1 ==============================
2 LLVM Language Reference Manual
3 ==============================
12 This document is a reference manual for the LLVM assembly language. LLVM
13 is a Static Single Assignment (SSA) based representation that provides
14 type safety, low-level operations, flexibility, and the capability of
15 representing 'all' high-level languages cleanly. It is the common code
16 representation used throughout all phases of the LLVM compilation
22 The LLVM code representation is designed to be used in three different
23 forms: as an in-memory compiler IR, as an on-disk bitcode representation
24 (suitable for fast loading by a Just-In-Time compiler), and as a human
25 readable assembly language representation. This allows LLVM to provide a
26 powerful intermediate representation for efficient compiler
27 transformations and analysis, while providing a natural means to debug
28 and visualize the transformations. The three different forms of LLVM are
29 all equivalent. This document describes the human readable
30 representation and notation.
32 The LLVM representation aims to be light-weight and low-level while
33 being expressive, typed, and extensible at the same time. It aims to be
34 a "universal IR" of sorts, by being at a low enough level that
35 high-level ideas may be cleanly mapped to it (similar to how
36 microprocessors are "universal IR's", allowing many source languages to
37 be mapped to them). By providing type information, LLVM can be used as
38 the target of optimizations: for example, through pointer analysis, it
39 can be proven that a C automatic variable is never accessed outside of
40 the current function, allowing it to be promoted to a simple SSA value
41 instead of a memory location.
48 It is important to note that this document describes 'well formed' LLVM
49 assembly language. There is a difference between what the parser accepts
50 and what is considered 'well formed'. For example, the following
51 instruction is syntactically okay, but not well formed:
57 because the definition of ``%x`` does not dominate all of its uses. The
58 LLVM infrastructure provides a verification pass that may be used to
59 verify that an LLVM module is well formed. This pass is automatically
60 run by the parser after parsing input assembly and by the optimizer
61 before it outputs bitcode. The violations pointed out by the verifier
62 pass indicate bugs in transformation passes or input to the parser.
69 LLVM identifiers come in two basic types: global and local. Global
70 identifiers (functions, global variables) begin with the ``'@'``
71 character. Local identifiers (register names, types) begin with the
72 ``'%'`` character. Additionally, there are three different formats for
73 identifiers, for different purposes:
75 #. Named values are represented as a string of characters with their
76 prefix. For example, ``%foo``, ``@DivisionByZero``,
77 ``%a.really.long.identifier``. The actual regular expression used is
78 '``[%@][a-zA-Z$._][a-zA-Z$._0-9]*``'. Identifiers which require other
79 characters in their names can be surrounded with quotes. Special
80 characters may be escaped using ``"\xx"`` where ``xx`` is the ASCII
81 code for the character in hexadecimal. In this way, any character can
82 be used in a name value, even quotes themselves.
83 #. Unnamed values are represented as an unsigned numeric value with
84 their prefix. For example, ``%12``, ``@2``, ``%44``.
85 #. Constants, which are described in the section Constants_ below.
87 LLVM requires that values start with a prefix for two reasons: Compilers
88 don't need to worry about name clashes with reserved words, and the set
89 of reserved words may be expanded in the future without penalty.
90 Additionally, unnamed identifiers allow a compiler to quickly come up
91 with a temporary variable without having to avoid symbol table
94 Reserved words in LLVM are very similar to reserved words in other
95 languages. There are keywords for different opcodes ('``add``',
96 '``bitcast``', '``ret``', etc...), for primitive type names ('``void``',
97 '``i32``', etc...), and others. These reserved words cannot conflict
98 with variable names, because none of them start with a prefix character
101 Here is an example of LLVM code to multiply the integer variable
108 %result = mul i32 %X, 8
110 After strength reduction:
114 %result = shl i32 %X, 3
120 %0 = add i32 %X, %X ; yields {i32}:%0
121 %1 = add i32 %0, %0 ; yields {i32}:%1
122 %result = add i32 %1, %1
124 This last way of multiplying ``%X`` by 8 illustrates several important
125 lexical features of LLVM:
127 #. Comments are delimited with a '``;``' and go until the end of line.
128 #. Unnamed temporaries are created when the result of a computation is
129 not assigned to a named value.
130 #. Unnamed temporaries are numbered sequentially
132 It also shows a convention that we follow in this document. When
133 demonstrating instructions, we will follow an instruction with a comment
134 that defines the type and name of value produced.
142 LLVM programs are composed of ``Module``'s, each of which is a
143 translation unit of the input programs. Each module consists of
144 functions, global variables, and symbol table entries. Modules may be
145 combined together with the LLVM linker, which merges function (and
146 global variable) definitions, resolves forward declarations, and merges
147 symbol table entries. Here is an example of the "hello world" module:
151 ; Declare the string constant as a global constant.
152 @.str = private unnamed_addr constant [13 x i8] c"hello world\0A\00"
154 ; External declaration of the puts function
155 declare i32 @puts(i8* nocapture) nounwind
157 ; Definition of main function
158 define i32 @main() { ; i32()*
159 ; Convert [13 x i8]* to i8 *...
160 %cast210 = getelementptr [13 x i8]* @.str, i64 0, i64 0
162 ; Call puts function to write out the string to stdout.
163 call i32 @puts(i8* %cast210)
168 !1 = metadata !{i32 42}
171 This example is made up of a :ref:`global variable <globalvars>` named
172 "``.str``", an external declaration of the "``puts``" function, a
173 :ref:`function definition <functionstructure>` for "``main``" and
174 :ref:`named metadata <namedmetadatastructure>` "``foo``".
176 In general, a module is made up of a list of global values (where both
177 functions and global variables are global values). Global values are
178 represented by a pointer to a memory location (in this case, a pointer
179 to an array of char, and a pointer to a function), and have one of the
180 following :ref:`linkage types <linkage>`.
187 All Global Variables and Functions have one of the following types of
191 Global values with "``private``" linkage are only directly
192 accessible by objects in the current module. In particular, linking
193 code into a module with an private global value may cause the
194 private to be renamed as necessary to avoid collisions. Because the
195 symbol is private to the module, all references can be updated. This
196 doesn't show up in any symbol table in the object file.
198 Similar to ``private``, but the symbol is passed through the
199 assembler and evaluated by the linker. Unlike normal strong symbols,
200 they are removed by the linker from the final linked image
201 (executable or dynamic library).
202 ``linker_private_weak``
203 Similar to "``linker_private``", but the symbol is weak. Note that
204 ``linker_private_weak`` symbols are subject to coalescing by the
205 linker. The symbols are removed by the linker from the final linked
206 image (executable or dynamic library).
208 Similar to private, but the value shows as a local symbol
209 (``STB_LOCAL`` in the case of ELF) in the object file. This
210 corresponds to the notion of the '``static``' keyword in C.
211 ``available_externally``
212 Globals with "``available_externally``" linkage are never emitted
213 into the object file corresponding to the LLVM module. They exist to
214 allow inlining and other optimizations to take place given knowledge
215 of the definition of the global, which is known to be somewhere
216 outside the module. Globals with ``available_externally`` linkage
217 are allowed to be discarded at will, and are otherwise the same as
218 ``linkonce_odr``. This linkage type is only allowed on definitions,
221 Globals with "``linkonce``" linkage are merged with other globals of
222 the same name when linkage occurs. This can be used to implement
223 some forms of inline functions, templates, or other code which must
224 be generated in each translation unit that uses it, but where the
225 body may be overridden with a more definitive definition later.
226 Unreferenced ``linkonce`` globals are allowed to be discarded. Note
227 that ``linkonce`` linkage does not actually allow the optimizer to
228 inline the body of this function into callers because it doesn't
229 know if this definition of the function is the definitive definition
230 within the program or whether it will be overridden by a stronger
231 definition. To enable inlining and other optimizations, use
232 "``linkonce_odr``" linkage.
234 "``weak``" linkage has the same merging semantics as ``linkonce``
235 linkage, except that unreferenced globals with ``weak`` linkage may
236 not be discarded. This is used for globals that are declared "weak"
239 "``common``" linkage is most similar to "``weak``" linkage, but they
240 are used for tentative definitions in C, such as "``int X;``" at
241 global scope. Symbols with "``common``" linkage are merged in the
242 same way as ``weak symbols``, and they may not be deleted if
243 unreferenced. ``common`` symbols may not have an explicit section,
244 must have a zero initializer, and may not be marked
245 ':ref:`constant <globalvars>`'. Functions and aliases may not have
248 .. _linkage_appending:
251 "``appending``" linkage may only be applied to global variables of
252 pointer to array type. When two global variables with appending
253 linkage are linked together, the two global arrays are appended
254 together. This is the LLVM, typesafe, equivalent of having the
255 system linker append together "sections" with identical names when
258 The semantics of this linkage follow the ELF object file model: the
259 symbol is weak until linked, if not linked, the symbol becomes null
260 instead of being an undefined reference.
261 ``linkonce_odr``, ``weak_odr``
262 Some languages allow differing globals to be merged, such as two
263 functions with different semantics. Other languages, such as
264 ``C++``, ensure that only equivalent globals are ever merged (the
265 "one definition rule" --- "ODR"). Such languages can use the
266 ``linkonce_odr`` and ``weak_odr`` linkage types to indicate that the
267 global will only be merged with equivalent globals. These linkage
268 types are otherwise the same as their non-``odr`` versions.
269 ``linkonce_odr_auto_hide``
270 Similar to "``linkonce_odr``", but nothing in the translation unit
271 takes the address of this definition. For instance, functions that
272 had an inline definition, but the compiler decided not to inline it.
273 ``linkonce_odr_auto_hide`` may have only ``default`` visibility. The
274 symbols are removed by the linker from the final linked image
275 (executable or dynamic library).
277 If none of the above identifiers are used, the global is externally
278 visible, meaning that it participates in linkage and can be used to
279 resolve external symbol references.
281 The next two types of linkage are targeted for Microsoft Windows
282 platform only. They are designed to support importing (exporting)
283 symbols from (to) DLLs (Dynamic Link Libraries).
286 "``dllimport``" linkage causes the compiler to reference a function
287 or variable via a global pointer to a pointer that is set up by the
288 DLL exporting the symbol. On Microsoft Windows targets, the pointer
289 name is formed by combining ``__imp_`` and the function or variable
292 "``dllexport``" linkage causes the compiler to provide a global
293 pointer to a pointer in a DLL, so that it can be referenced with the
294 ``dllimport`` attribute. On Microsoft Windows targets, the pointer
295 name is formed by combining ``__imp_`` and the function or variable
298 For example, since the "``.LC0``" variable is defined to be internal, if
299 another module defined a "``.LC0``" variable and was linked with this
300 one, one of the two would be renamed, preventing a collision. Since
301 "``main``" and "``puts``" are external (i.e., lacking any linkage
302 declarations), they are accessible outside of the current module.
304 It is illegal for a function *declaration* to have any linkage type
305 other than ``external``, ``dllimport`` or ``extern_weak``.
307 Aliases can have only ``external``, ``internal``, ``weak`` or
308 ``weak_odr`` linkages.
315 LLVM :ref:`functions <functionstructure>`, :ref:`calls <i_call>` and
316 :ref:`invokes <i_invoke>` can all have an optional calling convention
317 specified for the call. The calling convention of any pair of dynamic
318 caller/callee must match, or the behavior of the program is undefined.
319 The following calling conventions are supported by LLVM, and more may be
322 "``ccc``" - The C calling convention
323 This calling convention (the default if no other calling convention
324 is specified) matches the target C calling conventions. This calling
325 convention supports varargs function calls and tolerates some
326 mismatch in the declared prototype and implemented declaration of
327 the function (as does normal C).
328 "``fastcc``" - The fast calling convention
329 This calling convention attempts to make calls as fast as possible
330 (e.g. by passing things in registers). This calling convention
331 allows the target to use whatever tricks it wants to produce fast
332 code for the target, without having to conform to an externally
333 specified ABI (Application Binary Interface). `Tail calls can only
334 be optimized when this, the GHC or the HiPE convention is
335 used. <CodeGenerator.html#id80>`_ This calling convention does not
336 support varargs and requires the prototype of all callees to exactly
337 match the prototype of the function definition.
338 "``coldcc``" - The cold calling convention
339 This calling convention attempts to make code in the caller as
340 efficient as possible under the assumption that the call is not
341 commonly executed. As such, these calls often preserve all registers
342 so that the call does not break any live ranges in the caller side.
343 This calling convention does not support varargs and requires the
344 prototype of all callees to exactly match the prototype of the
346 "``cc 10``" - GHC convention
347 This calling convention has been implemented specifically for use by
348 the `Glasgow Haskell Compiler (GHC) <http://www.haskell.org/ghc>`_.
349 It passes everything in registers, going to extremes to achieve this
350 by disabling callee save registers. This calling convention should
351 not be used lightly but only for specific situations such as an
352 alternative to the *register pinning* performance technique often
353 used when implementing functional programming languages. At the
354 moment only X86 supports this convention and it has the following
357 - On *X86-32* only supports up to 4 bit type parameters. No
358 floating point types are supported.
359 - On *X86-64* only supports up to 10 bit type parameters and 6
360 floating point parameters.
362 This calling convention supports `tail call
363 optimization <CodeGenerator.html#id80>`_ but requires both the
364 caller and callee are using it.
365 "``cc 11``" - The HiPE calling convention
366 This calling convention has been implemented specifically for use by
367 the `High-Performance Erlang
368 (HiPE) <http://www.it.uu.se/research/group/hipe/>`_ compiler, *the*
369 native code compiler of the `Ericsson's Open Source Erlang/OTP
370 system <http://www.erlang.org/download.shtml>`_. It uses more
371 registers for argument passing than the ordinary C calling
372 convention and defines no callee-saved registers. The calling
373 convention properly supports `tail call
374 optimization <CodeGenerator.html#id80>`_ but requires that both the
375 caller and the callee use it. It uses a *register pinning*
376 mechanism, similar to GHC's convention, for keeping frequently
377 accessed runtime components pinned to specific hardware registers.
378 At the moment only X86 supports this convention (both 32 and 64
380 "``cc <n>``" - Numbered convention
381 Any calling convention may be specified by number, allowing
382 target-specific calling conventions to be used. Target specific
383 calling conventions start at 64.
385 More calling conventions can be added/defined on an as-needed basis, to
386 support Pascal conventions or any other well-known target-independent
392 All Global Variables and Functions have one of the following visibility
395 "``default``" - Default style
396 On targets that use the ELF object file format, default visibility
397 means that the declaration is visible to other modules and, in
398 shared libraries, means that the declared entity may be overridden.
399 On Darwin, default visibility means that the declaration is visible
400 to other modules. Default visibility corresponds to "external
401 linkage" in the language.
402 "``hidden``" - Hidden style
403 Two declarations of an object with hidden visibility refer to the
404 same object if they are in the same shared object. Usually, hidden
405 visibility indicates that the symbol will not be placed into the
406 dynamic symbol table, so no other module (executable or shared
407 library) can reference it directly.
408 "``protected``" - Protected style
409 On ELF, protected visibility indicates that the symbol will be
410 placed in the dynamic symbol table, but that references within the
411 defining module will bind to the local symbol. That is, the symbol
412 cannot be overridden by another module.
417 LLVM IR allows you to specify name aliases for certain types. This can
418 make it easier to read the IR and make the IR more condensed
419 (particularly when recursive types are involved). An example of a name
424 %mytype = type { %mytype*, i32 }
426 You may give a name to any :ref:`type <typesystem>` except
427 ":ref:`void <t_void>`". Type name aliases may be used anywhere a type is
428 expected with the syntax "%mytype".
430 Note that type names are aliases for the structural type that they
431 indicate, and that you can therefore specify multiple names for the same
432 type. This often leads to confusing behavior when dumping out a .ll
433 file. Since LLVM IR uses structural typing, the name is not part of the
434 type. When printing out LLVM IR, the printer will pick *one name* to
435 render all types of a particular shape. This means that if you have code
436 where two different source types end up having the same LLVM type, that
437 the dumper will sometimes print the "wrong" or unexpected type. This is
438 an important design point and isn't going to change.
445 Global variables define regions of memory allocated at compilation time
446 instead of run-time. Global variables may optionally be initialized, may
447 have an explicit section to be placed in, and may have an optional
448 explicit alignment specified.
450 A variable may be defined as ``thread_local``, which means that it will
451 not be shared by threads (each thread will have a separated copy of the
452 variable). Not all targets support thread-local variables. Optionally, a
453 TLS model may be specified:
456 For variables that are only used within the current shared library.
458 For variables in modules that will not be loaded dynamically.
460 For variables defined in the executable and only used within it.
462 The models correspond to the ELF TLS models; see `ELF Handling For
463 Thread-Local Storage <http://people.redhat.com/drepper/tls.pdf>`_ for
464 more information on under which circumstances the different models may
465 be used. The target may choose a different TLS model if the specified
466 model is not supported, or if a better choice of model can be made.
468 A variable may be defined as a global ``constant``, which indicates that
469 the contents of the variable will **never** be modified (enabling better
470 optimization, allowing the global data to be placed in the read-only
471 section of an executable, etc). Note that variables that need runtime
472 initialization cannot be marked ``constant`` as there is a store to the
475 LLVM explicitly allows *declarations* of global variables to be marked
476 constant, even if the final definition of the global is not. This
477 capability can be used to enable slightly better optimization of the
478 program, but requires the language definition to guarantee that
479 optimizations based on the 'constantness' are valid for the translation
480 units that do not include the definition.
482 As SSA values, global variables define pointer values that are in scope
483 (i.e. they dominate) all basic blocks in the program. Global variables
484 always define a pointer to their "content" type because they describe a
485 region of memory, and all memory objects in LLVM are accessed through
488 Global variables can be marked with ``unnamed_addr`` which indicates
489 that the address is not significant, only the content. Constants marked
490 like this can be merged with other constants if they have the same
491 initializer. Note that a constant with significant address *can* be
492 merged with a ``unnamed_addr`` constant, the result being a constant
493 whose address is significant.
495 A global variable may be declared to reside in a target-specific
496 numbered address space. For targets that support them, address spaces
497 may affect how optimizations are performed and/or what target
498 instructions are used to access the variable. The default address space
499 is zero. The address space qualifier must precede any other attributes.
501 LLVM allows an explicit section to be specified for globals. If the
502 target supports it, it will emit globals to the section specified.
504 By default, LLVM optimizes global initializers by assuming that global
505 variables defined within the module are not modified from their
506 initial values before the start of the global initializer. This is
507 true even for variables potentially accessible from outside the
508 module, including those with external linkage or appearing in
511 An explicit alignment may be specified for a global, which must be a
512 power of 2. If not present, or if the alignment is set to zero, the
513 alignment of the global is set by the target to whatever it feels
514 convenient. If an explicit alignment is specified, the global is forced
515 to have exactly that alignment. Targets and optimizers are not allowed
516 to over-align the global if the global has an assigned section. In this
517 case, the extra alignment could be observable: for example, code could
518 assume that the globals are densely packed in their section and try to
519 iterate over them as an array, alignment padding would break this
522 For example, the following defines a global in a numbered address space
523 with an initializer, section, and alignment:
527 @G = addrspace(5) constant float 1.0, section "foo", align 4
529 The following example defines a thread-local global with the
530 ``initialexec`` TLS model:
534 @G = thread_local(initialexec) global i32 0, align 4
536 .. _functionstructure:
541 LLVM function definitions consist of the "``define``" keyword, an
542 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
543 style <visibility>`, an optional :ref:`calling convention <callingconv>`,
544 an optional ``unnamed_addr`` attribute, a return type, an optional
545 :ref:`parameter attribute <paramattrs>` for the return type, a function
546 name, a (possibly empty) argument list (each with optional :ref:`parameter
547 attributes <paramattrs>`), optional :ref:`function attributes <fnattrs>`,
548 an optional section, an optional alignment, an optional :ref:`garbage
549 collector name <gc>`, an opening curly brace, a list of basic blocks,
550 and a closing curly brace.
552 LLVM function declarations consist of the "``declare``" keyword, an
553 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
554 style <visibility>`, an optional :ref:`calling convention <callingconv>`,
555 an optional ``unnamed_addr`` attribute, a return type, an optional
556 :ref:`parameter attribute <paramattrs>` for the return type, a function
557 name, a possibly empty list of arguments, an optional alignment, and an
558 optional :ref:`garbage collector name <gc>`.
560 A function definition contains a list of basic blocks, forming the CFG
561 (Control Flow Graph) for the function. Each basic block may optionally
562 start with a label (giving the basic block a symbol table entry),
563 contains a list of instructions, and ends with a
564 :ref:`terminator <terminators>` instruction (such as a branch or function
567 The first basic block in a function is special in two ways: it is
568 immediately executed on entrance to the function, and it is not allowed
569 to have predecessor basic blocks (i.e. there can not be any branches to
570 the entry block of a function). Because the block can have no
571 predecessors, it also cannot have any :ref:`PHI nodes <i_phi>`.
573 LLVM allows an explicit section to be specified for functions. If the
574 target supports it, it will emit functions to the section specified.
576 An explicit alignment may be specified for a function. If not present,
577 or if the alignment is set to zero, the alignment of the function is set
578 by the target to whatever it feels convenient. If an explicit alignment
579 is specified, the function is forced to have at least that much
580 alignment. All alignments must be a power of 2.
582 If the ``unnamed_addr`` attribute is given, the address is know to not
583 be significant and two identical functions can be merged.
587 define [linkage] [visibility]
589 <ResultType> @<FunctionName> ([argument list])
590 [fn Attrs] [section "name"] [align N]
596 Aliases act as "second name" for the aliasee value (which can be either
597 function, global variable, another alias or bitcast of global value).
598 Aliases may have an optional :ref:`linkage type <linkage>`, and an optional
599 :ref:`visibility style <visibility>`.
603 @<Name> = alias [Linkage] [Visibility] <AliaseeTy> @<Aliasee>
605 .. _namedmetadatastructure:
610 Named metadata is a collection of metadata. :ref:`Metadata
611 nodes <metadata>` (but not metadata strings) are the only valid
612 operands for a named metadata.
616 ; Some unnamed metadata nodes, which are referenced by the named metadata.
617 !0 = metadata !{metadata !"zero"}
618 !1 = metadata !{metadata !"one"}
619 !2 = metadata !{metadata !"two"}
621 !name = !{!0, !1, !2}
628 The return type and each parameter of a function type may have a set of
629 *parameter attributes* associated with them. Parameter attributes are
630 used to communicate additional information about the result or
631 parameters of a function. Parameter attributes are considered to be part
632 of the function, not of the function type, so functions with different
633 parameter attributes can have the same function type.
635 Parameter attributes are simple keywords that follow the type specified.
636 If multiple parameter attributes are needed, they are space separated.
641 declare i32 @printf(i8* noalias nocapture, ...)
642 declare i32 @atoi(i8 zeroext)
643 declare signext i8 @returns_signed_char()
645 Note that any attributes for the function result (``nounwind``,
646 ``readonly``) come immediately after the argument list.
648 Currently, only the following parameter attributes are defined:
651 This indicates to the code generator that the parameter or return
652 value should be zero-extended to the extent required by the target's
653 ABI (which is usually 32-bits, but is 8-bits for a i1 on x86-64) by
654 the caller (for a parameter) or the callee (for a return value).
656 This indicates to the code generator that the parameter or return
657 value should be sign-extended to the extent required by the target's
658 ABI (which is usually 32-bits) by the caller (for a parameter) or
659 the callee (for a return value).
661 This indicates that this parameter or return value should be treated
662 in a special target-dependent fashion during while emitting code for
663 a function call or return (usually, by putting it in a register as
664 opposed to memory, though some targets use it to distinguish between
665 two different kinds of registers). Use of this attribute is
668 This indicates that the pointer parameter should really be passed by
669 value to the function. The attribute implies that a hidden copy of
670 the pointee is made between the caller and the callee, so the callee
671 is unable to modify the value in the caller. This attribute is only
672 valid on LLVM pointer arguments. It is generally used to pass
673 structs and arrays by value, but is also valid on pointers to
674 scalars. The copy is considered to belong to the caller not the
675 callee (for example, ``readonly`` functions should not write to
676 ``byval`` parameters). This is not a valid attribute for return
679 The byval attribute also supports specifying an alignment with the
680 align attribute. It indicates the alignment of the stack slot to
681 form and the known alignment of the pointer specified to the call
682 site. If the alignment is not specified, then the code generator
683 makes a target-specific assumption.
686 This indicates that the pointer parameter specifies the address of a
687 structure that is the return value of the function in the source
688 program. This pointer must be guaranteed by the caller to be valid:
689 loads and stores to the structure may be assumed by the callee
690 not to trap and to be properly aligned. This may only be applied to
691 the first parameter. This is not a valid attribute for return
694 This indicates that pointer values `*based* <pointeraliasing>` on
695 the argument or return value do not alias pointer values which are
696 not *based* on it, ignoring certain "irrelevant" dependencies. For a
697 call to the parent function, dependencies between memory references
698 from before or after the call and from those during the call are
699 "irrelevant" to the ``noalias`` keyword for the arguments and return
700 value used in that call. The caller shares the responsibility with
701 the callee for ensuring that these requirements are met. For further
702 details, please see the discussion of the NoAlias response in `alias
703 analysis <AliasAnalysis.html#MustMayNo>`_.
705 Note that this definition of ``noalias`` is intentionally similar
706 to the definition of ``restrict`` in C99 for function arguments,
707 though it is slightly weaker.
709 For function return values, C99's ``restrict`` is not meaningful,
710 while LLVM's ``noalias`` is.
712 This indicates that the callee does not make any copies of the
713 pointer that outlive the callee itself. This is not a valid
714 attribute for return values.
719 This indicates that the pointer parameter can be excised using the
720 :ref:`trampoline intrinsics <int_trampoline>`. This is not a valid
721 attribute for return values.
725 Garbage Collector Names
726 -----------------------
728 Each function may specify a garbage collector name, which is simply a
733 define void @f() gc "name" { ... }
735 The compiler declares the supported values of *name*. Specifying a
736 collector which will cause the compiler to alter its output in order to
737 support the named garbage collection algorithm.
744 Function attributes are set to communicate additional information about
745 a function. Function attributes are considered to be part of the
746 function, not of the function type, so functions with different function
747 attributes can have the same function type.
749 Function attributes are simple keywords that follow the type specified.
750 If multiple attributes are needed, they are space separated. For
755 define void @f() noinline { ... }
756 define void @f() alwaysinline { ... }
757 define void @f() alwaysinline optsize { ... }
758 define void @f() optsize { ... }
761 This attribute indicates that the address safety analysis is enabled
764 This attribute indicates that, when emitting the prologue and
765 epilogue, the backend should forcibly align the stack pointer.
766 Specify the desired alignment, which must be a power of two, in
769 This attribute indicates that the inliner should attempt to inline
770 this function into callers whenever possible, ignoring any active
771 inlining size threshold for this caller.
773 This attribute suppresses lazy symbol binding for the function. This
774 may make calls to the function faster, at the cost of extra program
775 startup time if the function is not called during program startup.
777 This attribute indicates that the source code contained a hint that
778 inlining this function is desirable (such as the "inline" keyword in
779 C/C++). It is just a hint; it imposes no requirements on the
782 This attribute disables prologue / epilogue emission for the
783 function. This can have very system-specific consequences.
785 This attributes disables implicit floating point instructions.
787 This attribute indicates that the inliner should never inline this
788 function in any situation. This attribute may not be used together
789 with the ``alwaysinline`` attribute.
791 This attribute indicates that the code generator should not use a
792 red zone, even if the target-specific ABI normally permits it.
794 This function attribute indicates that the function never returns
795 normally. This produces undefined behavior at runtime if the
796 function ever does dynamically return.
798 This function attribute indicates that the function never returns
799 with an unwind or exceptional control flow. If the function does
800 unwind, its runtime behavior is undefined.
802 This attribute suggests that optimization passes and code generator
803 passes make choices that keep the code size of this function low,
804 and otherwise do optimizations specifically to reduce code size.
806 This attribute indicates that the function computes its result (or
807 decides to unwind an exception) based strictly on its arguments,
808 without dereferencing any pointer arguments or otherwise accessing
809 any mutable state (e.g. memory, control registers, etc) visible to
810 caller functions. It does not write through any pointer arguments
811 (including ``byval`` arguments) and never changes any state visible
812 to callers. This means that it cannot unwind exceptions by calling
813 the ``C++`` exception throwing methods.
815 This attribute indicates that the function does not write through
816 any pointer arguments (including ``byval`` arguments) or otherwise
817 modify any state (e.g. memory, control registers, etc) visible to
818 caller functions. It may dereference pointer arguments and read
819 state that may be set in the caller. A readonly function always
820 returns the same value (or unwinds an exception identically) when
821 called with the same set of arguments and global state. It cannot
822 unwind an exception by calling the ``C++`` exception throwing
825 This attribute indicates that this function can return twice. The C
826 ``setjmp`` is an example of such a function. The compiler disables
827 some optimizations (like tail calls) in the caller of these
830 This attribute indicates that the function should emit a stack
831 smashing protector. It is in the form of a "canary" --- a random value
832 placed on the stack before the local variables that's checked upon
833 return from the function to see if it has been overwritten. A
834 heuristic is used to determine if a function needs stack protectors
835 or not. The heuristic used will enable protectors for functions with:
837 - Character arrays larger than ``ssp-buffer-size`` (default 8).
838 - Aggregates containing character arrays larger than ``ssp-buffer-size``.
839 - Calls to alloca() with variable sizes or constant sizes greater than
842 If a function that has an ``ssp`` attribute is inlined into a
843 function that doesn't have an ``ssp`` attribute, then the resulting
844 function will have an ``ssp`` attribute.
846 This attribute indicates that the function should *always* emit a
847 stack smashing protector. This overrides the ``ssp`` function
850 If a function that has an ``sspreq`` attribute is inlined into a
851 function that doesn't have an ``sspreq`` attribute or which has an
852 ``ssp`` or ``sspstrong`` attribute, then the resulting function will have
853 an ``sspreq`` attribute.
855 This attribute indicates that the function should emit a stack smashing
856 protector. This attribute causes a strong heuristic to be used when
857 determining if a function needs stack protectors. The strong heuristic
858 will enable protectors for functions with:
860 - Arrays of any size and type
861 - Aggregates containing an array of any size and type.
863 - Local variables that have had their address taken.
865 This overrides the ``ssp`` function attribute.
867 If a function that has an ``sspstrong`` attribute is inlined into a
868 function that doesn't have an ``sspstrong`` attribute, then the
869 resulting function will have an ``sspstrong`` attribute.
