code-block:: llvm
7725 %res = call {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
7726 %sum = extractvalue {i32, i1} %res, 0
7727 %obit = extractvalue {i32, i1} %res, 1
7728 br i1 %obit, label %overflow, label %normal
7730 '``llvm.umul.with.overflow.*``' Intrinsics
7731 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7736 This is an overloaded intrinsic. You can use ``llvm.umul.with.overflow``
7737 on any integer bit width.
7741 declare {i16, i1} @llvm.umul.with.overflow.i16(i16 %a, i16 %b)
7742 declare {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
7743 declare {i64, i1} @llvm.umul.with.overflow.i64(i64 %a, i64 %b)
7748 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
7749 a unsigned multiplication of the two arguments, and indicate whether an
7750 overflow occurred during the unsigned multiplication.
7755 The arguments (%a and %b) and the first element of the result structure
7756 may be of integer types of any bit width, but they must have the same
7757 bit width. The second element of the result structure must be of type
7758 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
7764 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
7765 an unsigned multiplication of the two arguments. They return a structure ---
7766 the first element of which is the multiplication, and the second
7767 element of which is a bit specifying if the unsigned multiplication
7768 resulted in an overflow.
7773 .. code-block:: llvm
7775 %res = call {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
7776 %sum = extractvalue {i32, i1} %res, 0
7777 %obit = extractvalue {i32, i1} %res, 1
7778 br i1 %obit, label %overflow, label %normal
7780 Specialised Arithmetic Intrinsics
7781 ---------------------------------
7783 '``llvm.fmuladd.*``' Intrinsic
7784 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7791 declare float @llvm.fmuladd.f32(float %a, float %b, float %c)
7792 declare double @llvm.fmuladd.f64(double %a, double %b, double %c)
7797 The '``llvm.fmuladd.*``' intrinsic functions represent multiply-add
7798 expressions that can be fused if the code generator determines that (a) the
7799 target instruction set has support for a fused operation, and (b) that the
7800 fused operation is more efficient than the equivalent, separate pair of mul
7801 and add instructions.
7806 The '``llvm.fmuladd.*``' intrinsics each take three arguments: two
7807 multiplicands, a and b, and an addend c.
7816 %0 = call float @llvm.fmuladd.f32(%a, %b, %c)
7818 is equivalent to the expression a \* b + c, except that rounding will
7819 not be performed between the multiplication and addition steps if the
7820 code generator fuses the operations. Fusion is not guaranteed, even if
7821 the target platform supports it. If a fused multiply-add is required the
7822 corresponding llvm.fma.\* intrinsic function should be used instead.
7827 .. code-block:: llvm
7829 %r2 = call float @llvm.fmuladd.f32(float %a, float %b, float %c) ; yields {float}:r2 = (a * b) + c
7831 Half Precision Floating Point Intrinsics
7832 ----------------------------------------
7834 For most target platforms, half precision floating point is a
7835 storage-only format. This means that it is a dense encoding (in memory)
7836 but does not support computation in the format.
7838 This means that code must first load the half-precision floating point
7839 value as an i16, then convert it to float with
7840 :ref:`llvm.convert.from.fp16 <int_convert_from_fp16>`. Computation can
7841 then be performed on the float value (including extending to double
7842 etc). To store the value back to memory, it is first converted to float
7843 if needed, then converted to i16 with
7844 :ref:`llvm.convert.to.fp16 <int_convert_to_fp16>`, then storing as an
7847 .. _int_convert_to_fp16:
7849 '``llvm.convert.to.fp16``' Intrinsic
7850 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7857 declare i16 @llvm.convert.to.fp16(f32 %a)
7862 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion
7863 from single precision floating point format to half precision floating
7869 The intrinsic function contains single argument - the value to be
7875 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion
7876 from single precision floating point format to half precision floating
7877 point format. The return value is an ``i16`` which contains the
7883 .. code-block:: llvm
7885 %res = call i16 @llvm.convert.to.fp16(f32 %a)
7886 store i16 %res, i16* @x, align 2
7888 .. _int_convert_from_fp16:
7890 '``llvm.convert.from.fp16``' Intrinsic
7891 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7898 declare f32 @llvm.convert.from.fp16(i16 %a)
7903 The '``llvm.convert.from.fp16``' intrinsic function performs a
7904 conversion from half precision floating point format to single precision
7905 floating point format.
