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 marked with ``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 An explicit alignment may be specified for a global, which must be a
505 power of 2. If not present, or if the alignment is set to zero, the
506 alignment of the global is set by the target to whatever it feels
507 convenient. If an explicit alignment is specified, the global is forced
508 to have exactly that alignment. Targets and optimizers are not allowed
509 to over-align the global if the global has an assigned section. In this
510 case, the extra alignment could be observable: for example, code could
511 assume that the globals are densely packed in their section and try to
512 iterate over them as an array, alignment padding would break this
515 For example, the following defines a global in a numbered address space
516 with an initializer, section, and alignment:
520 @G = addrspace(5) constant float 1.0, section "foo", align 4
522 The following example defines a thread-local global with the
523 ``initialexec`` TLS model:
527 @G = thread_local(initialexec) global i32 0, align 4
529 .. _functionstructure:
534 LLVM function definitions consist of the "``define``" keyword, an
535 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
536 style <visibility>`, an optional :ref:`calling convention <callingconv>`,
537 an optional ``unnamed_addr`` attribute, a return type, an optional
538 :ref:`parameter attribute <paramattrs>` for the return type, a function
539 name, a (possibly empty) argument list (each with optional :ref:`parameter
540 attributes <paramattrs>`), optional :ref:`function attributes <fnattrs>`,
541 an optional section, an optional alignment, an optional :ref:`garbage
542 collector name <gc>`, an opening curly brace, a list of basic blocks,
543 and a closing curly brace.
545 LLVM function declarations consist of the "``declare``" keyword, an
546 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
547 style <visibility>`, an optional :ref:`calling convention <callingconv>`,
548 an optional ``unnamed_addr`` attribute, a return type, an optional
549 :ref:`parameter attribute <paramattrs>` for the return type, a function
550 name, a possibly empty list of arguments, an optional alignment, and an
551 optional :ref:`garbage collector name <gc>`.
553 A function definition contains a list of basic blocks, forming the CFG
554 (Control Flow Graph) for the function. Each basic block may optionally
555 start with a label (giving the basic block a symbol table entry),
556 contains a list of instructions, and ends with a
557 :ref:`terminator <terminators>` instruction (such as a branch or function
560 The first basic block in a function is special in two ways: it is
561 immediately executed on entrance to the function, and it is not allowed
562 to have predecessor basic blocks (i.e. there can not be any branches to
563 the entry block of a function). Because the block can have no
564 predecessors, it also cannot have any :ref:`PHI nodes <i_phi>`.
566 LLVM allows an explicit section to be specified for functions. If the
567 target supports it, it will emit functions to the section specified.
569 An explicit alignment may be specified for a function. If not present,
570 or if the alignment is set to zero, the alignment of the function is set
571 by the target to whatever it feels convenient. If an explicit alignment
572 is specified, the function is forced to have at least that much
573 alignment. All alignments must be a power of 2.
575 If the ``unnamed_addr`` attribute is given, the address is know to not
576 be significant and two identical functions can be merged.
580 define [linkage] [visibility]
582 <ResultType> @<FunctionName> ([argument list])
583 [fn Attrs] [section "name"] [align N]
589 Aliases act as "second name" for the aliasee value (which can be either
590 function, global variable, another alias or bitcast of global value).
591 Aliases may have an optional :ref:`linkage type <linkage>`, and an optional
592 :ref:`visibility style <visibility>`.
596 @<Name> = alias [Linkage] [Visibility] <AliaseeTy> @<Aliasee>
598 .. _namedmetadatastructure:
603 Named metadata is a collection of metadata. :ref:`Metadata
604 nodes <metadata>` (but not metadata strings) are the only valid
605 operands for a named metadata.
609 ; Some unnamed metadata nodes, which are referenced by the named metadata.
610 !0 = metadata !{metadata !"zero"}
611 !1 = metadata !{metadata !"one"}
612 !2 = metadata !{metadata !"two"}
614 !name = !{!0, !1, !2}
621 The return type and each parameter of a function type may have a set of
622 *parameter attributes* associated with them. Parameter attributes are
623 used to communicate additional information about the result or
624 parameters of a function. Parameter attributes are considered to be part
625 of the function, not of the function type, so functions with different
626 parameter attributes can have the same function type.
628 Parameter attributes are simple keywords that follow the type specified.
629 If multiple parameter attributes are needed, they are space separated.
634 declare i32 @printf(i8* noalias nocapture, ...)
635 declare i32 @atoi(i8 zeroext)
636 declare signext i8 @returns_signed_char()
638 Note that any attributes for the function result (``nounwind``,
639 ``readonly``) come immediately after the argument list.
641 Currently, only the following parameter attributes are defined:
644 This indicates to the code generator that the parameter or return
645 value should be zero-extended to the extent required by the target's
646 ABI (which is usually 32-bits, but is 8-bits for a i1 on x86-64) by
647 the caller (for a parameter) or the callee (for a return value).
649 This indicates to the code generator that the parameter or return
650 value should be sign-extended to the extent required by the target's
651 ABI (which is usually 32-bits) by the caller (for a parameter) or
652 the callee (for a return value).
654 This indicates that this parameter or return value should be treated
655 in a special target-dependent fashion during while emitting code for
656 a function call or return (usually, by putting it in a register as
657 opposed to memory, though some targets use it to distinguish between
658 two different kinds of registers). Use of this attribute is
661 This indicates that the pointer parameter should really be passed by
662 value to the function. The attribute implies that a hidden copy of
663 the pointee is made between the caller and the callee, so the callee
664 is unable to modify the value in the caller. This attribute is only
665 valid on LLVM pointer arguments. It is generally used to pass
666 structs and arrays by value, but is also valid on pointers to
667 scalars. The copy is considered to belong to the caller not the
668 callee (for example, ``readonly`` functions should not write to
669 ``byval`` parameters). This is not a valid attribute for return
672 The byval attribute also supports specifying an alignment with the
673 align attribute. It indicates the alignment of the stack slot to
674 form and the known alignment of the pointer specified to the call
675 site. If the alignment is not specified, then the code generator
676 makes a target-specific assumption.
679 This indicates that the pointer parameter specifies the address of a
680 structure that is the return value of the function in the source
681 program. This pointer must be guaranteed by the caller to be valid:
682 loads and stores to the structure may be assumed by the callee
683 not to trap and to be properly aligned. This may only be applied to
684 the first parameter. This is not a valid attribute for return
687 This indicates that pointer values `*based* <pointeraliasing>` on
688 the argument or return value do not alias pointer values which are
689 not *based* on it, ignoring certain "irrelevant" dependencies. For a
690 call to the parent function, dependencies between memory references
691 from before or after the call and from those during the call are
692 "irrelevant" to the ``noalias`` keyword for the arguments and return
693 value used in that call. The caller shares the responsibility with
694 the callee for ensuring that these requirements are met. For further
695 details, please see the discussion of the NoAlias response in `alias
696 analysis <AliasAnalysis.html#MustMayNo>`_.
698 Note that this definition of ``noalias`` is intentionally similar
699 to the definition of ``restrict`` in C99 for function arguments,
700 though it is slightly weaker.
702 For function return values, C99's ``restrict`` is not meaningful,
703 while LLVM's ``noalias`` is.
705 This indicates that the callee does not make any copies of the
706 pointer that outlive the callee itself. This is not a valid
707 attribute for return values.
712 This indicates that the pointer parameter can be excised using the
713 :ref:`trampoline intrinsics <int_trampoline>`. This is not a valid
714 attribute for return values.
718 Garbage Collector Names
719 -----------------------
721 Each function may specify a garbage collector name, which is simply a
726 define void @f() gc "name" { ... }
728 The compiler declares the supported values of *name*. Specifying a
729 collector which will cause the compiler to alter its output in order to
730 support the named garbage collection algorithm.
737 Function attributes are set to communicate additional information about
738 a function. Function attributes are considered to be part of the
739 function, not of the function type, so functions with different function
740 attributes can have the same function type.
742 Function attributes are simple keywords that follow the type specified.
743 If multiple attributes are needed, they are space separated. For
748 define void @f() noinline { ... }
749 define void @f() alwaysinline { ... }
750 define void @f() alwaysinline optsize { ... }
751 define void @f() optsize { ... }
754 This attribute indicates that the address safety analysis is enabled
757 This attribute indicates that, when emitting the prologue and
758 epilogue, the backend should forcibly align the stack pointer.
759 Specify the desired alignment, which must be a power of two, in
762 This attribute indicates that the inliner should attempt to inline
763 this function into callers whenever possible, ignoring any active
764 inlining size threshold for this caller.
766 This attribute suppresses lazy symbol binding for the function. This
767 may make calls to the function faster, at the cost of extra program
768 startup time if the function is not called during program startup.
770 This attribute indicates that the source code contained a hint that
771 inlining this function is desirable (such as the "inline" keyword in
772 C/C++). It is just a hint; it imposes no requirements on the
775 This attribute disables prologue / epilogue emission for the
776 function. This can have very system-specific consequences.
778 This attributes disables implicit floating point instructions.
780 This attribute indicates that the inliner should never inline this
781 function in any situation. This attribute may not be used together
782 with the ``alwaysinline`` attribute.
784 This attribute indicates that the code generator should not use a
785 red zone, even if the target-specific ABI normally permits it.
787 This function attribute indicates that the function never returns
788 normally. This produces undefined behavior at runtime if the
789 function ever does dynamically return.
791 This function attribute indicates that the function never returns
792 with an unwind or exceptional control flow. If the function does
793 unwind, its runtime behavior is undefined.
795 This attribute suggests that optimization passes and code generator
796 passes make choices that keep the code size of this function low,
797 and otherwise do optimizations specifically to reduce code size.
799 This attribute indicates that the function computes its result (or
800 decides to unwind an exception) based strictly on its arguments,
801 without dereferencing any pointer arguments or otherwise accessing
802 any mutable state (e.g. memory, control registers, etc) visible to
803 caller functions. It does not write through any pointer arguments
804 (including ``byval`` arguments) and never changes any state visible
805 to callers. This means that it cannot unwind exceptions by calling
806 the ``C++`` exception throwing methods.
808 This attribute indicates that the function does not write through
809 any pointer arguments (including ``byval`` arguments) or otherwise
810 modify any state (e.g. memory, control registers, etc) visible to
811 caller functions. It may dereference pointer arguments and read
812 state that may be set in the caller. A readonly function always
813 returns the same value (or unwinds an exception identically) when
814 called with the same set of arguments and global state. It cannot
815 unwind an exception by calling the ``C++`` exception throwing
818 This attribute indicates that this function can return twice. The C
819 ``setjmp`` is an example of such a function. The compiler disables
820 some optimizations (like tail calls) in the caller of these
823 This attribute indicates that the function should emit a stack
824 smashing protector. It is in the form of a "canary" --- a random value
825 placed on the stack before the local variables that's checked upon
826 return from the function to see if it has been overwritten. A
827 heuristic is used to determine if a function needs stack protectors
828 or not. The heuristic used will enable protectors for functions with:
830 - Character arrays larger than ``ssp-buffer-size`` (default 8).
831 - Aggregates containing character arrays larger than ``ssp-buffer-size``.
832 - Calls to alloca() with variable sizes or constant sizes greater than
835 If a function that has an ``ssp`` attribute is inlined into a
836 function that doesn't have an ``ssp`` attribute, then the resulting
837 function will have an ``ssp`` attribute.
839 This attribute indicates that the function should *always* emit a
840 stack smashing protector. This overrides the ``ssp`` function
843 If a function that has an ``sspreq`` attribute is inlined into a
844 function that doesn't have an ``sspreq`` attribute or which has an
845 ``ssp`` or ``sspstrong`` attribute, then the resulting function will have
846 an ``sspreq`` attribute.
848 This attribute indicates that the function should emit a stack smashing
849 protector. This attribute causes a strong heuristic to be used when
850 determining if a function needs stack protectors. The strong heuristic
851 will enable protectors for functions with:
853 - Arrays of any size and type
854 - Aggregates containing an array of any size and type.
856 - Local variables that have had their address taken.
858 This overrides the ``ssp`` function attribute.
860 If a function that has an ``sspstrong`` attribute is inlined into a
861 function that doesn't have an ``sspstrong`` attribute, then the
862 resulting function will have an ``sspstrong`` attribute.
864 This attribute indicates that the ABI being targeted requires that
865 an unwind table entry be produce for this function even if we can
866 show that no exceptions passes by it. This is normally the case for
867 the ELF x86-64 abi, but it can be disabled for some compilation
870 This attribute indicates that calls to the function cannot be
871 duplicated. A call to a ``noduplicate`` function may be moved
872 within its parent function, but may not be duplicated within
875 A function containing a ``noduplicate`` call may still
876 be an inlining candidate, provided that the call is not
877 duplicated by inlining. That implies that the function has
878 internal linkage and only has one call site, so the original
879 call is dead after inlining.
883 Module-Level Inline Assembly
884 ----------------------------
886 Modules may contain "module-level inline asm" blocks, which corresponds
887 to the GCC "file scope inline asm" blocks. These blocks are internally
888 concatenated by LLVM and treated as a single unit, but may be separated
889 in the ``.ll`` file if desired. The syntax is very simple:
893 module asm "inline asm code goes here"
894 module asm "more can go here"
896 The strings can contain any character by escaping non-printable
897 characters. The escape sequence used is simply "\\xx" where "xx" is the
898 two digit hex code for the number.
900 The inline asm code is simply printed to the machine code .s file when
901 assembly code is generated.
906 A module may specify a target specific data layout string that specifies
907 how data is to be laid out in memory. The syntax for the data layout is
912 target datalayout = "layout specification"
914 The *layout specification* consists of a list of specifications
915 separated by the minus sign character ('-'). Each specification starts
916 with a letter and may include other information after the letter to
917 define some aspect of the data layout. The specifications accepted are
921 Specifies that the target lays out data in big-endian form. That is,
922 the bits with the most significance have the lowest address
925 Specifies that the target lays out data in little-endian form. That
926 is, the bits with the least significance have the lowest address
929 Specifies the natural alignment of the stack in bits. Alignment
930 promotion of stack variables is limited to the natural stack
931 alignment to avoid dynamic stack realignment. The stack alignment
932 must be a multiple of 8-bits. If omitted, the natural stack
933 alignment defaults to "unspecified", which does not prevent any
934 alignment promotions.
935 ``p[n]:<size>:<abi>:<pref>``
936 This specifies the *size* of a pointer and its ``<abi>`` and
937 ``<pref>``\erred alignments for address space ``n``. All sizes are in
938 bits. Specifying the ``<pref>`` alignment is optional. If omitted, the
939 preceding ``:`` should be omitted too. The address space, ``n`` is
940 optional, and if not specified, denotes the default address space 0.
941 The value of ``n`` must be in the range [1,2^23).
942 ``i<size>:<abi>:<pref>``
943 This specifies the alignment for an integer type of a given bit
944 ``<size>``. The value of ``<size>`` must be in the range [1,2^23).
945 ``v<size>:<abi>:<pref>``
946 This specifies the alignment for a vector type of a given bit
948 ``f<size>:<abi>:<pref>``
949 This specifies the alignment for a floating point type of a given bit
950 ``<size>``. Only values of ``<size>`` that are supported by the target
951 will work. 32 (float) and 64 (double) are supported on all targets; 80
952 or 128 (different flavors of long double) are also supported on some
954 ``a<size>:<abi>:<pref>``
955 This specifies the alignment for an aggregate type of a given bit
957 ``s<size>:<abi>:<pref>``
958 This specifies the alignment for a stack object of a given bit
960 ``n<size1>:<size2>:<size3>...``
961 This specifies a set of native integer widths for the target CPU in
962 bits. For example, it might contain ``n32`` for 32-bit PowerPC,
963 ``n32:64`` for PowerPC 64, or ``n8:16:32:64`` for X86-64. Elements of
964 this set are considered to support most general arithmetic operations
967 When constructing the data layout for a given target, LLVM starts with a
968 default set of specifications which are then (possibly) overridden by
969 the specifications in the ``datalayout`` keyword. The default
970 specifications are given in this list:
973 - ``p:64:64:64`` - 64-bit pointers with 64-bit alignment
974 - ``S0`` - natural stack alignment is unspecified
975 - ``i1:8:8`` - i1 is 8-bit (byte) aligned
976 - ``i8:8:8`` - i8 is 8-bit (byte) aligned
977 - ``i16:16:16`` - i16 is 16-bit aligned
978 - ``i32:32:32`` - i32 is 32-bit aligned
979 - ``i64:32:64`` - i64 has ABI alignment of 32-bits but preferred
981 - ``f16:16:16`` - half is 16-bit aligned
982 - ``f32:32:32`` - float is 32-bit aligned
983 - ``f64:64:64`` - double is 64-bit aligned
984 - ``f128:128:128`` - quad is 128-bit aligned
985 - ``v64:64:64`` - 64-bit vector is 64-bit aligned
986 - ``v128:128:128`` - 128-bit vector is 128-bit aligned
987 - ``a0:0:64`` - aggregates are 64-bit aligned
989 When LLVM is determining the alignment for a given type, it uses the
992 #. If the type sought is an exact match for one of the specifications,
993 that specification is used.
994 #. If no match is found, and the type sought is an integer type, then
995 the smallest integer type that is larger than the bitwidth of the
996 sought type is used. If none of the specifications are larger than
997 the bitwidth then the largest integer type is used. For example,
998 given the default specifications above, the i7 type will use the
999 alignment of i8 (next largest) while both i65 and i256 will use the
1000 alignment of i64 (largest specified).
1001 #. If no match is found, and the type sought is a vector type, then the
1002 largest vector type that is smaller than the sought vector type will
1003 be used as a fall back. This happens because <128 x double> can be
1004 implemented in terms of 64 <2 x double>, for example.
1006 The function of the data layout string may not be what you expect.
1007 Notably, this is not a specification from the frontend of what alignment
1008 the code generator should use.
1010 Instead, if specified, the target data layout is required to match what
1011 the ultimate *code generator* expects. This string is used by the
1012 mid-level optimizers to improve code, and this only works if it matches
1013 what the ultimate code generator uses. If you would like to generate IR
1014 that does not embed this target-specific detail into the IR, then you
1015 don't have to specify the string. This will disable some optimizations
1016 that require precise layout information, but this also prevents those
1017 optimizations from introducing target specificity into the IR.
1019 .. _pointeraliasing:
1021 Pointer Aliasing Rules
1022 ----------------------
1024 Any memory access must be done through a pointer value associated with
1025 an address range of the memory access, otherwise the behavior is
1026 undefined. Pointer values are associated with address ranges according
1027 to the following rules:
1029 - A pointer value is associated with the addresses associated with any
1030 value it is *based* on.
1031 - An address of a global variable is associated with the address range
1032 of the variable's storage.
1033 - The result value of an allocation instruction is associated with the
1034 address range of the allocated storage.
1035 - A null pointer in the default address-space is associated with no
1037 - An integer constant other than zero or a pointer value returned from
1038 a function not defined within LLVM may be associated with address
1039 ranges allocated through mechanisms other than those provided by
1040 LLVM. Such ranges shall not overlap with any ranges of addresses
1041 allocated by mechanisms provided by LLVM.
1043 A pointer value is *based* on another pointer value according to the
1046 - A pointer value formed from a ``getelementptr`` operation is *based*
1047 on the first operand of the ``getelementptr``.
1048 - The result value of a ``bitcast`` is *based* on the operand of the
1050 - A pointer value formed by an ``inttoptr`` is *based* on all pointer
1051 values that contribute (directly or indirectly) to the computation of
1052 the pointer's value.
1053 - The "*based* on" relationship is transitive.
1055 Note that this definition of *"based"* is intentionally similar to the
1056 definition of *"based"* in C99, though it is slightly weaker.
1058 LLVM IR does not associate types with memory. The result type of a
1059 ``load`` merely indicates the size and alignment of the memory from
1060 which to load, as well as the interpretation of the value. The first
1061 operand type of a ``store`` similarly only indicates the size and
1062 alignment of the store.
1064 Consequently, type-based alias analysis, aka TBAA, aka
1065 ``-fstrict-aliasing``, is not applicable to general unadorned LLVM IR.
1066 :ref:`Metadata <metadata>` may be used to encode additional information
1067 which specialized optimization passes may use to implement type-based
1072 Volatile Memory Accesses
1073 ------------------------
1075 Certain memory accesses, such as :ref:`load <i_load>`'s,
1076 :ref:`store <i_store>`'s, and :ref:`llvm.memcpy <int_memcpy>`'s may be
1077 marked ``volatile``. The optimizers must not change the number of
1078 volatile operations or change their order of execution relative to other
1079 volatile operations. The optimizers *may* change the order of volatile
1080 operations relative to non-volatile operations. This is not Java's
1081 "volatile" and has no cross-thread synchronization behavior.
1083 IR-level volatile loads and stores cannot safely be optimized into
1084 llvm.memcpy or llvm.memmove intrinsics even when those intrinsics are
1085 flagged volatile. Likewise, the backend should never split or merge
1086 target-legal volatile load/store instructions.
1088 .. admonition:: Rationale
1090 Platforms may rely on volatile loads and stores of natively supported
1091 data width to be executed as single instruction. For example, in C
1092 this holds for an l-value of volatile primitive type with native
1093 hardware support, but not necessarily for aggregate types. The
1094 frontend upholds these expectations, which are intentionally
1095 unspecified in the IR. The rules above ensure that IR transformation
1096 do not violate the frontend's contract with the language.
1100 Memory Model for Concurrent Operations
1101 --------------------------------------
1103 The LLVM IR does not define any way to start parallel threads of
1104 execution or to register signal handlers. Nonetheless, there are
1105 platform-specific ways to create them, and we define LLVM IR's behavior
1106 in their presence. This model is inspired by the C++0x memory model.
1108 For a more informal introduction to this model, see the :doc:`Atomics`.
1110 We define a *happens-before* partial order as the least partial order
1113 - Is a superset of single-thread program order, and
1114 - When a *synchronizes-with* ``b``, includes an edge from ``a`` to
1115 ``b``. *Synchronizes-with* pairs are introduced by platform-specific
1116 techniques, like pthread locks, thread creation, thread joining,
1117 etc., and by atomic instructions. (See also :ref:`Atomic Memory Ordering
1118 Constraints <ordering>`).
1120 Note that program order does not introduce *happens-before* edges
1121 between a thread and signals executing inside that thread.
1123 Every (defined) read operation (load instructions, memcpy, atomic
1124 loads/read-modify-writes, etc.) R reads a series of bytes written by
1125 (defined) write operations (store instructions, atomic
1126 stores/read-modify-writes, memcpy, etc.). For the purposes of this
1127 section, initialized globals are considered to have a write of the
1128 initializer which is atomic and happens before any other read or write
1129 of the memory in question. For each byte of a read R, R\ :sub:`byte`
1130 may see any write to the same byte, except:
1132 - If write\ :sub:`1` happens before write\ :sub:`2`, and
1133 write\ :sub:`2` happens before R\ :sub:`byte`, then
1134 R\ :sub:`byte` does not see write\ :sub:`1`.
1135 - If R\ :sub:`byte` happens before write\ :sub:`3`, then
1136 R\ :sub:`byte` does not see write\ :sub:`3`.
1138 Given that definition, R\ :sub:`byte` is defined as follows:
1140 - If R is volatile, the result is target-dependent. (Volatile is
1141 supposed to give guarantees which can support ``sig_atomic_t`` in
1142 C/C++, and may be used for accesses to addresses which do not behave
1143 like normal memory. It does not generally provide cross-thread
1145 - Otherwise, if there is no write to the same byte that happens before
1146 R\ :sub:`byte`, R\ :sub:`byte` returns ``undef`` for that byte.
1147 - Otherwise, if R\ :sub:`byte` may see exactly one write,
1148 R\ :sub:`byte` returns the value written by that write.
1149 - Otherwise, if R is atomic, and all the writes R\ :sub:`byte` may
1150 see are atomic, it chooses one of the values written. See the :ref:`Atomic
1151 Memory Ordering Constraints <ordering>` section for additional
1152 constraints on how the choice is made.
1153 - Otherwise R\ :sub:`byte` returns ``undef``.
1155 R returns the value composed of the series of bytes it read. This
1156 implies that some bytes within the value may be ``undef`` **without**
1157 the entire value being ``undef``. Note that this only defines the
1158 semantics of the operation; it doesn't mean that targets will emit more
1159 than one instruction to read the series of bytes.
1161 Note that in cases where none of the atomic intrinsics are used, this
1162 model places only one restriction on IR transformations on top of what
1163 is required for single-threaded execution: introducing a store to a byte
1164 which might not otherwise be stored is not allowed in general.
1165 (Specifically, in the case where another thread might write to and read
1166 from an address, introducing a store can change a load that may see
1167 exactly one write into a load that may see multiple writes.)
1171 Atomic Memory Ordering Constraints
1172 ----------------------------------
1174 Atomic instructions (:ref:`cmpxchg <i_cmpxchg>`,
1175 :ref:`atomicrmw <i_atomicrmw>`, :ref:`fence <i_fence>`,
1176 :ref:`atomic load <i_load>`, and :ref:`atomic store <i_store>`) take
1177 an ordering parameter that determines which other atomic instructions on
1178 the same address they *synchronize with*. These semantics are borrowed
1179 from Java and C++0x, but are somewhat more colloquial. If these
1180 descriptions aren't precise enough, check those specs (see spec
1181 references in the :doc:`atomics guide <Atomics>`).
1182 :ref:`fence <i_fence>` instructions treat these orderings somewhat
1183 differently since they don't take an address. See that instruction's
1184 documentation for details.
1186 For a simpler introduction to the ordering constraints, see the
1190 The set of values that can be read is governed by the happens-before
1191 partial order. A value cannot be read unless some operation wrote
1192 it. This is intended to provide a guarantee strong enough to model
1193 Java's non-volatile shared variables. This ordering cannot be
1194 specified for read-modify-write operations; it is not strong enough
1195 to make them atomic in any interesting way.
1197 In addition to the guarantees of ``unordered``, there is a single
1198 total order for modifications by ``monotonic`` operations on each
1199 address. All modification orders must be compatible with the
1200 happens-before order. There is no guarantee that the modification
1201 orders can be combined to a global total order for the whole program
1202 (and this often will not be possible). The read in an atomic
1203 read-modify-write operation (:ref:`cmpxchg <i_cmpxchg>` and
1204 :ref:`atomicrmw <i_atomicrmw>`) reads the value in the modification
1205 order immediately before the value it writes. If one atomic read
1206 happens before another atomic read of the same address, the later
1207 read must see the same value or a later value in the address's
1208 modification order. This disallows reordering of ``monotonic`` (or
1209 stronger) operations on the same address. If an address is written
1210 ``monotonic``-ally by one thread, and other threads ``monotonic``-ally
1211 read that address repeatedly, the other threads must eventually see
1212 the write. This corresponds to the C++0x/C1x
1213 ``memory_order_relaxed``.
1215 In addition to the guarantees of ``monotonic``, a
1216 *synchronizes-with* edge may be formed with a ``release`` operation.
1217 This is intended to model C++'s ``memory_order_acquire``.
1219 In addition to the guarantees of ``monotonic``, if this operation
1220 writes a value which is subsequently read by an ``acquire``
1221 operation, it *synchronizes-with* that operation. (This isn't a
1222 complete description; see the C++0x definition of a release
1223 sequence.) This corresponds to the C++0x/C1x
1224 ``memory_order_release``.
1225 ``acq_rel`` (acquire+release)
1226 Acts as both an ``acquire`` and ``release`` operation on its
1227 address. This corresponds to the C++0x/C1x ``memory_order_acq_rel``.
1228 ``seq_cst`` (sequentially consistent)
1229 In addition to the guarantees of ``acq_rel`` (``acquire`` for an
1230 operation which only reads, ``release`` for an operation which only
1231 writes), there is a global total order on all
1232 sequentially-consistent operations on all addresses, which is
1233 consistent with the *happens-before* partial order and with the
1234 modification orders of all the affected addresses. Each
1235 sequentially-consistent read sees the last preceding write to the
1236 same address in this global order. This corresponds to the C++0x/C1x
1237 ``memory_order_seq_cst`` and Java volatile.