871 This attribute indicates that the ABI being targeted requires that
872 an unwind table entry be produce for this function even if we can
873 show that no exceptions passes by it. This is normally the case for
874 the ELF x86-64 abi, but it can be disabled for some compilation
877 This attribute indicates that calls to the function cannot be
878 duplicated. A call to a ``noduplicate`` function may be moved
879 within its parent function, but may not be duplicated within
882 A function containing a ``noduplicate`` call may still
883 be an inlining candidate, provided that the call is not
884 duplicated by inlining. That implies that the function has
885 internal linkage and only has one call site, so the original
886 call is dead after inlining.
890 Module-Level Inline Assembly
891 ----------------------------
893 Modules may contain "module-level inline asm" blocks, which corresponds
894 to the GCC "file scope inline asm" blocks. These blocks are internally
895 concatenated by LLVM and treated as a single unit, but may be separated
896 in the ``.ll`` file if desired. The syntax is very simple:
900 module asm "inline asm code goes here"
901 module asm "more can go here"
903 The strings can contain any character by escaping non-printable
904 characters. The escape sequence used is simply "\\xx" where "xx" is the
905 two digit hex code for the number.
907 The inline asm code is simply printed to the machine code .s file when
908 assembly code is generated.
913 A module may specify a target specific data layout string that specifies
914 how data is to be laid out in memory. The syntax for the data layout is
919 target datalayout = "layout specification"
921 The *layout specification* consists of a list of specifications
922 separated by the minus sign character ('-'). Each specification starts
923 with a letter and may include other information after the letter to
924 define some aspect of the data layout. The specifications accepted are
928 Specifies that the target lays out data in big-endian form. That is,
929 the bits with the most significance have the lowest address
932 Specifies that the target lays out data in little-endian form. That
933 is, the bits with the least significance have the lowest address
936 Specifies the natural alignment of the stack in bits. Alignment
937 promotion of stack variables is limited to the natural stack
938 alignment to avoid dynamic stack realignment. The stack alignment
939 must be a multiple of 8-bits. If omitted, the natural stack
940 alignment defaults to "unspecified", which does not prevent any
941 alignment promotions.
942 ``p[n]:<size>:<abi>:<pref>``
943 This specifies the *size* of a pointer and its ``<abi>`` and
944 ``<pref>``\erred alignments for address space ``n``. All sizes are in
945 bits. Specifying the ``<pref>`` alignment is optional. If omitted, the
946 preceding ``:`` should be omitted too. The address space, ``n`` is
947 optional, and if not specified, denotes the default address space 0.
948 The value of ``n`` must be in the range [1,2^23).
949 ``i<size>:<abi>:<pref>``
950 This specifies the alignment for an integer type of a given bit
951 ``<size>``. The value of ``<size>`` must be in the range [1,2^23).
952 ``v<size>:<abi>:<pref>``
953 This specifies the alignment for a vector type of a given bit
955 ``f<size>:<abi>:<pref>``
956 This specifies the alignment for a floating point type of a given bit
957 ``<size>``. Only values of ``<size>`` that are supported by the target
958 will work. 32 (float) and 64 (double) are supported on all targets; 80
959 or 128 (different flavors of long double) are also supported on some
961 ``a<size>:<abi>:<pref>``
962 This specifies the alignment for an aggregate type of a given bit
964 ``s<size>:<abi>:<pref>``
965 This specifies the alignment for a stack object of a given bit
967 ``n<size1>:<size2>:<size3>...``
968 This specifies a set of native integer widths for the target CPU in
969 bits. For example, it might contain ``n32`` for 32-bit PowerPC,
970 ``n32:64`` for PowerPC 64, or ``n8:16:32:64`` for X86-64. Elements of
971 this set are considered to support most general arithmetic operations
974 When constructing the data layout for a given target, LLVM starts with a
975 default set of specifications which are then (possibly) overridden by
976 the specifications in the ``datalayout`` keyword. The default
977 specifications are given in this list:
980 - ``p:64:64:64`` - 64-bit pointers with 64-bit alignment
981 - ``S0`` - natural stack alignment is unspecified
982 - ``i1:8:8`` - i1 is 8-bit (byte) aligned
983 - ``i8:8:8`` - i8 is 8-bit (byte) aligned
984 - ``i16:16:16`` - i16 is 16-bit aligned
985 - ``i32:32:32`` - i32 is 32-bit aligned
986 - ``i64:32:64`` - i64 has ABI alignment of 32-bits but preferred
988 - ``f16:16:16`` - half is 16-bit aligned
989 - ``f32:32:32`` - float is 32-bit aligned
990 - ``f64:64:64`` - double is 64-bit aligned
991 - ``f128:128:128`` - quad is 128-bit aligned
992 - ``v64:64:64`` - 64-bit vector is 64-bit aligned
993 - ``v128:128:128`` - 128-bit vector is 128-bit aligned
994 - ``a0:0:64`` - aggregates are 64-bit aligned
996 When LLVM is determining the alignment for a given type, it uses the
999 #. If the type sought is an exact match for one of the specifications,
1000 that specification is used.
1001 #. If no match is found, and the type sought is an integer type, then
1002 the smallest integer type that is larger than the bitwidth of the
1003 sought type is used. If none of the specifications are larger than
1004 the bitwidth then the largest integer type is used. For example,
1005 given the default specifications above, the i7 type will use the
1006 alignment of i8 (next largest) while both i65 and i256 will use the
1007 alignment of i64 (largest specified).
1008 #. If no match is found, and the type sought is a vector type, then the
1009 largest vector type that is smaller than the sought vector type will
1010 be used as a fall back. This happens because <128 x double> can be
1011 implemented in terms of 64 <2 x double>, for example.
1013 The function of the data layout string may not be what you expect.
1014 Notably, this is not a specification from the frontend of what alignment
1015 the code generator should use.
1017 Instead, if specified, the target data layout is required to match what
1018 the ultimate *code generator* expects. This string is used by the
1019 mid-level optimizers to improve code, and this only works if it matches
1020 what the ultimate code generator uses. If you would like to generate IR
1021 that does not embed this target-specific detail into the IR, then you
1022 don't have to specify the string. This will disable some optimizations
1023 that require precise layout information, but this also prevents those
1024 optimizations from introducing target specificity into the IR.
1026 .. _pointeraliasing:
1028 Pointer Aliasing Rules
1029 ----------------------
1031 Any memory access must be done through a pointer value associated with
1032 an address range of the memory access, otherwise the behavior is
1033 undefined. Pointer values are associated with address ranges according
1034 to the following rules:
1036 - A pointer value is associated with the addresses associated with any
1037 value it is *based* on.
1038 - An address of a global variable is associated with the address range
1039 of the variable's storage.
1040 - The result value of an allocation instruction is associated with the
1041 address range of the allocated storage.
1042 - A null pointer in the default address-space is associated with no
1044 - An integer constant other than zero or a pointer value returned from
1045 a function not defined within LLVM may be associated with address
1046 ranges allocated through mechanisms other than those provided by
1047 LLVM. Such ranges shall not overlap with any ranges of addresses
1048 allocated by mechanisms provided by LLVM.
1050 A pointer value is *based* on another pointer value according to the
1053 - A pointer value formed from a ``getelementptr`` operation is *based*
1054 on the first operand of the ``getelementptr``.
1055 - The result value of a ``bitcast`` is *based* on the operand of the
1057 - A pointer value formed by an ``inttoptr`` is *based* on all pointer
1058 values that contribute (directly or indirectly) to the computation of
1059 the pointer's value.
1060 - The "*based* on" relationship is transitive.
1062 Note that this definition of *"based"* is intentionally similar to the
1063 definition of *"based"* in C99, though it is slightly weaker.
1065 LLVM IR does not associate types with memory. The result type of a
1066 ``load`` merely indicates the size and alignment of the memory from
1067 which to load, as well as the interpretation of the value. The first
1068 operand type of a ``store`` similarly only indicates the size and
1069 alignment of the store.
1071 Consequently, type-based alias analysis, aka TBAA, aka
1072 ``-fstrict-aliasing``, is not applicable to general unadorned LLVM IR.
1073 :ref:`Metadata <metadata>` may be used to encode additional information
1074 which specialized optimization passes may use to implement type-based
1079 Volatile Memory Accesses
1080 ------------------------
1082 Certain memory accesses, such as :ref:`load <i_load>`'s,
1083 :ref:`store <i_store>`'s, and :ref:`llvm.memcpy <int_memcpy>`'s may be
1084 marked ``volatile``. The optimizers must not change the number of
1085 volatile operations or change their order of execution relative to other
1086 volatile operations. The optimizers *may* change the order of volatile
1087 operations relative to non-volatile operations. This is not Java's
1088 "volatile" and has no cross-thread synchronization behavior.
1090 IR-level volatile loads and stores cannot safely be optimized into
1091 llvm.memcpy or llvm.memmove intrinsics even when those intrinsics are
1092 flagged volatile. Likewise, the backend should never split or merge
1093 target-legal volatile load/store instructions.
1095 .. admonition:: Rationale
1097 Platforms may rely on volatile loads and stores of natively supported
1098 data width to be executed as single instruction. For example, in C
1099 this holds for an l-value of volatile primitive type with native
1100 hardware support, but not necessarily for aggregate types. The
1101 frontend upholds these expectations, which are intentionally
1102 unspecified in the IR. The rules above ensure that IR transformation
1103 do not violate the frontend's contract with the language.
1107 Memory Model for Concurrent Operations
1108 --------------------------------------
1110 The LLVM IR does not define any way to start parallel threads of
1111 execution or to register signal handlers. Nonetheless, there are
1112 platform-specific ways to create them, and we define LLVM IR's behavior
1113 in their presence. This model is inspired by the C++0x memory model.
1115 For a more informal introduction to this model, see the :doc:`Atomics`.
1117 We define a *happens-before* partial order as the least partial order
1120 - Is a superset of single-thread program order, and
1121 - When a *synchronizes-with* ``b``, includes an edge from ``a`` to
1122 ``b``. *Synchronizes-with* pairs are introduced by platform-specific
1123 techniques, like pthread locks, thread creation, thread joining,
1124 etc., and by atomic instructions. (See also :ref:`Atomic Memory Ordering
1125 Constraints <ordering>`).
1127 Note that program order does not introduce *happens-before* edges
1128 between a thread and signals executing inside that thread.
1130 Every (defined) read operation (load instructions, memcpy, atomic
1131 loads/read-modify-writes, etc.) R reads a series of bytes written by
1132 (defined) write operations (store instructions, atomic
1133 stores/read-modify-writes, memcpy, etc.). For the purposes of this
1134 section, initialized globals are considered to have a write of the
1135 initializer which is atomic and happens before any other read or write
1136 of the memory in question. For each byte of a read R, R\ :sub:`byte`
1137 may see any write to the same byte, except:
1139 - If write\ :sub:`1` happens before write\ :sub:`2`, and
1140 write\ :sub:`2` happens before R\ :sub:`byte`, then
1141 R\ :sub:`byte` does not see write\ :sub:`1`.
1142 - If R\ :sub:`byte` happens before write\ :sub:`3`, then
1143 R\ :sub:`byte` does not see write\ :sub:`3`.
1145 Given that definition, R\ :sub:`byte` is defined as follows:
1147 - If R is volatile, the result is target-dependent. (Volatile is
1148 supposed to give guarantees which can support ``sig_atomic_t`` in
1149 C/C++, and may be used for accesses to addresses which do not behave
1150 like normal memory. It does not generally provide cross-thread
1152 - Otherwise, if there is no write to the same byte that happens before
1153 R\ :sub:`byte`, R\ :sub:`byte` returns ``undef`` for that byte.
1154 - Otherwise, if R\ :sub:`byte` may see exactly one write,
1155 R\ :sub:`byte` returns the value written by that write.
1156 - Otherwise, if R is atomic, and all the writes R\ :sub:`byte` may
1157 see are atomic, it chooses one of the values written. See the :ref:`Atomic
1158 Memory Ordering Constraints <ordering>` section for additional
1159 constraints on how the choice is made.
1160 - Otherwise R\ :sub:`byte` returns ``undef``.
1162 R returns the value composed of the series of bytes it read. This
1163 implies that some bytes within the value may be ``undef`` **without**
1164 the entire value being ``undef``. Note that this only defines the
1165 semantics of the operation; it doesn't mean that targets will emit more
1166 than one instruction to read the series of bytes.
1168 Note that in cases where none of the atomic intrinsics are used, this
1169 model places only one restriction on IR transformations on top of what
1170 is required for single-threaded execution: introducing a store to a byte
1171 which might not otherwise be stored is not allowed in general.
1172 (Specifically, in the case where another thread might write to and read
1173 from an address, introducing a store can change a load that may see
1174 exactly one write into a load that may see multiple writes.)
1178 Atomic Memory Ordering Constraints
1179 ----------------------------------
1181 Atomic instructions (:ref:`cmpxchg <i_cmpxchg>`,
1182 :ref:`atomicrmw <i_atomicrmw>`, :ref:`fence <i_fence>`,
1183 :ref:`atomic load <i_load>`, and :ref:`atomic store <i_store>`) take
1184 an ordering parameter that determines which other atomic instructions on
1185 the same address they *synchronize with*. These semantics are borrowed
1186 from Java and C++0x, but are somewhat more colloquial. If these
1187 descriptions aren't precise enough, check those specs (see spec
1188 references in the :doc:`atomics guide <Atomics>`).
1189 :ref:`fence <i_fence>` instructions treat these orderings somewhat
1190 differently since they don't take an address. See that instruction's
1191 documentation for details.
1193 For a simpler introduction to the ordering constraints, see the
1197 The set of values that can be read is governed by the happens-before
1198 partial order. A value cannot be read unless some operation wrote
1199 it. This is intended to provide a guarantee strong enough to model
1200 Java's non-volatile shared variables. This ordering cannot be
1201 specified for read-modify-write operations; it is not strong enough
1202 to make them atomic in any interesting way.
1204 In addition to the guarantees of ``unordered``, there is a single
1205 total order for modifications by ``monotonic`` operations on each
1206 address. All modification orders must be compatible with the
1207 happens-before order. There is no guarantee that the modification
1208 orders can be combined to a global total order for the whole program
1209 (and this often will not be possible). The read in an atomic
1210 read-modify-write operation (:ref:`cmpxchg <i_cmpxchg>` and
1211 :ref:`atomicrmw <i_atomicrmw>`) reads the value in the modification
1212 order immediately before the value it writes. If one atomic read
1213 happens before another atomic read of the same address, the later
1214 read must see the same value or a later value in the address's
1215 modification order. This disallows reordering of ``monotonic`` (or
1216 stronger) operations on the same address. If an address is written
1217 ``monotonic``-ally by one thread, and other threads ``monotonic``-ally
1218 read that address repeatedly, the other threads must eventually see
1219 the write. This corresponds to the C++0x/C1x
1220 ``memory_order_relaxed``.
1222 In addition to the guarantees of ``monotonic``, a
1223 *synchronizes-with* edge may be formed with a ``release`` operation.
1224 This is intended to model C++'s ``memory_order_acquire``.
1226 In addition to the guarantees of ``monotonic``, if this operation
1227 writes a value which is subsequently read by an ``acquire``
1228 operation, it *synchronizes-with* that operation. (This isn't a
1229 complete description; see the C++0x definition of a release
1230 sequence.) This corresponds to the C++0x/C1x
1231 ``memory_order_release``.
1232 ``acq_rel`` (acquire+release)
1233 Acts as both an ``acquire`` and ``release`` operation on its
1234 address. This corresponds to the C++0x/C1x ``memory_order_acq_rel``.
1235 ``seq_cst`` (sequentially consistent)
1236 In addition to the guarantees of ``acq_rel`` (``acquire`` for an
1237 operation which only reads, ``release`` for an operation which only
1238 writes), there is a global total order on all
1239 sequentially-consistent operations on all addresses, which is
1240 consistent with the *happens-before* partial order and with the
1241 modification orders of all the affected addresses. Each
1242 sequentially-consistent read sees the last preceding write to the
1243 same address in this global order. This corresponds to the C++0x/C1x
1244 ``memory_order_seq_cst`` and Java volatile.
1248 If an atomic operation is marked ``singlethread``, it only *synchronizes
1249 with* or participates in modification and seq\_cst total orderings with
1250 other operations running in the same thread (for example, in signal
1258 LLVM IR floating-point binary ops (:ref:`fadd <i_fadd>`,
1259 :ref:`fsub <i_fsub>`, :ref:`fmul <i_fmul>`, :ref:`fdiv <i_fdiv>`,
1260 :ref:`frem <i_frem>`) have the following flags that can set to enable
1261 otherwise unsafe floating point operations
1264 No NaNs - Allow optimizations to assume the arguments and result are not
1265 NaN. Such optimizations are required to retain defined behavior over
1266 NaNs, but the value of the result is undefined.
1269 No Infs - Allow optimizations to assume the arguments and result are not
1270 +/-Inf. Such optimizations are required to retain defined behavior over
1271 +/-Inf, but the value of the result is undefined.
1274 No Signed Zeros - Allow optimizations to treat the sign of a zero
1275 argument or result as insignificant.
1278 Allow Reciprocal - Allow optimizations to use the reciprocal of an
1279 argument rather than perform division.
1282 Fast - Allow algebraically equivalent transformations that may
1283 dramatically change results in floating point (e.g. reassociate). This
1284 flag implies all the others.
1291 The LLVM type system is one of the most important features of the
1292 intermediate representation. Being typed enables a number of
1293 optimizations to be performed on the intermediate representation
1294 directly, without having to do extra analyses on the side before the
1295 transformation. A strong type system makes it easier to read the
1296 generated code and enables novel analyses and transformations that are
1297 not feasible to perform on normal three address code representations.
1299 Type Classifications
1300 --------------------
1302 The types fall into a few useful classifications:
1311 * - :ref:`integer <t_integer>`
1312 - ``i1``, ``i2``, ``i3``, ... ``i8``, ... ``i16``, ... ``i32``, ...
1315 * - :ref:`floating point <t_floating>`
1316 - ``half``, ``float``, ``double``, ``x86_fp80``, ``fp128``,
1324 - :ref:`integer <t_integer>`, :ref:`floating point <t_floating>`,
1325 :ref:`pointer <t_pointer>`, :ref:`vector <t_vector>`,
1326 :ref:`structure <t_struct>`, :ref:`array <t_array>`,
1327 :ref:`label <t_label>`, :ref:`metadata <t_metadata>`.
1329 * - :ref:`primitive <t_primitive>`
1330 - :ref:`label <t_label>`,
1331 :ref:`void <t_void>`,
1332 :ref:`integer <t_integer>`,
1333 :ref:`floating point <t_floating>`,
1334 :ref:`x86mmx <t_x86mmx>`,
1335 :ref:`metadata <t_metadata>`.
1337 * - :ref:`derived <t_derived>`
1338 - :ref:`array <t_array>`,
1339 :ref:`function <t_function>`,
1340 :ref:`pointer <t_pointer>`,
1341 :ref:`structure <t_struct>`,
1342 :ref:`vector <t_vector>`,
1343 :ref:`opaque <t_opaque>`.
1345 The :ref:`first class <t_firstclass>` types are perhaps the most important.
1346 Values of these types are the only ones which can be produced by
1354 The primitive types are the fundamental building blocks of the LLVM
1365 The integer type is a very simple type that simply specifies an
1366 arbitrary bit width for the integer type desired. Any bit width from 1
1367 bit to 2\ :sup:`23`\ -1 (about 8 million) can be specified.
1376 The number of bits the integer will occupy is specified by the ``N``
1382 +----------------+------------------------------------------------+
1383 | ``i1`` | a single-bit integer. |
1384 +----------------+------------------------------------------------+
1385 | ``i32`` | a 32-bit integer. |
1386 +----------------+------------------------------------------------+
1387 | ``i1942652`` | a really big integer of over 1 million bits. |
1388 +----------------+------------------------------------------------+
1392 Floating Point Types
1393 ^^^^^^^^^^^^^^^^^^^^
1402 - 16-bit floating point value
1405 - 32-bit floating point value
1408 - 64-bit floating point value
1411 - 128-bit floating point value (112-bit mantissa)
1414 - 80-bit floating point value (X87)
1417 - 128-bit floating point value (two 64-bits)
1427 The x86mmx type represents a value held in an MMX register on an x86
1428 machine. The operations allowed on it are quite limited: parameters and
1429 return values, load and store, and bitcast. User-specified MMX
1430 instructions are represented as intrinsic or asm calls with arguments
1431 and/or results of this type. There are no arrays, vectors or constants
1449 The void type does not represent any value and has no size.
1466 The label type represents code labels.
1483 The metadata type represents embedded metadata. No derived types may be
1484 created from metadata except for :ref:`function <t_function>` arguments.
1498 The real power in LLVM comes from the derived types in the system. This
1499 is what allows a programmer to represent arrays, functions, pointers,
1500 and other useful types. Each of these types contain one or more element
1501 types which may be a primitive type, or another derived type. For
1502 example, it is possible to have a two dimensional array, using an array
1503 as the element type of another array.
1510 Aggregate Types are a subset of derived types that can contain multiple
1511 member types. :ref:`Arrays <t_array>` and :ref:`structs <t_struct>` are
1512 aggregate types. :ref:`Vectors <t_vector>` are not considered to be
1523 The array type is a very simple derived type that arranges elements
1524 sequentially in memory. The array type requires a size (number of
1525 elements) and an underlying data type.
1532 [<# elements> x <elementtype>]
1534 The number of elements is a constant integer value; ``elementtype`` may
1535 be any type with a size.
1540 +------------------+--------------------------------------+
1541 | ``[40 x i32]`` | Array of 40 32-bit integer values. |
1542 +------------------+--------------------------------------+
1543 | ``[41 x i32]`` | Array of 41 32-bit integer values. |
1544 +------------------+--------------------------------------+
1545 | ``[4 x i8]`` | Array of 4 8-bit integer values. |
1546 +------------------+--------------------------------------+
1548 Here are some examples of multidimensional arrays:
1550 +-----------------------------+----------------------------------------------------------+
1551 | ``[3 x [4 x i32]]`` | 3x4 array of 32-bit integer values. |
1552 +-----------------------------+----------------------------------------------------------+
1553 | ``[12 x [10 x float]]`` | 12x10 array of single precision floating point values. |
1554 +-----------------------------+----------------------------------------------------------+
1555 | ``[2 x [3 x [4 x i16]]]`` | 2x3x4 array of 16-bit integer values. |
1556 +-----------------------------+----------------------------------------------------------+
1558 There is no restriction on indexing beyond the end of the array implied
1559 by a static type (though there are restrictions on indexing beyond the
1560 bounds of an allocated object in some cases). This means that
1561 single-dimension 'variable sized array' addressing can be implemented in
1562 LLVM with a zero length array type. An implementation of 'pascal style
1563 arrays' in LLVM could use the type "``{ i32, [0 x float]}``", for
1574 The function type can be thought of as a function signature. It consists
1575 of a return type and a list of formal parameter types. The return type
1576 of a function type is a first class type or a void type.
1583 <returntype> (<parameter list>)
1585 ...where '``<parameter list>``' is a comma-separated list of type
1586 specifiers. Optionally, the parameter list may include a type ``...``,
1587 which indicates that the function takes a variable number of arguments.
1588 Variable argument functions can access their arguments with the
1589 :ref:`variable argument handling intrinsic <int_varargs>` functions.
1590 '``<returntype>``' is any type except :ref:`label <t_label>`.
1595 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1596 | ``i32 (i32)`` | function taking an ``i32``, returning an ``i32`` |
1597 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1598 | ``float (i16, i32 *) *`` | :ref:`Pointer <t_pointer>` to a function that takes an ``i16`` and a :ref:`pointer <t_pointer>` to ``i32``, returning ``float``. |
1599 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1600 | ``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. |
1601 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1602 | ``{i32, i32} (i32)`` | A function taking an ``i32``, returning a :ref:`structure <t_struct>` containing two ``i32`` values |
1603 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1613 The structure type is used to represent a collection of data members
1614 together in memory. The elements of a structure may be any type that has
1617 Structures in memory are accessed using '``load``' and '``store``' by
1618 getting a pointer to a field with the '``getelementptr``' instruction.
1619 Structures in registers are accessed using the '``extractvalue``' and
1620 '``insertvalue``' instructions.
1622 Structures may optionally be "packed" structures, which indicate that
1623 the alignment of the struct is one byte, and that there is no padding
1624 between the elements. In non-packed structs, padding between field types
1625 is inserted as defined by the DataLayout string in the module, which is
1626 required to match what the underlying code generator expects.
1628 Structures can either be "literal" or "identified". A literal structure
1629 is defined inline with other types (e.g. ``{i32, i32}*``) whereas
1630 identified types are always defined at the top level with a name.
1631 Literal types are uniqued by their contents and can never be recursive
1632 or opaque since there is no way to write one. Identified types can be
1633 recursive, can be opaqued, and are never uniqued.
1640 %T1 = type { <type list> } ; Identified normal struct type
1641 %T2 = type <{ <type list> }> ; Identified packed struct type
1646 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1647 | ``{ i32, i32, i32 }`` | A triple of three ``i32`` values |
1648 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1649 | ``{ 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``. |
1650 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1651 | ``<{ i8, i32 }>`` | A packed struct known to be 5 bytes in size. |
1652 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1656 Opaque Structure Types
1657 ^^^^^^^^^^^^^^^^^^^^^^
1662 Opaque structure types are used to represent named structure types that
1663 do not have a body specified. This corresponds (for example) to the C
1664 notion of a forward declared structure.
1677 +--------------+-------------------+
1678 | ``opaque`` | An opaque type. |
1679 +--------------+-------------------+
1689 The pointer type is used to specify memory locations. Pointers are
1690 commonly used to reference objects in memory.
1692 Pointer types may have an optional address space attribute defining the
1693 numbered address space where the pointed-to object resides. The default
1694 address space is number zero. The semantics of non-zero address spaces
1695 are target-specific.
1697 Note that LLVM does not permit pointers to void (``void*``) nor does it
1698 permit pointers to labels (``label*``). Use ``i8*`` instead.
1710 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1711 | ``[4 x i32]*`` | A :ref:`pointer <t_pointer>` to :ref:`array <t_array>` of four ``i32`` values. |
1712 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1713 | ``i32 (i32*) *`` | A :ref:`pointer <t_pointer>` to a :ref:`function <t_function>` that takes an ``i32*``, returning an ``i32``. |
1714 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1715 | ``i32 addrspace(5)*`` | A :ref:`pointer <t_pointer>` to an ``i32`` value that resides in address space #5. |
1716 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1726 A vector type is a simple derived type that represents a vector of
1727 elements. Vector types are used when multiple primitive data are
1728 operated in parallel using a single instruction (SIMD). A vector type
1729 requires a size (number of elements) and an underlying primitive data
1730 type. Vector types are considered :ref:`first class <t_firstclass>`.
1737 < <# elements> x <elementtype> >
1739 The number of elements is a constant integer value larger than 0;
1740 elementtype may be any integer or floating point type, or a pointer to
1741 these types. Vectors of size zero are not allowed.
1746 +-------------------+--------------------------------------------------+
1747 | ``<4 x i32>`` | Vector of 4 32-bit integer values. |
1748 +-------------------+--------------------------------------------------+
1749 | ``<8 x float>`` | Vector of 8 32-bit floating-point values. |
1750 +-------------------+--------------------------------------------------+
1751 | ``<2 x i64>`` | Vector of 2 64-bit integer values. |
1752 +-------------------+--------------------------------------------------+
1753 | ``<4 x i64*>`` | Vector of 4 pointers to 64-bit integer values. |
1754 +-------------------+--------------------------------------------------+
1759 LLVM has several different basic types of constants. This section
1760 describes them all and their syntax.
1765 **Boolean constants**
1766 The two strings '``true``' and '``false``' are both valid constants
1768 **Integer constants**
1769 Standard integers (such as '4') are constants of the
1770 :ref:`integer <t_integer>` type. Negative numbers may be used with
1772 **Floating point constants**
1773 Floating point constants use standard decimal notation (e.g.
1774 123.421), exponential notation (e.g. 1.23421e+2), or a more precise
1775 hexadecimal notation (see below). The assembler requires the exact
1776 decimal value of a floating-point constant. For example, the
1777 assembler accepts 1.25 but rejects 1.3 because 1.3 is a repeating
1778 decimal in binary. Floating point constants must have a :ref:`floating
1779 point <t_floating>` type.
1780 **Null pointer constants**
1781 The identifier '``null``' is recognized as a null pointer constant
1782 and must be of :ref:`pointer type <t_pointer>`.
1784 The one non-intuitive notation for constants is the hexadecimal form of
1785 floating point constants. For example, the form
1786 '``double 0x432ff973cafa8000``' is equivalent to (but harder to read
1787 than) '``double 4.5e+15``'. The only time hexadecimal floating point
1788 constants are required (and the only time that they are generated by the
1789 disassembler) is when a floating point constant must be emitted but it
1790 cannot be represented as a decimal floating point number in a reasonable
1791 number of digits. For example, NaN's, infinities, and other special
1792 values are represented in their IEEE hexadecimal format so that assembly
1793 and disassembly do not cause any bits to change in the constants.
1795 When using the hexadecimal form, constants of types half, float, and
1796 double are represented using the 16-digit form shown above (which
1797 matches the IEEE754 representation for double); half and float values
1798 must, however, be exactly representable as IEEE 754 half and single
1799 precision, respectively. Hexadecimal format is always used for long
1800 double, and there are three forms of long double. The 80-bit format used
1801 by x86 is represented as ``0xK`` followed by 20 hexadecimal digits. The
1802 128-bit format used by PowerPC (two adjacent doubles) is represented by
1803 ``0xM`` followed by 32 hexadecimal digits. The IEEE 128-bit format is
1804 represented by ``0xL`` followed by 32 hexadecimal digits; no currently
1805 supported target uses this format. Long doubles will only work if they
1806 match the long double format on your target. The IEEE 16-bit format
1807 (half precision) is represented by ``0xH`` followed by 4 hexadecimal
1808 digits. All hexadecimal formats are big-endian (sign bit at the left).
1810 There are no constants of type x86mmx.
1815 Complex constants are a (potentially recursive) combination of simple
1816 constants and smaller complex constants.
1818 **Structure constants**
1819 Structure constants are represented with notation similar to
1820 structure type definitions (a comma separated list of elements,
1821 surrounded by braces (``{}``)). For example:
1822 "``{ i32 4, float 17.0, i32* @G }``", where "``@G``" is declared as
1823 "``@G = external global i32``". Structure constants must have
1824 :ref:`structure type <t_struct>`, and the number and types of elements
1825 must match those specified by the type.