7910 The intrinsic function contains single argument - the value to be
7916 The '``llvm.convert.from.fp16``' intrinsic function performs a
7917 conversion from half single precision floating point format to single
7918 precision floating point format. The input half-float value is
7919 represented by an ``i16`` value.
7924 .. code-block:: llvm
7926 %a = load i16* @x, align 2
7927 %res = call f32 @llvm.convert.from.fp16(i16 %a)
7932 The LLVM debugger intrinsics (which all start with ``llvm.dbg.``
7933 prefix), are described in the `LLVM Source Level
7934 Debugging <SourceLevelDebugging.html#format_common_intrinsics>`_
7937 Exception Handling Intrinsics
7938 -----------------------------
7940 The LLVM exception handling intrinsics (which all start with
7941 ``llvm.eh.`` prefix), are described in the `LLVM Exception
7942 Handling <ExceptionHandling.html#format_common_intrinsics>`_ document.
7946 Trampoline Intrinsics
7947 ---------------------
7949 These intrinsics make it possible to excise one parameter, marked with
7950 the :ref:`nest <nest>` attribute, from a function. The result is a
7951 callable function pointer lacking the nest parameter - the caller does
7952 not need to provide a value for it. Instead, the value to use is stored
7953 in advance in a "trampoline", a block of memory usually allocated on the
7954 stack, which also contains code to splice the nest value into the
7955 argument list. This is used to implement the GCC nested function address
7958 For example, if the function is ``i32 f(i8* nest %c, i32 %x, i32 %y)``
7959 then the resulting function pointer has signature ``i32 (i32, i32)*``.
7960 It can be created as follows:
7962 .. code-block:: llvm
7964 %tramp = alloca [10 x i8], align 4 ; size and alignment only correct for X86
7965 %tramp1 = getelementptr [10 x i8]* %tramp, i32 0, i32 0
7966 call i8* @llvm.init.trampoline(i8* %tramp1, i8* bitcast (i32 (i8*, i32, i32)* @f to i8*), i8* %nval)
7967 %p = call i8* @llvm.adjust.trampoline(i8* %tramp1)
7968 %fp = bitcast i8* %p to i32 (i32, i32)*
7970 The call ``%val = call i32 %fp(i32 %x, i32 %y)`` is then equivalent to
7971 ``%val = call i32 %f(i8* %nval, i32 %x, i32 %y)``.
7975 '``llvm.init.trampoline``' Intrinsic
7976 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7983 declare void @llvm.init.trampoline(i8* <tramp>, i8* <func>, i8* <nval>)
7988 This fills the memory pointed to by ``tramp`` with executable code,
7989 turning it into a trampoline.
7994 The ``llvm.init.trampoline`` intrinsic takes three arguments, all
7995 pointers. The ``tramp`` argument must point to a sufficiently large and
7996 sufficiently aligned block of memory; this memory is written to by the
7997 intrinsic. Note that the size and the alignment are target-specific -
7998 LLVM currently provides no portable way of determining them, so a
7999 front-end that generates this intrinsic needs to have some
8000 target-specific knowledge. The ``func`` argument must hold a function
8001 bitcast to an ``i8*``.
8006 The block of memory pointed to by ``tramp`` is filled with target
8007 dependent code, turning it into a function. Then ``tramp`` needs to be
8008 passed to :ref:`llvm.adjust.trampoline <int_at>` to get a pointer which can
8009 be :ref:`bitcast (to a new function) and called <int_trampoline>`. The new
8010 function's signature is the same as that of ``func`` with any arguments
8011 marked with the ``nest`` attribute removed. At most one such ``nest``
8012 argument is allowed, and it must be of pointer type. Calling the new
8013 function is equivalent to calling ``func`` with the same argument list,
8014 but with ``nval`` used for the missing ``nest`` argument. If, after
8015 calling ``llvm.init.trampoline``, the memory pointed to by ``tramp`` is
8016 modified, then the effect of any later call to the returned function
8017 pointer is undefined.
8021 '``llvm.adjust.trampoline``' Intrinsic
8022 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8029 declare i8* @llvm.adjust.trampoline(i8* <tramp>)
8034 This performs any required machine-specific adjustment to the address of
8035 a trampoline (passed as ``tramp``).