1241 If an atomic operation is marked ``singlethread``, it only *synchronizes
1242 with* or participates in modification and seq\_cst total orderings with
1243 other operations running in the same thread (for example, in signal
1251 LLVM IR floating-point binary ops (:ref:`fadd <i_fadd>`,
1252 :ref:`fsub <i_fsub>`, :ref:`fmul <i_fmul>`, :ref:`fdiv <i_fdiv>`,
1253 :ref:`frem <i_frem>`) have the following flags that can set to enable
1254 otherwise unsafe floating point operations
1257 No NaNs - Allow optimizations to assume the arguments and result are not
1258 NaN. Such optimizations are required to retain defined behavior over
1259 NaNs, but the value of the result is undefined.
1262 No Infs - Allow optimizations to assume the arguments and result are not
1263 +/-Inf. Such optimizations are required to retain defined behavior over
1264 +/-Inf, but the value of the result is undefined.
1267 No Signed Zeros - Allow optimizations to treat the sign of a zero
1268 argument or result as insignificant.
1271 Allow Reciprocal - Allow optimizations to use the reciprocal of an
1272 argument rather than perform division.
1275 Fast - Allow algebraically equivalent transformations that may
1276 dramatically change results in floating point (e.g. reassociate). This
1277 flag implies all the others.
1284 The LLVM type system is one of the most important features of the
1285 intermediate representation. Being typed enables a number of
1286 optimizations to be performed on the intermediate representation
1287 directly, without having to do extra analyses on the side before the
1288 transformation. A strong type system makes it easier to read the
1289 generated code and enables novel analyses and transformations that are
1290 not feasible to perform on normal three address code representations.
1292 Type Classifications
1293 --------------------
1295 The types fall into a few useful classifications:
1304 * - :ref:`integer <t_integer>`
1305 - ``i1``, ``i2``, ``i3``, ... ``i8``, ... ``i16``, ... ``i32``, ...
1308 * - :ref:`floating point <t_floating>`
1309 - ``half``, ``float``, ``double``, ``x86_fp80``, ``fp128``,
1317 - :ref:`integer <t_integer>`, :ref:`floating point <t_floating>`,
1318 :ref:`pointer <t_pointer>`, :ref:`vector <t_vector>`,
1319 :ref:`structure <t_struct>`, :ref:`array <t_array>`,
1320 :ref:`label <t_label>`, :ref:`metadata <t_metadata>`.
1322 * - :ref:`primitive <t_primitive>`
1323 - :ref:`label <t_label>`,
1324 :ref:`void <t_void>`,
1325 :ref:`integer <t_integer>`,
1326 :ref:`floating point <t_floating>`,
1327 :ref:`x86mmx <t_x86mmx>`,
1328 :ref:`metadata <t_metadata>`.
1330 * - :ref:`derived <t_derived>`
1331 - :ref:`array <t_array>`,
1332 :ref:`function <t_function>`,
1333 :ref:`pointer <t_pointer>`,
1334 :ref:`structure <t_struct>`,
1335 :ref:`vector <t_vector>`,
1336 :ref:`opaque <t_opaque>`.
1338 The :ref:`first class <t_firstclass>` types are perhaps the most important.
1339 Values of these types are the only ones which can be produced by
1347 The primitive types are the fundamental building blocks of the LLVM
1358 The integer type is a very simple type that simply specifies an
1359 arbitrary bit width for the integer type desired. Any bit width from 1
1360 bit to 2\ :sup:`23`\ -1 (about 8 million) can be specified.
1369 The number of bits the integer will occupy is specified by the ``N``
1375 +----------------+------------------------------------------------+
1376 | ``i1`` | a single-bit integer. |
1377 +----------------+------------------------------------------------+
1378 | ``i32`` | a 32-bit integer. |
1379 +----------------+------------------------------------------------+
1380 | ``i1942652`` | a really big integer of over 1 million bits. |
1381 +----------------+------------------------------------------------+
1385 Floating Point Types
1386 ^^^^^^^^^^^^^^^^^^^^
1395 - 16-bit floating point value
1398 - 32-bit floating point value
1401 - 64-bit floating point value
1404 - 128-bit floating point value (112-bit mantissa)
1407 - 80-bit floating point value (X87)
1410 - 128-bit floating point value (two 64-bits)
1420 The x86mmx type represents a value held in an MMX register on an x86
1421 machine. The operations allowed on it are quite limited: parameters and
1422 return values, load and store, and bitcast. User-specified MMX
1423 instructions are represented as intrinsic or asm calls with arguments
1424 and/or results of this type. There are no arrays, vectors or constants
1442 The void type does not represent any value and has no size.
1459 The label type represents code labels.
1476 The metadata type represents embedded metadata. No derived types may be
1477 created from metadata except for :ref:`function <t_function>` arguments.
1491 The real power in LLVM comes from the derived types in the system. This
1492 is what allows a programmer to represent arrays, functions, pointers,
1493 and other useful types. Each of these types contain one or more element
1494 types which may be a primitive type, or another derived type. For
1495 example, it is possible to have a two dimensional array, using an array
1496 as the element type of another array.
1503 Aggregate Types are a subset of derived types that can contain multiple
1504 member types. :ref:`Arrays <t_array>` and :ref:`structs <t_struct>` are
1505 aggregate types. :ref:`Vectors <t_vector>` are not considered to be
1516 The array type is a very simple derived type that arranges elements
1517 sequentially in memory. The array type requires a size (number of
1518 elements) and an underlying data type.
1525 [<# elements> x <elementtype>]
1527 The number of elements is a constant integer value; ``elementtype`` may
1528 be any type with a size.
1533 +------------------+--------------------------------------+
1534 | ``[40 x i32]`` | Array of 40 32-bit integer values. |
1535 +------------------+--------------------------------------+
1536 | ``[41 x i32]`` | Array of 41 32-bit integer values. |
1537 +------------------+--------------------------------------+
1538 | ``[4 x i8]`` | Array of 4 8-bit integer values. |
1539 +------------------+--------------------------------------+
1541 Here are some examples of multidimensional arrays:
1543 +-----------------------------+----------------------------------------------------------+
1544 | ``[3 x [4 x i32]]`` | 3x4 array of 32-bit integer values. |
1545 +-----------------------------+----------------------------------------------------------+
1546 | ``[12 x [10 x float]]`` | 12x10 array of single precision floating point values. |
1547 +-----------------------------+----------------------------------------------------------+
1548 | ``[2 x [3 x [4 x i16]]]`` | 2x3x4 array of 16-bit integer values. |
1549 +-----------------------------+----------------------------------------------------------+
1551 There is no restriction on indexing beyond the end of the array implied
1552 by a static type (though there are restrictions on indexing beyond the
1553 bounds of an allocated object in some cases). This means that
1554 single-dimension 'variable sized array' addressing can be implemented in
1555 LLVM with a zero length array type. An implementation of 'pascal style
1556 arrays' in LLVM could use the type "``{ i32, [0 x float]}``", for
1567 The function type can be thought of as a function signature. It consists
1568 of a return type and a list of formal parameter types. The return type
1569 of a function type is a first class type or a void type.
1576 <returntype> (<parameter list>)
1578 ...where '``<parameter list>``' is a comma-separated list of type
1579 specifiers. Optionally, the parameter list may include a type ``...``,
1580 which indicates that the function takes a variable number of arguments.
1581 Variable argument functions can access their arguments with the
1582 :ref:`variable argument handling intrinsic <int_varargs>` functions.
1583 '``<returntype>``' is any type except :ref:`label <t_label>`.
1588 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1589 | ``i32 (i32)`` | function taking an ``i32``, returning an ``i32`` |
1590 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1591 | ``float (i16, i32 *) *`` | :ref:`Pointer <t_pointer>` to a function that takes an ``i16`` and a :ref:`pointer <t_pointer>` to ``i32``, returning ``float``. |
1592 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1593 | ``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. |
1594 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1595 | ``{i32, i32} (i32)`` | A function taking an ``i32``, returning a :ref:`structure <t_struct>` containing two ``i32`` values |
1596 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1606 The structure type is used to represent a collection of data members
1607 together in memory. The elements of a structure may be any type that has
1610 Structures in memory are accessed using '``load``' and '``store``' by
1611 getting a pointer to a field with the '``getelementptr``' instruction.
1612 Structures in registers are accessed using the '``extractvalue``' and
1613 '``insertvalue``' instructions.
1615 Structures may optionally be "packed" structures, which indicate that
1616 the alignment of the struct is one byte, and that there is no padding
1617 between the elements. In non-packed structs, padding between field types
1618 is inserted as defined by the DataLayout string in the module, which is
1619 required to match what the underlying code generator expects.
1621 Structures can either be "literal" or "identified". A literal structure
1622 is defined inline with other types (e.g. ``{i32, i32}*``) whereas
1623 identified types are always defined at the top level with a name.
1624 Literal types are uniqued by their contents and can never be recursive
1625 or opaque since there is no way to write one. Identified types can be
1626 recursive, can be opaqued, and are never uniqued.
1633 %T1 = type { <type list> } ; Identified normal struct type
1634 %T2 = type <{ <type list> }> ; Identified packed struct type
1639 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1640 | ``{ i32, i32, i32 }`` | A triple of three ``i32`` values |
1641 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1642 | ``{ 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``. |
1643 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1644 | ``<{ i8, i32 }>`` | A packed struct known to be 5 bytes in size. |
1645 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1649 Opaque Structure Types
1650 ^^^^^^^^^^^^^^^^^^^^^^
1655 Opaque structure types are used to represent named structure types that
1656 do not have a body specified. This corresponds (for example) to the C
1657 notion of a forward declared structure.
1670 +--------------+-------------------+
1671 | ``opaque`` | An opaque type. |
1672 +--------------+-------------------+
1682 The pointer type is used to specify memory locations. Pointers are
1683 commonly used to reference objects in memory.
1685 Pointer types may have an optional address space attribute defining the
1686 numbered address space where the pointed-to object resides. The default
1687 address space is number zero. The semantics of non-zero address spaces
1688 are target-specific.
1690 Note that LLVM does not permit pointers to void (``void*``) nor does it
1691 permit pointers to labels (``label*``). Use ``i8*`` instead.
1703 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1704 | ``[4 x i32]*`` | A :ref:`pointer <t_pointer>` to :ref:`array <t_array>` of four ``i32`` values. |
1705 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1706 | ``i32 (i32*) *`` | A :ref:`pointer <t_pointer>` to a :ref:`function <t_function>` that takes an ``i32*``, returning an ``i32``. |
1707 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1708 | ``i32 addrspace(5)*`` | A :ref:`pointer <t_pointer>` to an ``i32`` value that resides in address space #5. |
1709 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1719 A vector type is a simple derived type that represents a vector of
1720 elements. Vector types are used when multiple primitive data are
1721 operated in parallel using a single instruction (SIMD). A vector type
1722 requires a size (number of elements) and an underlying primitive data
1723 type. Vector types are considered :ref:`first class <t_firstclass>`.
1730 < <# elements> x <elementtype> >
1732 The number of elements is a constant integer value larger than 0;
1733 elementtype may be any integer or floating point type, or a pointer to
1734 these types. Vectors of size zero are not allowed.
1739 +-------------------+--------------------------------------------------+
1740 | ``<4 x i32>`` | Vector of 4 32-bit integer values. |
1741 +-------------------+--------------------------------------------------+
1742 | ``<8 x float>`` | Vector of 8 32-bit floating-point values. |
1743 +-------------------+--------------------------------------------------+
1744 | ``<2 x i64>`` | Vector of 2 64-bit integer values. |
1745 +-------------------+--------------------------------------------------+
1746 | ``<4 x i64*>`` | Vector of 4 pointers to 64-bit integer values. |
1747 +-------------------+--------------------------------------------------+
1752 LLVM has several different basic types of constants. This section
1753 describes them all and their syntax.
1758 **Boolean constants**
1759 The two strings '``true``' and '``false``' are both valid constants
1761 **Integer constants**
1762 Standard integers (such as '4') are constants of the
1763 :ref:`integer <t_integer>` type. Negative numbers may be used with
1765 **Floating point constants**
1766 Floating point constants use standard decimal notation (e.g.
1767 123.421), exponential notation (e.g. 1.23421e+2), or a more precise
1768 hexadecimal notation (see below). The assembler requires the exact
1769 decimal value of a floating-point constant. For example, the
1770 assembler accepts 1.25 but rejects 1.3 because 1.3 is a repeating
1771 decimal in binary. Floating point constants must have a :ref:`floating
1772 point <t_floating>` type.
1773 **Null pointer constants**
1774 The identifier '``null``' is recognized as a null pointer constant
1775 and must be of :ref:`pointer type <t_pointer>`.
1777 The one non-intuitive notation for constants is the hexadecimal form of
1778 floating point constants. For example, the form
1779 '``double 0x432ff973cafa8000``' is equivalent to (but harder to read
1780 than) '``double 4.5e+15``'. The only time hexadecimal floating point
1781 constants are required (and the only time that they are generated by the
1782 disassembler) is when a floating point constant must be emitted but it
1783 cannot be represented as a decimal floating point number in a reasonable
1784 number of digits. For example, NaN's, infinities, and other special
1785 values are represented in their IEEE hexadecimal format so that assembly
1786 and disassembly do not cause any bits to change in the constants.
1788 When using the hexadecimal form, constants of types half, float, and
1789 double are represented using the 16-digit form shown above (which
1790 matches the IEEE754 representation for double); half and float values
1791 must, however, be exactly representable as IEEE 754 half and single
1792 precision, respectively. Hexadecimal format is always used for long
1793 double, and there are three forms of long double. The 80-bit format used
1794 by x86 is represented as ``0xK`` followed by 20 hexadecimal digits. The
1795 128-bit format used by PowerPC (two adjacent doubles) is represented by
1796 ``0xM`` followed by 32 hexadecimal digits. The IEEE 128-bit format is
1797 represented by ``0xL`` followed by 32 hexadecimal digits; no currently
1798 supported target uses this format. Long doubles will only work if they
1799 match the long double format on your target. The IEEE 16-bit format
1800 (half precision) is represented by ``0xH`` followed by 4 hexadecimal
1801 digits. All hexadecimal formats are big-endian (sign bit at the left).
1803 There are no constants of type x86mmx.
1808 Complex constants are a (potentially recursive) combination of simple
1809 constants and smaller complex constants.
1811 **Structure constants**
1812 Structure constants are represented with notation similar to
1813 structure type definitions (a comma separated list of elements,
1814 surrounded by braces (``{}``)). For example:
1815 "``{ i32 4, float 17.0, i32* @G }``", where "``@G``" is declared as
1816 "``@G = external global i32``". Structure constants must have
1817 :ref:`structure type <t_struct>`, and the number and types of elements
1818 must match those specified by the type.
1820 Array constants are represented with notation similar to array type
1821 definitions (a comma separated list of elements, surrounded by
1822 square brackets (``[]``)). For example:
1823 "``[ i32 42, i32 11, i32 74 ]``". Array constants must have
1824 :ref:`array type <t_array>`, and the number and types of elements must
1825 match those specified by the type.
1826 **Vector constants**
1827 Vector constants are represented with notation similar to vector
1828 type definitions (a comma separated list of elements, surrounded by
1829 less-than/greater-than's (``<>``)). For example:
1830 "``< i32 42, i32 11, i32 74, i32 100 >``". Vector constants
1831 must have :ref:`vector type <t_vector>`, and the number and types of
1832 elements must match those specified by the type.
1833 **Zero initialization**
1834 The string '``zeroinitializer``' can be used to zero initialize a
1835 value to zero of *any* type, including scalar and
1836 :ref:`aggregate <t_aggregate>` types. This is often used to avoid
1837 having to print large zero initializers (e.g. for large arrays) and
1838 is always exactly equivalent to using explicit zero initializers.
1840 A metadata node is a structure-like constant with :ref:`metadata
1841 type <t_metadata>`. For example:
1842 "``metadata !{ i32 0, metadata !"test" }``". Unlike other
1843 constants that are meant to be interpreted as part of the
1844 instruction stream, metadata is a place to attach additional
1845 information such as debug info.
1847 Global Variable and Function Addresses
1848 --------------------------------------
1850 The addresses of :ref:`global variables <globalvars>` and
1851 :ref:`functions <functionstructure>` are always implicitly valid
1852 (link-time) constants. These constants are explicitly referenced when
1853 the :ref:`identifier for the global <identifiers>` is used and always have
1854 :ref:`pointer <t_pointer>` type. For example, the following is a legal LLVM
1857 .. code-block:: llvm
1861 @Z = global [2 x i32*] [ i32* @X, i32* @Y ]
1868 The string '``undef``' can be used anywhere a constant is expected, and
1869 indicates that the user of the value may receive an unspecified
1870 bit-pattern. Undefined values may be of any type (other than '``label``'
1871 or '``void``') and be used anywhere a constant is permitted.
1873 Undefined values are useful because they indicate to the compiler that
1874 the program is well defined no matter what value is used. This gives the
1875 compiler more freedom to optimize. Here are some examples of
1876 (potentially surprising) transformations that are valid (in pseudo IR):
1878 .. code-block:: llvm
1888 This is safe because all of the output bits are affected by the undef
1889 bits. Any output bit can have a zero or one depending on the input bits.
1891 .. code-block:: llvm
1902 These logical operations have bits that are not always affected by the
1903 input. For example, if ``%X`` has a zero bit, then the output of the
1904 '``and``' operation will always be a zero for that bit, no matter what
1905 the corresponding bit from the '``undef``' is. As such, it is unsafe to
1906 optimize or assume that the result of the '``and``' is '``undef``'.
1907 However, it is safe to assume that all bits of the '``undef``' could be
1908 0, and optimize the '``and``' to 0. Likewise, it is safe to assume that
1909 all the bits of the '``undef``' operand to the '``or``' could be set,
1910 allowing the '``or``' to be folded to -1.
1912 .. code-block:: llvm
1914 %A = select undef, %X, %Y
1915 %B = select undef, 42, %Y
1916 %C = select %X, %Y, undef
1926 This set of examples shows that undefined '``select``' (and conditional
1927 branch) conditions can go *either way*, but they have to come from one
1928 of the two operands. In the ``%A`` example, if ``%X`` and ``%Y`` were
1929 both known to have a clear low bit, then ``%A`` would have to have a
1930 cleared low bit. However, in the ``%C`` example, the optimizer is
1931 allowed to assume that the '``undef``' operand could be the same as
1932 ``%Y``, allowing the whole '``select``' to be eliminated.
1934 .. code-block:: llvm
1936 %A = xor undef, undef
1953 This example points out that two '``undef``' operands are not
1954 necessarily the same. This can be surprising to people (and also matches
1955 C semantics) where they assume that "``X^X``" is always zero, even if
1956 ``X`` is undefined. This isn't true for a number of reasons, but the
1957 short answer is that an '``undef``' "variable" can arbitrarily change
1958 its value over its "live range". This is true because the variable
1959 doesn't actually *have a live range*. Instead, the value is logically
1960 read from arbitrary registers that happen to be around when needed, so
1961 the value is not necessarily consistent over time. In fact, ``%A`` and
1962 ``%C`` need to have the same semantics or the core LLVM "replace all
1963 uses with" concept would not hold.
1965 .. code-block:: llvm
1973 These examples show the crucial difference between an *undefined value*
1974 and *undefined behavior*. An undefined value (like '``undef``') is
1975 allowed to have an arbitrary bit-pattern. This means that the ``%A``
1976 operation can be constant folded to '``undef``', because the '``undef``'
1977 could be an SNaN, and ``fdiv`` is not (currently) defined on SNaN's.
1978 However, in the second example, we can make a more aggressive
1979 assumption: because the ``undef`` is allowed to be an arbitrary value,
1980 we are allowed to assume that it could be zero. Since a divide by zero
1981 has *undefined behavior*, we are allowed to assume that the operation
1982 does not execute at all. This allows us to delete the divide and all
1983 code after it. Because the undefined operation "can't happen", the
1984 optimizer can assume that it occurs in dead code.
1986 .. code-block:: llvm
1988 a: store undef -> %X
1989 b: store %X -> undef
1994 These examples reiterate the ``fdiv`` example: a store *of* an undefined
1995 value can be assumed to not have any effect; we can assume that the
1996 value is overwritten with bits that happen to match what was already
1997 there. However, a store *to* an undefined location could clobber
1998 arbitrary memory, therefore, it has undefined behavior.
2005 Poison values are similar to :ref:`undef values <undefvalues>`, however
2006 they also represent the fact that an instruction or constant expression
2007 which cannot evoke side effects has nevertheless detected a condition
2008 which results in undefined behavior.
2010 There is currently no way of representing a poison value in the IR; they
2011 only exist when produced by operations such as :ref:`add <i_add>` with
2014 Poison value behavior is defined in terms of value *dependence*:
2016 - Values other than :ref:`phi <i_phi>` nodes depend on their operands.
2017 - :ref:`Phi <i_phi>` nodes depend on the operand corresponding to
2018 their dynamic predecessor basic block.
2019 - Function arguments depend on the corresponding actual argument values
2020 in the dynamic callers of their functions.
2021 - :ref:`Call <i_call>` instructions depend on the :ref:`ret <i_ret>`
2022 instructions that dynamically transfer control back to them.
2023 - :ref:`Invoke <i_invoke>` instructions depend on the
2024 :ref:`ret <i_ret>`, :ref:`resume <i_resume>`, or exception-throwing
2025 call instructions that dynamically transfer control back to them.
2026 - Non-volatile loads and stores depend on the most recent stores to all
2027 of the referenced memory addresses, following the order in the IR
2028 (including loads and stores implied by intrinsics such as
2029 :ref:`@llvm.memcpy <int_memcpy>`.)
2030 - An instruction with externally visible side effects depends on the
2031 most recent preceding instruction with externally visible side
2032 effects, following the order in the IR. (This includes :ref:`volatile
2033 operations <volatile>`.)
2034 - An instruction *control-depends* on a :ref:`terminator
2035 instruction <terminators>` if the terminator instruction has
2036 multiple successors and the instruction is always executed when
2037 control transfers to one of the successors, and may not be executed
2038 when control is transferred to another.
2039 - Additionally, an instruction also *control-depends* on a terminator
2040 instruction if the set of instructions it otherwise depends on would
2041 be different if the terminator had transferred control to a different
2043 - Dependence is transitive.
2045 Poison Values have the same behavior as :ref:`undef values <undefvalues>`,
2046 with the additional affect that any instruction which has a *dependence*
2047 on a poison value has undefined behavior.
2049 Here are some examples:
2051 .. code-block:: llvm
2054 %poison = sub nuw i32 0, 1 ; Results in a poison value.
2055 %still_poison = and i32 %poison, 0 ; 0, but also poison.
2056 %poison_yet_again = getelementptr i32* @h, i32 %still_poison
2057 store i32 0, i32* %poison_yet_again ; memory at @h[0] is poisoned
2059 store i32 %poison, i32* @g ; Poison value stored to memory.
2060 %poison2 = load i32* @g ; Poison value loaded back from memory.
2062 store volatile i32 %poison, i32* @g ; External observation; undefined behavior.
2064 %narrowaddr = bitcast i32* @g to i16*
2065 %wideaddr = bitcast i32* @g to i64*
2066 %poison3 = load i16* %narrowaddr ; Returns a poison value.
2067 %poison4 = load i64* %wideaddr ; Returns a poison value.
2069 %cmp = icmp slt i32 %poison, 0 ; Returns a poison value.
2070 br i1 %cmp, label %true, label %end ; Branch to either destination.
2073 store volatile i32 0, i32* @g ; This is control-dependent on %cmp, so
2074 ; it has undefined behavior.
2078 %p = phi i32 [ 0, %entry ], [ 1, %true ]
2079 ; Both edges into this PHI are
2080 ; control-dependent on %cmp, so this
2081 ; always results in a poison value.
2083 store volatile i32 0, i32* @g ; This would depend on the store in %true
2084 ; if %cmp is true, or the store in %entry
2085 ; otherwise, so this is undefined behavior.
2087 br i1 %cmp, label %second_true, label %second_end
2088 ; The same branch again, but this time the
2089 ; true block doesn't have side effects.
2096 store volatile i32 0, i32* @g ; This time, the instruction always depends
2097 ; on the store in %end. Also, it is
2098 ; control-equivalent to %end, so this is
2099 ; well-defined (ignoring earlier undefined
2100 ; behavior in this example).
2104 Addresses of Basic Blocks
2105 -------------------------
2107 ``blockaddress(@function, %block)``
2109 The '``blockaddress``' constant computes the address of the specified
2110 basic block in the specified function, and always has an ``i8*`` type.
2111 Taking the address of the entry block is illegal.
2113 This value only has defined behavior when used as an operand to the
2114 ':ref:`indirectbr <i_indirectbr>`' instruction, or for comparisons
2115 against null. Pointer equality tests between labels addresses results in
2116 undefined behavior --- though, again, comparison against null is ok, and
2117 no label is equal to the null pointer. This may be passed around as an
2118 opaque pointer sized value as long as the bits are not inspected. This
2119 allows ``ptrtoint`` and arithmetic to be performed on these values so
2120 long as the original value is reconstituted before the ``indirectbr``
2123 Finally, some targets may provide defined semantics when using the value
2124 as the operand to an inline assembly, but that is target specific.
2126 Constant Expressions
2127 --------------------
2129 Constant expressions are used to allow expressions involving other
2130 constants to be used as constants. Constant expressions may be of any
2131 :ref:`first class <t_firstclass>` type and may involve any LLVM operation
2132 that does not have side effects (e.g. load and call are not supported).
2133 The following is the syntax for constant expressions:
2135 ``trunc (CST to TYPE)``
2136 Truncate a constant to another type. The bit size of CST must be
2137 larger than the bit size of TYPE. Both types must be integers.
2138 ``zext (CST to TYPE)``
2139 Zero extend a constant to another type. The bit size of CST must be
2140 smaller than the bit size of TYPE. Both types must be integers.
2141 ``sext (CST to TYPE)``
2142 Sign extend a constant to another type. The bit size of CST must be
2143 smaller than the bit size of TYPE. Both types must be integers.
2144 ``fptrunc (CST to TYPE)``
2145 Truncate a floating point constant to another floating point type.
2146 The size of CST must be larger than the size of TYPE. Both types
2147 must be floating point.
2148 ``fpext (CST to TYPE)``
2149 Floating point extend a constant to another type. The size of CST
2150 must be smaller or equal to the size of TYPE. Both types must be
2152 ``fptoui (CST to TYPE)``
2153 Convert a floating point constant to the corresponding unsigned
2154 integer constant. TYPE must be a scalar or vector integer type. CST
2155 must be of scalar or vector floating point type. Both CST and TYPE
2156 must be scalars, or vectors of the same number of elements. If the
2157 value won't fit in the integer type, the results are undefined.
2158 ``fptosi (CST to TYPE)``
2159 Convert a floating point constant to the corresponding signed
2160 integer constant. TYPE must be a scalar or vector integer type. CST
2161 must be of scalar or vector floating point type. Both CST and TYPE
2162 must be scalars, or vectors of the same number of elements. If the
2163 value won't fit in the integer type, the results are undefined.
2164 ``uitofp (CST to TYPE)``
2165 Convert an unsigned integer constant to the corresponding floating
2166 point constant. TYPE must be a scalar or vector floating point type.
2167 CST must be of scalar or vector integer type. Both CST and TYPE must
2168 be scalars, or vectors of the same number of elements. If the value
2169 won't fit in the floating point type, the results are undefined.
2170 ``sitofp (CST to TYPE)``
2171 Convert a signed integer constant to the corresponding floating
2172 point constant. TYPE must be a scalar or vector floating point type.
2173 CST must be of scalar or vector integer type. Both CST and TYPE must
2174 be scalars, or vectors of the same number of elements. If the value
2175 won't fit in the floating point type, the results are undefined.
2176 ``ptrtoint (CST to TYPE)``
2177 Convert a pointer typed constant to the corresponding integer
2178 constant ``TYPE`` must be an integer type. ``CST`` must be of
2179 pointer type. The ``CST`` value is zero extended, truncated, or
2180 unchanged to make it fit in ``TYPE``.
2181 ``inttoptr (CST to TYPE)``
2182 Convert an integer constant to a pointer constant. TYPE must be a
2183 pointer type. CST must be of integer type. The CST value is zero
2184 extended, truncated, or unchanged to make it fit in a pointer size.
2185 This one is *really* dangerous!
2186 ``bitcast (CST to TYPE)``
2187 Convert a constant, CST, to another TYPE. The constraints of the
2188 operands are the same as those for the :ref:`bitcast
2189 instruction <i_bitcast>`.