1827 Array constants are represented with notation similar to array type
1828 definitions (a comma separated list of elements, surrounded by
1829 square brackets (``[]``)). For example:
1830 "``[ i32 42, i32 11, i32 74 ]``". Array constants must have
1831 :ref:`array type <t_array>`, and the number and types of elements must
1832 match those specified by the type.
1833 **Vector constants**
1834 Vector constants are represented with notation similar to vector
1835 type definitions (a comma separated list of elements, surrounded by
1836 less-than/greater-than's (``<>``)). For example:
1837 "``< i32 42, i32 11, i32 74, i32 100 >``". Vector constants
1838 must have :ref:`vector type <t_vector>`, and the number and types of
1839 elements must match those specified by the type.
1840 **Zero initialization**
1841 The string '``zeroinitializer``' can be used to zero initialize a
1842 value to zero of *any* type, including scalar and
1843 :ref:`aggregate <t_aggregate>` types. This is often used to avoid
1844 having to print large zero initializers (e.g. for large arrays) and
1845 is always exactly equivalent to using explicit zero initializers.
1847 A metadata node is a structure-like constant with :ref:`metadata
1848 type <t_metadata>`. For example:
1849 "``metadata !{ i32 0, metadata !"test" }``". Unlike other
1850 constants that are meant to be interpreted as part of the
1851 instruction stream, metadata is a place to attach additional
1852 information such as debug info.
1854 Global Variable and Function Addresses
1855 --------------------------------------
1857 The addresses of :ref:`global variables <globalvars>` and
1858 :ref:`functions <functionstructure>` are always implicitly valid
1859 (link-time) constants. These constants are explicitly referenced when
1860 the :ref:`identifier for the global <identifiers>` is used and always have
1861 :ref:`pointer <t_pointer>` type. For example, the following is a legal LLVM
1864 .. code-block:: llvm
1868 @Z = global [2 x i32*] [ i32* @X, i32* @Y ]
1875 The string '``undef``' can be used anywhere a constant is expected, and
1876 indicates that the user of the value may receive an unspecified
1877 bit-pattern. Undefined values may be of any type (other than '``label``'
1878 or '``void``') and be used anywhere a constant is permitted.
1880 Undefined values are useful because they indicate to the compiler that
1881 the program is well defined no matter what value is used. This gives the
1882 compiler more freedom to optimize. Here are some examples of
1883 (potentially surprising) transformations that are valid (in pseudo IR):
1885 .. code-block:: llvm
1895 This is safe because all of the output bits are affected by the undef
1896 bits. Any output bit can have a zero or one depending on the input bits.
1898 .. code-block:: llvm
1909 These logical operations have bits that are not always affected by the
1910 input. For example, if ``%X`` has a zero bit, then the output of the
1911 '``and``' operation will always be a zero for that bit, no matter what
1912 the corresponding bit from the '``undef``' is. As such, it is unsafe to
1913 optimize or assume that the result of the '``and``' is '``undef``'.
1914 However, it is safe to assume that all bits of the '``undef``' could be
1915 0, and optimize the '``and``' to 0. Likewise, it is safe to assume that
1916 all the bits of the '``undef``' operand to the '``or``' could be set,
1917 allowing the '``or``' to be folded to -1.
1919 .. code-block:: llvm
1921 %A = select undef, %X, %Y
1922 %B = select undef, 42, %Y
1923 %C = select %X, %Y, undef
1933 This set of examples shows that undefined '``select``' (and conditional
1934 branch) conditions can go *either way*, but they have to come from one
1935 of the two operands. In the ``%A`` example, if ``%X`` and ``%Y`` were
1936 both known to have a clear low bit, then ``%A`` would have to have a
1937 cleared low bit. However, in the ``%C`` example, the optimizer is
1938 allowed to assume that the '``undef``' operand could be the same as
1939 ``%Y``, allowing the whole '``select``' to be eliminated.
1941 .. code-block:: llvm
1943 %A = xor undef, undef
1960 This example points out that two '``undef``' operands are not
1961 necessarily the same. This can be surprising to people (and also matches
1962 C semantics) where they assume that "``X^X``" is always zero, even if
1963 ``X`` is undefined. This isn't true for a number of reasons, but the
1964 short answer is that an '``undef``' "variable" can arbitrarily change
1965 its value over its "live range". This is true because the variable
1966 doesn't actually *have a live range*. Instead, the value is logically
1967 read from arbitrary registers that happen to be around when needed, so
1968 the value is not necessarily consistent over time. In fact, ``%A`` and
1969 ``%C`` need to have the same semantics or the core LLVM "replace all
1970 uses with" concept would not hold.
1972 .. code-block:: llvm
1980 These examples show the crucial difference between an *undefined value*
1981 and *undefined behavior*. An undefined value (like '``undef``') is
1982 allowed to have an arbitrary bit-pattern. This means that the ``%A``
1983 operation can be constant folded to '``undef``', because the '``undef``'
1984 could be an SNaN, and ``fdiv`` is not (currently) defined on SNaN's.
1985 However, in the second example, we can make a more aggressive
1986 assumption: because the ``undef`` is allowed to be an arbitrary value,
1987 we are allowed to assume that it could be zero. Since a divide by zero
1988 has *undefined behavior*, we are allowed to assume that the operation
1989 does not execute at all. This allows us to delete the divide and all
1990 code after it. Because the undefined operation "can't happen", the
1991 optimizer can assume that it occurs in dead code.
1993 .. code-block:: llvm
1995 a: store undef -> %X
1996 b: store %X -> undef
2001 These examples reiterate the ``fdiv`` example: a store *of* an undefined
2002 value can be assumed to not have any effect; we can assume that the
2003 value is overwritten with bits that happen to match what was already
2004 there. However, a store *to* an undefined location could clobber
2005 arbitrary memory, therefore, it has undefined behavior.
2012 Poison values are similar to :ref:`undef values <undefvalues>`, however
2013 they also represent the fact that an instruction or constant expression
2014 which cannot evoke side effects has nevertheless detected a condition
2015 which results in undefined behavior.
2017 There is currently no way of representing a poison value in the IR; they
2018 only exist when produced by operations such as :ref:`add <i_add>` with
2021 Poison value behavior is defined in terms of value *dependence*:
2023 - Values other than :ref:`phi <i_phi>` nodes depend on their operands.
2024 - :ref:`Phi <i_phi>` nodes depend on the operand corresponding to
2025 their dynamic predecessor basic block.
2026 - Function arguments depend on the corresponding actual argument values
2027 in the dynamic callers of their functions.
2028 - :ref:`Call <i_call>` instructions depend on the :ref:`ret <i_ret>`
2029 instructions that dynamically transfer control back to them.
2030 - :ref:`Invoke <i_invoke>` instructions depend on the
2031 :ref:`ret <i_ret>`, :ref:`resume <i_resume>`, or exception-throwing
2032 call instructions that dynamically transfer control back to them.
2033 - Non-volatile loads and stores depend on the most recent stores to all
2034 of the referenced memory addresses, following the order in the IR
2035 (including loads and stores implied by intrinsics such as
2036 :ref:`@llvm.memcpy <int_memcpy>`.)
2037 - An instruction with externally visible side effects depends on the
2038 most recent preceding instruction with externally visible side
2039 effects, following the order in the IR. (This includes :ref:`volatile
2040 operations <volatile>`.)
2041 - An instruction *control-depends* on a :ref:`terminator
2042 instruction <terminators>` if the terminator instruction has
2043 multiple successors and the instruction is always executed when
2044 control transfers to one of the successors, and may not be executed
2045 when control is transferred to another.
2046 - Additionally, an instruction also *control-depends* on a terminator
2047 instruction if the set of instructions it otherwise depends on would
2048 be different if the terminator had transferred control to a different
2050 - Dependence is transitive.
2052 Poison Values have the same behavior as :ref:`undef values <undefvalues>`,
2053 with the additional affect that any instruction which has a *dependence*
2054 on a poison value has undefined behavior.
2056 Here are some examples:
2058 .. code-block:: llvm
2061 %poison = sub nuw i32 0, 1 ; Results in a poison value.
2062 %still_poison = and i32 %poison, 0 ; 0, but also poison.
2063 %poison_yet_again = getelementptr i32* @h, i32 %still_poison
2064 store i32 0, i32* %poison_yet_again ; memory at @h[0] is poisoned
2066 store i32 %poison, i32* @g ; Poison value stored to memory.
2067 %poison2 = load i32* @g ; Poison value loaded back from memory.
2069 store volatile i32 %poison, i32* @g ; External observation; undefined behavior.
2071 %narrowaddr = bitcast i32* @g to i16*
2072 %wideaddr = bitcast i32* @g to i64*
2073 %poison3 = load i16* %narrowaddr ; Returns a poison value.
2074 %poison4 = load i64* %wideaddr ; Returns a poison value.
2076 %cmp = icmp slt i32 %poison, 0 ; Returns a poison value.
2077 br i1 %cmp, label %true, label %end ; Branch to either destination.
2080 store volatile i32 0, i32* @g ; This is control-dependent on %cmp, so
2081 ; it has undefined behavior.
2085 %p = phi i32 [ 0, %entry ], [ 1, %true ]
2086 ; Both edges into this PHI are
2087 ; control-dependent on %cmp, so this
2088 ; always results in a poison value.
2090 store volatile i32 0, i32* @g ; This would depend on the store in %true
2091 ; if %cmp is true, or the store in %entry
2092 ; otherwise, so this is undefined behavior.
2094 br i1 %cmp, label %second_true, label %second_end
2095 ; The same branch again, but this time the
2096 ; true block doesn't have side effects.
2103 store volatile i32 0, i32* @g ; This time, the instruction always depends
2104 ; on the store in %end. Also, it is
2105 ; control-equivalent to %end, so this is
2106 ; well-defined (ignoring earlier undefined
2107 ; behavior in this example).
2111 Addresses of Basic Blocks
2112 -------------------------
2114 ``blockaddress(@function, %block)``
2116 The '``blockaddress``' constant computes the address of the specified
2117 basic block in the specified function, and always has an ``i8*`` type.
2118 Taking the address of the entry block is illegal.
2120 This value only has defined behavior when used as an operand to the
2121 ':ref:`indirectbr <i_indirectbr>`' instruction, or for comparisons
2122 against null. Pointer equality tests between labels addresses results in
2123 undefined behavior --- though, again, comparison against null is ok, and
2124 no label is equal to the null pointer. This may be passed around as an
2125 opaque pointer sized value as long as the bits are not inspected. This
2126 allows ``ptrtoint`` and arithmetic to be performed on these values so
2127 long as the original value is reconstituted before the ``indirectbr``
2130 Finally, some targets may provide defined semantics when using the value
2131 as the operand to an inline assembly, but that is target specific.
2133 Constant Expressions
2134 --------------------
2136 Constant expressions are used to allow expressions involving other
2137 constants to be used as constants. Constant expressions may be of any
2138 :ref:`first class <t_firstclass>` type and may involve any LLVM operation
2139 that does not have side effects (e.g. load and call are not supported).
2140 The following is the syntax for constant expressions:
2142 ``trunc (CST to TYPE)``
2143 Truncate a constant to another type. The bit size of CST must be
2144 larger than the bit size of TYPE. Both types must be integers.
2145 ``zext (CST to TYPE)``
2146 Zero extend a constant to another type. The bit size of CST must be
2147 smaller than the bit size of TYPE. Both types must be integers.
2148 ``sext (CST to TYPE)``
2149 Sign extend a constant to another type. The bit size of CST must be
2150 smaller than the bit size of TYPE. Both types must be integers.
2151 ``fptrunc (CST to TYPE)``
2152 Truncate a floating point constant to another floating point type.
2153 The size of CST must be larger than the size of TYPE. Both types
2154 must be floating point.
2155 ``fpext (CST to TYPE)``
2156 Floating point extend a constant to another type. The size of CST
2157 must be smaller or equal to the size of TYPE. Both types must be
2159 ``fptoui (CST to TYPE)``
2160 Convert a floating point constant to the corresponding unsigned
2161 integer constant. TYPE must be a scalar or vector integer type. CST
2162 must be of scalar or vector floating point type. Both CST and TYPE
2163 must be scalars, or vectors of the same number of elements. If the
2164 value won't fit in the integer type, the results are undefined.
2165 ``fptosi (CST to TYPE)``
2166 Convert a floating point constant to the corresponding signed
2167 integer constant. TYPE must be a scalar or vector integer type. CST
2168 must be of scalar or vector floating point type. Both CST and TYPE
2169 must be scalars, or vectors of the same number of elements. If the
2170 value won't fit in the integer type, the results are undefined.
2171 ``uitofp (CST to TYPE)``
2172 Convert an unsigned integer constant to the corresponding floating
2173 point constant. TYPE must be a scalar or vector floating point type.
2174 CST must be of scalar or vector integer type. Both CST and TYPE must
2175 be scalars, or vectors of the same number of elements. If the value
2176 won't fit in the floating point type, the results are undefined.
2177 ``sitofp (CST to TYPE)``
2178 Convert a signed integer constant to the corresponding floating
2179 point constant. TYPE must be a scalar or vector floating point type.
2180 CST must be of scalar or vector integer type. Both CST and TYPE must
2181 be scalars, or vectors of the same number of elements. If the value
2182 won't fit in the floating point type, the results are undefined.
2183 ``ptrtoint (CST to TYPE)``
2184 Convert a pointer typed constant to the corresponding integer
2185 constant ``TYPE`` must be an integer type. ``CST`` must be of
2186 pointer type. The ``CST`` value is zero extended, truncated, or
2187 unchanged to make it fit in ``TYPE``.
2188 ``inttoptr (CST to TYPE)``
2189 Convert an integer constant to a pointer constant. TYPE must be a
2190 pointer type. CST must be of integer type. The CST value is zero
2191 extended, truncated, or unchanged to make it fit in a pointer size.
2192 This one is *really* dangerous!
2193 ``bitcast (CST to TYPE)``
2194 Convert a constant, CST, to another TYPE. The constraints of the
2195 operands are the same as those for the :ref:`bitcast
2196 instruction <i_bitcast>`.
2197 ``getelementptr (CSTPTR, IDX0, IDX1, ...)``, ``getelementptr inbounds (CSTPTR, IDX0, IDX1, ...)``
2198 Perform the :ref:`getelementptr operation <i_getelementptr>` on
2199 constants. As with the :ref:`getelementptr <i_getelementptr>`
2200 instruction, the index list may have zero or more indexes, which are
2201 required to make sense for the type of "CSTPTR".
2202 ``select (COND, VAL1, VAL2)``
2203 Perform the :ref:`select operation <i_select>` on constants.
2204 ``icmp COND (VAL1, VAL2)``
2205 Performs the :ref:`icmp operation <i_icmp>` on constants.
2206 ``fcmp COND (VAL1, VAL2)``
2207 Performs the :ref:`fcmp operation <i_fcmp>` on constants.
2208 ``extractelement (VAL, IDX)``
2209 Perform the :ref:`extractelement operation <i_extractelement>` on
2211 ``insertelement (VAL, ELT, IDX)``
2212 Perform the :ref:`insertelement operation <i_insertelement>` on
2214 ``shufflevector (VEC1, VEC2, IDXMASK)``
2215 Perform the :ref:`shufflevector operation <i_shufflevector>` on
2217 ``extractvalue (VAL, IDX0, IDX1, ...)``
2218 Perform the :ref:`extractvalue operation <i_extractvalue>` on
2219 constants. The index list is interpreted in a similar manner as
2220 indices in a ':ref:`getelementptr <i_getelementptr>`' operation. At
2221 least one index value must be specified.
2222 ``insertvalue (VAL, ELT, IDX0, IDX1, ...)``
2223 Perform the :ref:`insertvalue operation <i_insertvalue>` on constants.
2224 The index list is interpreted in a similar manner as indices in a
2225 ':ref:`getelementptr <i_getelementptr>`' operation. At least one index
2226 value must be specified.
2227 ``OPCODE (LHS, RHS)``
2228 Perform the specified operation of the LHS and RHS constants. OPCODE
2229 may be any of the :ref:`binary <binaryops>` or :ref:`bitwise
2230 binary <bitwiseops>` operations. The constraints on operands are
2231 the same as those for the corresponding instruction (e.g. no bitwise
2232 operations on floating point values are allowed).
2237 Inline Assembler Expressions
2238 ----------------------------
2240 LLVM supports inline assembler expressions (as opposed to :ref:`Module-Level
2241 Inline Assembly <moduleasm>`) through the use of a special value. This
2242 value represents the inline assembler as a string (containing the
2243 instructions to emit), a list of operand constraints (stored as a
2244 string), a flag that indicates whether or not the inline asm expression
2245 has side effects, and a flag indicating whether the function containing
2246 the asm needs to align its stack conservatively. An example inline
2247 assembler expression is:
2249 .. code-block:: llvm
2251 i32 (i32) asm "bswap $0", "=r,r"
2253 Inline assembler expressions may **only** be used as the callee operand
2254 of a :ref:`call <i_call>` or an :ref:`invoke <i_invoke>` instruction.
2255 Thus, typically we have:
2257 .. code-block:: llvm
2259 %X = call i32 asm "bswap $0", "=r,r"(i32 %Y)
2261 Inline asms with side effects not visible in the constraint list must be
2262 marked as having side effects. This is done through the use of the
2263 '``sideeffect``' keyword, like so:
2265 .. code-block:: llvm
2267 call void asm sideeffect "eieio", ""()
2269 In some cases inline asms will contain code that will not work unless
2270 the stack is aligned in some way, such as calls or SSE instructions on
2271 x86, yet will not contain code that does that alignment within the asm.
2272 The compiler should make conservative assumptions about what the asm
2273 might contain and should generate its usual stack alignment code in the
2274 prologue if the '``alignstack``' keyword is present:
2276 .. code-block:: llvm
2278 call void asm alignstack "eieio", ""()
2280 Inline asms also support using non-standard assembly dialects. The
2281 assumed dialect is ATT. When the '``inteldialect``' keyword is present,
2282 the inline asm is using the Intel dialect. Currently, ATT and Intel are
2283 the only supported dialects. An example is:
2285 .. code-block:: llvm
2287 call void asm inteldialect "eieio", ""()
2289 If multiple keywords appear the '``sideeffect``' keyword must come
2290 first, the '``alignstack``' keyword second and the '``inteldialect``'
2296 The call instructions that wrap inline asm nodes may have a
2297 "``!srcloc``" MDNode attached to it that contains a list of constant
2298 integers. If present, the code generator will use the integer as the
2299 location cookie value when report errors through the ``LLVMContext``
2300 error reporting mechanisms. This allows a front-end to correlate backend
2301 errors that occur with inline asm back to the source code that produced
2304 .. code-block:: llvm
2306 call void asm sideeffect "something bad", ""(), !srcloc !42
2308 !42 = !{ i32 1234567 }
2310 It is up to the front-end to make sense of the magic numbers it places
2311 in the IR. If the MDNode contains multiple constants, the code generator
2312 will use the one that corresponds to the line of the asm that the error
2317 Metadata Nodes and Metadata Strings
2318 -----------------------------------
2320 LLVM IR allows metadata to be attached to instructions in the program
2321 that can convey extra information about the code to the optimizers and
2322 code generator. One example application of metadata is source-level
2323 debug information. There are two metadata primitives: strings and nodes.
2324 All metadata has the ``metadata`` type and is identified in syntax by a
2325 preceding exclamation point ('``!``').
2327 A metadata string is a string surrounded by double quotes. It can
2328 contain any character by escaping non-printable characters with
2329 "``\xx``" where "``xx``" is the two digit hex code. For example:
2332 Metadata nodes are represented with notation similar to structure
2333 constants (a comma separated list of elements, surrounded by braces and
2334 preceded by an exclamation point). Metadata nodes can have any values as
2335 their operand. For example:
2337 .. code-block:: llvm
2339 !{ metadata !"test\00", i32 10}
2341 A :ref:`named metadata <namedmetadatastructure>` is a collection of
2342 metadata nodes, which can be looked up in the module symbol table. For
2345 .. code-block:: llvm
2347 !foo = metadata !{!4, !3}
2349 Metadata can be used as function arguments. Here ``llvm.dbg.value``
2350 function is using two metadata arguments:
2352 .. code-block:: llvm
2354 call void @llvm.dbg.value(metadata !24, i64 0, metadata !25)
2356 Metadata can be attached with an instruction. Here metadata ``!21`` is
2357 attached to the ``add`` instruction using the ``!dbg`` identifier:
2359 .. code-block:: llvm
2361 %indvar.next = add i64 %indvar, 1, !dbg !21
2363 More information about specific metadata nodes recognized by the
2364 optimizers and code generator is found below.
2369 In LLVM IR, memory does not have types, so LLVM's own type system is not
2370 suitable for doing TBAA. Instead, metadata is added to the IR to
2371 describe a type system of a higher level language. This can be used to
2372 implement typical C/C++ TBAA, but it can also be used to implement
2373 custom alias analysis behavior for other languages.
2375 The current metadata format is very simple. TBAA metadata nodes have up
2376 to three fields, e.g.:
2378 .. code-block:: llvm
2380 !0 = metadata !{ metadata !"an example type tree" }
2381 !1 = metadata !{ metadata !"int", metadata !0 }
2382 !2 = metadata !{ metadata !"float", metadata !0 }
2383 !3 = metadata !{ metadata !"const float", metadata !2, i64 1 }
2385 The first field is an identity field. It can be any value, usually a
2386 metadata string, which uniquely identifies the type. The most important
2387 name in the tree is the name of the root node. Two trees with different
2388 root node names are entirely disjoint, even if they have leaves with
2391 The second field identifies the type's parent node in the tree, or is
2392 null or omitted for a root node. A type is considered to alias all of
2393 its descendants and all of its ancestors in the tree. Also, a type is
2394 considered to alias all types in other trees, so that bitcode produced
2395 from multiple front-ends is handled conservatively.
2397 If the third field is present, it's an integer which if equal to 1
2398 indicates that the type is "constant" (meaning
2399 ``pointsToConstantMemory`` should return true; see `other useful
2400 AliasAnalysis methods <AliasAnalysis.html#OtherItfs>`_).
2402 '``tbaa.struct``' Metadata
2403 ^^^^^^^^^^^^^^^^^^^^^^^^^^
2405 The :ref:`llvm.memcpy <int_memcpy>` is often used to implement
2406 aggregate assignment operations in C and similar languages, however it
2407 is defined to copy a contiguous region of memory, which is more than
2408 strictly necessary for aggregate types which contain holes due to
2409 padding. Also, it doesn't contain any TBAA information about the fields
2412 ``!tbaa.struct`` metadata can describe which memory subregions in a
2413 memcpy are padding and what the TBAA tags of the struct are.
2415 The current metadata format is very simple. ``!tbaa.struct`` metadata
2416 nodes are a list of operands which are in conceptual groups of three.
2417 For each group of three, the first operand gives the byte offset of a
2418 field in bytes, the second gives its size in bytes, and the third gives
2421 .. code-block:: llvm
2423 !4 = metadata !{ i64 0, i64 4, metadata !1, i64 8, i64 4, metadata !2 }
2425 This describes a struct with two fields. The first is at offset 0 bytes
2426 with size 4 bytes, and has tbaa tag !1. The second is at offset 8 bytes
2427 and has size 4 bytes and has tbaa tag !2.
2429 Note that the fields need not be contiguous. In this example, there is a
2430 4 byte gap between the two fields. This gap represents padding which
2431 does not carry useful data and need not be preserved.
2433 '``fpmath``' Metadata
2434 ^^^^^^^^^^^^^^^^^^^^^
2436 ``fpmath`` metadata may be attached to any instruction of floating point
2437 type. It can be used to express the maximum acceptable error in the
2438 result of that instruction, in ULPs, thus potentially allowing the
2439 compiler to use a more efficient but less accurate method of computing
2440 it. ULP is defined as follows:
2442 If ``x`` is a real number that lies between two finite consecutive
2443 floating-point numbers ``a`` and ``b``, without being equal to one
2444 of them, then ``ulp(x) = |b - a|``, otherwise ``ulp(x)`` is the
2445 distance between the two non-equal finite floating-point numbers
2446 nearest ``x``. Moreover, ``ulp(NaN)`` is ``NaN``.
2448 The metadata node shall consist of a single positive floating point
2449 number representing the maximum relative error, for example:
2451 .. code-block:: llvm
2453 !0 = metadata !{ float 2.5 } ; maximum acceptable inaccuracy is 2.5 ULPs
2455 '``range``' Metadata
2456 ^^^^^^^^^^^^^^^^^^^^
2458 ``range`` metadata may be attached only to loads of integer types. It
2459 expresses the possible ranges the loaded value is in. The ranges are
2460 represented with a flattened list of integers. The loaded value is known
2461 to be in the union of the ranges defined by each consecutive pair. Each
2462 pair has the following properties:
2464 - The type must match the type loaded by the instruction.
2465 - The pair ``a,b`` represents the range ``[a,b)``.
2466 - Both ``a`` and ``b`` are constants.
2467 - The range is allowed to wrap.
2468 - The range should not represent the full or empty set. That is,
2471 In addition, the pairs must be in signed order of the lower bound and
2472 they must be non-contiguous.
2476 .. code-block:: llvm
2478 %a = load i8* %x, align 1, !range !0 ; Can only be 0 or 1
2479 %b = load i8* %y, align 1, !range !1 ; Can only be 255 (-1), 0 or 1
2480 %c = load i8* %z, align 1, !range !2 ; Can only be 0, 1, 3, 4 or 5
2481 %d = load i8* %z, align 1, !range !3 ; Can only be -2, -1, 3, 4 or 5
2483 !0 = metadata !{ i8 0, i8 2 }
2484 !1 = metadata !{ i8 255, i8 2 }
2485 !2 = metadata !{ i8 0, i8 2, i8 3, i8 6 }
2486 !3 = metadata !{ i8 -2, i8 0, i8 3, i8 6 }
2488 Module Flags Metadata
2489 =====================
2491 Information about the module as a whole is difficult to convey to LLVM's
2492 subsystems. The LLVM IR isn't sufficient to transmit this information.
2493 The ``llvm.module.flags`` named metadata exists in order to facilitate
2494 this. These flags are in the form of key / value pairs --- much like a
2495 dictionary --- making it easy for any subsystem who cares about a flag to
2498 The ``llvm.module.flags`` metadata contains a list of metadata triplets.
2499 Each triplet has the following form:
2501 - The first element is a *behavior* flag, which specifies the behavior
2502 when two (or more) modules are merged together, and it encounters two
2503 (or more) metadata with the same ID. The supported behaviors are
2505 - The second element is a metadata string that is a unique ID for the
2506 metadata. Each module may only have one flag entry for each unique ID (not
2507 including entries with the **Require** behavior).
2508 - The third element is the value of the flag.
2510 When two (or more) modules are merged together, the resulting
2511 ``llvm.module.flags`` metadata is the union of the modules' flags. That is, for
2512 each unique metadata ID string, there will be exactly one entry in the merged
2513 modules ``llvm.module.flags`` metadata table, and the value for that entry will
2514 be determined by the merge behavior flag, as described below. The only exception
2515 is that entries with the *Require* behavior are always preserved.
2517 The following behaviors are supported:
2528 Emits an error if two values disagree, otherwise the resulting value
2529 is that of the operands.
2533 Emits a warning if two values disagree. The result value will be the
2534 operand for the flag from the first module being linked.
2538 Adds a requirement that another module flag be present and have a
2539 specified value after linking is performed. The value must be a
2540 metadata pair, where the first element of the pair is the ID of the
2541 module flag to be restricted, and the second element of the pair is
2542 the value the module flag should be restricted to. This behavior can
2543 be used to restrict the allowable results (via triggering of an
2544 error) of linking IDs with the **Override** behavior.
2548 Uses the specified value, regardless of the behavior or value of the
2549 other module. If both modules specify **Override**, but the values
2550 differ, an error will be emitted.
2554 Appends the two values, which are required to be metadata nodes.
2558 Appends the two values, which are required to be metadata
2559 nodes. However, duplicate entries in the second list are dropped
2560 during the append operation.
2562 It is an error for a particular unique flag ID to have multiple behaviors,
2563 except in the case of **Require** (which adds restrictions on another metadata
2564 value) or **Override**.
2566 An example of module flags:
2568 .. code-block:: llvm
2570 !0 = metadata !{ i32 1, metadata !"foo", i32 1 }
2571 !1 = metadata !{ i32 4, metadata !"bar", i32 37 }
2572 !2 = metadata !{ i32 2, metadata !"qux", i32 42 }
2573 !3 = metadata !{ i32 3, metadata !"qux",
2575 metadata !"foo", i32 1
2578 !llvm.module.flags = !{ !0, !1, !2, !3 }
2580 - Metadata ``!0`` has the ID ``!"foo"`` and the value '1'. The behavior
2581 if two or more ``!"foo"`` flags are seen is to emit an error if their
2582 values are not equal.
2584 - Metadata ``!1`` has the ID ``!"bar"`` and the value '37'. The
2585 behavior if two or more ``!"bar"`` flags are seen is to use the value
2588 - Metadata ``!2`` has the ID ``!"qux"`` and the value '42'. The
2589 behavior if two or more ``!"qux"`` flags are seen is to emit a
2590 warning if their values are not equal.
2592 - Metadata ``!3`` has the ID ``!"qux"`` and the value:
2596 metadata !{ metadata !"foo", i32 1 }
2598 The behavior is to emit an error if the ``llvm.module.flags`` does not
2599 contain a flag with the ID ``!"foo"`` that has the value '1' after linking is
2602 Objective-C Garbage Collection Module Flags Metadata
2603 ----------------------------------------------------
2605 On the Mach-O platform, Objective-C stores metadata about garbage
2606 collection in a special section called "image info". The metadata
2607 consists of a version number and a bitmask specifying what types of
2608 garbage collection are supported (if any) by the file. If two or more
2609 modules are linked together their garbage collection metadata needs to
2610 be merged rather than appended together.
2612 The Objective-C garbage collection module flags metadata consists of the
2613 following key-value pairs:
2622 * - ``Objective-C Version``
2623 - **[Required]** --- The Objective-C ABI version. Valid values are 1 and 2.
2625 * - ``Objective-C Image Info Version``
2626 - **[Required]** --- The version of the image info section. Currently
2629 * - ``Objective-C Image Info Section``
2630 - **[Required]** --- The section to place the metadata. Valid values are
2631 ``"__OBJC, __image_info, regular"`` for Objective-C ABI version 1, and
2632 ``"__DATA,__objc_imageinfo, regular, no_dead_strip"`` for
2633 Objective-C ABI version 2.
2635 * - ``Objective-C Garbage Collection``
2636 - **[Required]** --- Specifies whether garbage collection is supported or
2637 not. Valid values are 0, for no garbage collection, and 2, for garbage
2638 collection supported.