8040 ``tramp`` must point to a block of memory which already has trampoline
8041 code filled in by a previous call to
8042 :ref:`llvm.init.trampoline <int_it>`.
8047 On some architectures the address of the code to be executed needs to be
8048 different to the address where the trampoline is actually stored. This
8049 intrinsic returns the executable address corresponding to ``tramp``
8050 after performing the required machine specific adjustments. The pointer
8051 returned can then be :ref:`bitcast and executed <int_trampoline>`.
8056 This class of intrinsics exists to information about the lifetime of
8057 memory objects and ranges where variables are immutable.
8059 '``llvm.lifetime.start``' Intrinsic
8060 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8067 declare void @llvm.lifetime.start(i64 <size>, i8* nocapture <ptr>)
8072 The '``llvm.lifetime.start``' intrinsic specifies the start of a memory
8078 The first argument is a constant integer representing the size of the
8079 object, or -1 if it is variable sized. The second argument is a pointer
8085 This intrinsic indicates that before this point in the code, the value
8086 of the memory pointed to by ``ptr`` is dead. This means that it is known
8087 to never be used and has an undefined value. A load from the pointer
8088 that precedes this intrinsic can be replaced with ``'undef'``.
8090 '``llvm.lifetime.end``' Intrinsic
8091 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8098 declare void @llvm.lifetime.end(i64 <size>, i8* nocapture <ptr>)
8103 The '``llvm.lifetime.end``' intrinsic specifies the end of a memory
8109 The first argument is a constant integer representing the size of the
8110 object, or -1 if it is variable sized. The second argument is a pointer
8116 This intrinsic indicates that after this point in the code, the value of
8117 the memory pointed to by ``ptr`` is dead. This means that it is known to
8118 never be used and has an undefined value. Any stores into the memory
8119 object following this intrinsic may be removed as dead.
8121 '``llvm.invariant.start``' Intrinsic
8122 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8129 declare {}* @llvm.invariant.start(i64 <size>, i8* nocapture <ptr>)
8134 The '``llvm.invariant.start``' intrinsic specifies that the contents of
8135 a memory object will not change.
8140 The first argument is a constant integer representing the size of the
8141 object, or -1 if it is variable sized. The second argument is a pointer
8147 This intrinsic indicates that until an ``llvm.invariant.end`` that uses
8148 the return value, the referenced memory location is constant and
8151 '``llvm.invariant.end``' Intrinsic
8152 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8159 declare void @llvm.invariant.end({}* <start>, i64 <size>, i8* nocapture <ptr>)
8164 The '``llvm.invariant.end``' intrinsic specifies that the contents of a
8165 memory object are mutable.
8170 The first argument is the matching ``llvm.invariant.start`` intrinsic.
8171 The second argument is a constant integer representing the size of the
8172 object, or -1 if it is variable sized and the third argument is a
8173 pointer to the object.
8178 This intrinsic indicates that the memory is mutable again.
8183 This class of intrinsics is designed to be generic and has no specific
8186 '``llvm.var.annotation``' Intrinsic
8187 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8194 declare void @llvm.var.annotation(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
8199 The '``llvm.var.annotation``' intrinsic.
8204 The first argument is a pointer to a value, the second is a pointer to a
8205 global string, the third is a pointer to a global string which is the
8206 source file name, and the last argument is the line number.
8211 This intrinsic allows annotation of local variables with arbitrary
8212 strings. This can be useful for special purpose optimizations that want
8213 to look for these annotations. These have no other defined use; they are
8214 ignored by code generation and optimization.
8216 '``llvm.annotation.*``' Intrinsic
8217 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8222 This is an overloaded intrinsic. You can use '``llvm.annotation``' on
8223 any integer bit width.
8227 declare i8 @llvm.annotation.i8(i8 <val>, i8* <str>, i8* <str>, i32 <int>)
8228 declare i16 @llvm.annotation.i16(i16 <val>, i8* <str>, i8* <str>, i32 <int>)
8229 declare i32 @llvm.annotation.i32(i32 <val>, i8* <str>, i8* <str>, i32 <int>)
8230 declare i64 @llvm.annotation.i64(i64 <val>, i8* <str>, i8* <str>, i32 <int>)
8231 declare i256 @llvm.annotation.i256(i256 <val>, i8* <str>, i8* <str>, i32 <int>)
8236 The '``llvm.annotation``' intrinsic.
8241 The first argument is an integer value (result of some expression), the
8242 second is a pointer to a global string, the third is a pointer to a
8243 global string which is the source file name, and the last argument is
8244 the line number. It returns the value of the first argument.