2190 ``getelementptr (CSTPTR, IDX0, IDX1, ...)``, ``getelementptr inbounds (CSTPTR, IDX0, IDX1, ...)``
2191 Perform the :ref:`getelementptr operation <i_getelementptr>` on
2192 constants. As with the :ref:`getelementptr <i_getelementptr>`
2193 instruction, the index list may have zero or more indexes, which are
2194 required to make sense for the type of "CSTPTR".
2195 ``select (COND, VAL1, VAL2)``
2196 Perform the :ref:`select operation <i_select>` on constants.
2197 ``icmp COND (VAL1, VAL2)``
2198 Performs the :ref:`icmp operation <i_icmp>` on constants.
2199 ``fcmp COND (VAL1, VAL2)``
2200 Performs the :ref:`fcmp operation <i_fcmp>` on constants.
2201 ``extractelement (VAL, IDX)``
2202 Perform the :ref:`extractelement operation <i_extractelement>` on
2204 ``insertelement (VAL, ELT, IDX)``
2205 Perform the :ref:`insertelement operation <i_insertelement>` on
2207 ``shufflevector (VEC1, VEC2, IDXMASK)``
2208 Perform the :ref:`shufflevector operation <i_shufflevector>` on
2210 ``extractvalue (VAL, IDX0, IDX1, ...)``
2211 Perform the :ref:`extractvalue operation <i_extractvalue>` on
2212 constants. The index list is interpreted in a similar manner as
2213 indices in a ':ref:`getelementptr <i_getelementptr>`' operation. At
2214 least one index value must be specified.
2215 ``insertvalue (VAL, ELT, IDX0, IDX1, ...)``
2216 Perform the :ref:`insertvalue operation <i_insertvalue>` on constants.
2217 The index list is interpreted in a similar manner as indices in a
2218 ':ref:`getelementptr <i_getelementptr>`' operation. At least one index
2219 value must be specified.
2220 ``OPCODE (LHS, RHS)``
2221 Perform the specified operation of the LHS and RHS constants. OPCODE
2222 may be any of the :ref:`binary <binaryops>` or :ref:`bitwise
2223 binary <bitwiseops>` operations. The constraints on operands are
2224 the same as those for the corresponding instruction (e.g. no bitwise
2225 operations on floating point values are allowed).
2230 Inline Assembler Expressions
2231 ----------------------------
2233 LLVM supports inline assembler expressions (as opposed to :ref:`Module-Level
2234 Inline Assembly <moduleasm>`) through the use of a special value. This
2235 value represents the inline assembler as a string (containing the
2236 instructions to emit), a list of operand constraints (stored as a
2237 string), a flag that indicates whether or not the inline asm expression
2238 has side effects, and a flag indicating whether the function containing
2239 the asm needs to align its stack conservatively. An example inline
2240 assembler expression is:
2242 .. code-block:: llvm
2244 i32 (i32) asm "bswap $0", "=r,r"
2246 Inline assembler expressions may **only** be used as the callee operand
2247 of a :ref:`call <i_call>` or an :ref:`invoke <i_invoke>` instruction.
2248 Thus, typically we have:
2250 .. code-block:: llvm
2252 %X = call i32 asm "bswap $0", "=r,r"(i32 %Y)
2254 Inline asms with side effects not visible in the constraint list must be
2255 marked as having side effects. This is done through the use of the
2256 '``sideeffect``' keyword, like so:
2258 .. code-block:: llvm
2260 call void asm sideeffect "eieio", ""()
2262 In some cases inline asms will contain code that will not work unless
2263 the stack is aligned in some way, such as calls or SSE instructions on
2264 x86, yet will not contain code that does that alignment within the asm.
2265 The compiler should make conservative assumptions about what the asm
2266 might contain and should generate its usual stack alignment code in the
2267 prologue if the '``alignstack``' keyword is present:
2269 .. code-block:: llvm
2271 call void asm alignstack "eieio", ""()
2273 Inline asms also support using non-standard assembly dialects. The
2274 assumed dialect is ATT. When the '``inteldialect``' keyword is present,
2275 the inline asm is using the Intel dialect. Currently, ATT and Intel are
2276 the only supported dialects. An example is:
2278 .. code-block:: llvm
2280 call void asm inteldialect "eieio", ""()
2282 If multiple keywords appear the '``sideeffect``' keyword must come
2283 first, the '``alignstack``' keyword second and the '``inteldialect``'
2289 The call instructions that wrap inline asm nodes may have a
2290 "``!srcloc``" MDNode attached to it that contains a list of constant
2291 integers. If present, the code generator will use the integer as the
2292 location cookie value when report errors through the ``LLVMContext``
2293 error reporting mechanisms. This allows a front-end to correlate backend
2294 errors that occur with inline asm back to the source code that produced
2297 .. code-block:: llvm
2299 call void asm sideeffect "something bad", ""(), !srcloc !42
2301 !42 = !{ i32 1234567 }
2303 It is up to the front-end to make sense of the magic numbers it places
2304 in the IR. If the MDNode contains multiple constants, the code generator
2305 will use the one that corresponds to the line of the asm that the error
2310 Metadata Nodes and Metadata Strings
2311 -----------------------------------
2313 LLVM IR allows metadata to be attached to instructions in the program
2314 that can convey extra information about the code to the optimizers and
2315 code generator. One example application of metadata is source-level
2316 debug information. There are two metadata primitives: strings and nodes.
2317 All metadata has the ``metadata`` type and is identified in syntax by a
2318 preceding exclamation point ('``!``').
2320 A metadata string is a string surrounded by double quotes. It can
2321 contain any character by escaping non-printable characters with
2322 "``\xx``" where "``xx``" is the two digit hex code. For example:
2325 Metadata nodes are represented with notation similar to structure
2326 constants (a comma separated list of elements, surrounded by braces and
2327 preceded by an exclamation point). Metadata nodes can have any values as
2328 their operand. For example:
2330 .. code-block:: llvm
2332 !{ metadata !"test\00", i32 10}
2334 A :ref:`named metadata <namedmetadatastructure>` is a collection of
2335 metadata nodes, which can be looked up in the module symbol table. For
2338 .. code-block:: llvm
2340 !foo = metadata !{!4, !3}
2342 Metadata can be used as function arguments. Here ``llvm.dbg.value``
2343 function is using two metadata arguments:
2345 .. code-block:: llvm
2347 call void @llvm.dbg.value(metadata !24, i64 0, metadata !25)
2349 Metadata can be attached with an instruction. Here metadata ``!21`` is
2350 attached to the ``add`` instruction using the ``!dbg`` identifier:
2352 .. code-block:: llvm
2354 %indvar.next = add i64 %indvar, 1, !dbg !21
2356 More information about specific metadata nodes recognized by the
2357 optimizers and code generator is found below.
2362 In LLVM IR, memory does not have types, so LLVM's own type system is not
2363 suitable for doing TBAA. Instead, metadata is added to the IR to
2364 describe a type system of a higher level language. This can be used to
2365 implement typical C/C++ TBAA, but it can also be used to implement
2366 custom alias analysis behavior for other languages.
2368 The current metadata format is very simple. TBAA metadata nodes have up
2369 to three fields, e.g.:
2371 .. code-block:: llvm
2373 !0 = metadata !{ metadata !"an example type tree" }
2374 !1 = metadata !{ metadata !"int", metadata !0 }
2375 !2 = metadata !{ metadata !"float", metadata !0 }
2376 !3 = metadata !{ metadata !"const float", metadata !2, i64 1 }
2378 The first field is an identity field. It can be any value, usually a
2379 metadata string, which uniquely identifies the type. The most important
2380 name in the tree is the name of the root node. Two trees with different
2381 root node names are entirely disjoint, even if they have leaves with
2384 The second field identifies the type's parent node in the tree, or is
2385 null or omitted for a root node. A type is considered to alias all of
2386 its descendants and all of its ancestors in the tree. Also, a type is
2387 considered to alias all types in other trees, so that bitcode produced
2388 from multiple front-ends is handled conservatively.
2390 If the third field is present, it's an integer which if equal to 1
2391 indicates that the type is "constant" (meaning
2392 ``pointsToConstantMemory`` should return true; see `other useful
2393 AliasAnalysis methods <AliasAnalysis.html#OtherItfs>`_).
2395 '``tbaa.struct``' Metadata
2396 ^^^^^^^^^^^^^^^^^^^^^^^^^^
2398 The :ref:`llvm.memcpy <int_memcpy>` is often used to implement
2399 aggregate assignment operations in C and similar languages, however it
2400 is defined to copy a contiguous region of memory, which is more than
2401 strictly necessary for aggregate types which contain holes due to
2402 padding. Also, it doesn't contain any TBAA information about the fields
2405 ``!tbaa.struct`` metadata can describe which memory subregions in a
2406 memcpy are padding and what the TBAA tags of the struct are.
2408 The current metadata format is very simple. ``!tbaa.struct`` metadata
2409 nodes are a list of operands which are in conceptual groups of three.
2410 For each group of three, the first operand gives the byte offset of a
2411 field in bytes, the second gives its size in bytes, and the third gives
2414 .. code-block:: llvm
2416 !4 = metadata !{ i64 0, i64 4, metadata !1, i64 8, i64 4, metadata !2 }
2418 This describes a struct with two fields. The first is at offset 0 bytes
2419 with size 4 bytes, and has tbaa tag !1. The second is at offset 8 bytes
2420 and has size 4 bytes and has tbaa tag !2.
2422 Note that the fields need not be contiguous. In this example, there is a
2423 4 byte gap between the two fields. This gap represents padding which
2424 does not carry useful data and need not be preserved.
2426 '``fpmath``' Metadata
2427 ^^^^^^^^^^^^^^^^^^^^^
2429 ``fpmath`` metadata may be attached to any instruction of floating point
2430 type. It can be used to express the maximum acceptable error in the
2431 result of that instruction, in ULPs, thus potentially allowing the
2432 compiler to use a more efficient but less accurate method of computing
2433 it. ULP is defined as follows:
2435 If ``x`` is a real number that lies between two finite consecutive
2436 floating-point numbers ``a`` and ``b``, without being equal to one
2437 of them, then ``ulp(x) = |b - a|``, otherwise ``ulp(x)`` is the
2438 distance between the two non-equal finite floating-point numbers
2439 nearest ``x``. Moreover, ``ulp(NaN)`` is ``NaN``.
2441 The metadata node shall consist of a single positive floating point
2442 number representing the maximum relative error, for example:
2444 .. code-block:: llvm
2446 !0 = metadata !{ float 2.5 } ; maximum acceptable inaccuracy is 2.5 ULPs
2448 '``range``' Metadata
2449 ^^^^^^^^^^^^^^^^^^^^
2451 ``range`` metadata may be attached only to loads of integer types. It
2452 expresses the possible ranges the loaded value is in. The ranges are
2453 represented with a flattened list of integers. The loaded value is known
2454 to be in the union of the ranges defined by each consecutive pair. Each
2455 pair has the following properties:
2457 - The type must match the type loaded by the instruction.
2458 - The pair ``a,b`` represents the range ``[a,b)``.
2459 - Both ``a`` and ``b`` are constants.
2460 - The range is allowed to wrap.
2461 - The range should not represent the full or empty set. That is,
2464 In addition, the pairs must be in signed order of the lower bound and
2465 they must be non-contiguous.
2469 .. code-block:: llvm
2471 %a = load i8* %x, align 1, !range !0 ; Can only be 0 or 1
2472 %b = load i8* %y, align 1, !range !1 ; Can only be 255 (-1), 0 or 1
2473 %c = load i8* %z, align 1, !range !2 ; Can only be 0, 1, 3, 4 or 5
2474 %d = load i8* %z, align 1, !range !3 ; Can only be -2, -1, 3, 4 or 5
2476 !0 = metadata !{ i8 0, i8 2 }
2477 !1 = metadata !{ i8 255, i8 2 }
2478 !2 = metadata !{ i8 0, i8 2, i8 3, i8 6 }
2479 !3 = metadata !{ i8 -2, i8 0, i8 3, i8 6 }
2481 Module Flags Metadata
2482 =====================
2484 Information about the module as a whole is difficult to convey to LLVM's
2485 subsystems. The LLVM IR isn't sufficient to transmit this information.
2486 The ``llvm.module.flags`` named metadata exists in order to facilitate
2487 this. These flags are in the form of key / value pairs --- much like a
2488 dictionary --- making it easy for any subsystem who cares about a flag to
2491 The ``llvm.module.flags`` metadata contains a list of metadata triplets.
2492 Each triplet has the following form:
2494 - The first element is a *behavior* flag, which specifies the behavior
2495 when two (or more) modules are merged together, and it encounters two
2496 (or more) metadata with the same ID. The supported behaviors are
2498 - The second element is a metadata string that is a unique ID for the
2499 metadata. Each module may only have one flag entry for each unique ID (not
2500 including entries with the **Require** behavior).
2501 - The third element is the value of the flag.
2503 When two (or more) modules are merged together, the resulting
2504 ``llvm.module.flags`` metadata is the union of the modules' flags. That is, for
2505 each unique metadata ID string, there will be exactly one entry in the merged
2506 modules ``llvm.module.flags`` metadata table, and the value for that entry will
2507 be determined by the merge behavior flag, as described below. The only exception
2508 is that entries with the *Require* behavior are always preserved.
2510 The following behaviors are supported:
2521 Emits an error if two values disagree, otherwise the resulting value
2522 is that of the operands.
2526 Emits a warning if two values disagree. The result value will be the
2527 operand for the flag from the first module being linked.
2531 Adds a requirement that another module flag be present and have a
2532 specified value after linking is performed. The value must be a
2533 metadata pair, where the first element of the pair is the ID of the
2534 module flag to be restricted, and the second element of the pair is
2535 the value the module flag should be restricted to. This behavior can
2536 be used to restrict the allowable results (via triggering of an
2537 error) of linking IDs with the **Override** behavior.
2541 Uses the specified value, regardless of the behavior or value of the
2542 other module. If both modules specify **Override**, but the values
2543 differ, an error will be emitted.
2547 Appends the two values, which are required to be metadata nodes.
2551 Appends the two values, which are required to be metadata
2552 nodes. However, duplicate entries in the second list are dropped
2553 during the append operation.
2555 It is an error for a particular unique flag ID to have multiple behaviors,
2556 except in the case of **Require** (which adds restrictions on another metadata
2557 value) or **Override**.
2559 An example of module flags:
2561 .. code-block:: llvm
2563 !0 = metadata !{ i32 1, metadata !"foo", i32 1 }
2564 !1 = metadata !{ i32 4, metadata !"bar", i32 37 }
2565 !2 = metadata !{ i32 2, metadata !"qux", i32 42 }
2566 !3 = metadata !{ i32 3, metadata !"qux",
2568 metadata !"foo", i32 1
2571 !llvm.module.flags = !{ !0, !1, !2, !3 }
2573 - Metadata ``!0`` has the ID ``!"foo"`` and the value '1'. The behavior
2574 if two or more ``!"foo"`` flags are seen is to emit an error if their
2575 values are not equal.
2577 - Metadata ``!1`` has the ID ``!"bar"`` and the value '37'. The
2578 behavior if two or more ``!"bar"`` flags are seen is to use the value
2581 - Metadata ``!2`` has the ID ``!"qux"`` and the value '42'. The
2582 behavior if two or more ``!"qux"`` flags are seen is to emit a
2583 warning if their values are not equal.
2585 - Metadata ``!3`` has the ID ``!"qux"`` and the value:
2589 metadata !{ metadata !"foo", i32 1 }
2591 The behavior is to emit an error if the ``llvm.module.flags`` does not
2592 contain a flag with the ID ``!"foo"`` that has the value '1' after linking is
2595 Objective-C Garbage Collection Module Flags Metadata
2596 ----------------------------------------------------
2598 On the Mach-O platform, Objective-C stores metadata about garbage
2599 collection in a special section called "image info". The metadata
2600 consists of a version number and a bitmask specifying what types of
2601 garbage collection are supported (if any) by the file. If two or more
2602 modules are linked together their garbage collection metadata needs to
2603 be merged rather than appended together.
2605 The Objective-C garbage collection module flags metadata consists of the
2606 following key-value pairs:
2615 * - ``Objective-C Version``
2616 - **[Required]** --- The Objective-C ABI version. Valid values are 1 and 2.
2618 * - ``Objective-C Image Info Version``
2619 - **[Required]** --- The version of the image info section. Currently
2622 * - ``Objective-C Image Info Section``
2623 - **[Required]** --- The section to place the metadata. Valid values are
2624 ``"__OBJC, __image_info, regular"`` for Objective-C ABI version 1, and
2625 ``"__DATA,__objc_imageinfo, regular, no_dead_strip"`` for
2626 Objective-C ABI version 2.
2628 * - ``Objective-C Garbage Collection``
2629 - **[Required]** --- Specifies whether garbage collection is supported or
2630 not. Valid values are 0, for no garbage collection, and 2, for garbage
2631 collection supported.
2633 * - ``Objective-C GC Only``
2634 - **[Optional]** --- Specifies that only garbage collection is supported.
2635 If present, its value must be 6. This flag requires that the
2636 ``Objective-C Garbage Collection`` flag have the value 2.
2638 Some important flag interactions:
2640 - If a module with ``Objective-C Garbage Collection`` set to 0 is
2641 merged with a module with ``Objective-C Garbage Collection`` set to
2642 2, then the resulting module has the
2643 ``Objective-C Garbage Collection`` flag set to 0.
2644 - A module with ``Objective-C Garbage Collection`` set to 0 cannot be
2645 merged with a module with ``Objective-C GC Only`` set to 6.
2647 Automatic Linker Flags Module Flags Metadata
2648 --------------------------------------------
2650 Some targets support embedding flags to the linker inside individual object
2651 files. Typically this is used in conjunction with language extensions which
2652 allow source files to explicitly declare the libraries they depend on, and have
2653 these automatically be transmitted to the linker via object files.
2655 These flags are encoded in the IR using metadata in the module flags section,
2656 using the ``Linker Options`` key. The merge behavior for this flag is required
2657 to be ``AppendUnique``, and the value for the key is expected to be a metadata
2658 node which should be a list of other metadata nodes, each of which should be a
2659 list of metadata strings defining linker options.
2661 For example, the following metadata section specifies two separate sets of
2662 linker options, presumably to link against ``libz`` and the ``Cocoa``
2665 !0 = metadata !{ i32 6, metadata !"Linker Options",
2667 metadata !{ metadata !"-lz" },
2668 metadata !{ metadata !"-framework", metadata !"Cocoa" } } }
2669 !llvm.module.flags = !{ !0 }
2671 The metadata encoding as lists of lists of options, as opposed to a collapsed
2672 list of options, is chosen so that the IR encoding can use multiple option
2673 strings to specify e.g., a single library, while still having that specifier be
2674 preserved as an atomic element that can be recognized by a target specific
2675 assembly writer or object file emitter.
2677 Each individual option is required to be either a valid option for the target's
2678 linker, or an option that is reserved by the target specific assembly writer or
2679 object file emitter. No other aspect of these options is defined by the IR.
2681 Intrinsic Global Variables
2682 ==========================
2684 LLVM has a number of "magic" global variables that contain data that
2685 affect code generation or other IR semantics. These are documented here.
2686 All globals of this sort should have a section specified as
2687 "``llvm.metadata``". This section and all globals that start with
2688 "``llvm.``" are reserved for use by LLVM.
2690 The '``llvm.used``' Global Variable
2691 -----------------------------------
2693 The ``@llvm.used`` global is an array with i8\* element type which has
2694 :ref:`appending linkage <linkage_appending>`. This array contains a list of
2695 pointers to global variables and functions which may optionally have a
2696 pointer cast formed of bitcast or getelementptr. For example, a legal
2699 .. code-block:: llvm
2704 @llvm.used = appending global [2 x i8*] [
2706 i8* bitcast (i32* @Y to i8*)
2707 ], section "llvm.metadata"
2709 If a global variable appears in the ``@llvm.used`` list, then the
2710 compiler, assembler, and linker are required to treat the symbol as if
2711 there is a reference to the global that it cannot see. For example, if a
2712 variable has internal linkage and no references other than that from the
2713 ``@llvm.used`` list, it cannot be deleted. This is commonly used to
2714 represent references from inline asms and other things the compiler
2715 cannot "see", and corresponds to "``attribute((used))``" in GNU C.
2717 On some targets, the code generator must emit a directive to the
2718 assembler or object file to prevent the assembler and linker from
2719 molesting the symbol.
2721 The '``llvm.compiler.used``' Global Variable
2722 --------------------------------------------
2724 The ``@llvm.compiler.used`` directive is the same as the ``@llvm.used``
2725 directive, except that it only prevents the compiler from touching the
2726 symbol. On targets that support it, this allows an intelligent linker to
2727 optimize references to the symbol without being impeded as it would be
2730 This is a rare construct that should only be used in rare circumstances,
2731 and should not be exposed to source languages.
2733 The '``llvm.global_ctors``' Global Variable
2734 -------------------------------------------
2736 .. code-block:: llvm
2738 %0 = type { i32, void ()* }
2739 @llvm.global_ctors = appending global [1 x %0] [%0 { i32 65535, void ()* @ctor }]
2741 The ``@llvm.global_ctors`` array contains a list of constructor
2742 functions and associated priorities. The functions referenced by this
2743 array will be called in ascending order of priority (i.e. lowest first)
2744 when the module is loaded. The order of functions with the same priority
2747 The '``llvm.global_dtors``' Global Variable
2748 -------------------------------------------
2750 .. code-block:: llvm
2752 %0 = type { i32, void ()* }
2753 @llvm.global_dtors = appending global [1 x %0] [%0 { i32 65535, void ()* @dtor }]
2755 The ``@llvm.global_dtors`` array contains a list of destructor functions
2756 and associated priorities. The functions referenced by this array will
2757 be called in descending order of priority (i.e. highest first) when the
2758 module is loaded. The order of functions with the same priority is not
2761 Instruction Reference
2762 =====================
2764 The LLVM instruction set consists of several different classifications
2765 of instructions: :ref:`terminator instructions <terminators>`, :ref:`binary
2766 instructions <binaryops>`, :ref:`bitwise binary
2767 instructions <bitwiseops>`, :ref:`memory instructions <memoryops>`, and
2768 :ref:`other instructions <otherops>`.
2772 Terminator Instructions
2773 -----------------------
2775 As mentioned :ref:`previously <functionstructure>`, every basic block in a
2776 program ends with a "Terminator" instruction, which indicates which
2777 block should be executed after the current block is finished. These
2778 terminator instructions typically yield a '``void``' value: they produce
2779 control flow, not values (the one exception being the
2780 ':ref:`invoke <i_invoke>`' instruction).
2782 The terminator instructions are: ':ref:`ret <i_ret>`',
2783 ':ref:`br <i_br>`', ':ref:`switch <i_switch>`',
2784 ':ref:`indirectbr <i_indirectbr>`', ':ref:`invoke <i_invoke>`',
2785 ':ref:`resume <i_resume>`', and ':ref:`unreachable <i_unreachable>`'.
2789 '``ret``' Instruction
2790 ^^^^^^^^^^^^^^^^^^^^^
2797 ret <type> <value> ; Return a value from a non-void function
2798 ret void ; Return from void function
2803 The '``ret``' instruction is used to return control flow (and optionally
2804 a value) from a function back to the caller.
2806 There are two forms of the '``ret``' instruction: one that returns a
2807 value and then causes control flow, and one that just causes control
2813 The '``ret``' instruction optionally accepts a single argument, the
2814 return value. The type of the return value must be a ':ref:`first
2815 class <t_firstclass>`' type.
2817 A function is not :ref:`well formed <wellformed>` if it it has a non-void
2818 return type and contains a '``ret``' instruction with no return value or
2819 a return value with a type that does not match its type, or if it has a
2820 void return type and contains a '``ret``' instruction with a return
2826 When the '``ret``' instruction is executed, control flow returns back to
2827 the calling function's context. If the caller is a
2828 ":ref:`call <i_call>`" instruction, execution continues at the
2829 instruction after the call. If the caller was an
2830 ":ref:`invoke <i_invoke>`" instruction, execution continues at the
2831 beginning of the "normal" destination block. If the instruction returns
2832 a value, that value shall set the call or invoke instruction's return
2838 .. code-block:: llvm
2840 ret i32 5 ; Return an integer value of 5
2841 ret void ; Return from a void function
2842 ret { i32, i8 } { i32 4, i8 2 } ; Return a struct of values 4 and 2
2846 '``br``' Instruction
2847 ^^^^^^^^^^^^^^^^^^^^
2854 br i1 <cond>, label <iftrue>, label <iffalse>
2855 br label <dest> ; Unconditional branch
2860 The '``br``' instruction is used to cause control flow to transfer to a
2861 different basic block in the current function. There are two forms of
2862 this instruction, corresponding to a conditional branch and an
2863 unconditional branch.
2868 The conditional branch form of the '``br``' instruction takes a single
2869 '``i1``' value and two '``label``' values. The unconditional form of the
2870 '``br``' instruction takes a single '``label``' value as a target.
2875 Upon execution of a conditional '``br``' instruction, the '``i1``'
2876 argument is evaluated. If the value is ``true``, control flows to the
2877 '``iftrue``' ``label`` argument. If "cond" is ``false``, control flows
2878 to the '``iffalse``' ``label`` argument.
2883 .. code-block:: llvm
2886 %cond = icmp eq i32 %a, %b
2887 br i1 %cond, label %IfEqual, label %IfUnequal
2895 '``switch``' Instruction
2896 ^^^^^^^^^^^^^^^^^^^^^^^^
2903 switch <intty> <value>, label <defaultdest> [ <intty> <val>, label <dest> ... ]
2908 The '``switch``' instruction is used to transfer control flow to one of
2909 several different places. It is a generalization of the '``br``'
2910 instruction, allowing a branch to occur to one of many possible
2916 The '``switch``' instruction uses three parameters: an integer
2917 comparison value '``value``', a default '``label``' destination, and an
2918 array of pairs of comparison value constants and '``label``'s. The table
2919 is not allowed to contain duplicate constant entries.
2924 The ``switch`` instruction specifies a table of values and destinations.
2925 When the '``switch``' instruction is executed, this table is searched
2926 for the given value. If the value is found, control flow is transferred
2927 to the corresponding destination; otherwise, control flow is transferred
2928 to the default destination.
2933 Depending on properties of the target machine and the particular
2934 ``switch`` instruction, this instruction may be code generated in
2935 different ways. For example, it could be generated as a series of
2936 chained conditional branches or with a lookup table.
2941 .. code-block:: llvm
2943 ; Emulate a conditional br instruction
2944 %Val = zext i1 %value to i32
2945 switch i32 %Val, label %truedest [ i32 0, label %falsedest ]
2947 ; Emulate an unconditional br instruction
2948 switch i32 0, label %dest [ ]
2950 ; Implement a jump table:
2951 switch i32 %val, label %otherwise [ i32 0, label %onzero
2953 i32 2, label %ontwo ]
2957 '``indirectbr``' Instruction
2958 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2965 indirectbr <somety>* <address>, [ label <dest1>, label <dest2>, ... ]
2970 The '``indirectbr``' instruction implements an indirect branch to a
2971 label within the current function, whose address is specified by
2972 "``address``". Address must be derived from a
2973 :ref:`blockaddress <blockaddress>` constant.
2978 The '``address``' argument is the address of the label to jump to. The
2979 rest of the arguments indicate the full set of possible destinations
2980 that the address may point to. Blocks are allowed to occur multiple
2981 times in the destination list, though this isn't particularly useful.
2983 This destination list is required so that dataflow analysis has an
2984 accurate understanding of the CFG.
2989 Control transfers to the block specified in the address argument. All
2990 possible destination blocks must be listed in the label list, otherwise
2991 this instruction has undefined behavior. This implies that jumps to
2992 labels defined in other functions have undefined behavior as well.
2997 This is typically implemented with a jump through a register.
3002 .. code-block:: llvm
3004 indirectbr i8* %Addr, [ label %bb1, label %bb2, label %bb3 ]
3008 '``invoke``' Instruction
3009 ^^^^^^^^^^^^^^^^^^^^^^^^
3016 <result> = invoke [cconv] [ret attrs] <ptr to function ty> <function ptr val>(<function args>) [fn attrs]
3017 to label <normal label> unwind label <exception label>
3022 The '``invoke``' instruction causes control to transfer to a specified
3023 function, with the possibility of control flow transfer to either the
3024 '``normal``' label or the '``exception``' label. If the callee function
3025 returns with the "``ret``" instruction, control flow will return to the
3026 "normal" label. If the callee (or any indirect callees) returns via the
3027 ":ref:`resume <i_resume>`" instruction or other exception handling
3028 mechanism, control is interrupted and continued at the dynamically
3029 nearest "exception" label.