2640 * - ``Objective-C GC Only``
2641 - **[Optional]** --- Specifies that only garbage collection is supported.
2642 If present, its value must be 6. This flag requires that the
2643 ``Objective-C Garbage Collection`` flag have the value 2.
2645 Some important flag interactions:
2647 - If a module with ``Objective-C Garbage Collection`` set to 0 is
2648 merged with a module with ``Objective-C Garbage Collection`` set to
2649 2, then the resulting module has the
2650 ``Objective-C Garbage Collection`` flag set to 0.
2651 - A module with ``Objective-C Garbage Collection`` set to 0 cannot be
2652 merged with a module with ``Objective-C GC Only`` set to 6.
2654 Automatic Linker Flags Module Flags Metadata
2655 --------------------------------------------
2657 Some targets support embedding flags to the linker inside individual object
2658 files. Typically this is used in conjunction with language extensions which
2659 allow source files to explicitly declare the libraries they depend on, and have
2660 these automatically be transmitted to the linker via object files.
2662 These flags are encoded in the IR using metadata in the module flags section,
2663 using the ``Linker Options`` key. The merge behavior for this flag is required
2664 to be ``AppendUnique``, and the value for the key is expected to be a metadata
2665 node which should be a list of other metadata nodes, each of which should be a
2666 list of metadata strings defining linker options.
2668 For example, the following metadata section specifies two separate sets of
2669 linker options, presumably to link against ``libz`` and the ``Cocoa``
2672 !0 = metadata !{ i32 6, metadata !"Linker Options",
2674 metadata !{ metadata !"-lz" },
2675 metadata !{ metadata !"-framework", metadata !"Cocoa" } } }
2676 !llvm.module.flags = !{ !0 }
2678 The metadata encoding as lists of lists of options, as opposed to a collapsed
2679 list of options, is chosen so that the IR encoding can use multiple option
2680 strings to specify e.g., a single library, while still having that specifier be
2681 preserved as an atomic element that can be recognized by a target specific
2682 assembly writer or object file emitter.
2684 Each individual option is required to be either a valid option for the target's
2685 linker, or an option that is reserved by the target specific assembly writer or
2686 object file emitter. No other aspect of these options is defined by the IR.
2688 Intrinsic Global Variables
2689 ==========================
2691 LLVM has a number of "magic" global variables that contain data that
2692 affect code generation or other IR semantics. These are documented here.
2693 All globals of this sort should have a section specified as
2694 "``llvm.metadata``". This section and all globals that start with
2695 "``llvm.``" are reserved for use by LLVM.
2697 The '``llvm.used``' Global Variable
2698 -----------------------------------
2700 The ``@llvm.used`` global is an array with i8\* element type which has
2701 :ref:`appending linkage <linkage_appending>`. This array contains a list of
2702 pointers to global variables and functions which may optionally have a
2703 pointer cast formed of bitcast or getelementptr. For example, a legal
2706 .. code-block:: llvm
2711 @llvm.used = appending global [2 x i8*] [
2713 i8* bitcast (i32* @Y to i8*)
2714 ], section "llvm.metadata"
2716 If a global variable appears in the ``@llvm.used`` list, then the
2717 compiler, assembler, and linker are required to treat the symbol as if
2718 there is a reference to the global that it cannot see. For example, if a
2719 variable has internal linkage and no references other than that from the
2720 ``@llvm.used`` list, it cannot be deleted. This is commonly used to
2721 represent references from inline asms and other things the compiler
2722 cannot "see", and corresponds to "``attribute((used))``" in GNU C.
2724 On some targets, the code generator must emit a directive to the
2725 assembler or object file to prevent the assembler and linker from
2726 molesting the symbol.
2728 The '``llvm.compiler.used``' Global Variable
2729 --------------------------------------------
2731 The ``@llvm.compiler.used`` directive is the same as the ``@llvm.used``
2732 directive, except that it only prevents the compiler from touching the
2733 symbol. On targets that support it, this allows an intelligent linker to
2734 optimize references to the symbol without being impeded as it would be
2737 This is a rare construct that should only be used in rare circumstances,
2738 and should not be exposed to source languages.
2740 The '``llvm.global_ctors``' Global Variable
2741 -------------------------------------------
2743 .. code-block:: llvm
2745 %0 = type { i32, void ()* }
2746 @llvm.global_ctors = appending global [1 x %0] [%0 { i32 65535, void ()* @ctor }]
2748 The ``@llvm.global_ctors`` array contains a list of constructor
2749 functions and associated priorities. The functions referenced by this
2750 array will be called in ascending order of priority (i.e. lowest first)
2751 when the module is loaded. The order of functions with the same priority
2754 The '``llvm.global_dtors``' Global Variable
2755 -------------------------------------------
2757 .. code-block:: llvm
2759 %0 = type { i32, void ()* }
2760 @llvm.global_dtors = appending global [1 x %0] [%0 { i32 65535, void ()* @dtor }]
2762 The ``@llvm.global_dtors`` array contains a list of destructor functions
2763 and associated priorities. The functions referenced by this array will
2764 be called in descending order of priority (i.e. highest first) when the
2765 module is loaded. The order of functions with the same priority is not
2768 Instruction Reference
2769 =====================
2771 The LLVM instruction set consists of several different classifications
2772 of instructions: :ref:`terminator instructions <terminators>`, :ref:`binary
2773 instructions <binaryops>`, :ref:`bitwise binary
2774 instructions <bitwiseops>`, :ref:`memory instructions <memoryops>`, and
2775 :ref:`other instructions <otherops>`.
2779 Terminator Instructions
2780 -----------------------
2782 As mentioned :ref:`previously <functionstructure>`, every basic block in a
2783 program ends with a "Terminator" instruction, which indicates which
2784 block should be executed after the current block is finished. These
2785 terminator instructions typically yield a '``void``' value: they produce
2786 control flow, not values (the one exception being the
2787 ':ref:`invoke <i_invoke>`' instruction).
2789 The terminator instructions are: ':ref:`ret <i_ret>`',
2790 ':ref:`br <i_br>`', ':ref:`switch <i_switch>`',
2791 ':ref:`indirectbr <i_indirectbr>`', ':ref:`invoke <i_invoke>`',
2792 ':ref:`resume <i_resume>`', and ':ref:`unreachable <i_unreachable>`'.
2796 '``ret``' Instruction
2797 ^^^^^^^^^^^^^^^^^^^^^
2804 ret <type> <value> ; Return a value from a non-void function
2805 ret void ; Return from void function
2810 The '``ret``' instruction is used to return control flow (and optionally
2811 a value) from a function back to the caller.
2813 There are two forms of the '``ret``' instruction: one that returns a
2814 value and then causes control flow, and one that just causes control
2820 The '``ret``' instruction optionally accepts a single argument, the
2821 return value. The type of the return value must be a ':ref:`first
2822 class <t_firstclass>`' type.
2824 A function is not :ref:`well formed <wellformed>` if it it has a non-void
2825 return type and contains a '``ret``' instruction with no return value or
2826 a return value with a type that does not match its type, or if it has a
2827 void return type and contains a '``ret``' instruction with a return
2833 When the '``ret``' instruction is executed, control flow returns back to
2834 the calling function's context. If the caller is a
2835 ":ref:`call <i_call>`" instruction, execution continues at the
2836 instruction after the call. If the caller was an
2837 ":ref:`invoke <i_invoke>`" instruction, execution continues at the
2838 beginning of the "normal" destination block. If the instruction returns
2839 a value, that value shall set the call or invoke instruction's return
2845 .. code-block:: llvm
2847 ret i32 5 ; Return an integer value of 5
2848 ret void ; Return from a void function
2849 ret { i32, i8 } { i32 4, i8 2 } ; Return a struct of values 4 and 2
2853 '``br``' Instruction
2854 ^^^^^^^^^^^^^^^^^^^^
2861 br i1 <cond>, label <iftrue>, label <iffalse>
2862 br label <dest> ; Unconditional branch
2867 The '``br``' instruction is used to cause control flow to transfer to a
2868 different basic block in the current function. There are two forms of
2869 this instruction, corresponding to a conditional branch and an
2870 unconditional branch.
2875 The conditional branch form of the '``br``' instruction takes a single
2876 '``i1``' value and two '``label``' values. The unconditional form of the
2877 '``br``' instruction takes a single '``label``' value as a target.
2882 Upon execution of a conditional '``br``' instruction, the '``i1``'
2883 argument is evaluated. If the value is ``true``, control flows to the
2884 '``iftrue``' ``label`` argument. If "cond" is ``false``, control flows
2885 to the '``iffalse``' ``label`` argument.
2890 .. code-block:: llvm
2893 %cond = icmp eq i32 %a, %b
2894 br i1 %cond, label %IfEqual, label %IfUnequal
2902 '``switch``' Instruction
2903 ^^^^^^^^^^^^^^^^^^^^^^^^
2910 switch <intty> <value>, label <defaultdest> [ <intty> <val>, label <dest> ... ]
2915 The '``switch``' instruction is used to transfer control flow to one of
2916 several different places. It is a generalization of the '``br``'
2917 instruction, allowing a branch to occur to one of many possible
2923 The '``switch``' instruction uses three parameters: an integer
2924 comparison value '``value``', a default '``label``' destination, and an
2925 array of pairs of comparison value constants and '``label``'s. The table
2926 is not allowed to contain duplicate constant entries.
2931 The ``switch`` instruction specifies a table of values and destinations.
2932 When the '``switch``' instruction is executed, this table is searched
2933 for the given value. If the value is found, control flow is transferred
2934 to the corresponding destination; otherwise, control flow is transferred
2935 to the default destination.
2940 Depending on properties of the target machine and the particular
2941 ``switch`` instruction, this instruction may be code generated in
2942 different ways. For example, it could be generated as a series of
2943 chained conditional branches or with a lookup table.
2948 .. code-block:: llvm
2950 ; Emulate a conditional br instruction
2951 %Val = zext i1 %value to i32
2952 switch i32 %Val, label %truedest [ i32 0, label %falsedest ]
2954 ; Emulate an unconditional br instruction
2955 switch i32 0, label %dest [ ]
2957 ; Implement a jump table:
2958 switch i32 %val, label %otherwise [ i32 0, label %onzero
2960 i32 2, label %ontwo ]
2964 '``indirectbr``' Instruction
2965 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2972 indirectbr <somety>* <address>, [ label <dest1>, label <dest2>, ... ]
2977 The '``indirectbr``' instruction implements an indirect branch to a
2978 label within the current function, whose address is specified by
2979 "``address``". Address must be derived from a
2980 :ref:`blockaddress <blockaddress>` constant.
2985 The '``address``' argument is the address of the label to jump to. The
2986 rest of the arguments indicate the full set of possible destinations
2987 that the address may point to. Blocks are allowed to occur multiple
2988 times in the destination list, though this isn't particularly useful.
2990 This destination list is required so that dataflow analysis has an
2991 accurate understanding of the CFG.
2996 Control transfers to the block specified in the address argument. All
2997 possible destination blocks must be listed in the label list, otherwise
2998 this instruction has undefined behavior. This implies that jumps to
2999 labels defined in other functions have undefined behavior as well.
3004 This is typically implemented with a jump through a register.
3009 .. code-block:: llvm
3011 indirectbr i8* %Addr, [ label %bb1, label %bb2, label %bb3 ]
3015 '``invoke``' Instruction
3016 ^^^^^^^^^^^^^^^^^^^^^^^^
3023 <result> = invoke [cconv] [ret attrs] <ptr to function ty> <function ptr val>(<function args>) [fn attrs]
3024 to label <normal label> unwind label <exception label>
3029 The '``invoke``' instruction causes control to transfer to a specified
3030 function, with the possibility of control flow transfer to either the
3031 '``normal``' label or the '``exception``' label. If the callee function
3032 returns with the "``ret``" instruction, control flow will return to the
3033 "normal" label. If the callee (or any indirect callees) returns via the
3034 ":ref:`resume <i_resume>`" instruction or other exception handling
3035 mechanism, control is interrupted and continued at the dynamically
3036 nearest "exception" label.
3038 The '``exception``' label is a `landing
3039 pad <ExceptionHandling.html#overview>`_ for the exception. As such,
3040 '``exception``' label is required to have the
3041 ":ref:`landingpad <i_landingpad>`" instruction, which contains the
3042 information about the behavior of the program after unwinding happens,
3043 as its first non-PHI instruction. The restrictions on the
3044 "``landingpad``" instruction's tightly couples it to the "``invoke``"
3045 instruction, so that the important information contained within the
3046 "``landingpad``" instruction can't be lost through normal code motion.
3051 This instruction requires several arguments:
3053 #. The optional "cconv" marker indicates which :ref:`calling
3054 convention <callingconv>` the call should use. If none is
3055 specified, the call defaults to using C calling conventions.
3056 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
3057 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
3059 #. '``ptr to function ty``': shall be the signature of the pointer to
3060 function value being invoked. In most cases, this is a direct
3061 function invocation, but indirect ``invoke``'s are just as possible,
3062 branching off an arbitrary pointer to function value.
3063 #. '``function ptr val``': An LLVM value containing a pointer to a
3064 function to be invoked.
3065 #. '``function args``': argument list whose types match the function
3066 signature argument types and parameter attributes. All arguments must
3067 be of :ref:`first class <t_firstclass>` type. If the function signature
3068 indicates the function accepts a variable number of arguments, the
3069 extra arguments can be specified.
3070 #. '``normal label``': the label reached when the called function
3071 executes a '``ret``' instruction.
3072 #. '``exception label``': the label reached when a callee returns via
3073 the :ref:`resume <i_resume>` instruction or other exception handling
3075 #. The optional :ref:`function attributes <fnattrs>` list. Only
3076 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
3077 attributes are valid here.
3082 This instruction is designed to operate as a standard '``call``'
3083 instruction in most regards. The primary difference is that it
3084 establishes an association with a label, which is used by the runtime
3085 library to unwind the stack.
3087 This instruction is used in languages with destructors to ensure that
3088 proper cleanup is performed in the case of either a ``longjmp`` or a
3089 thrown exception. Additionally, this is important for implementation of
3090 '``catch``' clauses in high-level languages that support them.
3092 For the purposes of the SSA form, the definition of the value returned
3093 by the '``invoke``' instruction is deemed to occur on the edge from the
3094 current block to the "normal" label. If the callee unwinds then no
3095 return value is available.
3100 .. code-block:: llvm
3102 %retval = invoke i32 @Test(i32 15) to label %Continue
3103 unwind label %TestCleanup ; {i32}:retval set
3104 %retval = invoke coldcc i32 %Testfnptr(i32 15) to label %Continue
3105 unwind label %TestCleanup ; {i32}:retval set
3109 '``resume``' Instruction
3110 ^^^^^^^^^^^^^^^^^^^^^^^^
3117 resume <type> <value>
3122 The '``resume``' instruction is a terminator instruction that has no
3128 The '``resume``' instruction requires one argument, which must have the
3129 same type as the result of any '``landingpad``' instruction in the same
3135 The '``resume``' instruction resumes propagation of an existing
3136 (in-flight) exception whose unwinding was interrupted with a
3137 :ref:`landingpad <i_landingpad>` instruction.
3142 .. code-block:: llvm
3144 resume { i8*, i32 } %exn
3148 '``unreachable``' Instruction
3149 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3161 The '``unreachable``' instruction has no defined semantics. This
3162 instruction is used to inform the optimizer that a particular portion of
3163 the code is not reachable. This can be used to indicate that the code
3164 after a no-return function cannot be reached, and other facts.
3169 The '``unreachable``' instruction has no defined semantics.
3176 Binary operators are used to do most of the computation in a program.
3177 They require two operands of the same type, execute an operation on
3178 them, and produce a single value. The operands might represent multiple
3179 data, as is the case with the :ref:`vector <t_vector>` data type. The
3180 result value has the same type as its operands.
3182 There are several different binary operators:
3186 '``add``' Instruction
3187 ^^^^^^^^^^^^^^^^^^^^^
3194 <result> = add <ty> <op1>, <op2> ; yields {ty}:result
3195 <result> = add nuw <ty> <op1>, <op2> ; yields {ty}:result
3196 <result> = add nsw <ty> <op1>, <op2> ; yields {ty}:result
3197 <result> = add nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3202 The '``add``' instruction returns the sum of its two operands.
3207 The two arguments to the '``add``' instruction must be
3208 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3209 arguments must have identical types.
3214 The value produced is the integer sum of the two operands.
3216 If the sum has unsigned overflow, the result returned is the
3217 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
3220 Because LLVM integers use a two's complement representation, this
3221 instruction is appropriate for both signed and unsigned integers.
3223 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3224 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3225 result value of the ``add`` is a :ref:`poison value <poisonvalues>` if
3226 unsigned and/or signed overflow, respectively, occurs.
3231 .. code-block:: llvm
3233 <result> = add i32 4, %var ; yields {i32}:result = 4 + %var
3237 '``fadd``' Instruction
3238 ^^^^^^^^^^^^^^^^^^^^^^
3245 <result> = fadd [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3250 The '``fadd``' instruction returns the sum of its two operands.
3255 The two arguments to the '``fadd``' instruction must be :ref:`floating
3256 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3257 Both arguments must have identical types.
3262 The value produced is the floating point sum of the two operands. This
3263 instruction can also take any number of :ref:`fast-math flags <fastmath>`,
3264 which are optimization hints to enable otherwise unsafe floating point
3270 .. code-block:: llvm
3272 <result> = fadd float 4.0, %var ; yields {float}:result = 4.0 + %var
3274 '``sub``' Instruction
3275 ^^^^^^^^^^^^^^^^^^^^^
3282 <result> = sub <ty> <op1>, <op2> ; yields {ty}:result
3283 <result> = sub nuw <ty> <op1>, <op2> ; yields {ty}:result
3284 <result> = sub nsw <ty> <op1>, <op2> ; yields {ty}:result
3285 <result> = sub nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3290 The '``sub``' instruction returns the difference of its two operands.
3292 Note that the '``sub``' instruction is used to represent the '``neg``'
3293 instruction present in most other intermediate representations.
3298 The two arguments to the '``sub``' instruction must be
3299 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3300 arguments must have identical types.
3305 The value produced is the integer difference of the two operands.
3307 If the difference has unsigned overflow, the result returned is the
3308 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
3311 Because LLVM integers use a two's complement representation, this
3312 instruction is appropriate for both signed and unsigned integers.
3314 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3315 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3316 result value of the ``sub`` is a :ref:`poison value <poisonvalues>` if
3317 unsigned and/or signed overflow, respectively, occurs.
3322 .. code-block:: llvm
3324 <result> = sub i32 4, %var ; yields {i32}:result = 4 - %var
3325 <result> = sub i32 0, %val ; yields {i32}:result = -%var
3329 '``fsub``' Instruction
3330 ^^^^^^^^^^^^^^^^^^^^^^
3337 <result> = fsub [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3342 The '``fsub``' instruction returns the difference of its two operands.
3344 Note that the '``fsub``' instruction is used to represent the '``fneg``'
3345 instruction present in most other intermediate representations.
3350 The two arguments to the '``fsub``' instruction must be :ref:`floating
3351 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3352 Both arguments must have identical types.
3357 The value produced is the floating point difference of the two operands.
3358 This instruction can also take any number of :ref:`fast-math
3359 flags <fastmath>`, which are optimization hints to enable otherwise
3360 unsafe floating point optimizations:
3365 .. code-block:: llvm
3367 <result> = fsub float 4.0, %var ; yields {float}:result = 4.0 - %var
3368 <result> = fsub float -0.0, %val ; yields {float}:result = -%var
3370 '``mul``' Instruction
3371 ^^^^^^^^^^^^^^^^^^^^^
3378 <result> = mul <ty> <op1>, <op2> ; yields {ty}:result
3379 <result> = mul nuw <ty> <op1>, <op2> ; yields {ty}:result
3380 <result> = mul nsw <ty> <op1>, <op2> ; yields {ty}:result
3381 <result> = mul nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3386 The '``mul``' instruction returns the product of its two operands.
3391 The two arguments to the '``mul``' instruction must be
3392 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3393 arguments must have identical types.
3398 The value produced is the integer product of the two operands.
3400 If the result of the multiplication has unsigned overflow, the result
3401 returned is the mathematical result modulo 2\ :sup:`n`\ , where n is the
3402 bit width of the result.
3404 Because LLVM integers use a two's complement representation, and the
3405 result is the same width as the operands, this instruction returns the
3406 correct result for both signed and unsigned integers. If a full product
3407 (e.g. ``i32`` * ``i32`` -> ``i64``) is needed, the operands should be
3408 sign-extended or zero-extended as appropriate to the width of the full
3411 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3412 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3413 result value of the ``mul`` is a :ref:`poison value <poisonvalues>` if
3414 unsigned and/or signed overflow, respectively, occurs.
3419 .. code-block:: llvm
3421 <result> = mul i32 4, %var ; yields {i32}:result = 4 * %var
3425 '``fmul``' Instruction
3426 ^^^^^^^^^^^^^^^^^^^^^^
3433 <result> = fmul [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3438 The '``fmul``' instruction returns the product of its two operands.
3443 The two arguments to the '``fmul``' instruction must be :ref:`floating
3444 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3445 Both arguments must have identical types.
3450 The value produced is the floating point product of the two operands.
3451 This instruction can also take any number of :ref:`fast-math
3452 flags <fastmath>`, which are optimization hints to enable otherwise
3453 unsafe floating point optimizations:
3458 .. code-block:: llvm
3460 <result> = fmul float 4.0, %var ; yields {float}:result = 4.0 * %var
3462 '``udiv``' Instruction
3463 ^^^^^^^^^^^^^^^^^^^^^^
3470 <result> = udiv <ty> <op1>, <op2> ; yields {ty}:result
3471 <result> = udiv exact <ty> <op1>, <op2> ; yields {ty}:result
3476 The '``udiv``' instruction returns the quotient of its two operands.
3481 The two arguments to the '``udiv``' instruction must be
3482 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3483 arguments must have identical types.
3488 The value produced is the unsigned integer quotient of the two operands.
3490 Note that unsigned integer division and signed integer division are
3491 distinct operations; for signed integer division, use '``sdiv``'.
3493 Division by zero leads to undefined behavior.
3495 If the ``exact`` keyword is present, the result value of the ``udiv`` is
3496 a :ref:`poison value <poisonvalues>` if %op1 is not a multiple of %op2 (as
3497 such, "((a udiv exact b) mul b) == a").
3502 .. code-block:: llvm
3504 <result> = udiv i32 4, %var ; yields {i32}:result = 4 / %var
3506 '``sdiv``' Instruction
3507 ^^^^^^^^^^^^^^^^^^^^^^
3514 <result> = sdiv <ty> <op1>, <op2> ; yields {ty}:result
3515 <result> = sdiv exact <ty> <op1>, <op2> ; yields {ty}:result
3520 The '``sdiv``' instruction returns the quotient of its two operands.
3525 The two arguments to the '``sdiv``' instruction must be
3526 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3527 arguments must have identical types.
3532 The value produced is the signed integer quotient of the two operands
3533 rounded towards zero.
3535 Note that signed integer division and unsigned integer division are
3536 distinct operations; for unsigned integer division, use '``udiv``'.
3538 Division by zero leads to undefined behavior. Overflow also leads to
3539 undefined behavior; this is a rare case, but can occur, for example, by
3540 doing a 32-bit division of -2147483648 by -1.
3542 If the ``exact`` keyword is present, the result value of the ``sdiv`` is
3543 a :ref:`poison value <poisonvalues>` if the result would be rounded.
3548 .. code-block:: llvm
3550 <result> = sdiv i32 4, %var ; yields {i32}:result = 4 / %var
3554 '``fdiv``' Instruction
3555 ^^^^^^^^^^^^^^^^^^^^^^
3562 <result> = fdiv [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3567 The '``fdiv``' instruction returns the quotient of its two operands.
3572 The two arguments to the '``fdiv``' instruction must be :ref:`floating
3573 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3574 Both arguments must have identical types.
3579 The value produced is the floating point quotient of the two operands.
3580 This instruction can also take any number of :ref:`fast-math
3581 flags <fastmath>`, which are optimization hints to enable otherwise
3582 unsafe floating point optimizations:
3587 .. code-block:: llvm
3589 <result> = fdiv float 4.0, %var ; yields {float}:result = 4.0 / %var
3591 '``urem``' Instruction
3592 ^^^^^^^^^^^^^^^^^^^^^^
3599 <result> = urem <ty> <op1>, <op2> ; yields {ty}:result
3604 The '``urem``' instruction returns the remainder from the unsigned
3605 division of its two arguments.
3610 The two arguments to the '``urem``' instruction must be
3611 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3612 arguments must have identical types.
3617 This instruction returns the unsigned integer *remainder* of a division.
3618 This instruction always performs an unsigned division to get the
3621 Note that unsigned integer remainder and signed integer remainder are
3622 distinct operations; for signed integer remainder, use '``srem``'.
3624 Taking the remainder of a division by zero leads to undefined behavior.
3629 .. code-block:: llvm
3631 <result> = urem i32 4, %var ; yields {i32}:result = 4 % %var
3633 '``srem``' Instruction
3634 ^^^^^^^^^^^^^^^^^^^^^^
3641 <result> = srem <ty> <op1>, <op2> ; yields {ty}:result
3646 The '``srem``' instruction returns the remainder from the signed
3647 division of its two operands. This instruction can also take
3648 :ref:`vector <t_vector>` versions of the values in which case the elements
3654 The two arguments to the '``srem``' instruction must be
3655 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3656 arguments must have identical types.
3661 This instruction returns the *remainder* of a division (where the result
3662 is either zero or has the same sign as the dividend, ``op1``), not the
3663 *modulo* operator (where the result is either zero or has the same sign
3664 as the divisor, ``op2``) of a value. For more information about the
3665 difference, see `The Math
3666 Forum <http://mathforum.org/dr.math/problems/anne.4.28.99.html>`_. For a
3667 table of how this is implemented in various languages, please see
3669 operation <http://en.wikipedia.org/wiki/Modulo_operation>`_.
3671 Note that signed integer remainder and unsigned integer remainder are
3672 distinct operations; for unsigned integer remainder, use '``urem``'.
3674 Taking the remainder of a division by zero leads to undefined behavior.
3675 Overflow also leads to undefined behavior; this is a rare case, but can
3676 occur, for example, by taking the remainder of a 32-bit division of
3677 -2147483648 by -1. (The remainder doesn't actually overflow, but this
3678 rule lets srem be implemented using instructions that return both the
3679 result of the division and the remainder.)
3684 .. code-block:: llvm
3686 <result> = srem i32 4, %var ; yields {i32}:result = 4 % %var
3690 '``frem``' Instruction
3691 ^^^^^^^^^^^^^^^^^^^^^^
3698 <result> = frem [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3703 The '``frem``' instruction returns the remainder from the division of
3709 The two arguments to the '``frem``' instruction must be :ref:`floating
3710 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3711 Both arguments must have identical types.
3716 This instruction returns the *remainder* of a division. The remainder
3717 has the same sign as the dividend. This instruction can also take any
3718 number of :ref:`fast-math flags <fastmath>`, which are optimization hints
3719 to enable otherwise unsafe floating point optimizations:
3724 .. code-block:: llvm
3726 <result> = frem float 4.0, %var ; yields {float}:result = 4.0 % %var
3730 Bitwise Binary Operations
3731 -------------------------
3733 Bitwise binary operators are used to do various forms of bit-twiddling
3734 in a program. They are generally very efficient instructions and can
3735 commonly be strength reduced from other instructions. They require two
3736 operands of the same type, execute an operation on them, and produce a
3737 single value. The resulting value is the same type as its operands.
3739 '``shl``' Instruction
3740 ^^^^^^^^^^^^^^^^^^^^^
3747 <result> = shl <ty> <op1>, <op2> ; yields {ty}:result
3748 <result> = shl nuw <ty> <op1>, <op2> ; yields {ty}:result
3749 <result> = shl nsw <ty> <op1>, <op2> ; yields {ty}:result
3750 <result> = shl nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3755 The '``shl``' instruction returns the first operand shifted to the left
3756 a specified number of bits.
3761 Both arguments to the '``shl``' instruction must be the same
3762 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
3763 '``op2``' is treated as an unsigned value.
3768 The value produced is ``op1`` \* 2\ :sup:`op2` mod 2\ :sup:`n`,
3769 where ``n`` is the width of the result. If ``op2`` is (statically or
3770 dynamically) negative or equal to or larger than the number of bits in
3771 ``op1``, the result is undefined. If the arguments are vectors, each
3772 vector element of ``op1`` is shifted by the corresponding shift amount
3775 If the ``nuw`` keyword is present, then the shift produces a :ref:`poison
3776 value <poisonvalues>` if it shifts out any non-zero bits. If the
3777 ``nsw`` keyword is present, then the shift produces a :ref:`poison
3778 value <poisonvalues>` if it shifts out any bits that disagree with the
3779 resultant sign bit. As such, NUW/NSW have the same semantics as they
3780 would if the shift were expressed as a mul instruction with the same
3781 nsw/nuw bits in (mul %op1, (shl 1, %op2)).
3786 .. code-block:: llvm
3788 <result> = shl i32 4, %var ; yields {i32}: 4 << %var
3789 <result> = shl i32 4, 2 ; yields {i32}: 16
3790 <result> = shl i32 1, 10 ; yields {i32}: 1024
3791 <result> = shl i32 1, 32 ; undefined
3792 <result> = shl <2 x i32> < i32 1, i32 1>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 2, i32 4>
3794 '``lshr``' Instruction
3795 ^^^^^^^^^^^^^^^^^^^^^^
3802 <result> = lshr <ty> <op1>, <op2> ; yields {ty}:result
3803 <result> = lshr exact <ty> <op1>, <op2> ; yields {ty}:result
3808 The '``lshr``' instruction (logical shift right) returns the first
3809 operand shifted to the right a specified number of bits with zero fill.
3814 Both arguments to the '``lshr``' instruction must be the same
3815 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
3816 '``op2``' is treated as an unsigned value.
3821 This instruction always performs a logical shift right operation. The
3822 most significant bits of the result will be filled with zero bits after
3823 the shift. If ``op2`` is (statically or dynamically) equal to or larger
3824 than the number of bits in ``op1``, the result is undefined. If the
3825 arguments are vectors, each vector element of ``op1`` is shifted by the
3826 corresponding shift amount in ``op2``.