8249 This intrinsic allows annotations to be put on arbitrary expressions
8250 with arbitrary strings. This can be useful for special purpose
8251 optimizations that want to look for these annotations. These have no
8252 other defined use; they are ignored by code generation and optimization.
8254 '``llvm.trap``' Intrinsic
8255 ^^^^^^^^^^^^^^^^^^^^^^^^^
8262 declare void @llvm.trap() noreturn nounwind
8267 The '``llvm.trap``' intrinsic.
8277 This intrinsic is lowered to the target dependent trap instruction. If
8278 the target does not have a trap instruction, this intrinsic will be
8279 lowered to a call of the ``abort()`` function.
8281 '``llvm.debugtrap``' Intrinsic
8282 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8289 declare void @llvm.debugtrap() nounwind
8294 The '``llvm.debugtrap``' intrinsic.
8304 This intrinsic is lowered to code which is intended to cause an
8305 execution trap with the intention of requesting the attention of a
8308 '``llvm.stackprotector``' Intrinsic
8309 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8316 declare void @llvm.stackprotector(i8* <guard>, i8** <slot>)
8321 The ``llvm.stackprotector`` intrinsic takes the ``guard`` and stores it
8322 onto the stack at ``slot``. The stack slot is adjusted to ensure that it
8323 is placed on the stack before local variables.
8328 The ``llvm.stackprotector`` intrinsic requires two pointer arguments.
8329 The first argument is the value loaded from the stack guard
8330 ``@__stack_chk_guard``. The second variable is an ``alloca`` that has
8331 enough space to hold the value of the guard.
8336 This intrinsic causes the prologue/epilogue inserter to force the
8337 position of the ``AllocaInst`` stack slot to be before local variables
8338 on the stack. This is to ensure that if a local variable on the stack is
8339 overwritten, it will destroy the value of the guard. When the function
8340 exits, the guard on the stack is checked against the original guard. If
8341 they are different, then the program aborts by calling the
8342 ``__stack_chk_fail()`` function.
8344 '``llvm.objectsize``' Intrinsic
8345 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8352 declare i32 @llvm.objectsize.i32(i8* <object>, i1 <min>)
8353 declare i64 @llvm.objectsize.i64(i8* <object>, i1 <min>)
8358 The ``llvm.objectsize`` intrinsic is designed to provide information to
8359 the optimizers to determine at compile time whether a) an operation
8360 (like memcpy) will overflow a buffer that corresponds to an object, or
8361 b) that a runtime check for overflow isn't necessary. An object in this
8362 context means an allocation of a specific class, structure, array, or
8368 The ``llvm.objectsize`` intrinsic takes two arguments. The first
8369 argument is a pointer to or into the ``object``. The second argument is
8370 a boolean and determines whether ``llvm.objectsize`` returns 0 (if true)
8371 or -1 (if false) when the object size is unknown. The second argument
8372 only accepts constants.
8377 The ``llvm.objectsize`` intrinsic is lowered to a constant representing
8378 the size of the object concerned. If the size cannot be determined at
8379 compile time, ``llvm.objectsize`` returns ``i32/i64 -1 or 0`` (depending
8380 on the ``min`` argument).
8382 '``llvm.expect``' Intrinsic
8383 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8390 declare i32 @llvm.expect.i32(i32 <val>, i32 <expected_val>)
8391 declare i64 @llvm.expect.i64(i64 <val>, i64 <expected_val>)
8396 The ``llvm.expect`` intrinsic provides information about expected (the
8397 most probable) value of ``val``, which can be used by optimizers.
8402 The ``llvm.expect`` intrinsic takes two arguments. The first argument is
8403 a value. The second argument is an expected value, this needs to be a
8404 constant value, variables are not allowed.
8409 This intrinsic is lowered to the ``val``.
8411 '``llvm.donothing``' Intrinsic
8412 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8419 declare void @llvm.donothing() nounwind readnone
8424 The ``llvm.donothing`` intrinsic doesn't perform any operation. It's the
8425 only intrinsic that can be called with an invoke instruction.
8435 This intrinsic does nothing, and it's removed by optimizers and ignored