3031 The '``exception``' label is a `landing
3032 pad <ExceptionHandling.html#overview>`_ for the exception. As such,
3033 '``exception``' label is required to have the
3034 ":ref:`landingpad <i_landingpad>`" instruction, which contains the
3035 information about the behavior of the program after unwinding happens,
3036 as its first non-PHI instruction. The restrictions on the
3037 "``landingpad``" instruction's tightly couples it to the "``invoke``"
3038 instruction, so that the important information contained within the
3039 "``landingpad``" instruction can't be lost through normal code motion.
3044 This instruction requires several arguments:
3046 #. The optional "cconv" marker indicates which :ref:`calling
3047 convention <callingconv>` the call should use. If none is
3048 specified, the call defaults to using C calling conventions.
3049 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
3050 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
3052 #. '``ptr to function ty``': shall be the signature of the pointer to
3053 function value being invoked. In most cases, this is a direct
3054 function invocation, but indirect ``invoke``'s are just as possible,
3055 branching off an arbitrary pointer to function value.
3056 #. '``function ptr val``': An LLVM value containing a pointer to a
3057 function to be invoked.
3058 #. '``function args``': argument list whose types match the function
3059 signature argument types and parameter attributes. All arguments must
3060 be of :ref:`first class <t_firstclass>` type. If the function signature
3061 indicates the function accepts a variable number of arguments, the
3062 extra arguments can be specified.
3063 #. '``normal label``': the label reached when the called function
3064 executes a '``ret``' instruction.
3065 #. '``exception label``': the label reached when a callee returns via
3066 the :ref:`resume <i_resume>` instruction or other exception handling
3068 #. The optional :ref:`function attributes <fnattrs>` list. Only
3069 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
3070 attributes are valid here.
3075 This instruction is designed to operate as a standard '``call``'
3076 instruction in most regards. The primary difference is that it
3077 establishes an association with a label, which is used by the runtime
3078 library to unwind the stack.
3080 This instruction is used in languages with destructors to ensure that
3081 proper cleanup is performed in the case of either a ``longjmp`` or a
3082 thrown exception. Additionally, this is important for implementation of
3083 '``catch``' clauses in high-level languages that support them.
3085 For the purposes of the SSA form, the definition of the value returned
3086 by the '``invoke``' instruction is deemed to occur on the edge from the
3087 current block to the "normal" label. If the callee unwinds then no
3088 return value is available.
3093 .. code-block:: llvm
3095 %retval = invoke i32 @Test(i32 15) to label %Continue
3096 unwind label %TestCleanup ; {i32}:retval set
3097 %retval = invoke coldcc i32 %Testfnptr(i32 15) to label %Continue
3098 unwind label %TestCleanup ; {i32}:retval set
3102 '``resume``' Instruction
3103 ^^^^^^^^^^^^^^^^^^^^^^^^
3110 resume <type> <value>
3115 The '``resume``' instruction is a terminator instruction that has no
3121 The '``resume``' instruction requires one argument, which must have the
3122 same type as the result of any '``landingpad``' instruction in the same
3128 The '``resume``' instruction resumes propagation of an existing
3129 (in-flight) exception whose unwinding was interrupted with a
3130 :ref:`landingpad <i_landingpad>` instruction.
3135 .. code-block:: llvm
3137 resume { i8*, i32 } %exn
3141 '``unreachable``' Instruction
3142 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3154 The '``unreachable``' instruction has no defined semantics. This
3155 instruction is used to inform the optimizer that a particular portion of
3156 the code is not reachable. This can be used to indicate that the code
3157 after a no-return function cannot be reached, and other facts.
3162 The '``unreachable``' instruction has no defined semantics.
3169 Binary operators are used to do most of the computation in a program.
3170 They require two operands of the same type, execute an operation on
3171 them, and produce a single value. The operands might represent multiple
3172 data, as is the case with the :ref:`vector <t_vector>` data type. The
3173 result value has the same type as its operands.
3175 There are several different binary operators:
3179 '``add``' Instruction
3180 ^^^^^^^^^^^^^^^^^^^^^
3187 <result> = add <ty> <op1>, <op2> ; yields {ty}:result
3188 <result> = add nuw <ty> <op1>, <op2> ; yields {ty}:result
3189 <result> = add nsw <ty> <op1>, <op2> ; yields {ty}:result
3190 <result> = add nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3195 The '``add``' instruction returns the sum of its two operands.
3200 The two arguments to the '``add``' instruction must be
3201 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3202 arguments must have identical types.
3207 The value produced is the integer sum of the two operands.
3209 If the sum has unsigned overflow, the result returned is the
3210 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
3213 Because LLVM integers use a two's complement representation, this
3214 instruction is appropriate for both signed and unsigned integers.
3216 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3217 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3218 result value of the ``add`` is a :ref:`poison value <poisonvalues>` if
3219 unsigned and/or signed overflow, respectively, occurs.
3224 .. code-block:: llvm
3226 <result> = add i32 4, %var ; yields {i32}:result = 4 + %var
3230 '``fadd``' Instruction
3231 ^^^^^^^^^^^^^^^^^^^^^^
3238 <result> = fadd [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3243 The '``fadd``' instruction returns the sum of its two operands.
3248 The two arguments to the '``fadd``' instruction must be :ref:`floating
3249 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3250 Both arguments must have identical types.
3255 The value produced is the floating point sum of the two operands. This
3256 instruction can also take any number of :ref:`fast-math flags <fastmath>`,
3257 which are optimization hints to enable otherwise unsafe floating point
3263 .. code-block:: llvm
3265 <result> = fadd float 4.0, %var ; yields {float}:result = 4.0 + %var
3267 '``sub``' Instruction
3268 ^^^^^^^^^^^^^^^^^^^^^
3275 <result> = sub <ty> <op1>, <op2> ; yields {ty}:result
3276 <result> = sub nuw <ty> <op1>, <op2> ; yields {ty}:result
3277 <result> = sub nsw <ty> <op1>, <op2> ; yields {ty}:result
3278 <result> = sub nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3283 The '``sub``' instruction returns the difference of its two operands.
3285 Note that the '``sub``' instruction is used to represent the '``neg``'
3286 instruction present in most other intermediate representations.
3291 The two arguments to the '``sub``' instruction must be
3292 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3293 arguments must have identical types.
3298 The value produced is the integer difference of the two operands.
3300 If the difference has unsigned overflow, the result returned is the
3301 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
3304 Because LLVM integers use a two's complement representation, this
3305 instruction is appropriate for both signed and unsigned integers.
3307 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3308 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3309 result value of the ``sub`` is a :ref:`poison value <poisonvalues>` if
3310 unsigned and/or signed overflow, respectively, occurs.
3315 .. code-block:: llvm
3317 <result> = sub i32 4, %var ; yields {i32}:result = 4 - %var
3318 <result> = sub i32 0, %val ; yields {i32}:result = -%var
3322 '``fsub``' Instruction
3323 ^^^^^^^^^^^^^^^^^^^^^^
3330 <result> = fsub [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3335 The '``fsub``' instruction returns the difference of its two operands.
3337 Note that the '``fsub``' instruction is used to represent the '``fneg``'
3338 instruction present in most other intermediate representations.
3343 The two arguments to the '``fsub``' instruction must be :ref:`floating
3344 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3345 Both arguments must have identical types.
3350 The value produced is the floating point difference of the two operands.
3351 This instruction can also take any number of :ref:`fast-math
3352 flags <fastmath>`, which are optimization hints to enable otherwise
3353 unsafe floating point optimizations:
3358 .. code-block:: llvm
3360 <result> = fsub float 4.0, %var ; yields {float}:result = 4.0 - %var
3361 <result> = fsub float -0.0, %val ; yields {float}:result = -%var
3363 '``mul``' Instruction
3364 ^^^^^^^^^^^^^^^^^^^^^
3371 <result> = mul <ty> <op1>, <op2> ; yields {ty}:result
3372 <result> = mul nuw <ty> <op1>, <op2> ; yields {ty}:result
3373 <result> = mul nsw <ty> <op1>, <op2> ; yields {ty}:result
3374 <result> = mul nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3379 The '``mul``' instruction returns the product of its two operands.
3384 The two arguments to the '``mul``' instruction must be
3385 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3386 arguments must have identical types.
3391 The value produced is the integer product of the two operands.
3393 If the result of the multiplication has unsigned overflow, the result
3394 returned is the mathematical result modulo 2\ :sup:`n`\ , where n is the
3395 bit width of the result.
3397 Because LLVM integers use a two's complement representation, and the
3398 result is the same width as the operands, this instruction returns the
3399 correct result for both signed and unsigned integers. If a full product
3400 (e.g. ``i32`` * ``i32`` -> ``i64``) is needed, the operands should be
3401 sign-extended or zero-extended as appropriate to the width of the full
3404 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3405 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3406 result value of the ``mul`` is a :ref:`poison value <poisonvalues>` if
3407 unsigned and/or signed overflow, respectively, occurs.
3412 .. code-block:: llvm
3414 <result> = mul i32 4, %var ; yields {i32}:result = 4 * %var
3418 '``fmul``' Instruction
3419 ^^^^^^^^^^^^^^^^^^^^^^
3426 <result> = fmul [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3431 The '``fmul``' instruction returns the product of its two operands.
3436 The two arguments to the '``fmul``' instruction must be :ref:`floating
3437 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3438 Both arguments must have identical types.
3443 The value produced is the floating point product of the two operands.
3444 This instruction can also take any number of :ref:`fast-math
3445 flags <fastmath>`, which are optimization hints to enable otherwise
3446 unsafe floating point optimizations:
3451 .. code-block:: llvm
3453 <result> = fmul float 4.0, %var ; yields {float}:result = 4.0 * %var
3455 '``udiv``' Instruction
3456 ^^^^^^^^^^^^^^^^^^^^^^
3463 <result> = udiv <ty> <op1>, <op2> ; yields {ty}:result
3464 <result> = udiv exact <ty> <op1>, <op2> ; yields {ty}:result
3469 The '``udiv``' instruction returns the quotient of its two operands.
3474 The two arguments to the '``udiv``' instruction must be
3475 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3476 arguments must have identical types.
3481 The value produced is the unsigned integer quotient of the two operands.
3483 Note that unsigned integer division and signed integer division are
3484 distinct operations; for signed integer division, use '``sdiv``'.
3486 Division by zero leads to undefined behavior.
3488 If the ``exact`` keyword is present, the result value of the ``udiv`` is
3489 a :ref:`poison value <poisonvalues>` if %op1 is not a multiple of %op2 (as
3490 such, "((a udiv exact b) mul b) == a").
3495 .. code-block:: llvm
3497 <result> = udiv i32 4, %var ; yields {i32}:result = 4 / %var
3499 '``sdiv``' Instruction
3500 ^^^^^^^^^^^^^^^^^^^^^^
3507 <result> = sdiv <ty> <op1>, <op2> ; yields {ty}:result
3508 <result> = sdiv exact <ty> <op1>, <op2> ; yields {ty}:result
3513 The '``sdiv``' instruction returns the quotient of its two operands.
3518 The two arguments to the '``sdiv``' instruction must be
3519 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3520 arguments must have identical types.
3525 The value produced is the signed integer quotient of the two operands
3526 rounded towards zero.
3528 Note that signed integer division and unsigned integer division are
3529 distinct operations; for unsigned integer division, use '``udiv``'.
3531 Division by zero leads to undefined behavior. Overflow also leads to
3532 undefined behavior; this is a rare case, but can occur, for example, by
3533 doing a 32-bit division of -2147483648 by -1.
3535 If the ``exact`` keyword is present, the result value of the ``sdiv`` is
3536 a :ref:`poison value <poisonvalues>` if the result would be rounded.
3541 .. code-block:: llvm
3543 <result> = sdiv i32 4, %var ; yields {i32}:result = 4 / %var
3547 '``fdiv``' Instruction
3548 ^^^^^^^^^^^^^^^^^^^^^^
3555 <result> = fdiv [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3560 The '``fdiv``' instruction returns the quotient of its two operands.
3565 The two arguments to the '``fdiv``' instruction must be :ref:`floating
3566 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3567 Both arguments must have identical types.
3572 The value produced is the floating point quotient of the two operands.
3573 This instruction can also take any number of :ref:`fast-math
3574 flags <fastmath>`, which are optimization hints to enable otherwise
3575 unsafe floating point optimizations:
3580 .. code-block:: llvm
3582 <result> = fdiv float 4.0, %var ; yields {float}:result = 4.0 / %var
3584 '``urem``' Instruction
3585 ^^^^^^^^^^^^^^^^^^^^^^
3592 <result> = urem <ty> <op1>, <op2> ; yields {ty}:result
3597 The '``urem``' instruction returns the remainder from the unsigned
3598 division of its two arguments.
3603 The two arguments to the '``urem``' instruction must be
3604 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3605 arguments must have identical types.
3610 This instruction returns the unsigned integer *remainder* of a division.
3611 This instruction always performs an unsigned division to get the
3614 Note that unsigned integer remainder and signed integer remainder are
3615 distinct operations; for signed integer remainder, use '``srem``'.
3617 Taking the remainder of a division by zero leads to undefined behavior.
3622 .. code-block:: llvm
3624 <result> = urem i32 4, %var ; yields {i32}:result = 4 % %var
3626 '``srem``' Instruction
3627 ^^^^^^^^^^^^^^^^^^^^^^
3634 <result> = srem <ty> <op1>, <op2> ; yields {ty}:result
3639 The '``srem``' instruction returns the remainder from the signed
3640 division of its two operands. This instruction can also take
3641 :ref:`vector <t_vector>` versions of the values in which case the elements
3647 The two arguments to the '``srem``' instruction must be
3648 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3649 arguments must have identical types.
3654 This instruction returns the *remainder* of a division (where the result
3655 is either zero or has the same sign as the dividend, ``op1``), not the
3656 *modulo* operator (where the result is either zero or has the same sign
3657 as the divisor, ``op2``) of a value. For more information about the
3658 difference, see `The Math
3659 Forum <http://mathforum.org/dr.math/problems/anne.4.28.99.html>`_. For a
3660 table of how this is implemented in various languages, please see
3662 operation <http://en.wikipedia.org/wiki/Modulo_operation>`_.
3664 Note that signed integer remainder and unsigned integer remainder are
3665 distinct operations; for unsigned integer remainder, use '``urem``'.
3667 Taking the remainder of a division by zero leads to undefined behavior.
3668 Overflow also leads to undefined behavior; this is a rare case, but can
3669 occur, for example, by taking the remainder of a 32-bit division of
3670 -2147483648 by -1. (The remainder doesn't actually overflow, but this
3671 rule lets srem be implemented using instructions that return both the
3672 result of the division and the remainder.)
3677 .. code-block:: llvm
3679 <result> = srem i32 4, %var ; yields {i32}:result = 4 % %var
3683 '``frem``' Instruction
3684 ^^^^^^^^^^^^^^^^^^^^^^
3691 <result> = frem [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3696 The '``frem``' instruction returns the remainder from the division of
3702 The two arguments to the '``frem``' instruction must be :ref:`floating
3703 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3704 Both arguments must have identical types.
3709 This instruction returns the *remainder* of a division. The remainder
3710 has the same sign as the dividend. This instruction can also take any
3711 number of :ref:`fast-math flags <fastmath>`, which are optimization hints
3712 to enable otherwise unsafe floating point optimizations:
3717 .. code-block:: llvm
3719 <result> = frem float 4.0, %var ; yields {float}:result = 4.0 % %var
3723 Bitwise Binary Operations
3724 -------------------------
3726 Bitwise binary operators are used to do various forms of bit-twiddling
3727 in a program. They are generally very efficient instructions and can
3728 commonly be strength reduced from other instructions. They require two
3729 operands of the same type, execute an operation on them, and produce a
3730 single value. The resulting value is the same type as its operands.
3732 '``shl``' Instruction
3733 ^^^^^^^^^^^^^^^^^^^^^
3740 <result> = shl <ty> <op1>, <op2> ; yields {ty}:result
3741 <result> = shl nuw <ty> <op1>, <op2> ; yields {ty}:result
3742 <result> = shl nsw <ty> <op1>, <op2> ; yields {ty}:result
3743 <result> = shl nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3748 The '``shl``' instruction returns the first operand shifted to the left
3749 a specified number of bits.
3754 Both arguments to the '``shl``' instruction must be the same
3755 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
3756 '``op2``' is treated as an unsigned value.
3761 The value produced is ``op1`` \* 2\ :sup:`op2` mod 2\ :sup:`n`,
3762 where ``n`` is the width of the result. If ``op2`` is (statically or
3763 dynamically) negative or equal to or larger than the number of bits in
3764 ``op1``, the result is undefined. If the arguments are vectors, each
3765 vector element of ``op1`` is shifted by the corresponding shift amount
3768 If the ``nuw`` keyword is present, then the shift produces a :ref:`poison
3769 value <poisonvalues>` if it shifts out any non-zero bits. If the
3770 ``nsw`` keyword is present, then the shift produces a :ref:`poison
3771 value <poisonvalues>` if it shifts out any bits that disagree with the
3772 resultant sign bit. As such, NUW/NSW have the same semantics as they
3773 would if the shift were expressed as a mul instruction with the same
3774 nsw/nuw bits in (mul %op1, (shl 1, %op2)).
3779 .. code-block:: llvm
3781 <result> = shl i32 4, %var ; yields {i32}: 4 << %var
3782 <result> = shl i32 4, 2 ; yields {i32}: 16
3783 <result> = shl i32 1, 10 ; yields {i32}: 1024
3784 <result> = shl i32 1, 32 ; undefined
3785 <result> = shl <2 x i32> < i32 1, i32 1>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 2, i32 4>
3787 '``lshr``' Instruction
3788 ^^^^^^^^^^^^^^^^^^^^^^
3795 <result> = lshr <ty> <op1>, <op2> ; yields {ty}:result
3796 <result> = lshr exact <ty> <op1>, <op2> ; yields {ty}:result
3801 The '``lshr``' instruction (logical shift right) returns the first
3802 operand shifted to the right a specified number of bits with zero fill.
3807 Both arguments to the '``lshr``' instruction must be the same
3808 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
3809 '``op2``' is treated as an unsigned value.
3814 This instruction always performs a logical shift right operation. The
3815 most significant bits of the result will be filled with zero bits after
3816 the shift. If ``op2`` is (statically or dynamically) equal to or larger
3817 than the number of bits in ``op1``, the result is undefined. If the
3818 arguments are vectors, each vector element of ``op1`` is shifted by the
3819 corresponding shift amount in ``op2``.
3821 If the ``exact`` keyword is present, the result value of the ``lshr`` is
3822 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
3828 .. code-block:: llvm
3830 <result> = lshr i32 4, 1 ; yields {i32}:result = 2
3831 <result> = lshr i32 4, 2 ; yields {i32}:result = 1
3832 <result> = lshr i8 4, 3 ; yields {i8}:result = 0
3833 <result> = lshr i8 -2, 1 ; yields {i8}:result = 0x7FFFFFFF
3834 <result> = lshr i32 1, 32 ; undefined
3835 <result> = lshr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 0x7FFFFFFF, i32 1>
3837 '``ashr``' Instruction
3838 ^^^^^^^^^^^^^^^^^^^^^^
3845 <result> = ashr <ty> <op1>, <op2> ; yields {ty}:result
3846 <result> = ashr exact <ty> <op1>, <op2> ; yields {ty}:result
3851 The '``ashr``' instruction (arithmetic shift right) returns the first
3852 operand shifted to the right a specified number of bits with sign
3858 Both arguments to the '``ashr``' instruction must be the same
3859 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
3860 '``op2``' is treated as an unsigned value.
3865 This instruction always performs an arithmetic shift right operation,
3866 The most significant bits of the result will be filled with the sign bit
3867 of ``op1``. If ``op2`` is (statically or dynamically) equal to or larger
3868 than the number of bits in ``op1``, the result is undefined. If the
3869 arguments are vectors, each vector element of ``op1`` is shifted by the
3870 corresponding shift amount in ``op2``.
3872 If the ``exact`` keyword is present, the result value of the ``ashr`` is
3873 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
3879 .. code-block:: llvm
3881 <result> = ashr i32 4, 1 ; yields {i32}:result = 2
3882 <result> = ashr i32 4, 2 ; yields {i32}:result = 1
3883 <result> = ashr i8 4, 3 ; yields {i8}:result = 0
3884 <result> = ashr i8 -2, 1 ; yields {i8}:result = -1
3885 <result> = ashr i32 1, 32 ; undefined
3886 <result> = ashr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 3> ; yields: result=<2 x i32> < i32 -1, i32 0>
3888 '``and``' Instruction
3889 ^^^^^^^^^^^^^^^^^^^^^
3896 <result> = and <ty> <op1>, <op2> ; yields {ty}:result
3901 The '``and``' instruction returns the bitwise logical and of its two
3907 The two arguments to the '``and``' instruction must be
3908 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3909 arguments must have identical types.
3914 The truth table used for the '``and``' instruction is:
3931 .. code-block:: llvm
3933 <result> = and i32 4, %var ; yields {i32}:result = 4 & %var
3934 <result> = and i32 15, 40 ; yields {i32}:result = 8
3935 <result> = and i32 4, 8 ; yields {i32}:result = 0
3937 '``or``' Instruction
3938 ^^^^^^^^^^^^^^^^^^^^
3945 <result> = or <ty> <op1>, <op2> ; yields {ty}:result
3950 The '``or``' instruction returns the bitwise logical inclusive or of its
3956 The two arguments to the '``or``' instruction must be
3957 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3958 arguments must have identical types.
3963 The truth table used for the '``or``' instruction is:
3982 <result> = or i32 4, %var ; yields {i32}:result = 4 | %var
3983 <result> = or i32 15, 40 ; yields {i32}:result = 47
3984 <result> = or i32 4, 8 ; yields {i32}:result = 12
3986 '``xor``' Instruction
3987 ^^^^^^^^^^^^^^^^^^^^^
3994 <result> = xor <ty> <op1>, <op2> ; yields {ty}:result
3999 The '``xor``' instruction returns the bitwise logical exclusive or of
4000 its two operands. The ``xor`` is used to implement the "one's
4001 complement" operation, which is the "~" operator in C.
4006 The two arguments to the '``xor``' instruction must be
4007 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4008 arguments must have identical types.
4013 The truth table used for the '``xor``' instruction is:
4030 .. code-block:: llvm
4032 <result> = xor i32 4, %var ; yields {i32}:result = 4 ^ %var
4033 <result> = xor i32 15, 40 ; yields {i32}:result = 39
4034 <result> = xor i32 4, 8 ; yields {i32}:result = 12
4035 <result> = xor i32 %V, -1 ; yields {i32}:result = ~%V
4040 LLVM supports several instructions to represent vector operations in a
4041 target-independent manner. These instructions cover the element-access
4042 and vector-specific operations needed to process vectors effectively.
4043 While LLVM does directly support these vector operations, many
4044 sophisticated algorithms will want to use target-specific intrinsics to
4045 take full advantage of a specific target.
4047 .. _i_extractelement:
4049 '``extractelement``' Instruction
4050 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4057 <result> = extractelement <n x <ty>> <val>, i32 <idx> ; yields <ty>
4062 The '``extractelement``' instruction extracts a single scalar element
4063 from a vector at a specified index.
4068 The first operand of an '``extractelement``' instruction is a value of
4069 :ref:`vector <t_vector>` type. The second operand is an index indicating
4070 the position from which to extract the element. The index may be a
4076 The result is a scalar of the same type as the element type of ``val``.
4077 Its value is the value at position ``idx`` of ``val``. If ``idx``
4078 exceeds the length of ``val``, the results are undefined.
4083 .. code-block:: llvm
4085 <result> = extractelement <4 x i32> %vec, i32 0 ; yields i32
4087 .. _i_insertelement:
4089 '``insertelement``' Instruction
4090 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4097 <result> = insertelement <n x <ty>> <val>, <ty> <elt>, i32 <idx> ; yields <n x <ty>>
4102 The '``insertelement``' instruction inserts a scalar element into a
4103 vector at a specified index.
4108 The first operand of an '``insertelement``' instruction is a value of
4109 :ref:`vector <t_vector>` type. The second operand is a scalar value whose
4110 type must equal the element type of the first operand. The third operand
4111 is an index indicating the position at which to insert the value. The
4112 index may be a variable.
4117 The result is a vector of the same type as ``val``. Its element values
4118 are those of ``val`` except at position ``idx``, where it gets the value
4119 ``elt``. If ``idx`` exceeds the length of ``val``, the results are
4125 .. code-block:: llvm
4127 <result> = insertelement <4 x i32> %vec, i32 1, i32 0 ; yields <4 x i32>
4129 .. _i_shufflevector:
4131 '``shufflevector``' Instruction
4132 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4139 <result> = shufflevector <n x <ty>> <v1>, <n x <ty>> <v2>, <m x i32> <mask> ; yields <m x <ty>>
4144 The '``shufflevector``' instruction constructs a permutation of elements
4145 from two input vectors, returning a vector with the same element type as
4146 the input and length that is the same as the shuffle mask.
4151 The first two operands of a '``shufflevector``' instruction are vectors
4152 with the same type. The third argument is a shuffle mask whose element
4153 type is always 'i32'. The result of the instruction is a vector whose
4154 length is the same as the shuffle mask and whose element type is the
4155 same as the element type of the first two operands.
4157 The shuffle mask operand is required to be a constant vector with either
4158 constant integer or undef values.
4163 The elements of the two input vectors are numbered from left to right
4164 across both of the vectors. The shuffle mask operand specifies, for each
4165 element of the result vector, which element of the two input vectors the
4166 result element gets. The element selector may be undef (meaning "don't
4167 care") and the second operand may be undef if performing a shuffle from
4173 .. code-block:: llvm
4175 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
4176 <4 x i32> <i32 0, i32 4, i32 1, i32 5> ; yields <4 x i32>
4177 <result> = shufflevector <4 x i32> %v1, <4 x i32> undef,
4178 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32> - Identity shuffle.
4179 <result> = shufflevector <8 x i32> %v1, <8 x i32> undef,
4180 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32>
4181 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
4182 <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7 > ; yields <8 x i32>
4184 Aggregate Operations
4185 --------------------
4187 LLVM supports several instructions for working with
4188 :ref:`aggregate <t_aggregate>` values.
4192 '``extractvalue``' Instruction
4193 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4200 <result> = extractvalue <aggregate type> <val>, <idx>{, <idx>}*
4205 The '``extractvalue``' instruction extracts the value of a member field
4206 from an :ref:`aggregate <t_aggregate>` value.
4211 The first operand of an '``extractvalue``' instruction is a value of
4212 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The operands are
4213 constant indices to specify which value to extract in a similar manner
4214 as indices in a '``getelementptr``' instruction.
4216 The major differences to ``getelementptr`` indexing are:
4218 - Since the value being indexed is not a pointer, the first index is
4219 omitted and assumed to be zero.
4220 - At least one index must be specified.
4221 - Not only struct indices but also array indices must be in bounds.
4226 The result is the value at the position in the aggregate specified by
4232 .. code-block:: llvm
4234 <result> = extractvalue {i32, float} %agg, 0 ; yields i32
4238 '``insertvalue``' Instruction
4239 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4246 <result> = insertvalue <aggregate type> <val>, <ty> <elt>, <idx>{, <idx>}* ; yields <aggregate type>
4251 The '``insertvalue``' instruction inserts a value into a member field in
4252 an :ref:`aggregate <t_aggregate>` value.
4257 The first operand of an '``insertvalue``' instruction is a value of
4258 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The second operand is
4259 a first-class value to insert. The following operands are constant
4260 indices indicating the position at which to insert the value in a
4261 similar manner as indices in a '``extractvalue``' instruction. The value
4262 to insert must have the same type as the value identified by the
4268 The result is an aggregate of the same type as ``val``. Its value is
4269 that of ``val`` except that the value at the position specified by the
4270 indices is that of ``elt``.