3828 If the ``exact`` keyword is present, the result value of the ``lshr`` is
3829 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
3835 .. code-block:: llvm
3837 <result> = lshr i32 4, 1 ; yields {i32}:result = 2
3838 <result> = lshr i32 4, 2 ; yields {i32}:result = 1
3839 <result> = lshr i8 4, 3 ; yields {i8}:result = 0
3840 <result> = lshr i8 -2, 1 ; yields {i8}:result = 0x7FFFFFFF
3841 <result> = lshr i32 1, 32 ; undefined
3842 <result> = lshr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 0x7FFFFFFF, i32 1>
3844 '``ashr``' Instruction
3845 ^^^^^^^^^^^^^^^^^^^^^^
3852 <result> = ashr <ty> <op1>, <op2> ; yields {ty}:result
3853 <result> = ashr exact <ty> <op1>, <op2> ; yields {ty}:result
3858 The '``ashr``' instruction (arithmetic shift right) returns the first
3859 operand shifted to the right a specified number of bits with sign
3865 Both arguments to the '``ashr``' instruction must be the same
3866 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
3867 '``op2``' is treated as an unsigned value.
3872 This instruction always performs an arithmetic shift right operation,
3873 The most significant bits of the result will be filled with the sign bit
3874 of ``op1``. If ``op2`` is (statically or dynamically) equal to or larger
3875 than the number of bits in ``op1``, the result is undefined. If the
3876 arguments are vectors, each vector element of ``op1`` is shifted by the
3877 corresponding shift amount in ``op2``.
3879 If the ``exact`` keyword is present, the result value of the ``ashr`` is
3880 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
3886 .. code-block:: llvm
3888 <result> = ashr i32 4, 1 ; yields {i32}:result = 2
3889 <result> = ashr i32 4, 2 ; yields {i32}:result = 1
3890 <result> = ashr i8 4, 3 ; yields {i8}:result = 0
3891 <result> = ashr i8 -2, 1 ; yields {i8}:result = -1
3892 <result> = ashr i32 1, 32 ; undefined
3893 <result> = ashr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 3> ; yields: result=<2 x i32> < i32 -1, i32 0>
3895 '``and``' Instruction
3896 ^^^^^^^^^^^^^^^^^^^^^
3903 <result> = and <ty> <op1>, <op2> ; yields {ty}:result
3908 The '``and``' instruction returns the bitwise logical and of its two
3914 The two arguments to the '``and``' instruction must be
3915 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3916 arguments must have identical types.
3921 The truth table used for the '``and``' instruction is:
3938 .. code-block:: llvm
3940 <result> = and i32 4, %var ; yields {i32}:result = 4 & %var
3941 <result> = and i32 15, 40 ; yields {i32}:result = 8
3942 <result> = and i32 4, 8 ; yields {i32}:result = 0
3944 '``or``' Instruction
3945 ^^^^^^^^^^^^^^^^^^^^
3952 <result> = or <ty> <op1>, <op2> ; yields {ty}:result
3957 The '``or``' instruction returns the bitwise logical inclusive or of its
3963 The two arguments to the '``or``' instruction must be
3964 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3965 arguments must have identical types.
3970 The truth table used for the '``or``' instruction is:
3989 <result> = or i32 4, %var ; yields {i32}:result = 4 | %var
3990 <result> = or i32 15, 40 ; yields {i32}:result = 47
3991 <result> = or i32 4, 8 ; yields {i32}:result = 12
3993 '``xor``' Instruction
3994 ^^^^^^^^^^^^^^^^^^^^^
4001 <result> = xor <ty> <op1>, <op2> ; yields {ty}:result
4006 The '``xor``' instruction returns the bitwise logical exclusive or of
4007 its two operands. The ``xor`` is used to implement the "one's
4008 complement" operation, which is the "~" operator in C.
4013 The two arguments to the '``xor``' instruction must be
4014 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4015 arguments must have identical types.
4020 The truth table used for the '``xor``' instruction is:
4037 .. code-block:: llvm
4039 <result> = xor i32 4, %var ; yields {i32}:result = 4 ^ %var
4040 <result> = xor i32 15, 40 ; yields {i32}:result = 39
4041 <result> = xor i32 4, 8 ; yields {i32}:result = 12
4042 <result> = xor i32 %V, -1 ; yields {i32}:result = ~%V
4047 LLVM supports several instructions to represent vector operations in a
4048 target-independent manner. These instructions cover the element-access
4049 and vector-specific operations needed to process vectors effectively.
4050 While LLVM does directly support these vector operations, many
4051 sophisticated algorithms will want to use target-specific intrinsics to
4052 take full advantage of a specific target.
4054 .. _i_extractelement:
4056 '``extractelement``' Instruction
4057 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4064 <result> = extractelement <n x <ty>> <val>, i32 <idx> ; yields <ty>
4069 The '``extractelement``' instruction extracts a single scalar element
4070 from a vector at a specified index.
4075 The first operand of an '``extractelement``' instruction is a value of
4076 :ref:`vector <t_vector>` type. The second operand is an index indicating
4077 the position from which to extract the element. The index may be a
4083 The result is a scalar of the same type as the element type of ``val``.
4084 Its value is the value at position ``idx`` of ``val``. If ``idx``
4085 exceeds the length of ``val``, the results are undefined.
4090 .. code-block:: llvm
4092 <result> = extractelement <4 x i32> %vec, i32 0 ; yields i32
4094 .. _i_insertelement:
4096 '``insertelement``' Instruction
4097 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4104 <result> = insertelement <n x <ty>> <val>, <ty> <elt>, i32 <idx> ; yields <n x <ty>>
4109 The '``insertelement``' instruction inserts a scalar element into a
4110 vector at a specified index.
4115 The first operand of an '``insertelement``' instruction is a value of
4116 :ref:`vector <t_vector>` type. The second operand is a scalar value whose
4117 type must equal the element type of the first operand. The third operand
4118 is an index indicating the position at which to insert the value. The
4119 index may be a variable.
4124 The result is a vector of the same type as ``val``. Its element values
4125 are those of ``val`` except at position ``idx``, where it gets the value
4126 ``elt``. If ``idx`` exceeds the length of ``val``, the results are
4132 .. code-block:: llvm
4134 <result> = insertelement <4 x i32> %vec, i32 1, i32 0 ; yields <4 x i32>
4136 .. _i_shufflevector:
4138 '``shufflevector``' Instruction
4139 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4146 <result> = shufflevector <n x <ty>> <v1>, <n x <ty>> <v2>, <m x i32> <mask> ; yields <m x <ty>>
4151 The '``shufflevector``' instruction constructs a permutation of elements
4152 from two input vectors, returning a vector with the same element type as
4153 the input and length that is the same as the shuffle mask.
4158 The first two operands of a '``shufflevector``' instruction are vectors
4159 with the same type. The third argument is a shuffle mask whose element
4160 type is always 'i32'. The result of the instruction is a vector whose
4161 length is the same as the shuffle mask and whose element type is the
4162 same as the element type of the first two operands.
4164 The shuffle mask operand is required to be a constant vector with either
4165 constant integer or undef values.
4170 The elements of the two input vectors are numbered from left to right
4171 across both of the vectors. The shuffle mask operand specifies, for each
4172 element of the result vector, which element of the two input vectors the
4173 result element gets. The element selector may be undef (meaning "don't
4174 care") and the second operand may be undef if performing a shuffle from
4180 .. code-block:: llvm
4182 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
4183 <4 x i32> <i32 0, i32 4, i32 1, i32 5> ; yields <4 x i32>
4184 <result> = shufflevector <4 x i32> %v1, <4 x i32> undef,
4185 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32> - Identity shuffle.
4186 <result> = shufflevector <8 x i32> %v1, <8 x i32> undef,
4187 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32>
4188 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
4189 <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7 > ; yields <8 x i32>
4191 Aggregate Operations
4192 --------------------
4194 LLVM supports several instructions for working with
4195 :ref:`aggregate <t_aggregate>` values.
4199 '``extractvalue``' Instruction
4200 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4207 <result> = extractvalue <aggregate type> <val>, <idx>{, <idx>}*
4212 The '``extractvalue``' instruction extracts the value of a member field
4213 from an :ref:`aggregate <t_aggregate>` value.
4218 The first operand of an '``extractvalue``' instruction is a value of
4219 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The operands are
4220 constant indices to specify which value to extract in a similar manner
4221 as indices in a '``getelementptr``' instruction.
4223 The major differences to ``getelementptr`` indexing are:
4225 - Since the value being indexed is not a pointer, the first index is
4226 omitted and assumed to be zero.
4227 - At least one index must be specified.
4228 - Not only struct indices but also array indices must be in bounds.
4233 The result is the value at the position in the aggregate specified by
4239 .. code-block:: llvm
4241 <result> = extractvalue {i32, float} %agg, 0 ; yields i32
4245 '``insertvalue``' Instruction
4246 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4253 <result> = insertvalue <aggregate type> <val>, <ty> <elt>, <idx>{, <idx>}* ; yields <aggregate type>
4258 The '``insertvalue``' instruction inserts a value into a member field in
4259 an :ref:`aggregate <t_aggregate>` value.
4264 The first operand of an '``insertvalue``' instruction is a value of
4265 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The second operand is
4266 a first-class value to insert. The following operands are constant
4267 indices indicating the position at which to insert the value in a
4268 similar manner as indices in a '``extractvalue``' instruction. The value
4269 to insert must have the same type as the value identified by the
4275 The result is an aggregate of the same type as ``val``. Its value is
4276 that of ``val`` except that the value at the position specified by the
4277 indices is that of ``elt``.
4282 .. code-block:: llvm
4284 %agg1 = insertvalue {i32, float} undef, i32 1, 0 ; yields {i32 1, float undef}
4285 %agg2 = insertvalue {i32, float} %agg1, float %val, 1 ; yields {i32 1, float %val}
4286 %agg3 = insertvalue {i32, {float}} %agg1, float %val, 1, 0 ; yields {i32 1, float %val}
4290 Memory Access and Addressing Operations
4291 ---------------------------------------
4293 A key design point of an SSA-based representation is how it represents
4294 memory. In LLVM, no memory locations are in SSA form, which makes things
4295 very simple. This section describes how to read, write, and allocate
4300 '``alloca``' Instruction
4301 ^^^^^^^^^^^^^^^^^^^^^^^^
4308 <result> = alloca <type>[, <ty> <NumElements>][, align <alignment>] ; yields {type*}:result
4313 The '``alloca``' instruction allocates memory on the stack frame of the
4314 currently executing function, to be automatically released when this
4315 function returns to its caller. The object is always allocated in the
4316 generic address space (address space zero).
4321 The '``alloca``' instruction allocates ``sizeof(<type>)*NumElements``
4322 bytes of memory on the runtime stack, returning a pointer of the
4323 appropriate type to the program. If "NumElements" is specified, it is
4324 the number of elements allocated, otherwise "NumElements" is defaulted
4325 to be one. If a constant alignment is specified, the value result of the
4326 allocation is guaranteed to be aligned to at least that boundary. If not
4327 specified, or if zero, the target can choose to align the allocation on
4328 any convenient boundary compatible with the type.
4330 '``type``' may be any sized type.
4335 Memory is allocated; a pointer is returned. The operation is undefined
4336 if there is insufficient stack space for the allocation. '``alloca``'d
4337 memory is automatically released when the function returns. The
4338 '``alloca``' instruction is commonly used to represent automatic
4339 variables that must have an address available. When the function returns
4340 (either with the ``ret`` or ``resume`` instructions), the memory is
4341 reclaimed. Allocating zero bytes is legal, but the result is undefined.
4342 The order in which memory is allocated (ie., which way the stack grows)
4348 .. code-block:: llvm
4350 %ptr = alloca i32 ; yields {i32*}:ptr
4351 %ptr = alloca i32, i32 4 ; yields {i32*}:ptr
4352 %ptr = alloca i32, i32 4, align 1024 ; yields {i32*}:ptr
4353 %ptr = alloca i32, align 1024 ; yields {i32*}:ptr
4357 '``load``' Instruction
4358 ^^^^^^^^^^^^^^^^^^^^^^
4365 <result> = load [volatile] <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>][, !invariant.load !<index>]
4366 <result> = load atomic [volatile] <ty>* <pointer> [singlethread] <ordering>, align <alignment>
4367 !<index> = !{ i32 1 }
4372 The '``load``' instruction is used to read from memory.
4377 The argument to the '``load``' instruction specifies the memory address
4378 from which to load. The pointer must point to a :ref:`first
4379 class <t_firstclass>` type. If the ``load`` is marked as ``volatile``,
4380 then the optimizer is not allowed to modify the number or order of
4381 execution of this ``load`` with other :ref:`volatile
4382 operations <volatile>`.
4384 If the ``load`` is marked as ``atomic``, it takes an extra
4385 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
4386 ``release`` and ``acq_rel`` orderings are not valid on ``load``
4387 instructions. Atomic loads produce :ref:`defined <memmodel>` results
4388 when they may see multiple atomic stores. The type of the pointee must
4389 be an integer type whose bit width is a power of two greater than or
4390 equal to eight and less than or equal to a target-specific size limit.
4391 ``align`` must be explicitly specified on atomic loads, and the load has
4392 undefined behavior if the alignment is not set to a value which is at
4393 least the size in bytes of the pointee. ``!nontemporal`` does not have
4394 any defined semantics for atomic loads.
4396 The optional constant ``align`` argument specifies the alignment of the
4397 operation (that is, the alignment of the memory address). A value of 0
4398 or an omitted ``align`` argument means that the operation has the abi
4399 alignment for the target. It is the responsibility of the code emitter
4400 to ensure that the alignment information is correct. Overestimating the
4401 alignment results in undefined behavior. Underestimating the alignment
4402 may produce less efficient code. An alignment of 1 is always safe.
4404 The optional ``!nontemporal`` metadata must reference a single
4405 metatadata name <index> corresponding to a metadata node with one
4406 ``i32`` entry of value 1. The existence of the ``!nontemporal``
4407 metatadata on the instruction tells the optimizer and code generator
4408 that this load is not expected to be reused in the cache. The code
4409 generator may select special instructions to save cache bandwidth, such
4410 as the ``MOVNT`` instruction on x86.
4412 The optional ``!invariant.load`` metadata must reference a single
4413 metatadata name <index> corresponding to a metadata node with no
4414 entries. The existence of the ``!invariant.load`` metatadata on the
4415 instruction tells the optimizer and code generator that this load
4416 address points to memory which does not change value during program
4417 execution. The optimizer may then move this load around, for example, by
4418 hoisting it out of loops using loop invariant code motion.
4423 The location of memory pointed to is loaded. If the value being loaded
4424 is of scalar type then the number of bytes read does not exceed the
4425 minimum number of bytes needed to hold all bits of the type. For
4426 example, loading an ``i24`` reads at most three bytes. When loading a
4427 value of a type like ``i20`` with a size that is not an integral number
4428 of bytes, the result is undefined if the value was not originally
4429 written using a store of the same type.
4434 .. code-block:: llvm
4436 %ptr = alloca i32 ; yields {i32*}:ptr
4437 store i32 3, i32* %ptr ; yields {void}
4438 %val = load i32* %ptr ; yields {i32}:val = i32 3
4442 '``store``' Instruction
4443 ^^^^^^^^^^^^^^^^^^^^^^^
4450 store [volatile] <ty> <value>, <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>] ; yields {void}
4451 store atomic [volatile] <ty> <value>, <ty>* <pointer> [singlethread] <ordering>, align <alignment> ; yields {void}
4456 The '``store``' instruction is used to write to memory.
4461 There are two arguments to the '``store``' instruction: a value to store
4462 and an address at which to store it. The type of the '``<pointer>``'
4463 operand must be a pointer to the :ref:`first class <t_firstclass>` type of
4464 the '``<value>``' operand. If the ``store`` is marked as ``volatile``,
4465 then the optimizer is not allowed to modify the number or order of
4466 execution of this ``store`` with other :ref:`volatile
4467 operations <volatile>`.
4469 If the ``store`` is marked as ``atomic``, it takes an extra
4470 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
4471 ``acquire`` and ``acq_rel`` orderings aren't valid on ``store``
4472 instructions. Atomic loads produce :ref:`defined <memmodel>` results
4473 when they may see multiple atomic stores. The type of the pointee must
4474 be an integer type whose bit width is a power of two greater than or
4475 equal to eight and less than or equal to a target-specific size limit.
4476 ``align`` must be explicitly specified on atomic stores, and the store
4477 has undefined behavior if the alignment is not set to a value which is
4478 at least the size in bytes of the pointee. ``!nontemporal`` does not
4479 have any defined semantics for atomic stores.
4481 The optional constant "align" argument specifies the alignment of the
4482 operation (that is, the alignment of the memory address). A value of 0
4483 or an omitted "align" argument means that the operation has the abi
4484 alignment for the target. It is the responsibility of the code emitter
4485 to ensure that the alignment information is correct. Overestimating the
4486 alignment results in an undefined behavior. Underestimating the
4487 alignment may produce less efficient code. An alignment of 1 is always
4490 The optional !nontemporal metadata must reference a single metatadata
4491 name <index> corresponding to a metadata node with one i32 entry of
4492 value 1. The existence of the !nontemporal metatadata on the instruction
4493 tells the optimizer and code generator that this load is not expected to
4494 be reused in the cache. The code generator may select special
4495 instructions to save cache bandwidth, such as the MOVNT instruction on
4501 The contents of memory are updated to contain '``<value>``' at the
4502 location specified by the '``<pointer>``' operand. If '``<value>``' is
4503 of scalar type then the number of bytes written does not exceed the
4504 minimum number of bytes needed to hold all bits of the type. For
4505 example, storing an ``i24`` writes at most three bytes. When writing a
4506 value of a type like ``i20`` with a size that is not an integral number
4507 of bytes, it is unspecified what happens to the extra bits that do not
4508 belong to the type, but they will typically be overwritten.
4513 .. code-block:: llvm
4515 %ptr = alloca i32 ; yields {i32*}:ptr
4516 store i32 3, i32* %ptr ; yields {void}
4517 %val = load i32* %ptr ; yields {i32}:val = i32 3
4521 '``fence``' Instruction
4522 ^^^^^^^^^^^^^^^^^^^^^^^
4529 fence [singlethread] <ordering> ; yields {void}
4534 The '``fence``' instruction is used to introduce happens-before edges
4540 '``fence``' instructions take an :ref:`ordering <ordering>` argument which
4541 defines what *synchronizes-with* edges they add. They can only be given
4542 ``acquire``, ``release``, ``acq_rel``, and ``seq_cst`` orderings.
4547 A fence A which has (at least) ``release`` ordering semantics
4548 *synchronizes with* a fence B with (at least) ``acquire`` ordering
4549 semantics if and only if there exist atomic operations X and Y, both
4550 operating on some atomic object M, such that A is sequenced before X, X
4551 modifies M (either directly or through some side effect of a sequence
4552 headed by X), Y is sequenced before B, and Y observes M. This provides a
4553 *happens-before* dependency between A and B. Rather than an explicit
4554 ``fence``, one (but not both) of the atomic operations X or Y might
4555 provide a ``release`` or ``acquire`` (resp.) ordering constraint and
4556 still *synchronize-with* the explicit ``fence`` and establish the
4557 *happens-before* edge.
4559 A ``fence`` which has ``seq_cst`` ordering, in addition to having both
4560 ``acquire`` and ``release`` semantics specified above, participates in
4561 the global program order of other ``seq_cst`` operations and/or fences.
4563 The optional ":ref:`singlethread <singlethread>`" argument specifies
4564 that the fence only synchronizes with other fences in the same thread.
4565 (This is useful for interacting with signal handlers.)
4570 .. code-block:: llvm
4572 fence acquire ; yields {void}
4573 fence singlethread seq_cst ; yields {void}
4577 '``cmpxchg``' Instruction
4578 ^^^^^^^^^^^^^^^^^^^^^^^^^
4585 cmpxchg [volatile] <ty>* <pointer>, <ty> <cmp>, <ty> <new> [singlethread] <ordering> ; yields {ty}
4590 The '``cmpxchg``' instruction is used to atomically modify memory. It
4591 loads a value in memory and compares it to a given value. If they are
4592 equal, it stores a new value into the memory.
4597 There are three arguments to the '``cmpxchg``' instruction: an address
4598 to operate on, a value to compare to the value currently be at that
4599 address, and a new value to place at that address if the compared values
4600 are equal. The type of '<cmp>' must be an integer type whose bit width
4601 is a power of two greater than or equal to eight and less than or equal
4602 to a target-specific size limit. '<cmp>' and '<new>' must have the same
4603 type, and the type of '<pointer>' must be a pointer to that type. If the
4604 ``cmpxchg`` is marked as ``volatile``, then the optimizer is not allowed
4605 to modify the number or order of execution of this ``cmpxchg`` with
4606 other :ref:`volatile operations <volatile>`.
4608 The :ref:`ordering <ordering>` argument specifies how this ``cmpxchg``
4609 synchronizes with other atomic operations.
4611 The optional "``singlethread``" argument declares that the ``cmpxchg``
4612 is only atomic with respect to code (usually signal handlers) running in
4613 the same thread as the ``cmpxchg``. Otherwise the cmpxchg is atomic with
4614 respect to all other code in the system.
4616 The pointer passed into cmpxchg must have alignment greater than or
4617 equal to the size in memory of the operand.
4622 The contents of memory at the location specified by the '``<pointer>``'
4623 operand is read and compared to '``<cmp>``'; if the read value is the
4624 equal, '``<new>``' is written. The original value at the location is
4627 A successful ``cmpxchg`` is a read-modify-write instruction for the purpose
4628 of identifying release sequences. A failed ``cmpxchg`` is equivalent to an
4629 atomic load with an ordering parameter determined by dropping any
4630 ``release`` part of the ``cmpxchg``'s ordering.
4635 .. code-block:: llvm
4638 %orig = atomic load i32* %ptr unordered ; yields {i32}
4642 %cmp = phi i32 [ %orig, %entry ], [%old, %loop]
4643 %squared = mul i32 %cmp, %cmp
4644 %old = cmpxchg i32* %ptr, i32 %cmp, i32 %squared ; yields {i32}
4645 %success = icmp eq i32 %cmp, %old
4646 br i1 %success, label %done, label %loop
4653 '``atomicrmw``' Instruction
4654 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
4661 atomicrmw [volatile] <operation> <ty>* <pointer>, <ty> <value> [singlethread] <ordering> ; yields {ty}
4666 The '``atomicrmw``' instruction is used to atomically modify memory.
4671 There are three arguments to the '``atomicrmw``' instruction: an
4672 operation to apply, an address whose value to modify, an argument to the
4673 operation. The operation must be one of the following keywords:
4687 The type of '<value>' must be an integer type whose bit width is a power
4688 of two greater than or equal to eight and less than or equal to a
4689 target-specific size limit. The type of the '``<pointer>``' operand must
4690 be a pointer to that type. If the ``atomicrmw`` is marked as
4691 ``volatile``, then the optimizer is not allowed to modify the number or
4692 order of execution of this ``atomicrmw`` with other :ref:`volatile
4693 operations <volatile>`.
4698 The contents of memory at the location specified by the '``<pointer>``'
4699 operand are atomically read, modified, and written back. The original
4700 value at the location is returned. The modification is specified by the
4703 - xchg: ``*ptr = val``
4704 - add: ``*ptr = *ptr + val``
4705 - sub: ``*ptr = *ptr - val``
4706 - and: ``*ptr = *ptr & val``
4707 - nand: ``*ptr = ~(*ptr & val)``
4708 - or: ``*ptr = *ptr | val``
4709 - xor: ``*ptr = *ptr ^ val``
4710 - max: ``*ptr = *ptr > val ? *ptr : val`` (using a signed comparison)
4711 - min: ``*ptr = *ptr < val ? *ptr : val`` (using a signed comparison)
4712 - umax: ``*ptr = *ptr > val ? *ptr : val`` (using an unsigned
4714 - umin: ``*ptr = *ptr < val ? *ptr : val`` (using an unsigned
4720 .. code-block:: llvm
4722 %old = atomicrmw add i32* %ptr, i32 1 acquire ; yields {i32}
4724 .. _i_getelementptr:
4726 '``getelementptr``' Instruction
4727 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4734 <result> = getelementptr <pty>* <ptrval>{, <ty> <idx>}*
4735 <result> = getelementptr inbounds <pty>* <ptrval>{, <ty> <idx>}*
4736 <result> = getelementptr <ptr vector> ptrval, <vector index type> idx
4741 The '``getelementptr``' instruction is used to get the address of a
4742 subelement of an :ref:`aggregate <t_aggregate>` data structure. It performs
4743 address calculation only and does not access memory.
4748 The first argument is always a pointer or a vector of pointers, and
4749 forms the basis of the calculation. The remaining arguments are indices
4750 that indicate which of the elements of the aggregate object are indexed.
4751 The interpretation of each index is dependent on the type being indexed
4752 into. The first index always indexes the pointer value given as the
4753 first argument, the second index indexes a value of the type pointed to
4754 (not necessarily the value directly pointed to, since the first index
4755 can be non-zero), etc. The first type indexed into must be a pointer
4756 value, subsequent types can be arrays, vectors, and structs. Note that
4757 subsequent types being indexed into can never be pointers, since that
4758 would require loading the pointer before continuing calculation.
4760 The type of each index argument depends on the type it is indexing into.
4761 When indexing into a (optionally packed) structure, only ``i32`` integer
4762 **constants** are allowed (when using a vector of indices they must all
4763 be the **same** ``i32`` integer constant). When indexing into an array,
4764 pointer or vector, integers of any width are allowed, and they are not
4765 required to be constant. These integers are treated as signed values
4768 For example, let's consider a C code fragment and how it gets compiled
4784 int *foo(struct ST *s) {
4785 return &s[1].Z.B[5][13];
4788 The LLVM code generated by Clang is:
4790 .. code-block:: llvm
4792 %struct.RT = type { i8, [10 x [20 x i32]], i8 }
4793 %struct.ST = type { i32, double, %struct.RT }
4795 define i32* @foo(%struct.ST* %s) nounwind uwtable readnone optsize ssp {
4797 %arrayidx = getelementptr inbounds %struct.ST* %s, i64 1, i32 2, i32 1, i64 5, i64 13
4804 In the example above, the first index is indexing into the
4805 '``%struct.ST*``' type, which is a pointer, yielding a '``%struct.ST``'
4806 = '``{ i32, double, %struct.RT }``' type, a structure. The second index
4807 indexes into the third element of the structure, yielding a
4808 '``%struct.RT``' = '``{ i8 , [10 x [20 x i32]], i8 }``' type, another
4809 structure. The third index indexes into the second element of the
4810 structure, yielding a '``[10 x [20 x i32]]``' type, an array. The two
4811 dimensions of the array are subscripted into, yielding an '``i32``'
4812 type. The '``getelementptr``' instruction returns a pointer to this
4813 element, thus computing a value of '``i32*``' type.
4815 Note that it is perfectly legal to index partially through a structure,
4816 returning a pointer to an inner element. Because of this, the LLVM code
4817 for the given testcase is equivalent to:
4819 .. code-block:: llvm
4821 define i32* @foo(%struct.ST* %s) {
4822 %t1 = getelementptr %struct.ST* %s, i32 1 ; yields %struct.ST*:%t1
4823 %t2 = getelementptr %struct.ST* %t1, i32 0, i32 2 ; yields %struct.RT*:%t2
4824 %t3 = getelementptr %struct.RT* %t2, i32 0, i32 1 ; yields [10 x [20 x i32]]*:%t3
4825 %t4 = getelementptr [10 x [20 x i32]]* %t3, i32 0, i32 5 ; yields [20 x i32]*:%t4
4826 %t5 = getelementptr [20 x i32]* %t4, i32 0, i32 13 ; yields i32*:%t5
4830 If the ``inbounds`` keyword is present, the result value of the
4831 ``getelementptr`` is a :ref:`poison value <poisonvalues>` if the base
4832 pointer is not an *in bounds* address of an allocated object, or if any
4833 of the addresses that would be formed by successive addition of the
4834 offsets implied by the indices to the base address with infinitely
4835 precise signed arithmetic are not an *in bounds* address of that
4836 allocated object. The *in bounds* addresses for an allocated object are
4837 all the addresses that point into the object, plus the address one byte
4838 past the end. In cases where the base is a vector of pointers the
4839 ``inbounds`` keyword applies to each of the computations element-wise.
4841 If the ``inbounds`` keyword is not present, the offsets are added to the
4842 base address with silently-wrapping two's complement arithmetic. If the
4843 offsets have a different width from the pointer, they are sign-extended
4844 or truncated to the width of the pointer. The result value of the
4845 ``getelementptr`` may be outside the object pointed to by the base
4846 pointer. The result value may not necessarily be used to access memory
4847 though, even if it happens to point into allocated storage. See the
4848 :ref:`Pointer Aliasing Rules <pointeraliasing>` section for more
4851 The getelementptr instruction is often confusing. For some more insight
4852 into how it works, see :doc:`the getelementptr FAQ <GetElementPtr>`.
4857 .. code-block:: llvm
4859 ; yields [12 x i8]*:aptr
4860 %aptr = getelementptr {i32, [12 x i8]}* %saptr, i64 0, i32 1
4862 %vptr = getelementptr {i32, <2 x i8>}* %svptr, i64 0, i32 1, i32 1
4864 %eptr = getelementptr [12 x i8]* %aptr, i64 0, i32 1
4866 %iptr = getelementptr [10 x i32]* @arr, i16 0, i16 0
4868 In cases where the pointer argument is a vector of pointers, each index
4869 must be a vector with the same number of elements. For example:
4871 .. code-block:: llvm
4873 %A = getelementptr <4 x i8*> %ptrs, <4 x i64> %offsets,
4875 Conversion Operations
4876 ---------------------
4878 The instructions in this category are the conversion instructions
4879 (casting) which all take a single operand and a type. They perform
4880 various bit conversions on the operand.
4882 '``trunc .. to``' Instruction
4883 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4890 <result> = trunc <ty> <value> to <ty2> ; yields ty2
4895 The '``trunc``' instruction truncates its operand to the type ``ty2``.
4900 The '``trunc``' instruction takes a value to trunc, and a type to trunc
4901 it to. Both types must be of :ref:`integer <t_integer>` types, or vectors
4902 of the same number of integers. The bit size of the ``value`` must be
4903 larger than the bit size of the destination type, ``ty2``. Equal sized
4904 types are not allowed.
4909 The '``trunc``' instruction truncates the high order bits in ``value``
4910 and converts the remaining bits to ``ty2``. Since the source size must
4911 be larger than the destination size, ``trunc`` cannot be a *no-op cast*.
4912 It will always truncate bits.
4917 .. code-block:: llvm
4919 %X = trunc i32 257 to i8 ; yields i8:1
4920 %Y = trunc i32 123 to i1 ; yields i1:true
4921 %Z = trunc i32 122 to i1 ; yields i1:false
4922 %W = trunc <2 x i16> <i16 8, i16 7> to <2 x i8> ; yields <i8 8, i8 7>
4924 '``zext .. to``' Instruction
4925 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4932 <result> = zext <ty> <value> to <ty2> ; yields ty2
4937 The '``zext``' instruction zero extends its operand to type ``ty2``.
4942 The '``zext``' instruction takes a value to cast, and a type to cast it
4943 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
4944 the same number of integers. The bit size of the ``value`` must be
4945 smaller than the bit size of the destination type, ``ty2``.
4950 The ``zext`` fills the high order bits of the ``value`` with zero bits
4951 until it reaches the size of the destination type, ``ty2``.
4953 When zero extending from i1, the result will always be either 0 or 1.
4958 .. code-block:: llvm
4960 %X = zext i32 257 to i64 ; yields i64:257
4961 %Y = zext i1 true to i32 ; yields i32:1
4962 %Z = zext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
4964 '``sext .. to``' Instruction
4965 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4972 <result> = sext <ty> <value> to <ty2> ; yields ty2
4977 The '``sext``' sign extends ``value`` to the type ``ty2``.
4982 The '``sext``' instruction takes a value to cast, and a type to cast it
4983 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
4984 the same number of integers. The bit size of the ``value`` must be
4985 smaller than the bit size of the destination type, ``ty2``.
4990 The '``sext``' instruction performs a sign extension by copying the sign
4991 bit (highest order bit) of the ``value`` until it reaches the bit size
4992 of the type ``ty2``.
4994 When sign extending from i1, the extension always results in -1 or 0.
4999 .. code-block:: llvm
5001 %X = sext i8 -1 to i16 ; yields i16 :65535
5002 %Y = sext i1 true to i32 ; yields i32:-1
5003 %Z = sext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
5005 '``fptrunc .. to``' Instruction
5006 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5013 <result> = fptrunc <ty> <value> to <ty2> ; yields ty2
5018 The '``fptrunc``' instruction truncates ``value`` to type ``ty2``.
5023 The '``fptrunc``' instruction takes a :ref:`floating point <t_floating>`
5024 value to cast and a :ref:`floating point <t_floating>` type to cast it to.
5025 The size of ``value`` must be larger than the size of ``ty2``. This
5026 implies that ``fptrunc`` cannot be used to make a *no-op cast*.
5031 The '``fptrunc``' instruction truncates a ``value`` from a larger
5032 :ref:`floating point <t_floating>` type to a smaller :ref:`floating
5033 point <t_floating>` type. If the value cannot fit within the
5034 destination type, ``ty2``, then the results are undefined.
5039 .. code-block:: llvm
5041 %X = fptrunc double 123.0 to float ; yields float:123.0
5042 %Y = fptrunc double 1.0E+300 to float ; yields undefined
5044 '``fpext .. to``' Instruction
5045 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5052 <result> = fpext <ty> <value> to <ty2> ; yields ty2
5057 The '``fpext``' extends a floating point ``value`` to a larger floating
5063 The '``fpext``' instruction takes a :ref:`floating point <t_floating>`
5064 ``value`` to cast, and a :ref:`floating point <t_floating>` type to cast it
5065 to. The source type must be smaller than the destination type.
5070 The '``fpext``' instruction extends the ``value`` from a smaller
5071 :ref:`floating point <t_floating>` type to a larger :ref:`floating
5072 point <t_floating>` type. The ``fpext`` cannot be used to make a
5073 *no-op cast* because it always changes bits. Use ``bitcast`` to make a
5074 *no-op cast* for a floating point cast.
5079 .. code-block:: llvm
5081 %X = fpext float 3.125 to double ; yields double:3.125000e+00
5082 %Y = fpext double %X to fp128 ; yields fp128:0xL00000000000000004000900000000000
5084 '``fptoui .. to``' Instruction
5085 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5092 <result> = fptoui <ty> <value> to <ty2> ; yields ty2
5097 The '``fptoui``' converts a floating point ``value`` to its unsigned
5098 integer equivalent of type ``ty2``.
5103 The '``fptoui``' instruction takes a value to cast, which must be a
5104 scalar or vector :ref:`floating point <t_floating>` value, and a type to
5105 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
5106 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
5107 type with the same number of elements as ``ty``
5112 The '``fptoui``' instruction converts its :ref:`floating
5113 point <t_floating>` operand into the nearest (rounding towards zero)
5114 unsigned integer value. If the value cannot fit in ``ty2``, the results
5120 .. code-block:: llvm
5122 %X = fptoui double 123.0 to i32 ; yields i32:123
5123 %Y = fptoui float 1.0E+300 to i1 ; yields undefined:1
5124 %Z = fptoui float 1.04E+17 to i8 ; yields undefined:1
5126 '``fptosi .. to``' Instruction
5127 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5134 <result> = fptosi <ty> <value> to <ty2> ; yields ty2
5139 The '``fptosi``' instruction converts :ref:`floating point <t_floating>`
5140 ``value`` to type ``ty2``.
5145 The '``fptosi``' instruction takes a value to cast, which must be a
5146 scalar or vector :ref:`floating point <t_floating>` value, and a type to
5147 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
5148 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
5149 type with the same number of elements as ``ty``
5154 The '``fptosi``' instruction converts its :ref:`floating
5155 point <t_floating>` operand into the nearest (rounding towards zero)
5156 signed integer value. If the value cannot fit in ``ty2``, the results
5162 .. code-block:: llvm
5164 %X = fptosi double -123.0 to i32 ; yields i32:-123
5165 %Y = fptosi float 1.0E-247 to i1 ; yields undefined:1
5166 %Z = fptosi float 1.04E+17 to i8 ; yields undefined:1
5168 '``uitofp .. to``' Instruction
5169 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5176 <result> = uitofp <ty> <value> to <ty2> ; yields ty2
5181 The '``uitofp``' instruction regards ``value`` as an unsigned integer
5182 and converts that value to the ``ty2`` type.
5187 The '``uitofp``' instruction takes a value to cast, which must be a
5188 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
5189 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
5190 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
5191 type with the same number of elements as ``ty``
5196 The '``uitofp``' instruction interprets its operand as an unsigned
5197 integer quantity and converts it to the corresponding floating point
5198 value. If the value cannot fit in the floating point value, the results
5204 .. code-block:: llvm
5206 %X = uitofp i32 257 to float ; yields float:257.0
5207 %Y = uitofp i8 -1 to double ; yields double:255.0
5209 '``sitofp .. to``' Instruction
5210 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5217 <result> = sitofp <ty> <value> to <ty2> ; yields ty2
5222 The '``sitofp``' instruction regards ``value`` as a signed integer and
5223 converts that value to the ``ty2`` type.
5228 The '``sitofp``' instruction takes a value to cast, which must be a
5229 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
5230 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
5231 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
5232 type with the same number of elements as ``ty``
5237 The '``sitofp``' instruction interprets its operand as a signed integer
5238 quantity and converts it to the corresponding floating point value. If
5239 the value cannot fit in the floating point value, the results are
5245 .. code-block:: llvm
5247 %X = sitofp i32 257 to float ; yields float:257.0
5248 %Y = sitofp i8 -1 to double ; yields double:-1.0
5252 '``ptrtoint .. to``' Instruction
5253 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5260 <result> = ptrtoint <ty> <value> to <ty2> ; yields ty2
5265 The '``ptrtoint``' instruction converts the pointer or a vector of
5266 pointers ``value`` to the integer (or vector of integers) type ``ty2``.
5271 The '``ptrtoint``' instruction takes a ``value`` to cast, which must be
5272 a a value of type :ref:`pointer <t_pointer>` or a vector of pointers, and a
5273 type to cast it to ``ty2``, which must be an :ref:`integer <t_integer>` or
5274 a vector of integers type.
5279 The '``ptrtoint``' instruction converts ``value`` to integer type
5280 ``ty2`` by interpreting the pointer value as an integer and either
5281 truncating or zero extending that value to the size of the integer type.
5282 If ``value`` is smaller than ``ty2`` then a zero extension is done. If
5283 ``value`` is larger than ``ty2`` then a truncation is done. If they are
5284 the same size, then nothing is done (*no-op cast*) other than a type
5290 .. code-block:: llvm
5292 %X = ptrtoint i32* %P to i8 ; yields truncation on 32-bit architecture
5293 %Y = ptrtoint i32* %P to i64 ; yields zero extension on 32-bit architecture
5294 %Z = ptrtoint <4 x i32*> %P to <4 x i64>; yields vector zero extension for a vector of addresses on 32-bit architecture
5298 '``inttoptr .. to``' Instruction
5299 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5306 <result> = inttoptr <ty> <value> to <ty2> ; yields ty2
5311 The '``inttoptr``' instruction converts an integer ``value`` to a
5312 pointer type, ``ty2``.
5317 The '``inttoptr``' instruction takes an :ref:`integer <t_integer>` value to
5318 cast, and a type to cast it to, which must be a :ref:`pointer <t_pointer>`
5324 The '``inttoptr``' instruction converts ``value`` to type ``ty2`` by
5325 applying either a zero extension or a truncation depending on the size
5326 of the integer ``value``. If ``value`` is larger than the size of a
5327 pointer then a truncation is done. If ``value`` is smaller than the size
5328 of a pointer then a zero extension is done. If they are the same size,
5329 nothing is done (*no-op cast*).
5334 .. code-block:: llvm
5336 %X = inttoptr i32 255 to i32* ; yields zero extension on 64-bit architecture
5337 %Y = inttoptr i32 255 to i32* ; yields no-op on 32-bit architecture
5338 %Z = inttoptr i64 0 to i32* ; yields truncation on 32-bit architecture
5339 %Z = inttoptr <4 x i32> %G to <4 x i8*>; yields truncation of vector G to four pointers
5343 '``bitcast .. to``' Instruction
5344 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5351 <result> = bitcast <ty> <value> to <ty2> ; yields ty2
5356 The '``bitcast``' instruction converts ``value`` to type ``ty2`` without
5362 The '``bitcast``' instruction takes a value to cast, which must be a
5363 non-aggregate first class value, and a type to cast it to, which must
5364 also be a non-aggregate :ref:`first class <t_firstclass>` type. The bit
5365 sizes of ``value`` and the destination type, ``ty2``, must be identical.
5366 If the source type is a pointer, the destination type must also be a
5367 pointer. This instruction supports bitwise conversion of vectors to
5368 integers and to vectors of other types (as long as they have the same
5374 The '``bitcast``' instruction converts ``value`` to type ``ty2``. It is
5375 always a *no-op cast* because no bits change with this conversion. The
5376 conversion is done as if the ``value`` had been stored to memory and
5377 read back as type ``ty2``. Pointer (or vector of pointers) types may
5378 only be converted to other pointer (or vector of pointers) types with
5379 this instruction. To convert pointers to other types, use the
5380 :ref:`inttoptr <i_inttoptr>` or :ref:`ptrtoint <i_ptrtoint>` instructions
5386 .. code-block:: llvm
5388 %X = bitcast i8 255 to i8 ; yields i8 :-1
5389 %Y = bitcast i32* %x to sint* ; yields sint*:%x
5390 %Z = bitcast <2 x int> %V to i64; ; yields i64: %V
5391 %Z = bitcast <2 x i32*> %V to <2 x i64*> ; yields <2 x i64*>
5398 The instructions in this category are the "miscellaneous" instructions,
5399 which defy better classification.
5403 '``icmp``' Instruction
5404 ^^^^^^^^^^^^^^^^^^^^^^
5411 <result> = icmp <cond> <ty> <op1>, <op2> ; yields {i1} or {<N x i1>}:result
5416 The '``icmp``' instruction returns a boolean value or a vector of
5417 boolean values based on comparison of its two integer, integer vector,
5418 pointer, or pointer vector operands.
5423 The '``icmp``' instruction takes three operands. The first operand is
5424 the condition code indicating the kind of comparison to perform. It is
5425 not a value, just a keyword. The possible condition code are:
5428 #. ``ne``: not equal
5429 #. ``ugt``: unsigned greater than
5430 #. ``uge``: unsigned greater or equal
5431 #. ``ult``: unsigned less than
5432 #. ``ule``: unsigned less or equal
5433 #. ``sgt``: signed greater than
5434 #. ``sge``: signed greater or equal
5435 #. ``slt``: signed less than
5436 #. ``sle``: signed less or equal
5438 The remaining two arguments must be :ref:`integer <t_integer>` or
5439 :ref:`pointer <t_pointer>` or integer :ref:`vector <t_vector>` typed. They
5440 must also be identical types.
5445 The '``icmp``' compares ``op1`` and ``op2`` according to the condition
5446 code given as ``cond``. The comparison performed always yields either an
5447 :ref:`i1 <t_integer>` or vector of ``i1`` result, as follows:
5449 #. ``eq``: yields ``true`` if the operands are equal, ``false``
5450 otherwise. No sign interpretation is necessary or performed.
5451 #. ``ne``: yields ``true`` if the operands are unequal, ``false``
5452 otherwise. No sign interpretation is necessary or performed.
5453 #. ``ugt``: interprets the operands as unsigned values and yields
5454 ``true`` if ``op1`` is greater than ``op2``.
5455 #. ``uge``: interprets the operands as unsigned values and yields
5456 ``true`` if ``op1`` is greater than or equal to ``op2``.
5457 #. ``ult``: interprets the operands as unsigned values and yields
5458 ``true`` if ``op1`` is less than ``op2``.
5459 #. ``ule``: interprets the operands as unsigned values and yields
5460 ``true`` if ``op1`` is less than or equal to ``op2``.
5461 #. ``sgt``: interprets the operands as signed values and yields ``true``
5462 if ``op1`` is greater than ``op2``.
5463 #. ``sge``: interprets the operands as signed values and yields ``true``
5464 if ``op1`` is greater than or equal to ``op2``.
5465 #. ``slt``: interprets the operands as signed values and yields ``true``
5466 if ``op1`` is less than ``op2``.
5467 #. ``sle``: interprets the operands as signed values and yields ``true``
5468 if ``op1`` is less than or equal to ``op2``.
5470 If the operands are :ref:`pointer <t_pointer>` typed, the pointer values
5471 are compared as if they were integers.
5473 If the operands are integer vectors, then they are compared element by
5474 element. The result is an ``i1`` vector with the same number of elements
5475 as the values being compared. Otherwise, the result is an ``i1``.
5480 .. code-block:: llvm
5482 <result> = icmp eq i32 4, 5 ; yields: result=false
5483 <result> = icmp ne float* %X, %X ; yields: result=false
5484 <result> = icmp ult i16 4, 5 ; yields: result=true
5485 <result> = icmp sgt i16 4, 5 ; yields: result=false
5486 <result> = icmp ule i16 -4, 5 ; yields: result=false
5487 <result> = icmp sge i16 4, 5 ; yields: result=false
5489 Note that the code generator does not yet support vector types with the
5490 ``icmp`` instruction.
5494 '``fcmp``' Instruction
5495 ^^^^^^^^^^^^^^^^^^^^^^
5502 <result> = fcmp <cond> <ty> <op1>, <op2> ; yields {i1} or {<N x i1>}:result
5507 The '``fcmp``' instruction returns a boolean value or vector of boolean
5508 values based on comparison of its operands.
5510 If the operands are floating point scalars, then the result type is a
5511 boolean (:ref:`i1 <t_integer>`).
5513 If the operands are floating point vectors, then the result type is a
5514 vector of boolean with the same number of elements as the operands being
5520 The '``fcmp``' instruction takes three operands. The first operand is
5521 the condition code indicating the kind of comparison to perform. It is
5522 not a value, just a keyword. The possible condition code are:
5524 #. ``false``: no comparison, always returns false
5525 #. ``oeq``: ordered and equal
5526 #. ``ogt``: ordered and greater than
5527 #. ``oge``: ordered and greater than or equal
5528 #. ``olt``: ordered and less than
5529 #. ``ole``: ordered and less than or equal
5530 #. ``one``: ordered and not equal
5531 #. ``ord``: ordered (no nans)
5532 #. ``ueq``: unordered or equal
5533 #. ``ugt``: unordered or greater than
5534 #. ``uge``: unordered or greater than or equal
5535 #. ``ult``: unordered or less than
5536 #. ``ule``: unordered or less than or equal
5537 #. ``une``: unordered or not equal
5538 #. ``uno``: unordered (either nans)
5539 #. ``true``: no comparison, always returns true
5541 *Ordered* means that neither operand is a QNAN while *unordered* means
5542 that either operand may be a QNAN.
5544 Each of ``val1`` and ``val2`` arguments must be either a :ref:`floating
5545 point <t_floating>` type or a :ref:`vector <t_vector>` of floating point
5546 type. They must have identical types.
5551 The '``fcmp``' instruction compares ``op1`` and ``op2`` according to the
5552 condition code given as ``cond``. If the operands are vectors, then the
5553 vectors are compared element by element. Each comparison performed
5554 always yields an :ref:`i1 <t_integer>` result, as follows:
5556 #. ``false``: always yields ``false``, regardless of operands.
5557 #. ``oeq``: yields ``true`` if both operands are not a QNAN and ``op1``
5558 is equal to ``op2``.
5559 #. ``ogt``: yields ``true`` if both operands are not a QNAN and ``op1``
5560 is greater than ``op2``.
5561 #. ``oge``: yields ``true`` if both operands are not a QNAN and ``op1``
5562 is greater than or equal to ``op2``.
5563 #. ``olt``: yields ``true`` if both operands are not a QNAN and ``op1``
5564 is less than ``op2``.
5565 #. ``ole``: yields ``true`` if both operands are not a QNAN and ``op1``
5566 is less than or equal to ``op2``.
5567 #. ``one``: yields ``true`` if both operands are not a QNAN and ``op1``
5568 is not equal to ``op2``.
5569 #. ``ord``: yields ``true`` if both operands are not a QNAN.
5570 #. ``ueq``: yields ``true`` if either operand is a QNAN or ``op1`` is
5572 #. ``ugt``: yields ``true`` if either operand is a QNAN or ``op1`` is
5573 greater than ``op2``.
5574 #. ``uge``: yields ``true`` if either operand is a QNAN or ``op1`` is
5575 greater than or equal to ``op2``.
5576 #. ``ult``: yields ``true`` if either operand is a QNAN or ``op1`` is
5578 #. ``ule``: yields ``true`` if either operand is a QNAN or ``op1`` is
5579 less than or equal to ``op2``.
5580 #. ``une``: yields ``true`` if either operand is a QNAN or ``op1`` is
5581 not equal to ``op2``.
5582 #. ``uno``: yields ``true`` if either operand is a QNAN.
5583 #. ``true``: always yields ``true``, regardless of operands.
5588 .. code-block:: llvm
5590 <result> = fcmp oeq float 4.0, 5.0 ; yields: result=false
5591 <result> = fcmp one float 4.0, 5.0 ; yields: result=true
5592 <result> = fcmp olt float 4.0, 5.0 ; yields: result=true
5593 <result> = fcmp ueq double 1.0, 2.0 ; yields: result=false
5595 Note that the code generator does not yet support vector types with the
5596 ``fcmp`` instruction.
5600 '``phi``' Instruction
5601 ^^^^^^^^^^^^^^^^^^^^^
5608 <result> = phi <ty> [ <val0>, <label0>], ...
5613 The '``phi``' instruction is used to implement the φ node in the SSA
5614 graph representing the function.
5619 The type of the incoming values is specified with the first type field.
5620 After this, the '``phi``' instruction takes a list of pairs as
5621 arguments, with one pair for each predecessor basic block of the current
5622 block. Only values of :ref:`first class <t_firstclass>` type may be used as
5623 the value arguments to the PHI node. Only labels may be used as the
5626 There must be no non-phi instructions between the start of a basic block
5627 and the PHI instructions: i.e. PHI instructions must be first in a basic
5630 For the purposes of the SSA form, the use of each incoming value is
5631 deemed to occur on the edge from the corresponding predecessor block to
5632 the current block (but after any definition of an '``invoke``'
5633 instruction's return value on the same edge).
5638 At runtime, the '``phi``' instruction logically takes on the value
5639 specified by the pair corresponding to the predecessor basic block that
5640 executed just prior to the current block.
5645 .. code-block:: llvm
5647 Loop: ; Infinite loop that counts from 0 on up...
5648 %indvar = phi i32 [ 0, %LoopHeader ], [ %nextindvar, %Loop ]
5649 %nextindvar = add i32 %indvar, 1
5654 '``select``' Instruction
5655 ^^^^^^^^^^^^^^^^^^^^^^^^
5662 <result> = select selty <cond>, <ty> <val1>, <ty> <val2> ; yields ty
5664 selty is either i1 or {<N x i1>}
5669 The '``select``' instruction is used to choose one value based on a
5670 condition, without branching.
5675 The '``select``' instruction requires an 'i1' value or a vector of 'i1'
5676 values indicating the condition, and two values of the same :ref:`first
5677 class <t_firstclass>` type. If the val1/val2 are vectors and the
5678 condition is a scalar, then entire vectors are selected, not individual
5684 If the condition is an i1 and it evaluates to 1, the instruction returns
5685 the first value argument; otherwise, it returns the second value
5688 If the condition is a vector of i1, then the value arguments must be
5689 vectors of the same size, and the selection is done element by element.
5694 .. code-block:: llvm
5696 %X = select i1 true, i8 17, i8 42 ; yields i8:17
5700 '``call``' Instruction
5701 ^^^^^^^^^^^^^^^^^^^^^^
5708 <result> = [tail] call [cconv] [ret attrs] <ty> [<fnty>*] <fnptrval>(<function args>) [fn attrs]
5713 The '``call``' instruction represents a simple function call.
5718 This instruction requires several arguments:
5720 #. The optional "tail" marker indicates that the callee function does
5721 not access any allocas or varargs in the caller. Note that calls may
5722 be marked "tail" even if they do not occur before a
5723 :ref:`ret <i_ret>` instruction. If the "tail" marker is present, the
5724 function call is eligible for tail call optimization, but `might not
5725 in fact be optimized into a jump <CodeGenerator.html#tailcallopt>`_.
5726 The code generator may optimize calls marked "tail" with either 1)
5727 automatic `sibling call
5728 optimization <CodeGenerator.html#sibcallopt>`_ when the caller and
5729 callee have matching signatures, or 2) forced tail call optimization
5730 when the following extra requirements are met:
5732 - Caller and callee both have the calling convention ``fastcc``.
5733 - The call is in tail position (ret immediately follows call and ret
5734 uses value of call or is void).
5735 - Option ``-tailcallopt`` is enabled, or
5736 ``llvm::GuaranteedTailCallOpt`` is ``true``.
5737 - `Platform specific constraints are
5738 met. <CodeGenerator.html#tailcallopt>`_
5740 #. The optional "cconv" marker indicates which :ref:`calling
5741 convention <callingconv>` the call should use. If none is
5742 specified, the call defaults to using C calling conventions. The
5743 calling convention of the call must match the calling convention of
5744 the target function, or else the behavior is undefined.
5745 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
5746 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
5748 #. '``ty``': the type of the call instruction itself which is also the
5749 type of the return value. Functions that return no value are marked
5751 #. '``fnty``': shall be the signature of the pointer to function value
5752 being invoked. The argument types must match the types implied by
5753 this signature. This type can be omitted if the function is not
5754 varargs and if the function type does not return a pointer to a
5756 #. '``fnptrval``': An LLVM value containing a pointer to a function to
5757 be invoked. In most cases, this is a direct function invocation, but
5758 indirect ``call``'s are just as possible, calling an arbitrary pointer
5760 #. '``function args``': argument list whose types match the function
5761 signature argument types and parameter attributes. All arguments must
5762 be of :ref:`first class <t_firstclass>` type. If the function signature
5763 indicates the function accepts a variable number of arguments, the
5764 extra arguments can be specified.
5765 #. The optional :ref:`function attributes <fnattrs>` list. Only
5766 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
5767 attributes are valid here.
5772 The '``call``' instruction is used to cause control flow to transfer to
5773 a specified function, with its incoming arguments bound to the specified
5774 values. Upon a '``ret``' instruction in the called function, control
5775 flow continues with the instruction after the function call, and the
5776 return value of the function is bound to the result argument.
5781 .. code-block:: llvm
5783 %retval = call i32 @test(i32 %argc)
5784 call i32 (i8*, ...)* @printf(i8* %msg, i32 12, i8 42) ; yields i32
5785 %X = tail call i32 @foo() ; yields i32
5786 %Y = tail call fastcc i32 @foo() ; yields i32
5787 call void %foo(i8 97 signext)
5789 %struct.A = type { i32, i8 }
5790 %r = call %struct.A @foo() ; yields { 32, i8 }
5791 %gr = extractvalue %struct.A %r, 0 ; yields i32
5792 %gr1 = extractvalue %struct.A %r, 1 ; yields i8
5793 %Z = call void @foo() noreturn ; indicates that %foo never returns normally
5794 %ZZ = call zeroext i32 @bar() ; Return value is %zero extended
5796 llvm treats calls to some functions with names and arguments that match
5797 the standard C99 library as being the C99 library functions, and may
5798 perform optimizations or generate code for them under that assumption.
5799 This is something we'd like to change in the future to provide better
5800 support for freestanding environments and non-C-based languages.
5804 '``va_arg``' Instruction
5805 ^^^^^^^^^^^^^^^^^^^^^^^^
5812 <resultval> = va_arg <va_list*> <arglist>, <argty>
5817 The '``va_arg``' instruction is used to access arguments passed through
5818 the "variable argument" area of a function call. It is used to implement
5819 the ``va_arg`` macro in C.
5824 This instruction takes a ``va_list*`` value and the type of the
5825 argument. It returns a value of the specified argument type and
5826 increments the ``va_list`` to point to the next argument. The actual
5827 type of ``va_list`` is target specific.
5832 The '``va_arg``' instruction loads an argument of the specified type
5833 from the specified ``va_list`` and causes the ``va_list`` to point to
5834 the next argument. For more information, see the variable argument
5835 handling :ref:`Intrinsic Functions <int_varargs>`.
5837 It is legal for this instruction to be called in a function which does
5838 not take a variable number of arguments, for example, the ``vfprintf``
5841 ``va_arg`` is an LLVM instruction instead of an :ref:`intrinsic
5842 function <intrinsics>` because it takes a type as an argument.
5847 See the :ref:`variable argument processing <int_varargs>` section.
5849 Note that the code generator does not yet fully support va\_arg on many
5850 targets. Also, it does not currently support va\_arg with aggregate
5851 types on any target.
5855 '``landingpad``' Instruction
5856 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5863 <resultval> = landingpad <resultty> personality <type> <pers_fn> <clause>+
5864 <resultval> = landingpad <resultty> personality <type> <pers_fn> cleanup <clause>*
5866 <clause> := catch <type> <value>
5867 <clause> := filter <array constant type> <array constant>
5872 The '``landingpad``' instruction is used by `LLVM's exception handling
5873 system <ExceptionHandling.html#overview>`_ to specify that a basic block
5874 is a landing pad --- one where the exception lands, and corresponds to the
5875 code found in the ``catch`` portion of a ``try``/``catch`` sequence. It
5876 defines values supplied by the personality function (``pers_fn``) upon
5877 re-entry to the function. The ``resultval`` has the type ``resultty``.
5882 This instruction takes a ``pers_fn`` value. This is the personality
5883 function associated with the unwinding mechanism. The optional
5884 ``cleanup`` flag indicates that the landing pad block is a cleanup.
5886 A ``clause`` begins with the clause type --- ``catch`` or ``filter`` --- and
5887 contains the global variable representing the "type" that may be caught
5888 or filtered respectively. Unlike the ``catch`` clause, the ``filter``
5889 clause takes an array constant as its argument. Use
5890 "``[0 x i8**] undef``" for a filter which cannot throw. The
5891 '``landingpad``' instruction must contain *at least* one ``clause`` or
5892 the ``cleanup`` flag.
5897 The '``landingpad``' instruction defines the values which are set by the
5898 personality function (``pers_fn``) upon re-entry to the function, and
5899 therefore the "result type" of the ``landingpad`` instruction. As with
5900 calling conventions, how the personality function results are
5901 represented in LLVM IR is target specific.
5903 The clauses are applied in order from top to bottom. If two
5904 ``landingpad`` instructions are merged together through inlining, the
5905 clauses from the calling function are appended to the list of clauses.
5906 When the call stack is being unwound due to an exception being thrown,
5907 the exception is compared against each ``clause`` in turn. If it doesn't
5908 match any of the clauses, and the ``cleanup`` flag is not set, then
5909 unwinding continues further up the call stack.
5911 The ``landingpad`` instruction has several restrictions:
5913 - A landing pad block is a basic block which is the unwind destination
5914 of an '``invoke``' instruction.
5915 - A landing pad block must have a '``landingpad``' instruction as its
5916 first non-PHI instruction.
5917 - There can be only one '``landingpad``' instruction within the landing
5919 - A basic block that is not a landing pad block may not include a
5920 '``landingpad``' instruction.
5921 - All '``landingpad``' instructions in a function must have the same
5922 personality function.
5927 .. code-block:: llvm
5929 ;; A landing pad which can catch an integer.
5930 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
5932 ;; A landing pad that is a cleanup.
5933 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
5935 ;; A landing pad which can catch an integer and can only throw a double.
5936 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
5938 filter [1 x i8**] [@_ZTId]
5945 LLVM supports the notion of an "intrinsic function". These functions
5946 have well known names and semantics and are required to follow certain
5947 restrictions. Overall, these intrinsics represent an extension mechanism
5948 for the LLVM language that does not require changing all of the
5949 transformations in LLVM when adding to the language (or the bitcode
5950 reader/writer, the parser, etc...).
5952 Intrinsic function names must all start with an "``llvm.``" prefix. This
5953 prefix is reserved in LLVM for intrinsic names; thus, function names may
5954 not begin with this prefix. Intrinsic functions must always be external
5955 functions: you cannot define the body of intrinsic functions. Intrinsic
5956 functions may only be used in call or invoke instructions: it is illegal
5957 to take the address of an intrinsic function. Additionally, because
5958 intrinsic functions are part of the LLVM language, it is required if any
5959 are added that they be documented here.
5961 Some intrinsic functions can be overloaded, i.e., the intrinsic
5962 represents a family of functions that perform the same operation but on
5963 different data types. Because LLVM can represent over 8 million
5964 different integer types, overloading is used commonly to allow an
5965 intrinsic function to operate on any integer type. One or more of the
5966 argument types or the result type can be overloaded to accept any
5967 integer type. Argument types may also be defined as exactly matching a
5968 previous argument's type or the result type. This allows an intrinsic
5969 function which accepts multiple arguments, but needs all of them to be
5970 of the same type, to only be overloaded with respect to a single
5971 argument or the result.