4275 .. code-block:: llvm
4277 %agg1 = insertvalue {i32, float} undef, i32 1, 0 ; yields {i32 1, float undef}
4278 %agg2 = insertvalue {i32, float} %agg1, float %val, 1 ; yields {i32 1, float %val}
4279 %agg3 = insertvalue {i32, {float}} %agg1, float %val, 1, 0 ; yields {i32 1, float %val}
4283 Memory Access and Addressing Operations
4284 ---------------------------------------
4286 A key design point of an SSA-based representation is how it represents
4287 memory. In LLVM, no memory locations are in SSA form, which makes things
4288 very simple. This section describes how to read, write, and allocate
4293 '``alloca``' Instruction
4294 ^^^^^^^^^^^^^^^^^^^^^^^^
4301 <result> = alloca <type>[, <ty> <NumElements>][, align <alignment>] ; yields {type*}:result
4306 The '``alloca``' instruction allocates memory on the stack frame of the
4307 currently executing function, to be automatically released when this
4308 function returns to its caller. The object is always allocated in the
4309 generic address space (address space zero).
4314 The '``alloca``' instruction allocates ``sizeof(<type>)*NumElements``
4315 bytes of memory on the runtime stack, returning a pointer of the
4316 appropriate type to the program. If "NumElements" is specified, it is
4317 the number of elements allocated, otherwise "NumElements" is defaulted
4318 to be one. If a constant alignment is specified, the value result of the
4319 allocation is guaranteed to be aligned to at least that boundary. If not
4320 specified, or if zero, the target can choose to align the allocation on
4321 any convenient boundary compatible with the type.
4323 '``type``' may be any sized type.
4328 Memory is allocated; a pointer is returned. The operation is undefined
4329 if there is insufficient stack space for the allocation. '``alloca``'d
4330 memory is automatically released when the function returns. The
4331 '``alloca``' instruction is commonly used to represent automatic
4332 variables that must have an address available. When the function returns
4333 (either with the ``ret`` or ``resume`` instructions), the memory is
4334 reclaimed. Allocating zero bytes is legal, but the result is undefined.
4335 The order in which memory is allocated (ie., which way the stack grows)
4341 .. code-block:: llvm
4343 %ptr = alloca i32 ; yields {i32*}:ptr
4344 %ptr = alloca i32, i32 4 ; yields {i32*}:ptr
4345 %ptr = alloca i32, i32 4, align 1024 ; yields {i32*}:ptr
4346 %ptr = alloca i32, align 1024 ; yields {i32*}:ptr
4350 '``load``' Instruction
4351 ^^^^^^^^^^^^^^^^^^^^^^
4358 <result> = load [volatile] <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>][, !invariant.load !<index>]
4359 <result> = load atomic [volatile] <ty>* <pointer> [singlethread] <ordering>, align <alignment>
4360 !<index> = !{ i32 1 }
4365 The '``load``' instruction is used to read from memory.
4370 The argument to the '``load``' instruction specifies the memory address
4371 from which to load. The pointer must point to a :ref:`first
4372 class <t_firstclass>` type. If the ``load`` is marked as ``volatile``,
4373 then the optimizer is not allowed to modify the number or order of
4374 execution of this ``load`` with other :ref:`volatile
4375 operations <volatile>`.
4377 If the ``load`` is marked as ``atomic``, it takes an extra
4378 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
4379 ``release`` and ``acq_rel`` orderings are not valid on ``load``
4380 instructions. Atomic loads produce :ref:`defined <memmodel>` results
4381 when they may see multiple atomic stores. The type of the pointee must
4382 be an integer type whose bit width is a power of two greater than or
4383 equal to eight and less than or equal to a target-specific size limit.
4384 ``align`` must be explicitly specified on atomic loads, and the load has
4385 undefined behavior if the alignment is not set to a value which is at
4386 least the size in bytes of the pointee. ``!nontemporal`` does not have
4387 any defined semantics for atomic loads.
4389 The optional constant ``align`` argument specifies the alignment of the
4390 operation (that is, the alignment of the memory address). A value of 0
4391 or an omitted ``align`` argument means that the operation has the abi
4392 alignment for the target. It is the responsibility of the code emitter
4393 to ensure that the alignment information is correct. Overestimating the
4394 alignment results in undefined behavior. Underestimating the alignment
4395 may produce less efficient code. An alignment of 1 is always safe.
4397 The optional ``!nontemporal`` metadata must reference a single
4398 metatadata name <index> corresponding to a metadata node with one
4399 ``i32`` entry of value 1. The existence of the ``!nontemporal``
4400 metatadata on the instruction tells the optimizer and code generator
4401 that this load is not expected to be reused in the cache. The code
4402 generator may select special instructions to save cache bandwidth, such
4403 as the ``MOVNT`` instruction on x86.
4405 The optional ``!invariant.load`` metadata must reference a single
4406 metatadata name <index> corresponding to a metadata node with no
4407 entries. The existence of the ``!invariant.load`` metatadata on the
4408 instruction tells the optimizer and code generator that this load
4409 address points to memory which does not change value during program
4410 execution. The optimizer may then move this load around, for example, by
4411 hoisting it out of loops using loop invariant code motion.
4416 The location of memory pointed to is loaded. If the value being loaded
4417 is of scalar type then the number of bytes read does not exceed the
4418 minimum number of bytes needed to hold all bits of the type. For
4419 example, loading an ``i24`` reads at most three bytes. When loading a
4420 value of a type like ``i20`` with a size that is not an integral number
4421 of bytes, the result is undefined if the value was not originally
4422 written using a store of the same type.
4427 .. code-block:: llvm
4429 %ptr = alloca i32 ; yields {i32*}:ptr
4430 store i32 3, i32* %ptr ; yields {void}
4431 %val = load i32* %ptr ; yields {i32}:val = i32 3
4435 '``store``' Instruction
4436 ^^^^^^^^^^^^^^^^^^^^^^^
4443 store [volatile] <ty> <value>, <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>] ; yields {void}
4444 store atomic [volatile] <ty> <value>, <ty>* <pointer> [singlethread] <ordering>, align <alignment> ; yields {void}
4449 The '``store``' instruction is used to write to memory.
4454 There are two arguments to the '``store``' instruction: a value to store
4455 and an address at which to store it. The type of the '``<pointer>``'
4456 operand must be a pointer to the :ref:`first class <t_firstclass>` type of
4457 the '``<value>``' operand. If the ``store`` is marked as ``volatile``,
4458 then the optimizer is not allowed to modify the number or order of
4459 execution of this ``store`` with other :ref:`volatile
4460 operations <volatile>`.
4462 If the ``store`` is marked as ``atomic``, it takes an extra
4463 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
4464 ``acquire`` and ``acq_rel`` orderings aren't valid on ``store``
4465 instructions. Atomic loads produce :ref:`defined <memmodel>` results
4466 when they may see multiple atomic stores. The type of the pointee must
4467 be an integer type whose bit width is a power of two greater than or
4468 equal to eight and less than or equal to a target-specific size limit.
4469 ``align`` must be explicitly specified on atomic stores, and the store
4470 has undefined behavior if the alignment is not set to a value which is
4471 at least the size in bytes of the pointee. ``!nontemporal`` does not
4472 have any defined semantics for atomic stores.
4474 The optional constant "align" argument specifies the alignment of the
4475 operation (that is, the alignment of the memory address). A value of 0
4476 or an omitted "align" argument means that the operation has the abi
4477 alignment for the target. It is the responsibility of the code emitter
4478 to ensure that the alignment information is correct. Overestimating the
4479 alignment results in an undefined behavior. Underestimating the
4480 alignment may produce less efficient code. An alignment of 1 is always
4483 The optional !nontemporal metadata must reference a single metatadata
4484 name <index> corresponding to a metadata node with one i32 entry of
4485 value 1. The existence of the !nontemporal metatadata on the instruction
4486 tells the optimizer and code generator that this load is not expected to
4487 be reused in the cache. The code generator may select special
4488 instructions to save cache bandwidth, such as the MOVNT instruction on
4494 The contents of memory are updated to contain '``<value>``' at the
4495 location specified by the '``<pointer>``' operand. If '``<value>``' is
4496 of scalar type then the number of bytes written does not exceed the
4497 minimum number of bytes needed to hold all bits of the type. For
4498 example, storing an ``i24`` writes at most three bytes. When writing a
4499 value of a type like ``i20`` with a size that is not an integral number
4500 of bytes, it is unspecified what happens to the extra bits that do not
4501 belong to the type, but they will typically be overwritten.
4506 .. code-block:: llvm
4508 %ptr = alloca i32 ; yields {i32*}:ptr
4509 store i32 3, i32* %ptr ; yields {void}
4510 %val = load i32* %ptr ; yields {i32}:val = i32 3
4514 '``fence``' Instruction
4515 ^^^^^^^^^^^^^^^^^^^^^^^
4522 fence [singlethread] <ordering> ; yields {void}
4527 The '``fence``' instruction is used to introduce happens-before edges
4533 '``fence``' instructions take an :ref:`ordering <ordering>` argument which
4534 defines what *synchronizes-with* edges they add. They can only be given
4535 ``acquire``, ``release``, ``acq_rel``, and ``seq_cst`` orderings.
4540 A fence A which has (at least) ``release`` ordering semantics
4541 *synchronizes with* a fence B with (at least) ``acquire`` ordering
4542 semantics if and only if there exist atomic operations X and Y, both
4543 operating on some atomic object M, such that A is sequenced before X, X
4544 modifies M (either directly or through some side effect of a sequence
4545 headed by X), Y is sequenced before B, and Y observes M. This provides a
4546 *happens-before* dependency between A and B. Rather than an explicit
4547 ``fence``, one (but not both) of the atomic operations X or Y might
4548 provide a ``release`` or ``acquire`` (resp.) ordering constraint and
4549 still *synchronize-with* the explicit ``fence`` and establish the
4550 *happens-before* edge.
4552 A ``fence`` which has ``seq_cst`` ordering, in addition to having both
4553 ``acquire`` and ``release`` semantics specified above, participates in
4554 the global program order of other ``seq_cst`` operations and/or fences.
4556 The optional ":ref:`singlethread <singlethread>`" argument specifies
4557 that the fence only synchronizes with other fences in the same thread.
4558 (This is useful for interacting with signal handlers.)
4563 .. code-block:: llvm
4565 fence acquire ; yields {void}
4566 fence singlethread seq_cst ; yields {void}
4570 '``cmpxchg``' Instruction
4571 ^^^^^^^^^^^^^^^^^^^^^^^^^
4578 cmpxchg [volatile] <ty>* <pointer>, <ty> <cmp>, <ty> <new> [singlethread] <ordering> ; yields {ty}
4583 The '``cmpxchg``' instruction is used to atomically modify memory. It
4584 loads a value in memory and compares it to a given value. If they are
4585 equal, it stores a new value into the memory.
4590 There are three arguments to the '``cmpxchg``' instruction: an address
4591 to operate on, a value to compare to the value currently be at that
4592 address, and a new value to place at that address if the compared values
4593 are equal. The type of '<cmp>' must be an integer type whose bit width
4594 is a power of two greater than or equal to eight and less than or equal
4595 to a target-specific size limit. '<cmp>' and '<new>' must have the same
4596 type, and the type of '<pointer>' must be a pointer to that type. If the
4597 ``cmpxchg`` is marked as ``volatile``, then the optimizer is not allowed
4598 to modify the number or order of execution of this ``cmpxchg`` with
4599 other :ref:`volatile operations <volatile>`.
4601 The :ref:`ordering <ordering>` argument specifies how this ``cmpxchg``
4602 synchronizes with other atomic operations.
4604 The optional "``singlethread``" argument declares that the ``cmpxchg``
4605 is only atomic with respect to code (usually signal handlers) running in
4606 the same thread as the ``cmpxchg``. Otherwise the cmpxchg is atomic with
4607 respect to all other code in the system.
4609 The pointer passed into cmpxchg must have alignment greater than or
4610 equal to the size in memory of the operand.
4615 The contents of memory at the location specified by the '``<pointer>``'
4616 operand is read and compared to '``<cmp>``'; if the read value is the
4617 equal, '``<new>``' is written. The original value at the location is
4620 A successful ``cmpxchg`` is a read-modify-write instruction for the purpose
4621 of identifying release sequences. A failed ``cmpxchg`` is equivalent to an
4622 atomic load with an ordering parameter determined by dropping any
4623 ``release`` part of the ``cmpxchg``'s ordering.
4628 .. code-block:: llvm
4631 %orig = atomic load i32* %ptr unordered ; yields {i32}
4635 %cmp = phi i32 [ %orig, %entry ], [%old, %loop]
4636 %squared = mul i32 %cmp, %cmp
4637 %old = cmpxchg i32* %ptr, i32 %cmp, i32 %squared ; yields {i32}
4638 %success = icmp eq i32 %cmp, %old
4639 br i1 %success, label %done, label %loop
4646 '``atomicrmw``' Instruction
4647 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
4654 atomicrmw [volatile] <operation> <ty>* <pointer>, <ty> <value> [singlethread] <ordering> ; yields {ty}
4659 The '``atomicrmw``' instruction is used to atomically modify memory.
4664 There are three arguments to the '``atomicrmw``' instruction: an
4665 operation to apply, an address whose value to modify, an argument to the
4666 operation. The operation must be one of the following keywords:
4680 The type of '<value>' must be an integer type whose bit width is a power
4681 of two greater than or equal to eight and less than or equal to a
4682 target-specific size limit. The type of the '``<pointer>``' operand must
4683 be a pointer to that type. If the ``atomicrmw`` is marked as
4684 ``volatile``, then the optimizer is not allowed to modify the number or
4685 order of execution of this ``atomicrmw`` with other :ref:`volatile
4686 operations <volatile>`.
4691 The contents of memory at the location specified by the '``<pointer>``'
4692 operand are atomically read, modified, and written back. The original
4693 value at the location is returned. The modification is specified by the
4696 - xchg: ``*ptr = val``
4697 - add: ``*ptr = *ptr + val``
4698 - sub: ``*ptr = *ptr - val``
4699 - and: ``*ptr = *ptr & val``
4700 - nand: ``*ptr = ~(*ptr & val)``
4701 - or: ``*ptr = *ptr | val``
4702 - xor: ``*ptr = *ptr ^ val``
4703 - max: ``*ptr = *ptr > val ? *ptr : val`` (using a signed comparison)
4704 - min: ``*ptr = *ptr < val ? *ptr : val`` (using a signed comparison)
4705 - umax: ``*ptr = *ptr > val ? *ptr : val`` (using an unsigned
4707 - umin: ``*ptr = *ptr < val ? *ptr : val`` (using an unsigned
4713 .. code-block:: llvm
4715 %old = atomicrmw add i32* %ptr, i32 1 acquire ; yields {i32}
4717 .. _i_getelementptr:
4719 '``getelementptr``' Instruction
4720 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4727 <result> = getelementptr <pty>* <ptrval>{, <ty> <idx>}*
4728 <result> = getelementptr inbounds <pty>* <ptrval>{, <ty> <idx>}*
4729 <result> = getelementptr <ptr vector> ptrval, <vector index type> idx
4734 The '``getelementptr``' instruction is used to get the address of a
4735 subelement of an :ref:`aggregate <t_aggregate>` data structure. It performs
4736 address calculation only and does not access memory.
4741 The first argument is always a pointer or a vector of pointers, and
4742 forms the basis of the calculation. The remaining arguments are indices
4743 that indicate which of the elements of the aggregate object are indexed.
4744 The interpretation of each index is dependent on the type being indexed
4745 into. The first index always indexes the pointer value given as the
4746 first argument, the second index indexes a value of the type pointed to
4747 (not necessarily the value directly pointed to, since the first index
4748 can be non-zero), etc. The first type indexed into must be a pointer
4749 value, subsequent types can be arrays, vectors, and structs. Note that
4750 subsequent types being indexed into can never be pointers, since that
4751 would require loading the pointer before continuing calculation.
4753 The type of each index argument depends on the type it is indexing into.
4754 When indexing into a (optionally packed) structure, only ``i32`` integer
4755 **constants** are allowed (when using a vector of indices they must all
4756 be the **same** ``i32`` integer constant). When indexing into an array,
4757 pointer or vector, integers of any width are allowed, and they are not
4758 required to be constant. These integers are treated as signed values
4761 For example, let's consider a C code fragment and how it gets compiled
4777 int *foo(struct ST *s) {
4778 return &s[1].Z.B[5][13];
4781 The LLVM code generated by Clang is:
4783 .. code-block:: llvm
4785 %struct.RT = type { i8, [10 x [20 x i32]], i8 }
4786 %struct.ST = type { i32, double, %struct.RT }
4788 define i32* @foo(%struct.ST* %s) nounwind uwtable readnone optsize ssp {
4790 %arrayidx = getelementptr inbounds %struct.ST* %s, i64 1, i32 2, i32 1, i64 5, i64 13
4797 In the example above, the first index is indexing into the
4798 '``%struct.ST*``' type, which is a pointer, yielding a '``%struct.ST``'
4799 = '``{ i32, double, %struct.RT }``' type, a structure. The second index
4800 indexes into the third element of the structure, yielding a
4801 '``%struct.RT``' = '``{ i8 , [10 x [20 x i32]], i8 }``' type, another
4802 structure. The third index indexes into the second element of the
4803 structure, yielding a '``[10 x [20 x i32]]``' type, an array. The two
4804 dimensions of the array are subscripted into, yielding an '``i32``'
4805 type. The '``getelementptr``' instruction returns a pointer to this
4806 element, thus computing a value of '``i32*``' type.
4808 Note that it is perfectly legal to index partially through a structure,
4809 returning a pointer to an inner element. Because of this, the LLVM code
4810 for the given testcase is equivalent to:
4812 .. code-block:: llvm
4814 define i32* @foo(%struct.ST* %s) {
4815 %t1 = getelementptr %struct.ST* %s, i32 1 ; yields %struct.ST*:%t1
4816 %t2 = getelementptr %struct.ST* %t1, i32 0, i32 2 ; yields %struct.RT*:%t2
4817 %t3 = getelementptr %struct.RT* %t2, i32 0, i32 1 ; yields [10 x [20 x i32]]*:%t3
4818 %t4 = getelementptr [10 x [20 x i32]]* %t3, i32 0, i32 5 ; yields [20 x i32]*:%t4
4819 %t5 = getelementptr [20 x i32]* %t4, i32 0, i32 13 ; yields i32*:%t5
4823 If the ``inbounds`` keyword is present, the result value of the
4824 ``getelementptr`` is a :ref:`poison value <poisonvalues>` if the base
4825 pointer is not an *in bounds* address of an allocated object, or if any
4826 of the addresses that would be formed by successive addition of the
4827 offsets implied by the indices to the base address with infinitely
4828 precise signed arithmetic are not an *in bounds* address of that
4829 allocated object. The *in bounds* addresses for an allocated object are
4830 all the addresses that point into the object, plus the address one byte
4831 past the end. In cases where the base is a vector of pointers the
4832 ``inbounds`` keyword applies to each of the computations element-wise.
4834 If the ``inbounds`` keyword is not present, the offsets are added to the
4835 base address with silently-wrapping two's complement arithmetic. If the
4836 offsets have a different width from the pointer, they are sign-extended
4837 or truncated to the width of the pointer. The result value of the
4838 ``getelementptr`` may be outside the object pointed to by the base
4839 pointer. The result value may not necessarily be used to access memory
4840 though, even if it happens to point into allocated storage. See the
4841 :ref:`Pointer Aliasing Rules <pointeraliasing>` section for more
4844 The getelementptr instruction is often confusing. For some more insight
4845 into how it works, see :doc:`the getelementptr FAQ <GetElementPtr>`.
4850 .. code-block:: llvm
4852 ; yields [12 x i8]*:aptr
4853 %aptr = getelementptr {i32, [12 x i8]}* %saptr, i64 0, i32 1
4855 %vptr = getelementptr {i32, <2 x i8>}* %svptr, i64 0, i32 1, i32 1
4857 %eptr = getelementptr [12 x i8]* %aptr, i64 0, i32 1
4859 %iptr = getelementptr [10 x i32]* @arr, i16 0, i16 0
4861 In cases where the pointer argument is a vector of pointers, each index
4862 must be a vector with the same number of elements. For example:
4864 .. code-block:: llvm
4866 %A = getelementptr <4 x i8*> %ptrs, <4 x i64> %offsets,
4868 Conversion Operations
4869 ---------------------
4871 The instructions in this category are the conversion instructions
4872 (casting) which all take a single operand and a type. They perform
4873 various bit conversions on the operand.
4875 '``trunc .. to``' Instruction
4876 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4883 <result> = trunc <ty> <value> to <ty2> ; yields ty2
4888 The '``trunc``' instruction truncates its operand to the type ``ty2``.
4893 The '``trunc``' instruction takes a value to trunc, and a type to trunc
4894 it to. Both types must be of :ref:`integer <t_integer>` types, or vectors
4895 of the same number of integers. The bit size of the ``value`` must be
4896 larger than the bit size of the destination type, ``ty2``. Equal sized
4897 types are not allowed.
4902 The '``trunc``' instruction truncates the high order bits in ``value``
4903 and converts the remaining bits to ``ty2``. Since the source size must
4904 be larger than the destination size, ``trunc`` cannot be a *no-op cast*.
4905 It will always truncate bits.
4910 .. code-block:: llvm
4912 %X = trunc i32 257 to i8 ; yields i8:1
4913 %Y = trunc i32 123 to i1 ; yields i1:true
4914 %Z = trunc i32 122 to i1 ; yields i1:false
4915 %W = trunc <2 x i16> <i16 8, i16 7> to <2 x i8> ; yields <i8 8, i8 7>
4917 '``zext .. to``' Instruction
4918 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4925 <result> = zext <ty> <value> to <ty2> ; yields ty2
4930 The '``zext``' instruction zero extends its operand to type ``ty2``.
4935 The '``zext``' instruction takes a value to cast, and a type to cast it
4936 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
4937 the same number of integers. The bit size of the ``value`` must be
4938 smaller than the bit size of the destination type, ``ty2``.
4943 The ``zext`` fills the high order bits of the ``value`` with zero bits
4944 until it reaches the size of the destination type, ``ty2``.
4946 When zero extending from i1, the result will always be either 0 or 1.
4951 .. code-block:: llvm
4953 %X = zext i32 257 to i64 ; yields i64:257
4954 %Y = zext i1 true to i32 ; yields i32:1
4955 %Z = zext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
4957 '``sext .. to``' Instruction
4958 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4965 <result> = sext <ty> <value> to <ty2> ; yields ty2
4970 The '``sext``' sign extends ``value`` to the type ``ty2``.
4975 The '``sext``' instruction takes a value to cast, and a type to cast it
4976 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
4977 the same number of integers. The bit size of the ``value`` must be
4978 smaller than the bit size of the destination type, ``ty2``.
4983 The '``sext``' instruction performs a sign extension by copying the sign
4984 bit (highest order bit) of the ``value`` until it reaches the bit size
4985 of the type ``ty2``.
4987 When sign extending from i1, the extension always results in -1 or 0.
4992 .. code-block:: llvm
4994 %X = sext i8 -1 to i16 ; yields i16 :65535
4995 %Y = sext i1 true to i32 ; yields i32:-1
4996 %Z = sext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
4998 '``fptrunc .. to``' Instruction
4999 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5006 <result> = fptrunc <ty> <value> to <ty2> ; yields ty2
5011 The '``fptrunc``' instruction truncates ``value`` to type ``ty2``.
5016 The '``fptrunc``' instruction takes a :ref:`floating point <t_floating>`
5017 value to cast and a :ref:`floating point <t_floating>` type to cast it to.
5018 The size of ``value`` must be larger than the size of ``ty2``. This
5019 implies that ``fptrunc`` cannot be used to make a *no-op cast*.
5024 The '``fptrunc``' instruction truncates a ``value`` from a larger
5025 :ref:`floating point <t_floating>` type to a smaller :ref:`floating
5026 point <t_floating>` type. If the value cannot fit within the
5027 destination type, ``ty2``, then the results are undefined.
5032 .. code-block:: llvm
5034 %X = fptrunc double 123.0 to float ; yields float:123.0
5035 %Y = fptrunc double 1.0E+300 to float ; yields undefined
5037 '``fpext .. to``' Instruction
5038 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5045 <result> = fpext <ty> <value> to <ty2> ; yields ty2
5050 The '``fpext``' extends a floating point ``value`` to a larger floating
5056 The '``fpext``' instruction takes a :ref:`floating point <t_floating>`
5057 ``value`` to cast, and a :ref:`floating point <t_floating>` type to cast it
5058 to. The source type must be smaller than the destination type.
5063 The '``fpext``' instruction extends the ``value`` from a smaller
5064 :ref:`floating point <t_floating>` type to a larger :ref:`floating
5065 point <t_floating>` type. The ``fpext`` cannot be used to make a
5066 *no-op cast* because it always changes bits. Use ``bitcast`` to make a
5067 *no-op cast* for a floating point cast.
5072 .. code-block:: llvm
5074 %X = fpext float 3.125 to double ; yields double:3.125000e+00
5075 %Y = fpext double %X to fp128 ; yields fp128:0xL00000000000000004000900000000000
5077 '``fptoui .. to``' Instruction
5078 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5085 <result> = fptoui <ty> <value> to <ty2> ; yields ty2
5090 The '``fptoui``' converts a floating point ``value`` to its unsigned
5091 integer equivalent of type ``ty2``.
5096 The '``fptoui``' instruction takes a value to cast, which must be a
5097 scalar or vector :ref:`floating point <t_floating>` value, and a type to
5098 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
5099 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
5100 type with the same number of elements as ``ty``
5105 The '``fptoui``' instruction converts its :ref:`floating
5106 point <t_floating>` operand into the nearest (rounding towards zero)
5107 unsigned integer value. If the value cannot fit in ``ty2``, the results
5113 .. code-block:: llvm
5115 %X = fptoui double 123.0 to i32 ; yields i32:123
5116 %Y = fptoui float 1.0E+300 to i1 ; yields undefined:1
5117 %Z = fptoui float 1.04E+17 to i8 ; yields undefined:1
5119 '``fptosi .. to``' Instruction
5120 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5127 <result> = fptosi <ty> <value> to <ty2> ; yields ty2
5132 The '``fptosi``' instruction converts :ref:`floating point <t_floating>`
5133 ``value`` to type ``ty2``.
5138 The '``fptosi``' instruction takes a value to cast, which must be a
5139 scalar or vector :ref:`floating point <t_floating>` value, and a type to
5140 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
5141 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
5142 type with the same number of elements as ``ty``
5147 The '``fptosi``' instruction converts its :ref:`floating
5148 point <t_floating>` operand into the nearest (rounding towards zero)
5149 signed integer value. If the value cannot fit in ``ty2``, the results
5155 .. code-block:: llvm
5157 %X = fptosi double -123.0 to i32 ; yields i32:-123
5158 %Y = fptosi float 1.0E-247 to i1 ; yields undefined:1
5159 %Z = fptosi float 1.04E+17 to i8 ; yields undefined:1
5161 '``uitofp .. to``' Instruction
5162 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5169 <result> = uitofp <ty> <value> to <ty2> ; yields ty2
5174 The '``uitofp``' instruction regards ``value`` as an unsigned integer
5175 and converts that value to the ``ty2`` type.
5180 The '``uitofp``' instruction takes a value to cast, which must be a
5181 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
5182 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
5183 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
5184 type with the same number of elements as ``ty``
5189 The '``uitofp``' instruction interprets its operand as an unsigned
5190 integer quantity and converts it to the corresponding floating point
5191 value. If the value cannot fit in the floating point value, the results
5197 .. code-block:: llvm
5199 %X = uitofp i32 257 to float ; yields float:257.0
5200 %Y = uitofp i8 -1 to double ; yields double:255.0
5202 '``sitofp .. to``' Instruction
5203 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5210 <result> = sitofp <ty> <value> to <ty2> ; yields ty2
5215 The '``sitofp``' instruction regards ``value`` as a signed integer and
5216 converts that value to the ``ty2`` type.
5221 The '``sitofp``' instruction takes a value to cast, which must be a
5222 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
5223 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
5224 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
5225 type with the same number of elements as ``ty``
5230 The '``sitofp``' instruction interprets its operand as a signed integer
5231 quantity and converts it to the corresponding floating point value. If
5232 the value cannot fit in the floating point value, the results are
5238 .. code-block:: llvm
5240 %X = sitofp i32 257 to float ; yields float:257.0
5241 %Y = sitofp i8 -1 to double ; yields double:-1.0
5245 '``ptrtoint .. to``' Instruction
5246 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5253 <result> = ptrtoint <ty> <value> to <ty2> ; yields ty2
5258 The '``ptrtoint``' instruction converts the pointer or a vector of
5259 pointers ``value`` to the integer (or vector of integers) type ``ty2``.
5264 The '``ptrtoint``' instruction takes a ``value`` to cast, which must be
5265 a a value of type :ref:`pointer <t_pointer>` or a vector of pointers, and a
5266 type to cast it to ``ty2``, which must be an :ref:`integer <t_integer>` or
5267 a vector of integers type.
5272 The '``ptrtoint``' instruction converts ``value`` to integer type
5273 ``ty2`` by interpreting the pointer value as an integer and either
5274 truncating or zero extending that value to the size of the integer type.
5275 If ``value`` is smaller than ``ty2`` then a zero extension is done. If
5276 ``value`` is larger than ``ty2`` then a truncation is done. If they are
5277 the same size, then nothing is done (*no-op cast*) other than a type
5283 .. code-block:: llvm
5285 %X = ptrtoint i32* %P to i8 ; yields truncation on 32-bit architecture
5286 %Y = ptrtoint i32* %P to i64 ; yields zero extension on 32-bit architecture
5287 %Z = ptrtoint <4 x i32*> %P to <4 x i64>; yields vector zero extension for a vector of addresses on 32-bit architecture
5291 '``inttoptr .. to``' Instruction
5292 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5299 <result> = inttoptr <ty> <value> to <ty2> ; yields ty2
5304 The '``inttoptr``' instruction converts an integer ``value`` to a
5305 pointer type, ``ty2``.
5310 The '``inttoptr``' instruction takes an :ref:`integer <t_integer>` value to
5311 cast, and a type to cast it to, which must be a :ref:`pointer <t_pointer>`
5317 The '``inttoptr``' instruction converts ``value`` to type ``ty2`` by
5318 applying either a zero extension or a truncation depending on the size
5319 of the integer ``value``. If ``value`` is larger than the size of a
5320 pointer then a truncation is done. If ``value`` is smaller than the size
5321 of a pointer then a zero extension is done. If they are the same size,
5322 nothing is done (*no-op cast*).
5327 .. code-block:: llvm
5329 %X = inttoptr i32 255 to i32* ; yields zero extension on 64-bit architecture
5330 %Y = inttoptr i32 255 to i32* ; yields no-op on 32-bit architecture
5331 %Z = inttoptr i64 0 to i32* ; yields truncation on 32-bit architecture
5332 %Z = inttoptr <4 x i32> %G to <4 x i8*>; yields truncation of vector G to four pointers
5336 '``bitcast .. to``' Instruction
5337 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5344 <result> = bitcast <ty> <value> to <ty2> ; yields ty2
5349 The '``bitcast``' instruction converts ``value`` to type ``ty2`` without
5355 The '``bitcast``' instruction takes a value to cast, which must be a
5356 non-aggregate first class value, and a type to cast it to, which must
5357 also be a non-aggregate :ref:`first class <t_firstclass>` type. The bit
5358 sizes of ``value`` and the destination type, ``ty2``, must be identical.
5359 If the source type is a pointer, the destination type must also be a
5360 pointer. This instruction supports bitwise conversion of vectors to
5361 integers and to vectors of other types (as long as they have the same
5367 The '``bitcast``' instruction converts ``value`` to type ``ty2``. It is
5368 always a *no-op cast* because no bits change with this conversion. The
5369 conversion is done as if the ``value`` had been stored to memory and
5370 read back as type ``ty2``. Pointer (or vector of pointers) types may
5371 only be converted to other pointer (or vector of pointers) types with
5372 this instruction. To convert pointers to other types, use the
5373 :ref:`inttoptr <i_inttoptr>` or :ref:`ptrtoint <i_ptrtoint>` instructions
5379 .. code-block:: llvm
5381 %X = bitcast i8 255 to i8 ; yields i8 :-1
5382 %Y = bitcast i32* %x to sint* ; yields sint*:%x
5383 %Z = bitcast <2 x int> %V to i64; ; yields i64: %V
5384 %Z = bitcast <2 x i32*> %V to <2 x i64*> ; yields <2 x i64*>
5391 The instructions in this category are the "miscellaneous" instructions,
5392 which defy better classification.
5396 '``icmp``' Instruction
5397 ^^^^^^^^^^^^^^^^^^^^^^
5404 <result> = icmp <cond> <ty> <op1>, <op2> ; yields {i1} or {<N x i1>}:result
5409 The '``icmp``' instruction returns a boolean value or a vector of
5410 boolean values based on comparison of its two integer, integer vector,
5411 pointer, or pointer vector operands.
5416 The '``icmp``' instruction takes three operands. The first operand is
5417 the condition code indicating the kind of comparison to perform. It is
5418 not a value, just a keyword. The possible condition code are:
5421 #. ``ne``: not equal
5422 #. ``ugt``: unsigned greater than
5423 #. ``uge``: unsigned greater or equal
5424 #. ``ult``: unsigned less than
5425 #. ``ule``: unsigned less or equal
5426 #. ``sgt``: signed greater than
5427 #. ``sge``: signed greater or equal
5428 #. ``slt``: signed less than
5429 #. ``sle``: signed less or equal
5431 The remaining two arguments must be :ref:`integer <t_integer>` or
5432 :ref:`pointer <t_pointer>` or integer :ref:`vector <t_vector>` typed. They
5433 must also be identical types.
5438 The '``icmp``' compares ``op1`` and ``op2`` according to the condition
5439 code given as ``cond``. The comparison performed always yields either an
5440 :ref:`i1 <t_integer>` or vector of ``i1`` result, as follows:
5442 #. ``eq``: yields ``true`` if the operands are equal, ``false``
5443 otherwise. No sign interpretation is necessary or performed.
5444 #. ``ne``: yields ``true`` if the operands are unequal, ``false``
5445 otherwise. No sign interpretation is necessary or performed.
5446 #. ``ugt``: interprets the operands as unsigned values and yields
5447 ``true`` if ``op1`` is greater than ``op2``.
5448 #. ``uge``: interprets the operands as unsigned values and yields
5449 ``true`` if ``op1`` is greater than or equal to ``op2``.
5450 #. ``ult``: interprets the operands as unsigned values and yields
5451 ``true`` if ``op1`` is less than ``op2``.
5452 #. ``ule``: interprets the operands as unsigned values and yields
5453 ``true`` if ``op1`` is less than or equal to ``op2``.
5454 #. ``sgt``: interprets the operands as signed values and yields ``true``
5455 if ``op1`` is greater than ``op2``.
5456 #. ``sge``: interprets the operands as signed values and yields ``true``
5457 if ``op1`` is greater than or equal to ``op2``.
5458 #. ``slt``: interprets the operands as signed values and yields ``true``
5459 if ``op1`` is less than ``op2``.
5460 #. ``sle``: interprets the operands as signed values and yields ``true``
5461 if ``op1`` is less than or equal to ``op2``.
5463 If the operands are :ref:`pointer <t_pointer>` typed, the pointer values
5464 are compared as if they were integers.
5466 If the operands are integer vectors, then they are compared element by
5467 element. The result is an ``i1`` vector with the same number of elements
5468 as the values being compared. Otherwise, the result is an ``i1``.
5473 .. code-block:: llvm
5475 <result> = icmp eq i32 4, 5 ; yields: result=false
5476 <result> = icmp ne float* %X, %X ; yields: result=false
5477 <result> = icmp ult i16 4, 5 ; yields: result=true
5478 <result> = icmp sgt i16 4, 5 ; yields: result=false
5479 <result> = icmp ule i16 -4, 5 ; yields: result=false
5480 <result> = icmp sge i16 4, 5 ; yields: result=false
5482 Note that the code generator does not yet support vector types with the
5483 ``icmp`` instruction.
5487 '``fcmp``' Instruction
5488 ^^^^^^^^^^^^^^^^^^^^^^
5495 <result> = fcmp <cond> <ty> <op1>, <op2> ; yields {i1} or {<N x i1>}:result
5500 The '``fcmp``' instruction returns a boolean value or vector of boolean
5501 values based on comparison of its operands.
5503 If the operands are floating point scalars, then the result type is a
5504 boolean (:ref:`i1 <t_integer>`).
5506 If the operands are floating point vectors, then the result type is a
5507 vector of boolean with the same number of elements as the operands being
5513 The '``fcmp``' instruction takes three operands. The first operand is
5514 the condition code indicating the kind of comparison to perform. It is
5515 not a value, just a keyword. The possible condition code are:
5517 #. ``false``: no comparison, always returns false
5518 #. ``oeq``: ordered and equal
5519 #. ``ogt``: ordered and greater than
5520 #. ``oge``: ordered and greater than or equal
5521 #. ``olt``: ordered and less than
5522 #. ``ole``: ordered and less than or equal
5523 #. ``one``: ordered and not equal
5524 #. ``ord``: ordered (no nans)
5525 #. ``ueq``: unordered or equal
5526 #. ``ugt``: unordered or greater than
5527 #. ``uge``: unordered or greater than or equal
5528 #. ``ult``: unordered or less than
5529 #. ``ule``: unordered or less than or equal
5530 #. ``une``: unordered or not equal
5531 #. ``uno``: unordered (either nans)
5532 #. ``true``: no comparison, always returns true
5534 *Ordered* means that neither operand is a QNAN while *unordered* means
5535 that either operand may be a QNAN.
5537 Each of ``val1`` and ``val2`` arguments must be either a :ref:`floating
5538 point <t_floating>` type or a :ref:`vector <t_vector>` of floating point
5539 type. They must have identical types.
5544 The '``fcmp``' instruction compares ``op1`` and ``op2`` according to the
5545 condition code given as ``cond``. If the operands are vectors, then the
5546 vectors are compared element by element. Each comparison performed
5547 always yields an :ref:`i1 <t_integer>` result, as follows:
5549 #. ``false``: always yields ``false``, regardless of operands.
5550 #. ``oeq``: yields ``true`` if both operands are not a QNAN and ``op1``
5551 is equal to ``op2``.
5552 #. ``ogt``: yields ``true`` if both operands are not a QNAN and ``op1``
5553 is greater than ``op2``.
5554 #. ``oge``: yields ``true`` if both operands are not a QNAN and ``op1``
5555 is greater than or equal to ``op2``.
5556 #. ``olt``: yields ``true`` if both operands are not a QNAN and ``op1``
5557 is less than ``op2``.
5558 #. ``ole``: yields ``true`` if both operands are not a QNAN and ``op1``
5559 is less than or equal to ``op2``.
5560 #. ``one``: yields ``true`` if both operands are not a QNAN and ``op1``
5561 is not equal to ``op2``.
5562 #. ``ord``: yields ``true`` if both operands are not a QNAN.
5563 #. ``ueq``: yields ``true`` if either operand is a QNAN or ``op1`` is
5565 #. ``ugt``: yields ``true`` if either operand is a QNAN or ``op1`` is
5566 greater than ``op2``.
5567 #. ``uge``: yields ``true`` if either operand is a QNAN or ``op1`` is
5568 greater than or equal to ``op2``.
5569 #. ``ult``: yields ``true`` if either operand is a QNAN or ``op1`` is
5571 #. ``ule``: yields ``true`` if either operand is a QNAN or ``op1`` is
5572 less than or equal to ``op2``.
5573 #. ``une``: yields ``true`` if either operand is a QNAN or ``op1`` is
5574 not equal to ``op2``.
5575 #. ``uno``: yields ``true`` if either operand is a QNAN.
5576 #. ``true``: always yields ``true``, regardless of operands.
5581 .. code-block:: llvm
5583 <result> = fcmp oeq float 4.0, 5.0 ; yields: result=false
5584 <result> = fcmp one float 4.0, 5.0 ; yields: result=true
5585 <result> = fcmp olt float 4.0, 5.0 ; yields: result=true
5586 <result> = fcmp ueq double 1.0, 2.0 ; yields: result=false
5588 Note that the code generator does not yet support vector types with the
5589 ``fcmp`` instruction.
5593 '``phi``' Instruction
5594 ^^^^^^^^^^^^^^^^^^^^^
5601 <result> = phi <ty> [ <val0>, <label0>], ...
5606 The '``phi``' instruction is used to implement the φ node in the SSA
5607 graph representing the function.
5612 The type of the incoming values is specified with the first type field.
5613 After this, the '``phi``' instruction takes a list of pairs as
5614 arguments, with one pair for each predecessor basic block of the current
5615 block. Only values of :ref:`first class <t_firstclass>` type may be used as
5616 the value arguments to the PHI node. Only labels may be used as the
5619 There must be no non-phi instructions between the start of a basic block
5620 and the PHI instructions: i.e. PHI instructions must be first in a basic
5623 For the purposes of the SSA form, the use of each incoming value is
5624 deemed to occur on the edge from the corresponding predecessor block to
5625 the current block (but after any definition of an '``invoke``'
5626 instruction's return value on the same edge).
5631 At runtime, the '``phi``' instruction logically takes on the value
5632 specified by the pair corresponding to the predecessor basic block that
5633 executed just prior to the current block.
5638 .. code-block:: llvm
5640 Loop: ; Infinite loop that counts from 0 on up...
5641 %indvar = phi i32 [ 0, %LoopHeader ], [ %nextindvar, %Loop ]
5642 %nextindvar = add i32 %indvar, 1
5647 '``select``' Instruction
5648 ^^^^^^^^^^^^^^^^^^^^^^^^
5655 <result> = select selty <cond>, <ty> <val1>, <ty> <val2> ; yields ty
5657 selty is either i1 or {<N x i1>}
5662 The '``select``' instruction is used to choose one value based on a
5663 condition, without branching.
5668 The '``select``' instruction requires an 'i1' value or a vector of 'i1'
5669 values indicating the condition, and two values of the same :ref:`first
5670 class <t_firstclass>` type. If the val1/val2 are vectors and the
5671 condition is a scalar, then entire vectors are selected, not individual
5677 If the condition is an i1 and it evaluates to 1, the instruction returns
5678 the first value argument; otherwise, it returns the second value
5681 If the condition is a vector of i1, then the value arguments must be
5682 vectors of the same size, and the selection is done element by element.
5687 .. code-block:: llvm
5689 %X = select i1 true, i8 17, i8 42 ; yields i8:17
5693 '``call``' Instruction
5694 ^^^^^^^^^^^^^^^^^^^^^^
5701 <result> = [tail] call [cconv] [ret attrs] <ty> [<fnty>*] <fnptrval>(<function args>) [fn attrs]
5706 The '``call``' instruction represents a simple function call.
5711 This instruction requires several arguments:
5713 #. The optional "tail" marker indicates that the callee function does
5714 not access any allocas or varargs in the caller. Note that calls may
5715 be marked "tail" even if they do not occur before a
5716 :ref:`ret <i_ret>` instruction. If the "tail" marker is present, the
5717 function call is eligible for tail call optimization, but `might not
5718 in fact be optimized into a jump <CodeGenerator.html#tailcallopt>`_.
5719 The code generator may optimize calls marked "tail" with either 1)
5720 automatic `sibling call
5721 optimization <CodeGenerator.html#sibcallopt>`_ when the caller and
5722 callee have matching signatures, or 2) forced tail call optimization
5723 when the following extra requirements are met:
5725 - Caller and callee both have the calling convention ``fastcc``.
5726 - The call is in tail position (ret immediately follows call and ret
5727 uses value of call or is void).
5728 - Option ``-tailcallopt`` is enabled, or
5729 ``llvm::GuaranteedTailCallOpt`` is ``true``.
5730 - `Platform specific constraints are
5731 met. <CodeGenerator.html#tailcallopt>`_
5733 #. The optional "cconv" marker indicates which :ref:`calling
5734 convention <callingconv>` the call should use. If none is
5735 specified, the call defaults to using C calling conventions. The
5736 calling convention of the call must match the calling convention of
5737 the target function, or else the behavior is undefined.
5738 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
5739 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
5741 #. '``ty``': the type of the call instruction itself which is also the
5742 type of the return value. Functions that return no value are marked
5744 #. '``fnty``': shall be the signature of the pointer to function value
5745 being invoked. The argument types must match the types implied by
5746 this signature. This type can be omitted if the function is not
5747 varargs and if the function type does not return a pointer to a
5749 #. '``fnptrval``': An LLVM value containing a pointer to a function to
5750 be invoked. In most cases, this is a direct function invocation, but
5751 indirect ``call``'s are just as possible, calling an arbitrary pointer
5753 #. '``function args``': argument list whose types match the function
5754 signature argument types and parameter attributes. All arguments must
5755 be of :ref:`first class <t_firstclass>` type. If the function signature
5756 indicates the function accepts a variable number of arguments, the
5757 extra arguments can be specified.
5758 #. The optional :ref:`function attributes <fnattrs>` list. Only
5759 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
5760 attributes are valid here.
5765 The '``call``' instruction is used to cause control flow to transfer to
5766 a specified function, with its incoming arguments bound to the specified
5767 values. Upon a '``ret``' instruction in the called function, control
5768 flow continues with the instruction after the function call, and the
5769 return value of the function is bound to the result argument.
5774 .. code-block:: llvm
5776 %retval = call i32 @test(i32 %argc)
5777 call i32 (i8*, ...)* @printf(i8* %msg, i32 12, i8 42) ; yields i32
5778 %X = tail call i32 @foo() ; yields i32
5779 %Y = tail call fastcc i32 @foo() ; yields i32
5780 call void %foo(i8 97 signext)
5782 %struct.A = type { i32, i8 }
5783 %r = call %struct.A @foo() ; yields { 32, i8 }
5784 %gr = extractvalue %struct.A %r, 0 ; yields i32
5785 %gr1 = extractvalue %struct.A %r, 1 ; yields i8
5786 %Z = call void @foo() noreturn ; indicates that %foo never returns normally
5787 %ZZ = call zeroext i32 @bar() ; Return value is %zero extended
5789 llvm treats calls to some functions with names and arguments that match
5790 the standard C99 library as being the C99 library functions, and may
5791 perform optimizations or generate code for them under that assumption.
5792 This is something we'd like to change in the future to provide better
5793 support for freestanding environments and non-C-based languages.
5797 '``va_arg``' Instruction
5798 ^^^^^^^^^^^^^^^^^^^^^^^^
5805 <resultval> = va_arg <va_list*> <arglist>, <argty>
5810 The '``va_arg``' instruction is used to access arguments passed through
5811 the "variable argument" area of a function call. It is used to implement
5812 the ``va_arg`` macro in C.
5817 This instruction takes a ``va_list*`` value and the type of the
5818 argument. It returns a value of the specified argument type and
5819 increments the ``va_list`` to point to the next argument. The actual
5820 type of ``va_list`` is target specific.
5825 The '``va_arg``' instruction loads an argument of the specified type
5826 from the specified ``va_list`` and causes the ``va_list`` to point to
5827 the next argument. For more information, see the variable argument
5828 handling :ref:`Intrinsic Functions <int_varargs>`.
5830 It is legal for this instruction to be called in a function which does
5831 not take a variable number of arguments, for example, the ``vfprintf``
5834 ``va_arg`` is an LLVM instruction instead of an :ref:`intrinsic
5835 function <intrinsics>` because it takes a type as an argument.
5840 See the :ref:`variable argument processing <int_varargs>` section.
5842 Note that the code generator does not yet fully support va\_arg on many
5843 targets. Also, it does not currently support va\_arg with aggregate
5844 types on any target.
5848 '``landingpad``' Instruction
5849 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5856 <resultval> = landingpad <resultty> personality <type> <pers_fn> <clause>+
5857 <resultval> = landingpad <resultty> personality <type> <pers_fn> cleanup <clause>*
5859 <clause> := catch <type> <value>
5860 <clause> := filter <array constant type> <array constant>
5865 The '``landingpad``' instruction is used by `LLVM's exception handling
5866 system <ExceptionHandling.html#overview>`_ to specify that a basic block
5867 is a landing pad --- one where the exception lands, and corresponds to the
5868 code found in the ``catch`` portion of a ``try``/``catch`` sequence. It
5869 defines values supplied by the personality function (``pers_fn``) upon
5870 re-entry to the function. The ``resultval`` has the type ``resultty``.
5875 This instruction takes a ``pers_fn`` value. This is the personality
5876 function associated with the unwinding mechanism. The optional
5877 ``cleanup`` flag indicates that the landing pad block is a cleanup.
5879 A ``clause`` begins with the clause type --- ``catch`` or ``filter`` --- and
5880 contains the global variable representing the "type" that may be caught
5881 or filtered respectively. Unlike the ``catch`` clause, the ``filter``
5882 clause takes an array constant as its argument. Use
5883 "``[0 x i8**] undef``" for a filter which cannot throw. The
5884 '``landingpad``' instruction must contain *at least* one ``clause`` or
5885 the ``cleanup`` flag.
5890 The '``landingpad``' instruction defines the values which are set by the
5891 personality function (``pers_fn``) upon re-entry to the function, and
5892 therefore the "result type" of the ``landingpad`` instruction. As with
5893 calling conventions, how the personality function results are
5894 represented in LLVM IR is target specific.
5896 The clauses are applied in order from top to bottom. If two
5897 ``landingpad`` instructions are merged together through inlining, the
5898 clauses from the calling function are appended to the list of clauses.
5899 When the call stack is being unwound due to an exception being thrown,
5900 the exception is compared against each ``clause`` in turn. If it doesn't
5901 match any of the clauses, and the ``cleanup`` flag is not set, then
5902 unwinding continues further up the call stack.
5904 The ``landingpad`` instruction has several restrictions:
5906 - A landing pad block is a basic block which is the unwind destination
5907 of an '``invoke``' instruction.
5908 - A landing pad block must have a '``landingpad``' instruction as its
5909 first non-PHI instruction.
5910 - There can be only one '``landingpad``' instruction within the landing
5912 - A basic block that is not a landing pad block may not include a
5913 '``landingpad``' instruction.
5914 - All '``landingpad``' instructions in a function must have the same
5915 personality function.
5920 .. code-block:: llvm
5922 ;; A landing pad which can catch an integer.
5923 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
5925 ;; A landing pad that is a cleanup.
5926 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
5928 ;; A landing pad which can catch an integer and can only throw a double.
5929 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
5931 filter [1 x i8**] [@_ZTId]
5938 LLVM supports the notion of an "intrinsic function". These functions
5939 have well known names and semantics and are required to follow certain
5940 restrictions. Overall, these intrinsics represent an extension mechanism
5941 for the LLVM language that does not require changing all of the
5942 transformations in LLVM when adding to the language (or the bitcode
5943 reader/writer, the parser, etc...).
5945 Intrinsic function names must all start with an "``llvm.``" prefix. This
5946 prefix is reserved in LLVM for intrinsic names; thus, function names may
5947 not begin with this prefix. Intrinsic functions must always be external
5948 functions: you cannot define the body of intrinsic functions. Intrinsic
5949 functions may only be used in call or invoke instructions: it is illegal
5950 to take the address of an intrinsic function. Additionally, because
5951 intrinsic functions are part of the LLVM language, it is required if any
5952 are added that they be documented here.
5954 Some intrinsic functions can be overloaded, i.e., the intrinsic
5955 represents a family of functions that perform the same operation but on
5956 different data types. Because LLVM can represent over 8 million
5957 different integer types, overloading is used commonly to allow an
5958 intrinsic function to operate on any integer type. One or more of the
5959 argument types or the result type can be overloaded to accept any
5960 integer type. Argument types may also be defined as exactly matching a
5961 previous argument's type or the result type. This allows an intrinsic
5962 function which accepts multiple arguments, but needs all of them to be
5963 of the same type, to only be overloaded with respect to a single
5964 argument or the result.
5966 Overloaded intrinsics will have the names of its overloaded argument
5967 types encoded into its function name, each preceded by a period. Only
5968 those types which are overloaded result in a name suffix. Arguments
5969 whose type is matched against another type do not. For example, the
5970 ``llvm.ctpop`` function can take an integer of any width and returns an
5971 integer of exactly the same integer width. This leads to a family of
5972 functions such as ``i8 @llvm.ctpop.i8(i8 %val)`` and
5973 ``i29 @llvm.ctpop.i29(i29 %val)``. Only one type, the return type, is
5974 overloaded, and only one type suffix is required. Because the argument's
5975 type is matched against the return type, it does not require its own
5978 To learn how to add an intrinsic function, please see the `Extending
5979 LLVM Guide <ExtendingLLVM.html>`_.
5983 Variable Argument Handling Intrinsics
5984 -------------------------------------
5986 Variable argument support is defined in LLVM with the
5987 :ref:`va_arg <i_va_arg>` instruction and these three intrinsic
5988 functions. These functions are related to the similarly named macros
5989 defined in the ``<stdarg.h>`` header file.
5991 All of these functions operate on arguments that use a target-specific
5992 value type "``va_list``". The LLVM assembly language reference manual
5993 does not define what this type is, so all transformations should be
5994 prepared to handle these functions regardless of the type used.
5996 This example shows how the :ref:`va_arg <i_va_arg>` instruction and the
5997 variable argument handling intrinsic functions are used.
5999 .. code-block:: llvm
6001 define i32 @test(i32 %X, ...) {
6002 ; Initialize variable argument processing
6004 %ap2 = bitcast i8** %ap to i8*
6005 call void @llvm.va_start(i8* %ap2)
6007 ; Read a single integer argument
6008 %tmp = va_arg i8** %ap, i32
6010 ; Demonstrate usage of llvm.va_copy and llvm.va_end
6012 %aq2 = bitcast i8** %aq to i8*
6013 call void @llvm.va_copy(i8* %aq2, i8* %ap2)
6014 call void @llvm.va_end(i8* %aq2)
6016 ; Stop processing of arguments.
6017 call void @llvm.va_end(i8* %ap2)
6021 declare void @llvm.va_start(i8*)
6022 declare void @llvm.va_copy(i8*, i8*)
6023 declare void @llvm.va_end(i8*)
6027 '``llvm.va_start``' Intrinsic
6028 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6035 declare void %llvm.va_start(i8* <arglist>)
6040 The '``llvm.va_start``' intrinsic initializes ``*<arglist>`` for
6041 subsequent use by ``va_arg``.
6046 The argument is a pointer to a ``va_list`` element to initialize.
6051 The '``llvm.va_start``' intrinsic works just like the ``va_start`` macro
6052 available in C. In a target-dependent way, it initializes the
6053 ``va_list`` element to which the argument points, so that the next call
6054 to ``va_arg`` will produce the first variable argument passed to the
6055 function. Unlike the C ``va_start`` macro, this intrinsic does not need
6056 to know the last argument of the function as the compiler can figure
6059 '``llvm.va_end``' Intrinsic
6060 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6067 declare void @llvm.va_end(i8* <arglist>)
6072 The '``llvm.va_end``' intrinsic destroys ``*<arglist>``, which has been
6073 initialized previously with ``llvm.va_start`` or ``llvm.va_copy``.
6078 The argument is a pointer to a ``va_list`` to destroy.
6083 The '``llvm.va_end``' intrinsic works just like the ``va_end`` macro
6084 available in C. In a target-dependent way, it destroys the ``va_list``
6085 element to which the argument points. Calls to
6086 :ref:`llvm.va_start <int_va_start>` and
6087 :ref:`llvm.va_copy <int_va_copy>` must be matched exactly with calls to
6092 '``llvm.va_copy``' Intrinsic
6093 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6100 declare void @llvm.va_copy(i8* <destarglist>, i8* <srcarglist>)
6105 The '``llvm.va_copy``' intrinsic copies the current argument position
6106 from the source argument list to the destination argument list.
6111 The first argument is a pointer to a ``va_list`` element to initialize.
6112 The second argument is a pointer to a ``va_list`` element to copy from.
6117 The '``llvm.va_copy``' intrinsic works just like the ``va_copy`` macro
6118 available in C. In a target-dependent way, it copies the source
6119 ``va_list`` element into the destination ``va_list`` element. This
6120 intrinsic is necessary because the `` llvm.va_start`` intrinsic may be
6121 arbitrarily complex and require, for example, memory allocation.
6123 Accurate Garbage Collection Intrinsics
6124 --------------------------------------
6126 LLVM support for `Accurate Garbage Collection <GarbageCollection.html>`_
6127 (GC) requires the implementation and generation of these intrinsics.