5973 Overloaded intrinsics will have the names of its overloaded argument
5974 types encoded into its function name, each preceded by a period. Only
5975 those types which are overloaded result in a name suffix. Arguments
5976 whose type is matched against another type do not. For example, the
5977 ``llvm.ctpop`` function can take an integer of any width and returns an
5978 integer of exactly the same integer width. This leads to a family of
5979 functions such as ``i8 @llvm.ctpop.i8(i8 %val)`` and
5980 ``i29 @llvm.ctpop.i29(i29 %val)``. Only one type, the return type, is
5981 overloaded, and only one type suffix is required. Because the argument's
5982 type is matched against the return type, it does not require its own
5985 To learn how to add an intrinsic function, please see the `Extending
5986 LLVM Guide <ExtendingLLVM.html>`_.
5990 Variable Argument Handling Intrinsics
5991 -------------------------------------
5993 Variable argument support is defined in LLVM with the
5994 :ref:`va_arg <i_va_arg>` instruction and these three intrinsic
5995 functions. These functions are related to the similarly named macros
5996 defined in the ``<stdarg.h>`` header file.
5998 All of these functions operate on arguments that use a target-specific
5999 value type "``va_list``". The LLVM assembly language reference manual
6000 does not define what this type is, so all transformations should be
6001 prepared to handle these functions regardless of the type used.
6003 This example shows how the :ref:`va_arg <i_va_arg>` instruction and the
6004 variable argument handling intrinsic functions are used.
6006 .. code-block:: llvm
6008 define i32 @test(i32 %X, ...) {
6009 ; Initialize variable argument processing
6011 %ap2 = bitcast i8** %ap to i8*
6012 call void @llvm.va_start(i8* %ap2)
6014 ; Read a single integer argument
6015 %tmp = va_arg i8** %ap, i32
6017 ; Demonstrate usage of llvm.va_copy and llvm.va_end
6019 %aq2 = bitcast i8** %aq to i8*
6020 call void @llvm.va_copy(i8* %aq2, i8* %ap2)
6021 call void @llvm.va_end(i8* %aq2)
6023 ; Stop processing of arguments.
6024 call void @llvm.va_end(i8* %ap2)
6028 declare void @llvm.va_start(i8*)
6029 declare void @llvm.va_copy(i8*, i8*)
6030 declare void @llvm.va_end(i8*)
6034 '``llvm.va_start``' Intrinsic
6035 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6042 declare void %llvm.va_start(i8* <arglist>)
6047 The '``llvm.va_start``' intrinsic initializes ``*<arglist>`` for
6048 subsequent use by ``va_arg``.
6053 The argument is a pointer to a ``va_list`` element to initialize.
6058 The '``llvm.va_start``' intrinsic works just like the ``va_start`` macro
6059 available in C. In a target-dependent way, it initializes the
6060 ``va_list`` element to which the argument points, so that the next call
6061 to ``va_arg`` will produce the first variable argument passed to the
6062 function. Unlike the C ``va_start`` macro, this intrinsic does not need
6063 to know the last argument of the function as the compiler can figure
6066 '``llvm.va_end``' Intrinsic
6067 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6074 declare void @llvm.va_end(i8* <arglist>)
6079 The '``llvm.va_end``' intrinsic destroys ``*<arglist>``, which has been
6080 initialized previously with ``llvm.va_start`` or ``llvm.va_copy``.
6085 The argument is a pointer to a ``va_list`` to destroy.
6090 The '``llvm.va_end``' intrinsic works just like the ``va_end`` macro
6091 available in C. In a target-dependent way, it destroys the ``va_list``
6092 element to which the argument points. Calls to
6093 :ref:`llvm.va_start <int_va_start>` and
6094 :ref:`llvm.va_copy <int_va_copy>` must be matched exactly with calls to
6099 '``llvm.va_copy``' Intrinsic
6100 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6107 declare void @llvm.va_copy(i8* <destarglist>, i8* <srcarglist>)
6112 The '``llvm.va_copy``' intrinsic copies the current argument position
6113 from the source argument list to the destination argument list.
6118 The first argument is a pointer to a ``va_list`` element to initialize.
6119 The second argument is a pointer to a ``va_list`` element to copy from.
6124 The '``llvm.va_copy``' intrinsic works just like the ``va_copy`` macro
6125 available in C. In a target-dependent way, it copies the source
6126 ``va_list`` element into the destination ``va_list`` element. This
6127 intrinsic is necessary because the `` llvm.va_start`` intrinsic may be
6128 arbitrarily complex and require, for example, memory allocation.
6130 Accurate Garbage Collection Intrinsics
6131 --------------------------------------
6133 LLVM support for `Accurate Garbage Collection <GarbageCollection.html>`_
6134 (GC) requires the implementation and generation of these intrinsics.
6135 These intrinsics allow identification of :ref:`GC roots on the
6136 stack <int_gcroot>`, as well as garbage collector implementations that
6137 require :ref:`read <int_gcread>` and :ref:`write <int_gcwrite>` barriers.
6138 Front-ends for type-safe garbage collected languages should generate
6139 these intrinsics to make use of the LLVM garbage collectors. For more
6140 details, see `Accurate Garbage Collection with
6141 LLVM <GarbageCollection.html>`_.
6143 The garbage collection intrinsics only operate on objects in the generic
6144 address space (address space zero).
6148 '``llvm.gcroot``' Intrinsic
6149 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6156 declare void @llvm.gcroot(i8** %ptrloc, i8* %metadata)
6161 The '``llvm.gcroot``' intrinsic declares the existence of a GC root to
6162 the code generator, and allows some metadata to be associated with it.
6167 The first argument specifies the address of a stack object that contains
6168 the root pointer. The second pointer (which must be either a constant or
6169 a global value address) contains the meta-data to be associated with the
6175 At runtime, a call to this intrinsic stores a null pointer into the
6176 "ptrloc" location. At compile-time, the code generator generates
6177 information to allow the runtime to find the pointer at GC safe points.
6178 The '``llvm.gcroot``' intrinsic may only be used in a function which
6179 :ref:`specifies a GC algorithm <gc>`.
6183 '``llvm.gcread``' Intrinsic
6184 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6191 declare i8* @llvm.gcread(i8* %ObjPtr, i8** %Ptr)
6196 The '``llvm.gcread``' intrinsic identifies reads of references from heap
6197 locations, allowing garbage collector implementations that require read
6203 The second argument is the address to read from, which should be an
6204 address allocated from the garbage collector. The first object is a
6205 pointer to the start of the referenced object, if needed by the language
6206 runtime (otherwise null).
6211 The '``llvm.gcread``' intrinsic has the same semantics as a load
6212 instruction, but may be replaced with substantially more complex code by
6213 the garbage collector runtime, as needed. The '``llvm.gcread``'
6214 intrinsic may only be used in a function which :ref:`specifies a GC
6219 '``llvm.gcwrite``' Intrinsic
6220 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6227 declare void @llvm.gcwrite(i8* %P1, i8* %Obj, i8** %P2)
6232 The '``llvm.gcwrite``' intrinsic identifies writes of references to heap
6233 locations, allowing garbage collector implementations that require write
6234 barriers (such as generational or reference counting collectors).
6239 The first argument is the reference to store, the second is the start of
6240 the object to store it to, and the third is the address of the field of
6241 Obj to store to. If the runtime does not require a pointer to the
6242 object, Obj may be null.
6247 The '``llvm.gcwrite``' intrinsic has the same semantics as a store
6248 instruction, but may be replaced with substantially more complex code by
6249 the garbage collector runtime, as needed. The '``llvm.gcwrite``'
6250 intrinsic may only be used in a function which :ref:`specifies a GC
6253 Code Generator Intrinsics
6254 -------------------------
6256 These intrinsics are provided by LLVM to expose special features that
6257 may only be implemented with code generator support.
6259 '``llvm.returnaddress``' Intrinsic
6260 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6267 declare i8 *@llvm.returnaddress(i32 <level>)
6272 The '``llvm.returnaddress``' intrinsic attempts to compute a
6273 target-specific value indicating the return address of the current
6274 function or one of its callers.
6279 The argument to this intrinsic indicates which function to return the
6280 address for. Zero indicates the calling function, one indicates its
6281 caller, etc. The argument is **required** to be a constant integer
6287 The '``llvm.returnaddress``' intrinsic either returns a pointer
6288 indicating the return address of the specified call frame, or zero if it
6289 cannot be identified. The value returned by this intrinsic is likely to
6290 be incorrect or 0 for arguments other than zero, so it should only be
6291 used for debugging purposes.
6293 Note that calling this intrinsic does not prevent function inlining or
6294 other aggressive transformations, so the value returned may not be that
6295 of the obvious source-language caller.
6297 '``llvm.frameaddress``' Intrinsic
6298 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6305 declare i8* @llvm.frameaddress(i32 <level>)
6310 The '``llvm.frameaddress``' intrinsic attempts to return the
6311 target-specific frame pointer value for the specified stack frame.
6316 The argument to this intrinsic indicates which function to return the
6317 frame pointer for. Zero indicates the calling function, one indicates
6318 its caller, etc. The argument is **required** to be a constant integer
6324 The '``llvm.frameaddress``' intrinsic either returns a pointer
6325 indicating the frame address of the specified call frame, or zero if it
6326 cannot be identified. The value returned by this intrinsic is likely to
6327 be incorrect or 0 for arguments other than zero, so it should only be
6328 used for debugging purposes.
6330 Note that calling this intrinsic does not prevent function inlining or
6331 other aggressive transformations, so the value returned may not be that
6332 of the obvious source-language caller.
6336 '``llvm.stacksave``' Intrinsic
6337 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6344 declare i8* @llvm.stacksave()
6349 The '``llvm.stacksave``' intrinsic is used to remember the current state
6350 of the function stack, for use with
6351 :ref:`llvm.stackrestore <int_stackrestore>`. This is useful for
6352 implementing language features like scoped automatic variable sized
6358 This intrinsic returns a opaque pointer value that can be passed to
6359 :ref:`llvm.stackrestore <int_stackrestore>`. When an
6360 ``llvm.stackrestore`` intrinsic is executed with a value saved from
6361 ``llvm.stacksave``, it effectively restores the state of the stack to
6362 the state it was in when the ``llvm.stacksave`` intrinsic executed. In
6363 practice, this pops any :ref:`alloca <i_alloca>` blocks from the stack that
6364 were allocated after the ``llvm.stacksave`` was executed.
6366 .. _int_stackrestore:
6368 '``llvm.stackrestore``' Intrinsic
6369 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6376 declare void @llvm.stackrestore(i8* %ptr)
6381 The '``llvm.stackrestore``' intrinsic is used to restore the state of
6382 the function stack to the state it was in when the corresponding
6383 :ref:`llvm.stacksave <int_stacksave>` intrinsic executed. This is
6384 useful for implementing language features like scoped automatic variable
6385 sized arrays in C99.
6390 See the description for :ref:`llvm.stacksave <int_stacksave>`.
6392 '``llvm.prefetch``' Intrinsic
6393 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6400 declare void @llvm.prefetch(i8* <address>, i32 <rw>, i32 <locality>, i32 <cache type>)
6405 The '``llvm.prefetch``' intrinsic is a hint to the code generator to
6406 insert a prefetch instruction if supported; otherwise, it is a noop.
6407 Prefetches have no effect on the behavior of the program but can change
6408 its performance characteristics.
6413 ``address`` is the address to be prefetched, ``rw`` is the specifier
6414 determining if the fetch should be for a read (0) or write (1), and
6415 ``locality`` is a temporal locality specifier ranging from (0) - no
6416 locality, to (3) - extremely local keep in cache. The ``cache type``
6417 specifies whether the prefetch is performed on the data (1) or
6418 instruction (0) cache. The ``rw``, ``locality`` and ``cache type``
6419 arguments must be constant integers.
6424 This intrinsic does not modify the behavior of the program. In
6425 particular, prefetches cannot trap and do not produce a value. On
6426 targets that support this intrinsic, the prefetch can provide hints to
6427 the processor cache for better performance.
6429 '``llvm.pcmarker``' Intrinsic
6430 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6437 declare void @llvm.pcmarker(i32 <id>)
6442 The '``llvm.pcmarker``' intrinsic is a method to export a Program
6443 Counter (PC) in a region of code to simulators and other tools. The
6444 method is target specific, but it is expected that the marker will use
6445 exported symbols to transmit the PC of the marker. The marker makes no
6446 guarantees that it will remain with any specific instruction after
6447 optimizations. It is possible that the presence of a marker will inhibit
6448 optimizations. The intended use is to be inserted after optimizations to
6449 allow correlations of simulation runs.
6454 ``id`` is a numerical id identifying the marker.
6459 This intrinsic does not modify the behavior of the program. Backends
6460 that do not support this intrinsic may ignore it.
6462 '``llvm.readcyclecounter``' Intrinsic
6463 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6470 declare i64 @llvm.readcyclecounter()
6475 The '``llvm.readcyclecounter``' intrinsic provides access to the cycle
6476 counter register (or similar low latency, high accuracy clocks) on those
6477 targets that support it. On X86, it should map to RDTSC. On Alpha, it
6478 should map to RPCC. As the backing counters overflow quickly (on the
6479 order of 9 seconds on alpha), this should only be used for small
6485 When directly supported, reading the cycle counter should not modify any
6486 memory. Implementations are allowed to either return a application
6487 specific value or a system wide value. On backends without support, this
6488 is lowered to a constant 0.
6490 Standard C Library Intrinsics
6491 -----------------------------
6493 LLVM provides intrinsics for a few important standard C library
6494 functions. These intrinsics allow source-language front-ends to pass
6495 information about the alignment of the pointer arguments to the code
6496 generator, providing opportunity for more efficient code generation.
6500 '``llvm.memcpy``' Intrinsic
6501 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6506 This is an overloaded intrinsic. You can use ``llvm.memcpy`` on any
6507 integer bit width and for different address spaces. Not all targets
6508 support all bit widths however.
6512 declare void @llvm.memcpy.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
6513 i32 <len>, i32 <align>, i1 <isvolatile>)
6514 declare void @llvm.memcpy.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
6515 i64 <len>, i32 <align>, i1 <isvolatile>)
6520 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
6521 source location to the destination location.
6523 Note that, unlike the standard libc function, the ``llvm.memcpy.*``
6524 intrinsics do not return a value, takes extra alignment/isvolatile
6525 arguments and the pointers can be in specified address spaces.
6530 The first argument is a pointer to the destination, the second is a
6531 pointer to the source. The third argument is an integer argument
6532 specifying the number of bytes to copy, the fourth argument is the
6533 alignment of the source and destination locations, and the fifth is a
6534 boolean indicating a volatile access.
6536 If the call to this intrinsic has an alignment value that is not 0 or 1,
6537 then the caller guarantees that both the source and destination pointers
6538 are aligned to that boundary.
6540 If the ``isvolatile`` parameter is ``true``, the ``llvm.memcpy`` call is
6541 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
6542 very cleanly specified and it is unwise to depend on it.
6547 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
6548 source location to the destination location, which are not allowed to
6549 overlap. It copies "len" bytes of memory over. If the argument is known
6550 to be aligned to some boundary, this can be specified as the fourth
6551 argument, otherwise it should be set to 0 or 1.
6553 '``llvm.memmove``' Intrinsic
6554 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6559 This is an overloaded intrinsic. You can use llvm.memmove on any integer
6560 bit width and for different address space. Not all targets support all
6565 declare void @llvm.memmove.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
6566 i32 <len>, i32 <align>, i1 <isvolatile>)
6567 declare void @llvm.memmove.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
6568 i64 <len>, i32 <align>, i1 <isvolatile>)
6573 The '``llvm.memmove.*``' intrinsics move a block of memory from the
6574 source location to the destination location. It is similar to the
6575 '``llvm.memcpy``' intrinsic but allows the two memory locations to
6578 Note that, unlike the standard libc function, the ``llvm.memmove.*``
6579 intrinsics do not return a value, takes extra alignment/isvolatile
6580 arguments and the pointers can be in specified address spaces.
6585 The first argument is a pointer to the destination, the second is a
6586 pointer to the source. The third argument is an integer argument
6587 specifying the number of bytes to copy, the fourth argument is the
6588 alignment of the source and destination locations, and the fifth is a
6589 boolean indicating a volatile access.
6591 If the call to this intrinsic has an alignment value that is not 0 or 1,
6592 then the caller guarantees that the source and destination pointers are
6593 aligned to that boundary.
6595 If the ``isvolatile`` parameter is ``true``, the ``llvm.memmove`` call
6596 is a :ref:`volatile operation <volatile>`. The detailed access behavior is
6597 not very cleanly specified and it is unwise to depend on it.
6602 The '``llvm.memmove.*``' intrinsics copy a block of memory from the
6603 source location to the destination location, which may overlap. It
6604 copies "len" bytes of memory over. If the argument is known to be
6605 aligned to some boundary, this can be specified as the fourth argument,
6606 otherwise it should be set to 0 or 1.
6608 '``llvm.memset.*``' Intrinsics
6609 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6614 This is an overloaded intrinsic. You can use llvm.memset on any integer
6615 bit width and for different address spaces. However, not all targets
6616 support all bit widths.
6620 declare void @llvm.memset.p0i8.i32(i8* <dest>, i8 <val>,
6621 i32 <len>, i32 <align>, i1 <isvolatile>)
6622 declare void @llvm.memset.p0i8.i64(i8* <dest>, i8 <val>,
6623 i64 <len>, i32 <align>, i1 <isvolatile>)
6628 The '``llvm.memset.*``' intrinsics fill a block of memory with a
6629 particular byte value.
6631 Note that, unlike the standard libc function, the ``llvm.memset``
6632 intrinsic does not return a value and takes extra alignment/volatile
6633 arguments. Also, the destination can be in an arbitrary address space.
6638 The first argument is a pointer to the destination to fill, the second
6639 is the byte value with which to fill it, the third argument is an
6640 integer argument specifying the number of bytes to fill, and the fourth
6641 argument is the known alignment of the destination location.
6643 If the call to this intrinsic has an alignment value that is not 0 or 1,
6644 then the caller guarantees that the destination pointer is aligned to
6647 If the ``isvolatile`` parameter is ``true``, the ``llvm.memset`` call is
6648 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
6649 very cleanly specified and it is unwise to depend on it.
6654 The '``llvm.memset.*``' intrinsics fill "len" bytes of memory starting
6655 at the destination location. If the argument is known to be aligned to
6656 some boundary, this can be specified as the fourth argument, otherwise
6657 it should be set to 0 or 1.
6659 '``llvm.sqrt.*``' Intrinsic
6660 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6665 This is an overloaded intrinsic. You can use ``llvm.sqrt`` on any
6666 floating point or vector of floating point type. Not all targets support
6671 declare float @llvm.sqrt.f32(float %Val)
6672 declare double @llvm.sqrt.f64(double %Val)
6673 declare x86_fp80 @llvm.sqrt.f80(x86_fp80 %Val)
6674 declare fp128 @llvm.sqrt.f128(fp128 %Val)
6675 declare ppc_fp128 @llvm.sqrt.ppcf128(ppc_fp128 %Val)
6680 The '``llvm.sqrt``' intrinsics return the sqrt of the specified operand,
6681 returning the same value as the libm '``sqrt``' functions would. Unlike
6682 ``sqrt`` in libm, however, ``llvm.sqrt`` has undefined behavior for
6683 negative numbers other than -0.0 (which allows for better optimization,
6684 because there is no need to worry about errno being set).
6685 ``llvm.sqrt(-0.0)`` is defined to return -0.0 like IEEE sqrt.
6690 The argument and return value are floating point numbers of the same
6696 This function returns the sqrt of the specified operand if it is a
6697 nonnegative floating point number.
6699 '``llvm.powi.*``' Intrinsic
6700 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6705 This is an overloaded intrinsic. You can use ``llvm.powi`` on any
6706 floating point or vector of floating point type. Not all targets support
6711 declare float @llvm.powi.f32(float %Val, i32 %power)
6712 declare double @llvm.powi.f64(double %Val, i32 %power)
6713 declare x86_fp80 @llvm.powi.f80(x86_fp80 %Val, i32 %power)
6714 declare fp128 @llvm.powi.f128(fp128 %Val, i32 %power)
6715 declare ppc_fp128 @llvm.powi.ppcf128(ppc_fp128 %Val, i32 %power)
6720 The '``llvm.powi.*``' intrinsics return the first operand raised to the
6721 specified (positive or negative) power. The order of evaluation of
6722 multiplications is not defined. When a vector of floating point type is
6723 used, the second argument remains a scalar integer value.
6728 The second argument is an integer power, and the first is a value to
6729 raise to that power.
6734 This function returns the first value raised to the second power with an
6735 unspecified sequence of rounding operations.
6737 '``llvm.sin.*``' Intrinsic
6738 ^^^^^^^^^^^^^^^^^^^^^^^^^^
6743 This is an overloaded intrinsic. You can use ``llvm.sin`` on any
6744 floating point or vector of floating point type. Not all targets support
6749 declare float @llvm.sin.f32(float %Val)
6750 declare double @llvm.sin.f64(double %Val)
6751 declare x86_fp80 @llvm.sin.f80(x86_fp80 %Val)
6752 declare fp128 @llvm.sin.f128(fp128 %Val)
6753 declare ppc_fp128 @llvm.sin.ppcf128(ppc_fp128 %Val)
6758 The '``llvm.sin.*``' intrinsics return the sine of the operand.
6763 The argument and return value are floating point numbers of the same
6769 This function returns the sine of the specified operand, returning the
6770 same values as the libm ``sin`` functions would, and handles error
6771 conditions in the same way.
6773 '``llvm.cos.*``' Intrinsic
6774 ^^^^^^^^^^^^^^^^^^^^^^^^^^
6779 This is an overloaded intrinsic. You can use ``llvm.cos`` on any
6780 floating point or vector of floating point type. Not all targets support
6785 declare float @llvm.cos.f32(float %Val)
6786 declare double @llvm.cos.f64(double %Val)
6787 declare x86_fp80 @llvm.cos.f80(x86_fp80 %Val)
6788 declare fp128 @llvm.cos.f128(fp128 %Val)
6789 declare ppc_fp128 @llvm.cos.ppcf128(ppc_fp128 %Val)
6794 The '``llvm.cos.*``' intrinsics return the cosine of the operand.
6799 The argument and return value are floating point numbers of the same
6805 This function returns the cosine of the specified operand, returning the
6806 same values as the libm ``cos`` functions would, and handles error
6807 conditions in the same way.
6809 '``llvm.pow.*``' Intrinsic
6810 ^^^^^^^^^^^^^^^^^^^^^^^^^^
6815 This is an overloaded intrinsic. You can use ``llvm.pow`` on any
6816 floating point or vector of floating point type. Not all targets support
6821 declare float @llvm.pow.f32(float %Val, float %Power)
6822 declare double @llvm.pow.f64(double %Val, double %Power)
6823 declare x86_fp80 @llvm.pow.f80(x86_fp80 %Val, x86_fp80 %Power)
6824 declare fp128 @llvm.pow.f128(fp128 %Val, fp128 %Power)
6825 declare ppc_fp128 @llvm.pow.ppcf128(ppc_fp128 %Val, ppc_fp128 Power)
6830 The '``llvm.pow.*``' intrinsics return the first operand raised to the
6831 specified (positive or negative) power.
6836 The second argument is a floating point power, and the first is a value
6837 to raise to that power.
6842 This function returns the first value raised to the second power,
6843 returning the same values as the libm ``pow`` functions would, and
6844 handles error conditions in the same way.
6846 '``llvm.exp.*``' Intrinsic
6847 ^^^^^^^^^^^^^^^^^^^^^^^^^^
6852 This is an overloaded intrinsic. You can use ``llvm.exp`` on any
6853 floating point or vector of floating point type. Not all targets support
6858 declare float @llvm.exp.f32(float %Val)
6859 declare double @llvm.exp.f64(double %Val)
6860 declare x86_fp80 @llvm.exp.f80(x86_fp80 %Val)
6861 declare fp128 @llvm.exp.f128(fp128 %Val)
6862 declare ppc_fp128 @llvm.exp.ppcf128(ppc_fp128 %Val)
6867 The '``llvm.exp.*``' intrinsics perform the exp function.
6872 The argument and return value are floating point numbers of the same
6878 This function returns the same values as the libm ``exp`` functions
6879 would, and handles error conditions in the same way.
6881 '``llvm.exp2.*``' Intrinsic
6882 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6887 This is an overloaded intrinsic. You can use ``llvm.exp2`` on any
6888 floating point or vector of floating point type. Not all targets support
6893 declare float @llvm.exp2.f32(float %Val)
6894 declare double @llvm.exp2.f64(double %Val)
6895 declare x86_fp80 @llvm.exp2.f80(x86_fp80 %Val)
6896 declare fp128 @llvm.exp2.f128(fp128 %Val)
6897 declare ppc_fp128 @llvm.exp2.ppcf128(ppc_fp128 %Val)
6902 The '``llvm.exp2.*``' intrinsics perform the exp2 function.
6907 The argument and return value are floating point numbers of the same
6913 This function returns the same values as the libm ``exp2`` functions
6914 would, and handles error conditions in the same way.
6916 '``llvm.log.*``' Intrinsic
6917 ^^^^^^^^^^^^^^^^^^^^^^^^^^
6922 This is an overloaded intrinsic. You can use ``llvm.log`` on any
6923 floating point or vector of floating point type. Not all targets support
6928 declare float @llvm.log.f32(float %Val)
6929 declare double @llvm.log.f64(double %Val)
6930 declare x86_fp80 @llvm.log.f80(x86_fp80 %Val)
6931 declare fp128 @llvm.log.f128(fp128 %Val)
6932 declare ppc_fp128 @llvm.log.ppcf128(ppc_fp128 %Val)
6937 The '``llvm.log.*``' intrinsics perform the log function.
6942 The argument and return value are floating point numbers of the same
6948 This function returns the same values as the libm ``log`` functions
6949 would, and handles error conditions in the same way.
6951 '``llvm.log10.*``' Intrinsic
6952 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6957 This is an overloaded intrinsic. You can use ``llvm.log10`` on any
6958 floating point or vector of floating point type. Not all targets support
6963 declare float @llvm.log10.f32(float %Val)
6964 declare double @llvm.log10.f64(double %Val)
6965 declare x86_fp80 @llvm.log10.f80(x86_fp80 %Val)
6966 declare fp128 @llvm.log10.f128(fp128 %Val)
6967 declare ppc_fp128 @llvm.log10.ppcf128(ppc_fp128 %Val)
6972 The '``llvm.log10.*``' intrinsics perform the log10 function.
6977 The argument and return value are floating point numbers of the same
6983 This function returns the same values as the libm ``log10`` functions
6984 would, and handles error conditions in the same way.
6986 '``llvm.log2.*``' Intrinsic
6987 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6992 This is an overloaded intrinsic. You can use ``llvm.log2`` on any
6993 floating point or vector of floating point type. Not all targets support
6998 declare float @llvm.log2.f32(float %Val)
6999 declare double @llvm.log2.f64(double %Val)
7000 declare x86_fp80 @llvm.log2.f80(x86_fp80 %Val)
7001 declare fp128 @llvm.log2.f128(fp128 %Val)
7002 declare ppc_fp128 @llvm.log2.ppcf128(ppc_fp128 %Val)
7007 The '``llvm.log2.*``' intrinsics perform the log2 function.
7012 The argument and return value are floating point numbers of the same
7018 This function returns the same values as the libm ``log2`` functions
7019 would, and handles error conditions in the same way.
7021 '``llvm.fma.*``' Intrinsic
7022 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7027 This is an overloaded intrinsic. You can use ``llvm.fma`` on any
7028 floating point or vector of floating point type. Not all targets support
7033 declare float @llvm.fma.f32(float %a, float %b, float %c)
7034 declare double @llvm.fma.f64(double %a, double %b, double %c)
7035 declare x86_fp80 @llvm.fma.f80(x86_fp80 %a, x86_fp80 %b, x86_fp80 %c)
7036 declare fp128 @llvm.fma.f128(fp128 %a, fp128 %b, fp128 %c)
7037 declare ppc_fp128 @llvm.fma.ppcf128(ppc_fp128 %a, ppc_fp128 %b, ppc_fp128 %c)
7042 The '``llvm.fma.*``' intrinsics perform the fused multiply-add
7048 The argument and return value are floating point numbers of the same
7054 This function returns the same values as the libm ``fma`` functions
7057 '``llvm.fabs.*``' Intrinsic
7058 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7063 This is an overloaded intrinsic. You can use ``llvm.fabs`` on any
7064 floating point or vector of floating point type. Not all targets support
7069 declare float @llvm.fabs.f32(float %Val)
7070 declare double @llvm.fabs.f64(double %Val)
7071 declare x86_fp80 @llvm.fabs.f80(x86_fp80 %Val)
7072 declare fp128 @llvm.fabs.f128(fp128 %Val)
7073 declare ppc_fp128 @llvm.fabs.ppcf128(ppc_fp128 %Val)
7078 The '``llvm.fabs.*``' intrinsics return the absolute value of the
7084 The argument and return value are floating point numbers of the same
7090 This function returns the same values as the libm ``fabs`` functions
7091 would, and handles error conditions in the same way.
7093 '``llvm.floor.*``' Intrinsic
7094 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7099 This is an overloaded intrinsic. You can use ``llvm.floor`` on any
7100 floating point or vector of floating point type. Not all targets support
7105 declare float @llvm.floor.f32(float %Val)
7106 declare double @llvm.floor.f64(double %Val)
7107 declare x86_fp80 @llvm.floor.f80(x86_fp80 %Val)
7108 declare fp128 @llvm.floor.f128(fp128 %Val)
7109 declare ppc_fp128 @llvm.floor.ppcf128(ppc_fp128 %Val)
7114 The '``llvm.floor.*``' intrinsics return the floor of the operand.
7119 The argument and return value are floating point numbers of the same
7125 This function returns the same values as the libm ``floor`` functions
7126 would, and handles error conditions in the same way.
7128 '``llvm.ceil.*``' Intrinsic
7129 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7134 This is an overloaded intrinsic. You can use ``llvm.ceil`` on any
7135 floating point or vector of floating point type. Not all targets support
7140 declare float @llvm.ceil.f32(float %Val)
7141 declare double @llvm.ceil.f64(double %Val)
7142 declare x86_fp80 @llvm.ceil.f80(x86_fp80 %Val)
7143 declare fp128 @llvm.ceil.f128(fp128 %Val)
7144 declare ppc_fp128 @llvm.ceil.ppcf128(ppc_fp128 %Val)
7149 The '``llvm.ceil.*``' intrinsics return the ceiling of the operand.
7154 The argument and return value are floating point numbers of the same
7160 This function returns the same values as the libm ``ceil`` functions
7161 would, and handles error conditions in the same way.