6128 These intrinsics allow identification of :ref:`GC roots on the
6129 stack <int_gcroot>`, as well as garbage collector implementations that
6130 require :ref:`read <int_gcread>` and :ref:`write <int_gcwrite>` barriers.
6131 Front-ends for type-safe garbage collected languages should generate
6132 these intrinsics to make use of the LLVM garbage collectors. For more
6133 details, see `Accurate Garbage Collection with
6134 LLVM <GarbageCollection.html>`_.
6136 The garbage collection intrinsics only operate on objects in the generic
6137 address space (address space zero).
6141 '``llvm.gcroot``' Intrinsic
6142 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6149 declare void @llvm.gcroot(i8** %ptrloc, i8* %metadata)
6154 The '``llvm.gcroot``' intrinsic declares the existence of a GC root to
6155 the code generator, and allows some metadata to be associated with it.
6160 The first argument specifies the address of a stack object that contains
6161 the root pointer. The second pointer (which must be either a constant or
6162 a global value address) contains the meta-data to be associated with the
6168 At runtime, a call to this intrinsic stores a null pointer into the
6169 "ptrloc" location. At compile-time, the code generator generates
6170 information to allow the runtime to find the pointer at GC safe points.
6171 The '``llvm.gcroot``' intrinsic may only be used in a function which
6172 :ref:`specifies a GC algorithm <gc>`.
6176 '``llvm.gcread``' Intrinsic
6177 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6184 declare i8* @llvm.gcread(i8* %ObjPtr, i8** %Ptr)
6189 The '``llvm.gcread``' intrinsic identifies reads of references from heap
6190 locations, allowing garbage collector implementations that require read
6196 The second argument is the address to read from, which should be an
6197 address allocated from the garbage collector. The first object is a
6198 pointer to the start of the referenced object, if needed by the language
6199 runtime (otherwise null).
6204 The '``llvm.gcread``' intrinsic has the same semantics as a load
6205 instruction, but may be replaced with substantially more complex code by
6206 the garbage collector runtime, as needed. The '``llvm.gcread``'
6207 intrinsic may only be used in a function which :ref:`specifies a GC
6212 '``llvm.gcwrite``' Intrinsic
6213 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6220 declare void @llvm.gcwrite(i8* %P1, i8* %Obj, i8** %P2)
6225 The '``llvm.gcwrite``' intrinsic identifies writes of references to heap
6226 locations, allowing garbage collector implementations that require write
6227 barriers (such as generational or reference counting collectors).
6232 The first argument is the reference to store, the second is the start of
6233 the object to store it to, and the third is the address of the field of
6234 Obj to store to. If the runtime does not require a pointer to the
6235 object, Obj may be null.
6240 The '``llvm.gcwrite``' intrinsic has the same semantics as a store
6241 instruction, but may be replaced with substantially more complex code by
6242 the garbage collector runtime, as needed. The '``llvm.gcwrite``'
6243 intrinsic may only be used in a function which :ref:`specifies a GC
6246 Code Generator Intrinsics
6247 -------------------------
6249 These intrinsics are provided by LLVM to expose special features that
6250 may only be implemented with code generator support.
6252 '``llvm.returnaddress``' Intrinsic
6253 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6260 declare i8 *@llvm.returnaddress(i32 <level>)
6265 The '``llvm.returnaddress``' intrinsic attempts to compute a
6266 target-specific value indicating the return address of the current
6267 function or one of its callers.
6272 The argument to this intrinsic indicates which function to return the
6273 address for. Zero indicates the calling function, one indicates its
6274 caller, etc. The argument is **required** to be a constant integer
6280 The '``llvm.returnaddress``' intrinsic either returns a pointer
6281 indicating the return address of the specified call frame, or zero if it
6282 cannot be identified. The value returned by this intrinsic is likely to
6283 be incorrect or 0 for arguments other than zero, so it should only be
6284 used for debugging purposes.
6286 Note that calling this intrinsic does not prevent function inlining or
6287 other aggressive transformations, so the value returned may not be that
6288 of the obvious source-language caller.
6290 '``llvm.frameaddress``' Intrinsic
6291 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6298 declare i8* @llvm.frameaddress(i32 <level>)
6303 The '``llvm.frameaddress``' intrinsic attempts to return the
6304 target-specific frame pointer value for the specified stack frame.
6309 The argument to this intrinsic indicates which function to return the
6310 frame pointer for. Zero indicates the calling function, one indicates
6311 its caller, etc. The argument is **required** to be a constant integer
6317 The '``llvm.frameaddress``' intrinsic either returns a pointer
6318 indicating the frame address of the specified call frame, or zero if it
6319 cannot be identified. The value returned by this intrinsic is likely to
6320 be incorrect or 0 for arguments other than zero, so it should only be
6321 used for debugging purposes.
6323 Note that calling this intrinsic does not prevent function inlining or
6324 other aggressive transformations, so the value returned may not be that
6325 of the obvious source-language caller.
6329 '``llvm.stacksave``' Intrinsic
6330 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6337 declare i8* @llvm.stacksave()
6342 The '``llvm.stacksave``' intrinsic is used to remember the current state
6343 of the function stack, for use with
6344 :ref:`llvm.stackrestore <int_stackrestore>`. This is useful for
6345 implementing language features like scoped automatic variable sized
6351 This intrinsic returns a opaque pointer value that can be passed to
6352 :ref:`llvm.stackrestore <int_stackrestore>`. When an
6353 ``llvm.stackrestore`` intrinsic is executed with a value saved from
6354 ``llvm.stacksave``, it effectively restores the state of the stack to
6355 the state it was in when the ``llvm.stacksave`` intrinsic executed. In
6356 practice, this pops any :ref:`alloca <i_alloca>` blocks from the stack that
6357 were allocated after the ``llvm.stacksave`` was executed.
6359 .. _int_stackrestore:
6361 '``llvm.stackrestore``' Intrinsic
6362 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6369 declare void @llvm.stackrestore(i8* %ptr)
6374 The '``llvm.stackrestore``' intrinsic is used to restore the state of
6375 the function stack to the state it was in when the corresponding
6376 :ref:`llvm.stacksave <int_stacksave>` intrinsic executed. This is
6377 useful for implementing language features like scoped automatic variable
6378 sized arrays in C99.
6383 See the description for :ref:`llvm.stacksave <int_stacksave>`.
6385 '``llvm.prefetch``' Intrinsic
6386 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6393 declare void @llvm.prefetch(i8* <address>, i32 <rw>, i32 <locality>, i32 <cache type>)
6398 The '``llvm.prefetch``' intrinsic is a hint to the code generator to
6399 insert a prefetch instruction if supported; otherwise, it is a noop.
6400 Prefetches have no effect on the behavior of the program but can change
6401 its performance characteristics.
6406 ``address`` is the address to be prefetched, ``rw`` is the specifier
6407 determining if the fetch should be for a read (0) or write (1), and
6408 ``locality`` is a temporal locality specifier ranging from (0) - no
6409 locality, to (3) - extremely local keep in cache. The ``cache type``
6410 specifies whether the prefetch is performed on the data (1) or
6411 instruction (0) cache. The ``rw``, ``locality`` and ``cache type``
6412 arguments must be constant integers.
6417 This intrinsic does not modify the behavior of the program. In
6418 particular, prefetches cannot trap and do not produce a value. On
6419 targets that support this intrinsic, the prefetch can provide hints to
6420 the processor cache for better performance.
6422 '``llvm.pcmarker``' Intrinsic
6423 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6430 declare void @llvm.pcmarker(i32 <id>)
6435 The '``llvm.pcmarker``' intrinsic is a method to export a Program
6436 Counter (PC) in a region of code to simulators and other tools. The
6437 method is target specific, but it is expected that the marker will use
6438 exported symbols to transmit the PC of the marker. The marker makes no
6439 guarantees that it will remain with any specific instruction after
6440 optimizations. It is possible that the presence of a marker will inhibit
6441 optimizations. The intended use is to be inserted after optimizations to
6442 allow correlations of simulation runs.
6447 ``id`` is a numerical id identifying the marker.
6452 This intrinsic does not modify the behavior of the program. Backends
6453 that do not support this intrinsic may ignore it.
6455 '``llvm.readcyclecounter``' Intrinsic
6456 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6463 declare i64 @llvm.readcyclecounter()
6468 The '``llvm.readcyclecounter``' intrinsic provides access to the cycle
6469 counter register (or similar low latency, high accuracy clocks) on those
6470 targets that support it. On X86, it should map to RDTSC. On Alpha, it
6471 should map to RPCC. As the backing counters overflow quickly (on the
6472 order of 9 seconds on alpha), this should only be used for small
6478 When directly supported, reading the cycle counter should not modify any
6479 memory. Implementations are allowed to either return a application
6480 specific value or a system wide value. On backends without support, this
6481 is lowered to a constant 0.
6483 Standard C Library Intrinsics
6484 -----------------------------
6486 LLVM provides intrinsics for a few important standard C library
6487 functions. These intrinsics allow source-language front-ends to pass
6488 information about the alignment of the pointer arguments to the code
6489 generator, providing opportunity for more efficient code generation.
6493 '``llvm.memcpy``' Intrinsic
6494 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6499 This is an overloaded intrinsic. You can use ``llvm.memcpy`` on any
6500 integer bit width and for different address spaces. Not all targets
6501 support all bit widths however.
6505 declare void @llvm.memcpy.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
6506 i32 <len>, i32 <align>, i1 <isvolatile>)
6507 declare void @llvm.memcpy.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
6508 i64 <len>, i32 <align>, i1 <isvolatile>)
6513 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
6514 source location to the destination location.
6516 Note that, unlike the standard libc function, the ``llvm.memcpy.*``
6517 intrinsics do not return a value, takes extra alignment/isvolatile
6518 arguments and the pointers can be in specified address spaces.
6523 The first argument is a pointer to the destination, the second is a
6524 pointer to the source. The third argument is an integer argument
6525 specifying the number of bytes to copy, the fourth argument is the
6526 alignment of the source and destination locations, and the fifth is a
6527 boolean indicating a volatile access.
6529 If the call to this intrinsic has an alignment value that is not 0 or 1,
6530 then the caller guarantees that both the source and destination pointers
6531 are aligned to that boundary.
6533 If the ``isvolatile`` parameter is ``true``, the ``llvm.memcpy`` call is
6534 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
6535 very cleanly specified and it is unwise to depend on it.
6540 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
6541 source location to the destination location, which are not allowed to
6542 overlap. It copies "len" bytes of memory over. If the argument is known
6543 to be aligned to some boundary, this can be specified as the fourth
6544 argument, otherwise it should be set to 0 or 1.
6546 '``llvm.memmove``' Intrinsic
6547 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6552 This is an overloaded intrinsic. You can use llvm.memmove on any integer
6553 bit width and for different address space. Not all targets support all
6558 declare void @llvm.memmove.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
6559 i32 <len>, i32 <align>, i1 <isvolatile>)
6560 declare void @llvm.memmove.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
6561 i64 <len>, i32 <align>, i1 <isvolatile>)
6566 The '``llvm.memmove.*``' intrinsics move a block of memory from the
6567 source location to the destination location. It is similar to the
6568 '``llvm.memcpy``' intrinsic but allows the two memory locations to
6571 Note that, unlike the standard libc function, the ``llvm.memmove.*``
6572 intrinsics do not return a value, takes extra alignment/isvolatile
6573 arguments and the pointers can be in specified address spaces.
6578 The first argument is a pointer to the destination, the second is a
6579 pointer to the source. The third argument is an integer argument
6580 specifying the number of bytes to copy, the fourth argument is the
6581 alignment of the source and destination locations, and the fifth is a
6582 boolean indicating a volatile access.
6584 If the call to this intrinsic has an alignment value that is not 0 or 1,
6585 then the caller guarantees that the source and destination pointers are
6586 aligned to that boundary.
6588 If the ``isvolatile`` parameter is ``true``, the ``llvm.memmove`` call
6589 is a :ref:`volatile operation <volatile>`. The detailed access behavior is
6590 not very cleanly specified and it is unwise to depend on it.
6595 The '``llvm.memmove.*``' intrinsics copy a block of memory from the
6596 source location to the destination location, which may overlap. It
6597 copies "len" bytes of memory over. If the argument is known to be
6598 aligned to some boundary, this can be specified as the fourth argument,
6599 otherwise it should be set to 0 or 1.
6601 '``llvm.memset.*``' Intrinsics
6602 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6607 This is an overloaded intrinsic. You can use llvm.memset on any integer
6608 bit width and for different address spaces. However, not all targets
6609 support all bit widths.
6613 declare void @llvm.memset.p0i8.i32(i8* <dest>, i8 <val>,
6614 i32 <len>, i32 <align>, i1 <isvolatile>)
6615 declare void @llvm.memset.p0i8.i64(i8* <dest>, i8 <val>,
6616 i64 <len>, i32 <align>, i1 <isvolatile>)
6621 The '``llvm.memset.*``' intrinsics fill a block of memory with a
6622 particular byte value.
6624 Note that, unlike the standard libc function, the ``llvm.memset``
6625 intrinsic does not return a value and takes extra alignment/volatile
6626 arguments. Also, the destination can be in an arbitrary address space.
6631 The first argument is a pointer to the destination to fill, the second
6632 is the byte value with which to fill it, the third argument is an
6633 integer argument specifying the number of bytes to fill, and the fourth
6634 argument is the known alignment of the destination location.
6636 If the call to this intrinsic has an alignment value that is not 0 or 1,
6637 then the caller guarantees that the destination pointer is aligned to
6640 If the ``isvolatile`` parameter is ``true``, the ``llvm.memset`` call is
6641 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
6642 very cleanly specified and it is unwise to depend on it.
6647 The '``llvm.memset.*``' intrinsics fill "len" bytes of memory starting
6648 at the destination location. If the argument is known to be aligned to
6649 some boundary, this can be specified as the fourth argument, otherwise
6650 it should be set to 0 or 1.
6652 '``llvm.sqrt.*``' Intrinsic
6653 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6658 This is an overloaded intrinsic. You can use ``llvm.sqrt`` on any
6659 floating point or vector of floating point type. Not all targets support
6664 declare float @llvm.sqrt.f32(float %Val)
6665 declare double @llvm.sqrt.f64(double %Val)
6666 declare x86_fp80 @llvm.sqrt.f80(x86_fp80 %Val)
6667 declare fp128 @llvm.sqrt.f128(fp128 %Val)
6668 declare ppc_fp128 @llvm.sqrt.ppcf128(ppc_fp128 %Val)
6673 The '``llvm.sqrt``' intrinsics return the sqrt of the specified operand,
6674 returning the same value as the libm '``sqrt``' functions would. Unlike
6675 ``sqrt`` in libm, however, ``llvm.sqrt`` has undefined behavior for
6676 negative numbers other than -0.0 (which allows for better optimization,
6677 because there is no need to worry about errno being set).
6678 ``llvm.sqrt(-0.0)`` is defined to return -0.0 like IEEE sqrt.
6683 The argument and return value are floating point numbers of the same
6689 This function returns the sqrt of the specified operand if it is a
6690 nonnegative floating point number.
6692 '``llvm.powi.*``' Intrinsic
6693 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6698 This is an overloaded intrinsic. You can use ``llvm.powi`` on any
6699 floating point or vector of floating point type. Not all targets support
6704 declare float @llvm.powi.f32(float %Val, i32 %power)
6705 declare double @llvm.powi.f64(double %Val, i32 %power)
6706 declare x86_fp80 @llvm.powi.f80(x86_fp80 %Val, i32 %power)
6707 declare fp128 @llvm.powi.f128(fp128 %Val, i32 %power)
6708 declare ppc_fp128 @llvm.powi.ppcf128(ppc_fp128 %Val, i32 %power)
6713 The '``llvm.powi.*``' intrinsics return the first operand raised to the
6714 specified (positive or negative) power. The order of evaluation of
6715 multiplications is not defined. When a vector of floating point type is
6716 used, the second argument remains a scalar integer value.
6721 The second argument is an integer power, and the first is a value to
6722 raise to that power.
6727 This function returns the first value raised to the second power with an
6728 unspecified sequence of rounding operations.
6730 '``llvm.sin.*``' Intrinsic
6731 ^^^^^^^^^^^^^^^^^^^^^^^^^^
6736 This is an overloaded intrinsic. You can use ``llvm.sin`` on any
6737 floating point or vector of floating point type. Not all targets support
6742 declare float @llvm.sin.f32(float %Val)
6743 declare double @llvm.sin.f64(double %Val)
6744 declare x86_fp80 @llvm.sin.f80(x86_fp80 %Val)
6745 declare fp128 @llvm.sin.f128(fp128 %Val)
6746 declare ppc_fp128 @llvm.sin.ppcf128(ppc_fp128 %Val)
6751 The '``llvm.sin.*``' intrinsics return the sine of the operand.
6756 The argument and return value are floating point numbers of the same
6762 This function returns the sine of the specified operand, returning the
6763 same values as the libm ``sin`` functions would, and handles error
6764 conditions in the same way.
6766 '``llvm.cos.*``' Intrinsic
6767 ^^^^^^^^^^^^^^^^^^^^^^^^^^
6772 This is an overloaded intrinsic. You can use ``llvm.cos`` on any
6773 floating point or vector of floating point type. Not all targets support
6778 declare float @llvm.cos.f32(float %Val)
6779 declare double @llvm.cos.f64(double %Val)
6780 declare x86_fp80 @llvm.cos.f80(x86_fp80 %Val)
6781 declare fp128 @llvm.cos.f128(fp128 %Val)
6782 declare ppc_fp128 @llvm.cos.ppcf128(ppc_fp128 %Val)
6787 The '``llvm.cos.*``' intrinsics return the cosine of the operand.
6792 The argument and return value are floating point numbers of the same
6798 This function returns the cosine of the specified operand, returning the
6799 same values as the libm ``cos`` functions would, and handles error
6800 conditions in the same way.
6802 '``llvm.pow.*``' Intrinsic
6803 ^^^^^^^^^^^^^^^^^^^^^^^^^^
6808 This is an overloaded intrinsic. You can use ``llvm.pow`` on any
6809 floating point or vector of floating point type. Not all targets support
6814 declare float @llvm.pow.f32(float %Val, float %Power)
6815 declare double @llvm.pow.f64(double %Val, double %Power)
6816 declare x86_fp80 @llvm.pow.f80(x86_fp80 %Val, x86_fp80 %Power)
6817 declare fp128 @llvm.pow.f128(fp128 %Val, fp128 %Power)
6818 declare ppc_fp128 @llvm.pow.ppcf128(ppc_fp128 %Val, ppc_fp128 Power)
6823 The '``llvm.pow.*``' intrinsics return the first operand raised to the
6824 specified (positive or negative) power.
6829 The second argument is a floating point power, and the first is a value
6830 to raise to that power.
6835 This function returns the first value raised to the second power,
6836 returning the same values as the libm ``pow`` functions would, and
6837 handles error conditions in the same way.
6839 '``llvm.exp.*``' Intrinsic
6840 ^^^^^^^^^^^^^^^^^^^^^^^^^^
6845 This is an overloaded intrinsic. You can use ``llvm.exp`` on any
6846 floating point or vector of floating point type. Not all targets support
6851 declare float @llvm.exp.f32(float %Val)
6852 declare double @llvm.exp.f64(double %Val)
6853 declare x86_fp80 @llvm.exp.f80(x86_fp80 %Val)
6854 declare fp128 @llvm.exp.f128(fp128 %Val)
6855 declare ppc_fp128 @llvm.exp.ppcf128(ppc_fp128 %Val)
6860 The '``llvm.exp.*``' intrinsics perform the exp function.
6865 The argument and return value are floating point numbers of the same
6871 This function returns the same values as the libm ``exp`` functions
6872 would, and handles error conditions in the same way.
6874 '``llvm.exp2.*``' Intrinsic
6875 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6880 This is an overloaded intrinsic. You can use ``llvm.exp2`` on any
6881 floating point or vector of floating point type. Not all targets support
6886 declare float @llvm.exp2.f32(float %Val)
6887 declare double @llvm.exp2.f64(double %Val)
6888 declare x86_fp80 @llvm.exp2.f80(x86_fp80 %Val)
6889 declare fp128 @llvm.exp2.f128(fp128 %Val)
6890 declare ppc_fp128 @llvm.exp2.ppcf128(ppc_fp128 %Val)
6895 The '``llvm.exp2.*``' intrinsics perform the exp2 function.
6900 The argument and return value are floating point numbers of the same
6906 This function returns the same values as the libm ``exp2`` functions
6907 would, and handles error conditions in the same way.
6909 '``llvm.log.*``' Intrinsic
6910 ^^^^^^^^^^^^^^^^^^^^^^^^^^
6915 This is an overloaded intrinsic. You can use ``llvm.log`` on any
6916 floating point or vector of floating point type. Not all targets support
6921 declare float @llvm.log.f32(float %Val)
6922 declare double @llvm.log.f64(double %Val)
6923 declare x86_fp80 @llvm.log.f80(x86_fp80 %Val)
6924 declare fp128 @llvm.log.f128(fp128 %Val)
6925 declare ppc_fp128 @llvm.log.ppcf128(ppc_fp128 %Val)
6930 The '``llvm.log.*``' intrinsics perform the log function.
6935 The argument and return value are floating point numbers of the same
6941 This function returns the same values as the libm ``log`` functions
6942 would, and handles error conditions in the same way.
6944 '``llvm.log10.*``' Intrinsic
6945 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6950 This is an overloaded intrinsic. You can use ``llvm.log10`` on any
6951 floating point or vector of floating point type. Not all targets support
6956 declare float @llvm.log10.f32(float %Val)
6957 declare double @llvm.log10.f64(double %Val)
6958 declare x86_fp80 @llvm.log10.f80(x86_fp80 %Val)
6959 declare fp128 @llvm.log10.f128(fp128 %Val)
6960 declare ppc_fp128 @llvm.log10.ppcf128(ppc_fp128 %Val)
6965 The '``llvm.log10.*``' intrinsics perform the log10 function.
6970 The argument and return value are floating point numbers of the same
6976 This function returns the same values as the libm ``log10`` functions
6977 would, and handles error conditions in the same way.
6979 '``llvm.log2.*``' Intrinsic
6980 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6985 This is an overloaded intrinsic. You can use ``llvm.log2`` on any
6986 floating point or vector of floating point type. Not all targets support
6991 declare float @llvm.log2.f32(float %Val)
6992 declare double @llvm.log2.f64(double %Val)
6993 declare x86_fp80 @llvm.log2.f80(x86_fp80 %Val)
6994 declare fp128 @llvm.log2.f128(fp128 %Val)
6995 declare ppc_fp128 @llvm.log2.ppcf128(ppc_fp128 %Val)
7000 The '``llvm.log2.*``' intrinsics perform the log2 function.
7005 The argument and return value are floating point numbers of the same
7011 This function returns the same values as the libm ``log2`` functions
7012 would, and handles error conditions in the same way.
7014 '``llvm.fma.*``' Intrinsic
7015 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7020 This is an overloaded intrinsic. You can use ``llvm.fma`` on any
7021 floating point or vector of floating point type. Not all targets support
7026 declare float @llvm.fma.f32(float %a, float %b, float %c)
7027 declare double @llvm.fma.f64(double %a, double %b, double %c)
7028 declare x86_fp80 @llvm.fma.f80(x86_fp80 %a, x86_fp80 %b, x86_fp80 %c)
7029 declare fp128 @llvm.fma.f128(fp128 %a, fp128 %b, fp128 %c)
7030 declare ppc_fp128 @llvm.fma.ppcf128(ppc_fp128 %a, ppc_fp128 %b, ppc_fp128 %c)
7035 The '``llvm.fma.*``' intrinsics perform the fused multiply-add
7041 The argument and return value are floating point numbers of the same
7047 This function returns the same values as the libm ``fma`` functions
7050 '``llvm.fabs.*``' Intrinsic
7051 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7056 This is an overloaded intrinsic. You can use ``llvm.fabs`` on any
7057 floating point or vector of floating point type. Not all targets support
7062 declare float @llvm.fabs.f32(float %Val)
7063 declare double @llvm.fabs.f64(double %Val)
7064 declare x86_fp80 @llvm.fabs.f80(x86_fp80 %Val)
7065 declare fp128 @llvm.fabs.f128(fp128 %Val)
7066 declare ppc_fp128 @llvm.fabs.ppcf128(ppc_fp128 %Val)
7071 The '``llvm.fabs.*``' intrinsics return the absolute value of the
7077 The argument and return value are floating point numbers of the same
7083 This function returns the same values as the libm ``fabs`` functions
7084 would, and handles error conditions in the same way.
7086 '``llvm.floor.*``' Intrinsic
7087 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7092 This is an overloaded intrinsic. You can use ``llvm.floor`` on any
7093 floating point or vector of floating point type. Not all targets support
7098 declare float @llvm.floor.f32(float %Val)
7099 declare double @llvm.floor.f64(double %Val)
7100 declare x86_fp80 @llvm.floor.f80(x86_fp80 %Val)
7101 declare fp128 @llvm.floor.f128(fp128 %Val)
7102 declare ppc_fp128 @llvm.floor.ppcf128(ppc_fp128 %Val)
7107 The '``llvm.floor.*``' intrinsics return the floor of the operand.
7112 The argument and return value are floating point numbers of the same
7118 This function returns the same values as the libm ``floor`` functions
7119 would, and handles error conditions in the same way.
7121 '``llvm.ceil.*``' Intrinsic
7122 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7127 This is an overloaded intrinsic. You can use ``llvm.ceil`` on any
7128 floating point or vector of floating point type. Not all targets support
7133 declare float @llvm.ceil.f32(float %Val)
7134 declare double @llvm.ceil.f64(double %Val)
7135 declare x86_fp80 @llvm.ceil.f80(x86_fp80 %Val)
7136 declare fp128 @llvm.ceil.f128(fp128 %Val)
7137 declare ppc_fp128 @llvm.ceil.ppcf128(ppc_fp128 %Val)
7142 The '``llvm.ceil.*``' intrinsics return the ceiling of the operand.
7147 The argument and return value are floating point numbers of the same
7153 This function returns the same values as the libm ``ceil`` functions
7154 would, and handles error conditions in the same way.
7156 '``llvm.trunc.*``' Intrinsic
7157 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7162 This is an overloaded intrinsic. You can use ``llvm.trunc`` on any
7163 floating point or vector of floating point type. Not all targets support
7168 declare float @llvm.trunc.f32(float %Val)
7169 declare double @llvm.trunc.f64(double %Val)
7170 declare x86_fp80 @llvm.trunc.f80(x86_fp80 %Val)
7171 declare fp128 @llvm.trunc.f128(fp128 %Val)
7172 declare ppc_fp128 @llvm.trunc.ppcf128(ppc_fp128 %Val)
7177 The '``llvm.trunc.*``' intrinsics returns the operand rounded to the
7178 nearest integer not larger in magnitude than the operand.
7183 The argument and return value are floating point numbers of the same
7189 This function returns the same values as the libm ``trunc`` functions
7190 would, and handles error conditions in the same way.
7192 '``llvm.rint.*``' Intrinsic
7193 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7198 This is an overloaded intrinsic. You can use ``llvm.rint`` on any
7199 floating point or vector of floating point type. Not all targets support
7204 declare float @llvm.rint.f32(float %Val)
7205 declare double @llvm.rint.f64(double %Val)
7206 declare x86_fp80 @llvm.rint.f80(x86_fp80 %Val)
7207 declare fp128 @llvm.rint.f128(fp128 %Val)
7208 declare ppc_fp128 @llvm.rint.ppcf128(ppc_fp128 %Val)
7213 The '``llvm.rint.*``' intrinsics returns the operand rounded to the
7214 nearest integer. It may raise an inexact floating-point exception if the
7215 operand isn't an integer.
7220 The argument and return value are floating point numbers of the same
7226 This function returns the same values as the libm ``rint`` functions
7227 would, and handles error conditions in the same way.
7229 '``llvm.nearbyint.*``' Intrinsic
7230 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7235 This is an overloaded intrinsic. You can use ``llvm.nearbyint`` on any
7236 floating point or vector of floating point type. Not all targets support
7241 declare float @llvm.nearbyint.f32(float %Val)
7242 declare double @llvm.nearbyint.f64(double %Val)
7243 declare x86_fp80 @llvm.nearbyint.f80(x86_fp80 %Val)
7244 declare fp128 @llvm.nearbyint.f128(fp128 %Val)
7245 declare ppc_fp128 @llvm.nearbyint.ppcf128(ppc_fp128 %Val)
7250 The '``llvm.nearbyint.*``' intrinsics returns the operand rounded to the
7256 The argument and return value are floating point numbers of the same
7262 This function returns the same values as the libm ``nearbyint``
7263 functions would, and handles error conditions in the same way.
7265 Bit Manipulation Intrinsics
7266 ---------------------------
7268 LLVM provides intrinsics for a few important bit manipulation
7269 operations. These allow efficient code generation for some algorithms.
7271 '``llvm.bswap.*``' Intrinsics
7272 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7277 This is an overloaded intrinsic function. You can use bswap on any
7278 integer type that is an even number of bytes (i.e. BitWidth % 16 == 0).
7282 declare i16 @llvm.bswap.i16(i16 <id>)
7283 declare i32 @llvm.bswap.i32(i32 <id>)
7284 declare i64 @llvm.bswap.i64(i64 <id>)
7289 The '``llvm.bswap``' family of intrinsics is used to byte swap integer
7290 values with an even number of bytes (positive multiple of 16 bits).
7291 These are useful for performing operations on data that is not in the
7292 target's native byte order.
7297 The ``llvm.bswap.i16`` intrinsic returns an i16 value that has the high
7298 and low byte of the input i16 swapped. Similarly, the ``llvm.bswap.i32``
7299 intrinsic returns an i32 value that has the four bytes of the input i32
7300 swapped, so that if the input bytes are numbered 0, 1, 2, 3 then the
7301 returned i32 will have its bytes in 3, 2, 1, 0 order. The
7302 ``llvm.bswap.i48``, ``llvm.bswap.i64`` and other intrinsics extend this
7303 concept to additional even-byte lengths (6 bytes, 8 bytes and more,
7306 '``llvm.ctpop.*``' Intrinsic
7307 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7312 This is an overloaded intrinsic. You can use llvm.ctpop on any integer
7313 bit width, or on any vector with integer elements. Not all targets
7314 support all bit widths or vector types, however.
7318 declare i8 @llvm.ctpop.i8(i8 <src>)
7319 declare i16 @llvm.ctpop.i16(i16 <src>)
7320 declare i32 @llvm.ctpop.i32(i32 <src>)
7321 declare i64 @llvm.ctpop.i64(i64 <src>)
7322 declare i256 @llvm.ctpop.i256(i256 <src>)
7323 declare <2 x i32> @llvm.ctpop.v2i32(<2 x i32> <src>)
7328 The '``llvm.ctpop``' family of intrinsics counts the number of bits set
7334 The only argument is the value to be counted. The argument may be of any
7335 integer type, or a vector with integer elements. The return type must
7336 match the argument type.
7341 The '``llvm.ctpop``' intrinsic counts the 1's in a variable, or within
7342 each element of a vector.
7344 '``llvm.ctlz.*``' Intrinsic
7345 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7350 This is an overloaded intrinsic. You can use ``llvm.ctlz`` on any
7351 integer bit width, or any vector whose elements are integers. Not all
7352 targets support all bit widths or vector types, however.
7356 declare i8 @llvm.ctlz.i8 (i8 <src>, i1 <is_zero_undef>)
7357 declare i16 @llvm.ctlz.i16 (i16 <src>, i1 <is_zero_undef>)
7358 declare i32 @llvm.ctlz.i32 (i32 <src>, i1 <is_zero_undef>)
7359 declare i64 @llvm.ctlz.i64 (i64 <src>, i1 <is_zero_undef>)
7360 declare i256 @llvm.ctlz.i256(i256 <src>, i1 <is_zero_undef>)
7361 declase <2 x i32> @llvm.ctlz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
7366 The '``llvm.ctlz``' family of intrinsic functions counts the number of
7367 leading zeros in a variable.
7372 The first argument is the value to be counted. This argument may be of
7373 any integer type, or a vectory with integer element type. The return
7374 type must match the first argument type.
7376 The second argument must be a constant and is a flag to indicate whether
7377 the intrinsic should ensure that a zero as the first argument produces a
7378 defined result. Historically some architectures did not provide a
7379 defined result for zero values as efficiently, and many algorithms are
7380 now predicated on avoiding zero-value inputs.
7385 The '``llvm.ctlz``' intrinsic counts the leading (most significant)
7386 zeros in a variable, or within each element of the vector. If
7387 ``src == 0`` then the result is the size in bits of the type of ``src``
7388 if ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
7389 ``llvm.ctlz(i32 2) = 30``.
7391 '``llvm.cttz.*``' Intrinsic
7392 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7397 This is an overloaded intrinsic. You can use ``llvm.cttz`` on any
7398 integer bit width, or any vector of integer elements. Not all targets
7399 support all bit widths or vector types, however.
7403 declare i8 @llvm.cttz.i8 (i8 <src>, i1 <is_zero_undef>)
7404 declare i16 @llvm.cttz.i16 (i16 <src>, i1 <is_zero_undef>)
7405 declare i32 @llvm.cttz.i32 (i32 <src>, i1 <is_zero_undef>)
7406 declare i64 @llvm.cttz.i64 (i64 <src>, i1 <is_zero_undef>)
7407 declare i256 @llvm.cttz.i256(i256 <src>, i1 <is_zero_undef>)
7408 declase <2 x i32> @llvm.cttz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
7413 The '``llvm.cttz``' family of intrinsic functions counts the number of
7419 The first argument is the value to be counted. This argument may be of
7420 any integer type, or a vectory with integer element type. The return
7421 type must match the first argument type.
7423 The second argument must be a constant and is a flag to indicate whether
7424 the intrinsic should ensure that a zero as the first argument produces a
7425 defined result. Historically some architectures did not provide a
7426 defined result for zero values as efficiently, and many algorithms are
7427 now predicated on avoiding zero-value inputs.
7432 The '``llvm.cttz``' intrinsic counts the trailing (least significant)
7433 zeros in a variable, or within each element of a vector. If ``src == 0``
7434 then the result is the size in bits of the type of ``src`` if
7435 ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
7436 ``llvm.cttz(2) = 1``.
7438 Arithmetic with Overflow Intrinsics
7439 -----------------------------------
7441 LLVM provides intrinsics for some arithmetic with overflow operations.
7443 '``llvm.sadd.with.overflow.*``' Intrinsics
7444 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7449 This is an overloaded intrinsic. You can use ``llvm.sadd.with.overflow``
7450 on any integer bit width.
7454 declare {i16, i1} @llvm.sadd.with.overflow.i16(i16 %a, i16 %b)
7455 declare {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
7456 declare {i64, i1} @llvm.sadd.with.overflow.i64(i64 %a, i64 %b)
7461 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
7462 a signed addition of the two arguments, and indicate whether an overflow
7463 occurred during the signed summation.
7468 The arguments (%a and %b) and the first element of the result structure
7469 may be of integer types of any bit width, but they must have the same
7470 bit width. The second element of the result structure must be of type
7471 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
7477 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
7478 a signed addition of the two variables. They return a structure --- the
7479 first element of which is the signed summation, and the second element
7480 of which is a bit specifying if the signed summation resulted in an
7486 .. code-block:: llvm
7488 %res = call {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
7489 %sum = extractvalue {i32, i1} %res, 0
7490 %obit = extractvalue {i32, i1} %res, 1
7491 br i1 %obit, label %overflow, label %normal
7493 '``llvm.uadd.with.overflow.*``' Intrinsics
7494 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7499 This is an overloaded intrinsic. You can use ``llvm.uadd.with.overflow``
7500 on any integer bit width.
7504 declare {i16, i1} @llvm.uadd.with.overflow.i16(i16 %a, i16 %b)
7505 declare {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
7506 declare {i64, i1} @llvm.uadd.with.overflow.i64(i64 %a, i64 %b)
7511 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
7512 an unsigned addition of the two arguments, and indicate whether a carry
7513 occurred during the unsigned summation.
7518 The arguments (%a and %b) and the first element of the result structure
7519 may be of integer types of any bit width, but they must have the same
7520 bit width. The second element of the result structure must be of type
7521 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
7527 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
7528 an unsigned addition of the two arguments. They return a structure --- the
7529 first element of which is the sum, and the second element of which is a
7530 bit specifying if the unsigned summation resulted in a carry.
7535 .. code-block:: llvm
7537 %res = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
7538 %sum = extractvalue {i32, i1} %res, 0
7539 %obit = extractvalue {i32, i1} %res, 1
7540 br i1 %obit, label %carry, label %normal
7542 '``llvm.ssub.with.overflow.*``' Intrinsics
7543 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7548 This is an overloaded intrinsic. You can use ``llvm.ssub.with.overflow``
7549 on any integer bit width.
7553 declare {i16, i1} @llvm.ssub.with.overflow.i16(i16 %a, i16 %b)
7554 declare {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
7555 declare {i64, i1} @llvm.ssub.with.overflow.i64(i64 %a, i64 %b)
7560 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
7561 a signed subtraction of the two arguments, and indicate whether an
7562 overflow occurred during the signed subtraction.
7567 The arguments (%a and %b) and the first element of the result structure
7568 may be of integer types of any bit width, but they must have the same
7569 bit width. The second element of the result structure must be of type
7570 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
7576 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
7577 a signed subtraction of the two arguments. They return a structure --- the
7578 first element of which is the subtraction, and the second element of
7579 which is a bit specifying if the signed subtraction resulted in an
7585 .. code-block:: llvm
7587 %res = call {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
7588 %sum = extractvalue {i32, i1} %res, 0
7589 %obit = extractvalue {i32, i1} %res, 1
7590 br i1 %obit, label %overflow, label %normal
7592 '``llvm.usub.with.overflow.*``' Intrinsics
7593 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7598 This is an overloaded intrinsic. You can use ``llvm.usub.with.overflow``
7599 on any integer bit width.
7603 declare {i16, i1} @llvm.usub.with.overflow.i16(i16 %a, i16 %b)
7604 declare {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
7605 declare {i64, i1} @llvm.usub.with.overflow.i64(i64 %a, i64 %b)
7610 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
7611 an unsigned subtraction of the two arguments, and indicate whether an
7612 overflow occurred during the unsigned subtraction.
7617 The arguments (%a and %b) and the first element of the result structure
7618 may be of integer types of any bit width, but they must have the same
7619 bit width. The second element of the result structure must be of type
7620 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
7626 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
7627 an unsigned subtraction of the two arguments. They return a structure ---
7628 the first element of which is the subtraction, and the second element of
7629 which is a bit specifying if the unsigned subtraction resulted in an
7635 .. code-block:: llvm
7637 %res = call {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
7638 %sum = extractvalue {i32, i1} %res, 0
7639 %obit = extractvalue {i32, i1} %res, 1
7640 br i1 %obit, label %overflow, label %normal
7642 '``llvm.smul.with.overflow.*``' Intrinsics
7643 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7648 This is an overloaded intrinsic. You can use ``llvm.smul.with.overflow``
7649 on any integer bit width.
7653 declare {i16, i1} @llvm.smul.with.overflow.i16(i16 %a, i16 %b)
7654 declare {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
7655 declare {i64, i1} @llvm.smul.with.overflow.i64(i64 %a, i64 %b)
7660 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
7661 a signed multiplication of the two arguments, and indicate whether an
7662 overflow occurred during the signed multiplication.
7667 The arguments (%a and %b) and the first element of the result structure
7668 may be of integer types of any bit width, but they must have the same
7669 bit width. The second element of the result structure must be of type
7670 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
7676 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
7677 a signed multiplication of the two arguments. They return a structure ---
7678 the first element of which is the multiplication, and the second element
7679 of which is a bit specifying if the signed multiplication resulted in an
7685 .. code-block:: llvm
7687 %res = call {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
7688 %sum = extractvalue {i32, i1} %res, 0
7689 %obit = extractvalue {i32, i1} %res, 1
7690 br i1 %obit, label %overflow, label %normal
7692 '``llvm.umul.with.overflow.*``' Intrinsics
7693 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7698 This is an overloaded intrinsic. You can use ``llvm.umul.with.overflow``
7699 on any integer bit width.
7703 declare {i16, i1} @llvm.umul.with.overflow.i16(i16 %a, i16 %b)
7704 declare {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
7705 declare {i64, i1} @llvm.umul.with.overflow.i64(i64 %a, i64 %b)
7710 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
7711 a unsigned multiplication of the two arguments, and indicate whether an
7712 overflow occurred during the unsigned multiplication.
7717 The arguments (%a and %b) and the first element of the result structure
7718 may be of integer types of any bit width, but they must have the same
7719 bit width. The second element of the result structure must be of type
7720 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
7726 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
7727 an unsigned multiplication of the two arguments. They return a structure ---
7728 the first element of which is the multiplication, and the second
7729 element of which is a bit specifying if the unsigned multiplication
7730 resulted in an overflow.
7735 .. code-block:: llvm
7737 %res = call {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
7738 %sum = extractvalue {i32, i1} %res, 0
7739 %obit = extractvalue {i32, i1} %res, 1
7740 br i1 %obit, label %overflow, label %normal
7742 Specialised Arithmetic Intrinsics
7743 ---------------------------------
7745 '``llvm.fmuladd.*``' Intrinsic
7746 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7753 declare float @llvm.fmuladd.f32(float %a, float %b, float %c)
7754 declare double @llvm.fmuladd.f64(double %a, double %b, double %c)
7759 The '``llvm.fmuladd.*``' intrinsic functions represent multiply-add
7760 expressions that can be fused if the code generator determines that (a) the
7761 target instruction set has support for a fused operation, and (b) that the
7762 fused operation is more efficient than the equivalent, separate pair of mul
7763 and add instructions.
7768 The '``llvm.fmuladd.*``' intrinsics each take three arguments: two
7769 multiplicands, a and b, and an addend c.
7778 %0 = call float @llvm.fmuladd.f32(%a, %b, %c)
7780 is equivalent to the expression a \* b + c, except that rounding will
7781 not be performed between the multiplication and addition steps if the
7782 code generator fuses the operations. Fusion is not guaranteed, even if
7783 the target platform supports it. If a fused multiply-add is required the
7784 corresponding llvm.fma.\* intrinsic function should be used instead.
7789 .. code-block:: llvm
7791 %r2 = call float @llvm.fmuladd.f32(float %a, float %b, float %c) ; yields {float}:r2 = (a * b) + c
7793 Half Precision Floating Point Intrinsics
7794 ----------------------------------------
7796 For most target platforms, half precision floating point is a
7797 storage-only format. This means that it is a dense encoding (in memory)
7798 but does not support computation in the format.
7800 This means that code must first load the half-precision floating point
7801 value as an i16, then convert it to float with
7802 :ref:`llvm.convert.from.fp16 <int_convert_from_fp16>`. Computation can
7803 then be performed on the float value (including extending to double
7804 etc). To store the value back to memory, it is first converted to float
7805 if needed, then converted to i16 with
7806 :ref:`llvm.convert.to.fp16 <int_convert_to_fp16>`, then storing as an
7809 .. _int_convert_to_fp16:
7811 '``llvm.convert.to.fp16``' Intrinsic
7812 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7819 declare i16 @llvm.convert.to.fp16(f32 %a)
7824 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion
7825 from single precision floating point format to half precision floating
7831 The intrinsic function contains single argument - the value to be
7837 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion
7838 from single precision floating point format to half precision floating
7839 point format. The return value is an ``i16`` which contains the
7845 .. code-block:: llvm
7847 %res = call i16 @llvm.convert.to.fp16(f32 %a)
7848 store i16 %res, i16* @x, align 2
7850 .. _int_convert_from_fp16:
7852 '``llvm.convert.from.fp16``' Intrinsic
7853 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7860 declare f32 @llvm.convert.from.fp16(i16 %a)
7865 The '``llvm.convert.from.fp16``' intrinsic function performs a
7866 conversion from half precision floating point format to single precision
7867 floating point format.
7872 The intrinsic function contains single argument - the value to be
7878 The '``llvm.convert.from.fp16``' intrinsic function performs a
7879 conversion from half single precision floating point format to single
7880 precision floating point format. The input half-float value is
7881 represented by an ``i16`` value.
7886 .. code-block:: llvm
7888 %a = load i16* @x, align 2
7889 %res = call f32 @llvm.convert.from.fp16(i16 %a)
7894 The LLVM debugger intrinsics (which all start with ``llvm.dbg.``
7895 prefix), are described in the `LLVM Source Level
7896 Debugging <SourceLevelDebugging.html#format_common_intrinsics>`_
7899 Exception Handling Intrinsics
7900 -----------------------------
7902 The LLVM exception handling intrinsics (which all start with
7903 ``llvm.eh.`` prefix), are described in the `LLVM Exception
7904 Handling <ExceptionHandling.html#format_common_intrinsics>`_ document.
7908 Trampoline Intrinsics
7909 ---------------------
7911 These intrinsics make it possible to excise one parameter, marked with
7912 the :ref:`nest <nest>` attribute, from a function. The result is a
7913 callable function pointer lacking the nest parameter - the caller does
7914 not need to provide a value for it. Instead, the value to use is stored
7915 in advance in a "trampoline", a block of memory usually allocated on the
7916 stack, which also contains code to splice the nest value into the
7917 argument list. This is used to implement the GCC nested function address
7920 For example, if the function is ``i32 f(i8* nest %c, i32 %x, i32 %y)``
7921 then the resulting function pointer has signature ``i32 (i32, i32)*``.
7922 It can be created as follows:
7924 .. code-block:: llvm
7926 %tramp = alloca [10 x i8], align 4 ; size and alignment only correct for X86
7927 %tramp1 = getelementptr [10 x i8]* %tramp, i32 0, i32 0
7928 call i8* @llvm.init.trampoline(i8* %tramp1, i8* bitcast (i32 (i8*, i32, i32)* @f to i8*), i8* %nval)
7929 %p = call i8* @llvm.adjust.trampoline(i8* %tramp1)
7930 %fp = bitcast i8* %p to i32 (i32, i32)*
7932 The call ``%val = call i32 %fp(i32 %x, i32 %y)`` is then equivalent to
7933 ``%val = call i32 %f(i8* %nval, i32 %x, i32 %y)``.
7937 '``llvm.init.trampoline``' Intrinsic
7938 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7945 declare void @llvm.init.trampoline(i8* <tramp>, i8* <func>, i8* <nval>)
7950 This fills the memory pointed to by ``tramp`` with executable code,
7951 turning it into a trampoline.
7956 The ``llvm.init.trampoline`` intrinsic takes three arguments, all
7957 pointers. The ``tramp`` argument must point to a sufficiently large and
7958 sufficiently aligned block of memory; this memory is written to by the
7959 intrinsic. Note that the size and the alignment are target-specific -
7960 LLVM currently provides no portable way of determining them, so a
7961 front-end that generates this intrinsic needs to have some
7962 target-specific knowledge. The ``func`` argument must hold a function
7963 bitcast to an ``i8*``.
7968 The block of memory pointed to by ``tramp`` is filled with target
7969 dependent code, turning it into a function. Then ``tramp`` needs to be
7970 passed to :ref:`llvm.adjust.trampoline <int_at>` to get a pointer which can
7971 be :ref:`bitcast (to a new function) and called <int_trampoline>`. The new
7972 function's signature is the same as that of ``func`` with any arguments
7973 marked with the ``nest`` attribute removed. At most one such ``nest``
7974 argument is allowed, and it must be of pointer type. Calling the new
7975 function is equivalent to calling ``func`` with the same argument list,
7976 but with ``nval`` used for the missing ``nest`` argument. If, after
7977 calling ``llvm.init.trampoline``, the memory pointed to by ``tramp`` is
7978 modified, then the effect of any later call to the returned function
7979 pointer is undefined.
7983 '``llvm.adjust.trampoline``' Intrinsic
7984 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7991 declare i8* @llvm.adjust.trampoline(i8* <tramp>)
7996 This performs any required machine-specific adjustment to the address of
7997 a trampoline (passed as ``tramp``).
8002 ``tramp`` must point to a block of memory which already has trampoline
8003 code filled in by a previous call to
8004 :ref:`llvm.init.trampoline <int_it>`.
8009 On some architectures the address of the code to be executed needs to be
8010 different to the address where the trampoline is actually stored. This
8011 intrinsic returns the executable address corresponding to ``tramp``
8012 after performing the required machine specific adjustments. The pointer
8013 returned can then be :ref:`bitcast and executed <int_trampoline>`.
8018 This class of intrinsics exists to information about the lifetime of
8019 memory objects and ranges where variables are immutable.
8021 '``llvm.lifetime.start``' Intrinsic
8022 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8029 declare void @llvm.lifetime.start(i64 <size>, i8* nocapture <ptr>)
8034 The '``llvm.lifetime.start``' intrinsic specifies the start of a memory
8040 The first argument is a constant integer representing the size of the
8041 object, or -1 if it is variable sized. The second argument is a pointer
8047 This intrinsic indicates that before this point in the code, the value
8048 of the memory pointed to by ``ptr`` is dead. This means that it is known
8049 to never be used and has an undefined value. A load from the pointer
8050 that precedes this intrinsic can be replaced with ``'undef'``.
8052 '``llvm.lifetime.end``' Intrinsic
8053 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8060 declare void @llvm.lifetime.end(i64 <size>, i8* nocapture <ptr>)
8065 The '``llvm.lifetime.end``' intrinsic specifies the end of a memory
8071 The first argument is a constant integer representing the size of the
8072 object, or -1 if it is variable sized. The second argument is a pointer
8078 This intrinsic indicates that after this point in the code, the value of
8079 the memory pointed to by ``ptr`` is dead. This means that it is known to
8080 never be used and has an undefined value. Any stores into the memory
8081 object following this intrinsic may be removed as dead.
8083 '``llvm.invariant.start``' Intrinsic
8084 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8091 declare {}* @llvm.invariant.start(i64 <size>, i8* nocapture <ptr>)
8096 The '``llvm.invariant.start``' intrinsic specifies that the contents of
8097 a memory object will not change.
8102 The first argument is a constant integer representing the size of the
8103 object, or -1 if it is variable sized. The second argument is a pointer
8109 This intrinsic indicates that until an ``llvm.invariant.end`` that uses
8110 the return value, the referenced memory location is constant and
8113 '``llvm.invariant.end``' Intrinsic
8114 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8121 declare void @llvm.invariant.end({}* <start>, i64 <size>, i8* nocapture <ptr>)
8126 The '``llvm.invariant.end``' intrinsic specifies that the contents of a
8127 memory object are mutable.
8132 The first argument is the matching ``llvm.invariant.start`` intrinsic.
8133 The second argument is a constant integer representing the size of the
8134 object, or -1 if it is variable sized and the third argument is a
8135 pointer to the object.
8140 This intrinsic indicates that the memory is mutable again.
8145 This class of intrinsics is designed to be generic and has no specific
8148 '``llvm.var.annotation``' Intrinsic
8149 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8156 declare void @llvm.var.annotation(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
8161 The '``llvm.var.annotation``' intrinsic.
8166 The first argument is a pointer to a value, the second is a pointer to a
8167 global string, the third is a pointer to a global string which is the
8168 source file name, and the last argument is the line number.
8173 This intrinsic allows annotation of local variables with arbitrary
8174 strings. This can be useful for special purpose optimizations that want
8175 to look for these annotations. These have no other defined use; they are
8176 ignored by code generation and optimization.
8178 '``llvm.annotation.*``' Intrinsic
8179 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8184 This is an overloaded intrinsic. You can use '``llvm.annotation``' on
8185 any integer bit width.
8189 declare i8 @llvm.annotation.i8(i8 <val>, i8* <str>, i8* <str>, i32 <int>)
8190 declare i16 @llvm.annotation.i16(i16 <val>, i8* <str>, i8* <str>, i32 <int>)
8191 declare i32 @llvm.annotation.i32(i32 <val>, i8* <str>, i8* <str>, i32 <int>)
8192 declare i64 @llvm.annotation.i64(i64 <val>, i8* <str>, i8* <str>, i32 <int>)
8193 declare i256 @llvm.annotation.i256(i256 <val>, i8* <str>, i8* <str>, i32 <int>)
8198 The '``llvm.annotation``' intrinsic.
8203 The first argument is an integer value (result of some expression), the
8204 second is a pointer to a global string, the third is a pointer to a
8205 global string which is the source file name, and the last argument is
8206 the line number. It returns the value of the first argument.
8211 This intrinsic allows annotations to be put on arbitrary expressions
8212 with arbitrary strings. This can be useful for special purpose
8213 optimizations that want to look for these annotations. These have no
8214 other defined use; they are ignored by code generation and optimization.
8216 '``llvm.trap``' Intrinsic
8217 ^^^^^^^^^^^^^^^^^^^^^^^^^
8224 declare void @llvm.trap() noreturn nounwind
8229 The '``llvm.trap``' intrinsic.
8239 This intrinsic is lowered to the target dependent trap instruction. If
8240 the target does not have a trap instruction, this intrinsic will be
8241 lowered to a call of the ``abort()`` function.
8243 '``llvm.debugtrap``' Intrinsic
8244 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8251 declare void @llvm.debugtrap() nounwind
8256 The '``llvm.debugtrap``' intrinsic.
8266 This intrinsic is lowered to code which is intended to cause an
8267 execution trap with the intention of requesting the attention of a
8270 '``llvm.stackprotector``' Intrinsic
8271 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8278 declare void @llvm.stackprotector(i8* <guard>, i8** <slot>)
8283 The ``llvm.stackprotector`` intrinsic takes the ``guard`` and stores it
8284 onto the stack at ``slot``. The stack slot is adjusted to ensure that it
8285 is placed on the stack before local variables.
8290 The ``llvm.stackprotector`` intrinsic requires two pointer arguments.
8291 The first argument is the value loaded from the stack guard
8292 ``@__stack_chk_guard``. The second variable is an ``alloca`` that has
8293 enough space to hold the value of the guard.
8298 This intrinsic causes the prologue/epilogue inserter to force the
8299 position of the ``AllocaInst`` stack slot to be before local variables
8300 on the stack. This is to ensure that if a local variable on the stack is
8301 overwritten, it will destroy the value of the guard. When the function
8302 exits, the guard on the stack is checked against the original guard. If
8303 they are different, then the program aborts by calling the
8304 ``__stack_chk_fail()`` function.
8306 '``llvm.objectsize``' Intrinsic
8307 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8314 declare i32 @llvm.objectsize.i32(i8* <object>, i1 <min>)
8315 declare i64 @llvm.objectsize.i64(i8* <object>, i1 <min>)
8320 The ``llvm.objectsize`` intrinsic is designed to provide information to
8321 the optimizers to determine at compile time whether a) an operation
8322 (like memcpy) will overflow a buffer that corresponds to an object, or
8323 b) that a runtime check for overflow isn't necessary. An object in this
8324 context means an allocation of a specific class, structure, array, or
8330 The ``llvm.objectsize`` intrinsic takes two arguments. The first
8331 argument is a pointer to or into the ``object``. The second argument is
8332 a boolean and determines whether ``llvm.objectsize`` returns 0 (if true)
8333 or -1 (if false) when the object size is unknown. The second argument
8334 only accepts constants.
8339 The ``llvm.objectsize`` intrinsic is lowered to a constant representing
8340 the size of the object concerned. If the size cannot be determined at
8341 compile time, ``llvm.objectsize`` returns ``i32/i64 -1 or 0`` (depending
8342 on the ``min`` argument).
8344 '``llvm.expect``' Intrinsic
8345 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8352 declare i32 @llvm.expect.i32(i32 <val>, i32 <expected_val>)
8353 declare i64 @llvm.expect.i64(i64 <val>, i64 <expected_val>)
8358 The ``llvm.expect`` intrinsic provides information about expected (the
8359 most probable) value of ``val``, which can be used by optimizers.
8364 The ``llvm.expect`` intrinsic takes two arguments. The first argument is
8365 a value. The second argument is an expected value, this needs to be a
8366 constant value, variables are not allowed.
8371 This intrinsic is lowered to the ``val``.
8373 '``llvm.donothing``' Intrinsic
8374 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8381 declare void @llvm.donothing() nounwind readnone
8386 The ``llvm.donothing`` intrinsic doesn't perform any operation. It's the
8387 only intrinsic that can be called with an invoke instruction.
8397 This intrinsic does nothing, and it's removed by optimizers and ignored