7163 '``llvm.trunc.*``' Intrinsic
7164 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7169 This is an overloaded intrinsic. You can use ``llvm.trunc`` on any
7170 floating point or vector of floating point type. Not all targets support
7175 declare float @llvm.trunc.f32(float %Val)
7176 declare double @llvm.trunc.f64(double %Val)
7177 declare x86_fp80 @llvm.trunc.f80(x86_fp80 %Val)
7178 declare fp128 @llvm.trunc.f128(fp128 %Val)
7179 declare ppc_fp128 @llvm.trunc.ppcf128(ppc_fp128 %Val)
7184 The '``llvm.trunc.*``' intrinsics returns the operand rounded to the
7185 nearest integer not larger in magnitude than the operand.
7190 The argument and return value are floating point numbers of the same
7196 This function returns the same values as the libm ``trunc`` functions
7197 would, and handles error conditions in the same way.
7199 '``llvm.rint.*``' Intrinsic
7200 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7205 This is an overloaded intrinsic. You can use ``llvm.rint`` on any
7206 floating point or vector of floating point type. Not all targets support
7211 declare float @llvm.rint.f32(float %Val)
7212 declare double @llvm.rint.f64(double %Val)
7213 declare x86_fp80 @llvm.rint.f80(x86_fp80 %Val)
7214 declare fp128 @llvm.rint.f128(fp128 %Val)
7215 declare ppc_fp128 @llvm.rint.ppcf128(ppc_fp128 %Val)
7220 The '``llvm.rint.*``' intrinsics returns the operand rounded to the
7221 nearest integer. It may raise an inexact floating-point exception if the
7222 operand isn't an integer.
7227 The argument and return value are floating point numbers of the same
7233 This function returns the same values as the libm ``rint`` functions
7234 would, and handles error conditions in the same way.
7236 '``llvm.nearbyint.*``' Intrinsic
7237 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7242 This is an overloaded intrinsic. You can use ``llvm.nearbyint`` on any
7243 floating point or vector of floating point type. Not all targets support
7248 declare float @llvm.nearbyint.f32(float %Val)
7249 declare double @llvm.nearbyint.f64(double %Val)
7250 declare x86_fp80 @llvm.nearbyint.f80(x86_fp80 %Val)
7251 declare fp128 @llvm.nearbyint.f128(fp128 %Val)
7252 declare ppc_fp128 @llvm.nearbyint.ppcf128(ppc_fp128 %Val)
7257 The '``llvm.nearbyint.*``' intrinsics returns the operand rounded to the
7263 The argument and return value are floating point numbers of the same
7269 This function returns the same values as the libm ``nearbyint``
7270 functions would, and handles error conditions in the same way.
7272 Bit Manipulation Intrinsics
7273 ---------------------------
7275 LLVM provides intrinsics for a few important bit manipulation
7276 operations. These allow efficient code generation for some algorithms.
7278 '``llvm.bswap.*``' Intrinsics
7279 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7284 This is an overloaded intrinsic function. You can use bswap on any
7285 integer type that is an even number of bytes (i.e. BitWidth % 16 == 0).
7289 declare i16 @llvm.bswap.i16(i16 <id>)
7290 declare i32 @llvm.bswap.i32(i32 <id>)
7291 declare i64 @llvm.bswap.i64(i64 <id>)
7296 The '``llvm.bswap``' family of intrinsics is used to byte swap integer
7297 values with an even number of bytes (positive multiple of 16 bits).
7298 These are useful for performing operations on data that is not in the
7299 target's native byte order.
7304 The ``llvm.bswap.i16`` intrinsic returns an i16 value that has the high
7305 and low byte of the input i16 swapped. Similarly, the ``llvm.bswap.i32``
7306 intrinsic returns an i32 value that has the four bytes of the input i32
7307 swapped, so that if the input bytes are numbered 0, 1, 2, 3 then the
7308 returned i32 will have its bytes in 3, 2, 1, 0 order. The
7309 ``llvm.bswap.i48``, ``llvm.bswap.i64`` and other intrinsics extend this
7310 concept to additional even-byte lengths (6 bytes, 8 bytes and more,
7313 '``llvm.ctpop.*``' Intrinsic
7314 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7319 This is an overloaded intrinsic. You can use llvm.ctpop on any integer
7320 bit width, or on any vector with integer elements. Not all targets
7321 support all bit widths or vector types, however.
7325 declare i8 @llvm.ctpop.i8(i8 <src>)
7326 declare i16 @llvm.ctpop.i16(i16 <src>)
7327 declare i32 @llvm.ctpop.i32(i32 <src>)
7328 declare i64 @llvm.ctpop.i64(i64 <src>)
7329 declare i256 @llvm.ctpop.i256(i256 <src>)
7330 declare <2 x i32> @llvm.ctpop.v2i32(<2 x i32> <src>)
7335 The '``llvm.ctpop``' family of intrinsics counts the number of bits set
7341 The only argument is the value to be counted. The argument may be of any
7342 integer type, or a vector with integer elements. The return type must
7343 match the argument type.
7348 The '``llvm.ctpop``' intrinsic counts the 1's in a variable, or within
7349 each element of a vector.
7351 '``llvm.ctlz.*``' Intrinsic
7352 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7357 This is an overloaded intrinsic. You can use ``llvm.ctlz`` on any
7358 integer bit width, or any vector whose elements are integers. Not all
7359 targets support all bit widths or vector types, however.
7363 declare i8 @llvm.ctlz.i8 (i8 <src>, i1 <is_zero_undef>)
7364 declare i16 @llvm.ctlz.i16 (i16 <src>, i1 <is_zero_undef>)
7365 declare i32 @llvm.ctlz.i32 (i32 <src>, i1 <is_zero_undef>)
7366 declare i64 @llvm.ctlz.i64 (i64 <src>, i1 <is_zero_undef>)
7367 declare i256 @llvm.ctlz.i256(i256 <src>, i1 <is_zero_undef>)
7368 declase <2 x i32> @llvm.ctlz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
7373 The '``llvm.ctlz``' family of intrinsic functions counts the number of
7374 leading zeros in a variable.
7379 The first argument is the value to be counted. This argument may be of
7380 any integer type, or a vectory with integer element type. The return
7381 type must match the first argument type.
7383 The second argument must be a constant and is a flag to indicate whether
7384 the intrinsic should ensure that a zero as the first argument produces a
7385 defined result. Historically some architectures did not provide a
7386 defined result for zero values as efficiently, and many algorithms are
7387 now predicated on avoiding zero-value inputs.
7392 The '``llvm.ctlz``' intrinsic counts the leading (most significant)
7393 zeros in a variable, or within each element of the vector. If
7394 ``src == 0`` then the result is the size in bits of the type of ``src``
7395 if ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
7396 ``llvm.ctlz(i32 2) = 30``.
7398 '``llvm.cttz.*``' Intrinsic
7399 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7404 This is an overloaded intrinsic. You can use ``llvm.cttz`` on any
7405 integer bit width, or any vector of integer elements. Not all targets
7406 support all bit widths or vector types, however.
7410 declare i8 @llvm.cttz.i8 (i8 <src>, i1 <is_zero_undef>)
7411 declare i16 @llvm.cttz.i16 (i16 <src>, i1 <is_zero_undef>)
7412 declare i32 @llvm.cttz.i32 (i32 <src>, i1 <is_zero_undef>)
7413 declare i64 @llvm.cttz.i64 (i64 <src>, i1 <is_zero_undef>)
7414 declare i256 @llvm.cttz.i256(i256 <src>, i1 <is_zero_undef>)
7415 declase <2 x i32> @llvm.cttz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
7420 The '``llvm.cttz``' family of intrinsic functions counts the number of
7426 The first argument is the value to be counted. This argument may be of
7427 any integer type, or a vectory with integer element type. The return
7428 type must match the first argument type.
7430 The second argument must be a constant and is a flag to indicate whether
7431 the intrinsic should ensure that a zero as the first argument produces a
7432 defined result. Historically some architectures did not provide a
7433 defined result for zero values as efficiently, and many algorithms are
7434 now predicated on avoiding zero-value inputs.
7439 The '``llvm.cttz``' intrinsic counts the trailing (least significant)
7440 zeros in a variable, or within each element of a vector. If ``src == 0``
7441 then the result is the size in bits of the type of ``src`` if
7442 ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
7443 ``llvm.cttz(2) = 1``.
7445 Arithmetic with Overflow Intrinsics
7446 -----------------------------------
7448 LLVM provides intrinsics for some arithmetic with overflow operations.
7450 '``llvm.sadd.with.overflow.*``' Intrinsics
7451 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7456 This is an overloaded intrinsic. You can use ``llvm.sadd.with.overflow``
7457 on any integer bit width.
7461 declare {i16, i1} @llvm.sadd.with.overflow.i16(i16 %a, i16 %b)
7462 declare {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
7463 declare {i64, i1} @llvm.sadd.with.overflow.i64(i64 %a, i64 %b)
7468 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
7469 a signed addition of the two arguments, and indicate whether an overflow
7470 occurred during the signed summation.
7475 The arguments (%a and %b) and the first element of the result structure
7476 may be of integer types of any bit width, but they must have the same
7477 bit width. The second element of the result structure must be of type
7478 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
7484 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
7485 a signed addition of the two variables. They return a structure --- the
7486 first element of which is the signed summation, and the second element
7487 of which is a bit specifying if the signed summation resulted in an
7493 .. code-block:: llvm
7495 %res = call {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
7496 %sum = extractvalue {i32, i1} %res, 0
7497 %obit = extractvalue {i32, i1} %res, 1
7498 br i1 %obit, label %overflow, label %normal
7500 '``llvm.uadd.with.overflow.*``' Intrinsics
7501 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7506 This is an overloaded intrinsic. You can use ``llvm.uadd.with.overflow``
7507 on any integer bit width.
7511 declare {i16, i1} @llvm.uadd.with.overflow.i16(i16 %a, i16 %b)
7512 declare {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
7513 declare {i64, i1} @llvm.uadd.with.overflow.i64(i64 %a, i64 %b)
7518 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
7519 an unsigned addition of the two arguments, and indicate whether a carry
7520 occurred during the unsigned summation.
7525 The arguments (%a and %b) and the first element of the result structure
7526 may be of integer types of any bit width, but they must have the same
7527 bit width. The second element of the result structure must be of type
7528 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
7534 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
7535 an unsigned addition of the two arguments. They return a structure --- the
7536 first element of which is the sum, and the second element of which is a
7537 bit specifying if the unsigned summation resulted in a carry.
7542 .. code-block:: llvm
7544 %res = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
7545 %sum = extractvalue {i32, i1} %res, 0
7546 %obit = extractvalue {i32, i1} %res, 1
7547 br i1 %obit, label %carry, label %normal
7549 '``llvm.ssub.with.overflow.*``' Intrinsics
7550 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7555 This is an overloaded intrinsic. You can use ``llvm.ssub.with.overflow``
7556 on any integer bit width.
7560 declare {i16, i1} @llvm.ssub.with.overflow.i16(i16 %a, i16 %b)
7561 declare {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
7562 declare {i64, i1} @llvm.ssub.with.overflow.i64(i64 %a, i64 %b)
7567 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
7568 a signed subtraction of the two arguments, and indicate whether an
7569 overflow occurred during the signed subtraction.
7574 The arguments (%a and %b) and the first element of the result structure
7575 may be of integer types of any bit width, but they must have the same
7576 bit width. The second element of the result structure must be of type
7577 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
7583 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
7584 a signed subtraction of the two arguments. They return a structure --- the
7585 first element of which is the subtraction, and the second element of
7586 which is a bit specifying if the signed subtraction resulted in an
7592 .. code-block:: llvm
7594 %res = call {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
7595 %sum = extractvalue {i32, i1} %res, 0
7596 %obit = extractvalue {i32, i1} %res, 1
7597 br i1 %obit, label %overflow, label %normal
7599 '``llvm.usub.with.overflow.*``' Intrinsics
7600 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7605 This is an overloaded intrinsic. You can use ``llvm.usub.with.overflow``
7606 on any integer bit width.
7610 declare {i16, i1} @llvm.usub.with.overflow.i16(i16 %a, i16 %b)
7611 declare {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
7612 declare {i64, i1} @llvm.usub.with.overflow.i64(i64 %a, i64 %b)
7617 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
7618 an unsigned subtraction of the two arguments, and indicate whether an
7619 overflow occurred during the unsigned subtraction.
7624 The arguments (%a and %b) and the first element of the result structure
7625 may be of integer types of any bit width, but they must have the same
7626 bit width. The second element of the result structure must be of type
7627 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
7633 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
7634 an unsigned subtraction of the two arguments. They return a structure ---
7635 the first element of which is the subtraction, and the second element of
7636 which is a bit specifying if the unsigned subtraction resulted in an
7642 .. code-block:: llvm
7644 %res = call {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
7645 %sum = extractvalue {i32, i1} %res, 0
7646 %obit = extractvalue {i32, i1} %res, 1
7647 br i1 %obit, label %overflow, label %normal
7649 '``llvm.smul.with.overflow.*``' Intrinsics
7650 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7655 This is an overloaded intrinsic. You can use ``llvm.smul.with.overflow``
7656 on any integer bit width.
7660 declare {i16, i1} @llvm.smul.with.overflow.i16(i16 %a, i16 %b)
7661 declare {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
7662 declare {i64, i1} @llvm.smul.with.overflow.i64(i64 %a, i64 %b)
7667 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
7668 a signed multiplication of the two arguments, and indicate whether an
7669 overflow occurred during the signed multiplication.
7674 The arguments (%a and %b) and the first element of the result structure
7675 may be of integer types of any bit width, but they must have the same
7676 bit width. The second element of the result structure must be of type
7677 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
7683 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
7684 a signed multiplication of the two arguments. They return a structure ---
7685 the first element of which is the multiplication, and the second element
7686 of which is a bit specifying if the signed multiplication resulted in an
7692 .. code-block:: llvm
7694 %res = call {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
7695 %sum = extractvalue {i32, i1} %res, 0
7696 %obit = extractvalue {i32, i1} %res, 1
7697 br i1 %obit, label %overflow, label %normal
7699 '``llvm.umul.with.overflow.*``' Intrinsics
7700 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7705 This is an overloaded intrinsic. You can use ``llvm.umul.with.overflow``
7706 on any integer bit width.
7710 declare {i16, i1} @llvm.umul.with.overflow.i16(i16 %a, i16 %b)
7711 declare {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
7712 declare {i64, i1} @llvm.umul.with.overflow.i64(i64 %a, i64 %b)
7717 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
7718 a unsigned multiplication of the two arguments, and indicate whether an
7719 overflow occurred during the unsigned multiplication.
7724 The arguments (%a and %b) and the first element of the result structure
7725 may be of integer types of any bit width, but they must have the same
7726 bit width. The second element of the result structure must be of type
7727 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
7733 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
7734 an unsigned multiplication of the two arguments. They return a structure ---
7735 the first element of which is the multiplication, and the second
7736 element of which is a bit specifying if the unsigned multiplication
7737 resulted in an overflow.
7742 .. code-block:: llvm
7744 %res = call {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
7745 %sum = extractvalue {i32, i1} %res, 0
7746 %obit = extractvalue {i32, i1} %res, 1
7747 br i1 %obit, label %overflow, label %normal
7749 Specialised Arithmetic Intrinsics
7750 ---------------------------------
7752 '``llvm.fmuladd.*``' Intrinsic
7753 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7760 declare float @llvm.fmuladd.f32(float %a, float %b, float %c)
7761 declare double @llvm.fmuladd.f64(double %a, double %b, double %c)
7766 The '``llvm.fmuladd.*``' intrinsic functions represent multiply-add
7767 expressions that can be fused if the code generator determines that (a) the
7768 target instruction set has support for a fused operation, and (b) that the
7769 fused operation is more efficient than the equivalent, separate pair of mul
7770 and add instructions.
7775 The '``llvm.fmuladd.*``' intrinsics each take three arguments: two
7776 multiplicands, a and b, and an addend c.
7785 %0 = call float @llvm.fmuladd.f32(%a, %b, %c)
7787 is equivalent to the expression a \* b + c, except that rounding will
7788 not be performed between the multiplication and addition steps if the
7789 code generator fuses the operations. Fusion is not guaranteed, even if
7790 the target platform supports it. If a fused multiply-add is required the
7791 corresponding llvm.fma.\* intrinsic function should be used instead.
7796 .. code-block:: llvm
7798 %r2 = call float @llvm.fmuladd.f32(float %a, float %b, float %c) ; yields {float}:r2 = (a * b) + c
7800 Half Precision Floating Point Intrinsics
7801 ----------------------------------------
7803 For most target platforms, half precision floating point is a
7804 storage-only format. This means that it is a dense encoding (in memory)
7805 but does not support computation in the format.
7807 This means that code must first load the half-precision floating point
7808 value as an i16, then convert it to float with
7809 :ref:`llvm.convert.from.fp16 <int_convert_from_fp16>`. Computation can
7810 then be performed on the float value (including extending to double
7811 etc). To store the value back to memory, it is first converted to float
7812 if needed, then converted to i16 with
7813 :ref:`llvm.convert.to.fp16 <int_convert_to_fp16>`, then storing as an
7816 .. _int_convert_to_fp16:
7818 '``llvm.convert.to.fp16``' Intrinsic
7819 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7826 declare i16 @llvm.convert.to.fp16(f32 %a)
7831 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion
7832 from single precision floating point format to half precision floating
7838 The intrinsic function contains single argument - the value to be
7844 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion
7845 from single precision floating point format to half precision floating
7846 point format. The return value is an ``i16`` which contains the
7852 .. code-block:: llvm
7854 %res = call i16 @llvm.convert.to.fp16(f32 %a)
7855 store i16 %res, i16* @x, align 2
7857 .. _int_convert_from_fp16:
7859 '``llvm.convert.from.fp16``' Intrinsic
7860 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7867 declare f32 @llvm.convert.from.fp16(i16 %a)
7872 The '``llvm.convert.from.fp16``' intrinsic function performs a
7873 conversion from half precision floating point format to single precision
7874 floating point format.
7879 The intrinsic function contains single argument - the value to be
7885 The '``llvm.convert.from.fp16``' intrinsic function performs a
7886 conversion from half single precision floating point format to single
7887 precision floating point format. The input half-float value is
7888 represented by an ``i16`` value.
7893 .. code-block:: llvm
7895 %a = load i16* @x, align 2
7896 %res = call f32 @llvm.convert.from.fp16(i16 %a)
7901 The LLVM debugger intrinsics (which all start with ``llvm.dbg.``
7902 prefix), are described in the `LLVM Source Level
7903 Debugging <SourceLevelDebugging.html#format_common_intrinsics>`_
7906 Exception Handling Intrinsics
7907 -----------------------------
7909 The LLVM exception handling intrinsics (which all start with
7910 ``llvm.eh.`` prefix), are described in the `LLVM Exception
7911 Handling <ExceptionHandling.html#format_common_intrinsics>`_ document.
7915 Trampoline Intrinsics
7916 ---------------------
7918 These intrinsics make it possible to excise one parameter, marked with
7919 the :ref:`nest <nest>` attribute, from a function. The result is a
7920 callable function pointer lacking the nest parameter - the caller does
7921 not need to provide a value for it. Instead, the value to use is stored
7922 in advance in a "trampoline", a block of memory usually allocated on the
7923 stack, which also contains code to splice the nest value into the
7924 argument list. This is used to implement the GCC nested function address
7927 For example, if the function is ``i32 f(i8* nest %c, i32 %x, i32 %y)``
7928 then the resulting function pointer has signature ``i32 (i32, i32)*``.
7929 It can be created as follows:
7931 .. code-block:: llvm
7933 %tramp = alloca [10 x i8], align 4 ; size and alignment only correct for X86
7934 %tramp1 = getelementptr [10 x i8]* %tramp, i32 0, i32 0
7935 call i8* @llvm.init.trampoline(i8* %tramp1, i8* bitcast (i32 (i8*, i32, i32)* @f to i8*), i8* %nval)
7936 %p = call i8* @llvm.adjust.trampoline(i8* %tramp1)
7937 %fp = bitcast i8* %p to i32 (i32, i32)*
7939 The call ``%val = call i32 %fp(i32 %x, i32 %y)`` is then equivalent to
7940 ``%val = call i32 %f(i8* %nval, i32 %x, i32 %y)``.
7944 '``llvm.init.trampoline``' Intrinsic
7945 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7952 declare void @llvm.init.trampoline(i8* <tramp>, i8* <func>, i8* <nval>)
7957 This fills the memory pointed to by ``tramp`` with executable code,
7958 turning it into a trampoline.
7963 The ``llvm.init.trampoline`` intrinsic takes three arguments, all
7964 pointers. The ``tramp`` argument must point to a sufficiently large and
7965 sufficiently aligned block of memory; this memory is written to by the
7966 intrinsic. Note that the size and the alignment are target-specific -
7967 LLVM currently provides no portable way of determining them, so a
7968 front-end that generates this intrinsic needs to have some
7969 target-specific knowledge. The ``func`` argument must hold a function
7970 bitcast to an ``i8*``.
7975 The block of memory pointed to by ``tramp`` is filled with target
7976 dependent code, turning it into a function. Then ``tramp`` needs to be
7977 passed to :ref:`llvm.adjust.trampoline <int_at>` to get a pointer which can
7978 be :ref:`bitcast (to a new function) and called <int_trampoline>`. The new
7979 function's signature is the same as that of ``func`` with any arguments
7980 marked with the ``nest`` attribute removed. At most one such ``nest``
7981 argument is allowed, and it must be of pointer type. Calling the new
7982 function is equivalent to calling ``func`` with the same argument list,
7983 but with ``nval`` used for the missing ``nest`` argument. If, after
7984 calling ``llvm.init.trampoline``, the memory pointed to by ``tramp`` is
7985 modified, then the effect of any later call to the returned function
7986 pointer is undefined.
7990 '``llvm.adjust.trampoline``' Intrinsic
7991 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7998 declare i8* @llvm.adjust.trampoline(i8* <tramp>)
8003 This performs any required machine-specific adjustment to the address of
8004 a trampoline (passed as ``tramp``).
8009 ``tramp`` must point to a block of memory which already has trampoline
8010 code filled in by a previous call to
8011 :ref:`llvm.init.trampoline <int_it>`.
8016 On some architectures the address of the code to be executed needs to be
8017 different to the address where the trampoline is actually stored. This
8018 intrinsic returns the executable address corresponding to ``tramp``
8019 after performing the required machine specific adjustments. The pointer
8020 returned can then be :ref:`bitcast and executed <int_trampoline>`.
8025 This class of intrinsics exists to information about the lifetime of
8026 memory objects and ranges where variables are immutable.
8028 '``llvm.lifetime.start``' Intrinsic
8029 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8036 declare void @llvm.lifetime.start(i64 <size>, i8* nocapture <ptr>)
8041 The '``llvm.lifetime.start``' intrinsic specifies the start of a memory
8047 The first argument is a constant integer representing the size of the
8048 object, or -1 if it is variable sized. The second argument is a pointer
8054 This intrinsic indicates that before this point in the code, the value
8055 of the memory pointed to by ``ptr`` is dead. This means that it is known
8056 to never be used and has an undefined value. A load from the pointer
8057 that precedes this intrinsic can be replaced with ``'undef'``.
8059 '``llvm.lifetime.end``' Intrinsic
8060 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8067 declare void @llvm.lifetime.end(i64 <size>, i8* nocapture <ptr>)
8072 The '``llvm.lifetime.end``' intrinsic specifies the end of a memory
8078 The first argument is a constant integer representing the size of the
8079 object, or -1 if it is variable sized. The second argument is a pointer
8085 This intrinsic indicates that after this point in the code, the value of
8086 the memory pointed to by ``ptr`` is dead. This means that it is known to
8087 never be used and has an undefined value. Any stores into the memory
8088 object following this intrinsic may be removed as dead.
8090 '``llvm.invariant.start``' Intrinsic
8091 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8098 declare {}* @llvm.invariant.start(i64 <size>, i8* nocapture <ptr>)
8103 The '``llvm.invariant.start``' intrinsic specifies that the contents of
8104 a memory object will not change.
8109 The first argument is a constant integer representing the size of the
8110 object, or -1 if it is variable sized. The second argument is a pointer
8116 This intrinsic indicates that until an ``llvm.invariant.end`` that uses
8117 the return value, the referenced memory location is constant and
8120 '``llvm.invariant.end``' Intrinsic
8121 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8128 declare void @llvm.invariant.end({}* <start>, i64 <size>, i8* nocapture <ptr>)
8133 The '``llvm.invariant.end``' intrinsic specifies that the contents of a
8134 memory object are mutable.
8139 The first argument is the matching ``llvm.invariant.start`` intrinsic.
8140 The second argument is a constant integer representing the size of the
8141 object, or -1 if it is variable sized and the third argument is a
8142 pointer to the object.
8147 This intrinsic indicates that the memory is mutable again.
8152 This class of intrinsics is designed to be generic and has no specific
8155 '``llvm.var.annotation``' Intrinsic
8156 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8163 declare void @llvm.var.annotation(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
8168 The '``llvm.var.annotation``' intrinsic.
8173 The first argument is a pointer to a value, the second is a pointer to a
8174 global string, the third is a pointer to a global string which is the
8175 source file name, and the last argument is the line number.
8180 This intrinsic allows annotation of local variables with arbitrary
8181 strings. This can be useful for special purpose optimizations that want
8182 to look for these annotations. These have no other defined use; they are
8183 ignored by code generation and optimization.
8185 '``llvm.annotation.*``' Intrinsic
8186 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8191 This is an overloaded intrinsic. You can use '``llvm.annotation``' on
8192 any integer bit width.
8196 declare i8 @llvm.annotation.i8(i8 <val>, i8* <str>, i8* <str>, i32 <int>)
8197 declare i16 @llvm.annotation.i16(i16 <val>, i8* <str>, i8* <str>, i32 <int>)
8198 declare i32 @llvm.annotation.i32(i32 <val>, i8* <str>, i8* <str>, i32 <int>)
8199 declare i64 @llvm.annotation.i64(i64 <val>, i8* <str>, i8* <str>, i32 <int>)
8200 declare i256 @llvm.annotation.i256(i256 <val>, i8* <str>, i8* <str>, i32 <int>)
8205 The '``llvm.annotation``' intrinsic.
8210 The first argument is an integer value (result of some expression), the
8211 second is a pointer to a global string, the third is a pointer to a
8212 global string which is the source file name, and the last argument is
8213 the line number. It returns the value of the first argument.
8218 This intrinsic allows annotations to be put on arbitrary expressions
8219 with arbitrary strings. This can be useful for special purpose
8220 optimizations that want to look for these annotations. These have no
8221 other defined use; they are ignored by code generation and optimization.
8223 '``llvm.trap``' Intrinsic
8224 ^^^^^^^^^^^^^^^^^^^^^^^^^
8231 declare void @llvm.trap() noreturn nounwind
8236 The '``llvm.trap``' intrinsic.
8246 This intrinsic is lowered to the target dependent trap instruction. If
8247 the target does not have a trap instruction, this intrinsic will be
8248 lowered to a call of the ``abort()`` function.
8250 '``llvm.debugtrap``' Intrinsic
8251 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8258 declare void @llvm.debugtrap() nounwind
8263 The '``llvm.debugtrap``' intrinsic.
8273 This intrinsic is lowered to code which is intended to cause an
8274 execution trap with the intention of requesting the attention of a
8277 '``llvm.stackprotector``' Intrinsic
8278 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8285 declare void @llvm.stackprotector(i8* <guard>, i8** <slot>)
8290 The ``llvm.stackprotector`` intrinsic takes the ``guard`` and stores it
8291 onto the stack at ``slot``. The stack slot is adjusted to ensure that it
8292 is placed on the stack before local variables.
8297 The ``llvm.stackprotector`` intrinsic requires two pointer arguments.
8298 The first argument is the value loaded from the stack guard
8299 ``@__stack_chk_guard``. The second variable is an ``alloca`` that has
8300 enough space to hold the value of the guard.
8305 This intrinsic causes the prologue/epilogue inserter to force the
8306 position of the ``AllocaInst`` stack slot to be before local variables
8307 on the stack. This is to ensure that if a local variable on the stack is
8308 overwritten, it will destroy the value of the guard. When the function
8309 exits, the guard on the stack is checked against the original guard. If
8310 they are different, then the program aborts by calling the
8311 ``__stack_chk_fail()`` function.
8313 '``llvm.objectsize``' Intrinsic
8314 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8321 declare i32 @llvm.objectsize.i32(i8* <object>, i1 <min>)
8322 declare i64 @llvm.objectsize.i64(i8* <object>, i1 <min>)
8327 The ``llvm.objectsize`` intrinsic is designed to provide information to
8328 the optimizers to determine at compile time whether a) an operation
8329 (like memcpy) will overflow a buffer that corresponds to an object, or
8330 b) that a runtime check for overflow isn't necessary. An object in this
8331 context means an allocation of a specific class, structure, array, or
8337 The ``llvm.objectsize`` intrinsic takes two arguments. The first
8338 argument is a pointer to or into the ``object``. The second argument is
8339 a boolean and determines whether ``llvm.objectsize`` returns 0 (if true)
8340 or -1 (if false) when the object size is unknown. The second argument
8341 only accepts constants.
8346 The ``llvm.objectsize`` intrinsic is lowered to a constant representing
8347 the size of the object concerned. If the size cannot be determined at
8348 compile time, ``llvm.objectsize`` returns ``i32/i64 -1 or 0`` (depending
8349 on the ``min`` argument).
8351 '``llvm.expect``' Intrinsic
8352 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8359 declare i32 @llvm.expect.i32(i32 <val>, i32 <expected_val>)
8360 declare i64 @llvm.expect.i64(i64 <val>, i64 <expected_val>)
8365 The ``llvm.expect`` intrinsic provides information about expected (the
8366 most probable) value of ``val``, which can be used by optimizers.
8371 The ``llvm.expect`` intrinsic takes two arguments. The first argument is
8372 a value. The second argument is an expected value, this needs to be a
8373 constant value, variables are not allowed.
8378 This intrinsic is lowered to the ``val``.
8380 '``llvm.donothing``' Intrinsic
8381 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8388 declare void @llvm.donothing() nounwind readnone
8393 The ``llvm.donothing`` intrinsic doesn't perform any operation. It's the
8394 only intrinsic that can be called with an invoke instruction.
8404 This intrinsic does nothing, and it's removed by optimizers and ignored