1 ==============================
2 LLVM Language Reference Manual
3 ==============================
12 This document is a reference manual for the LLVM assembly language. LLVM
13 is a Static Single Assignment (SSA) based representation that provides
14 type safety, low-level operations, flexibility, and the capability of
15 representing 'all' high-level languages cleanly. It is the common code
16 representation used throughout all phases of the LLVM compilation
22 The LLVM code representation is designed to be used in three different
23 forms: as an in-memory compiler IR, as an on-disk bitcode representation
24 (suitable for fast loading by a Just-In-Time compiler), and as a human
25 readable assembly language representation. This allows LLVM to provide a
26 powerful intermediate representation for efficient compiler
27 transformations and analysis, while providing a natural means to debug
28 and visualize the transformations. The three different forms of LLVM are
29 all equivalent. This document describes the human readable
30 representation and notation.
32 The LLVM representation aims to be light-weight and low-level while
33 being expressive, typed, and extensible at the same time. It aims to be
34 a "universal IR" of sorts, by being at a low enough level that
35 high-level ideas may be cleanly mapped to it (similar to how
36 microprocessors are "universal IR's", allowing many source languages to
37 be mapped to them). By providing type information, LLVM can be used as
38 the target of optimizations: for example, through pointer analysis, it
39 can be proven that a C automatic variable is never accessed outside of
40 the current function, allowing it to be promoted to a simple SSA value
41 instead of a memory location.
48 It is important to note that this document describes 'well formed' LLVM
49 assembly language. There is a difference between what the parser accepts
50 and what is considered 'well formed'. For example, the following
51 instruction is syntactically okay, but not well formed:
57 because the definition of ``%x`` does not dominate all of its uses. The
58 LLVM infrastructure provides a verification pass that may be used to
59 verify that an LLVM module is well formed. This pass is automatically
60 run by the parser after parsing input assembly and by the optimizer
61 before it outputs bitcode. The violations pointed out by the verifier
62 pass indicate bugs in transformation passes or input to the parser.
69 LLVM identifiers come in two basic types: global and local. Global
70 identifiers (functions, global variables) begin with the ``'@'``
71 character. Local identifiers (register names, types) begin with the
72 ``'%'`` character. Additionally, there are three different formats for
73 identifiers, for different purposes:
75 #. Named values are represented as a string of characters with their
76 prefix. For example, ``%foo``, ``@DivisionByZero``,
77 ``%a.really.long.identifier``. The actual regular expression used is
78 '``[%@][a-zA-Z$._][a-zA-Z$._0-9]*``'. Identifiers which require other
79 characters in their names can be surrounded with quotes. Special
80 characters may be escaped using ``"\xx"`` where ``xx`` is the ASCII
81 code for the character in hexadecimal. In this way, any character can
82 be used in a name value, even quotes themselves.
83 #. Unnamed values are represented as an unsigned numeric value with
84 their prefix. For example, ``%12``, ``@2``, ``%44``.
85 #. Constants, which are described in the section Constants_ below.
87 LLVM requires that values start with a prefix for two reasons: Compilers
88 don't need to worry about name clashes with reserved words, and the set
89 of reserved words may be expanded in the future without penalty.
90 Additionally, unnamed identifiers allow a compiler to quickly come up
91 with a temporary variable without having to avoid symbol table
94 Reserved words in LLVM are very similar to reserved words in other
95 languages. There are keywords for different opcodes ('``add``',
96 '``bitcast``', '``ret``', etc...), for primitive type names ('``void``',
97 '``i32``', etc...), and others. These reserved words cannot conflict
98 with variable names, because none of them start with a prefix character
101 Here is an example of LLVM code to multiply the integer variable
108 %result = mul i32 %X, 8
110 After strength reduction:
114 %result = shl i32 %X, 3
120 %0 = add i32 %X, %X ; yields {i32}:%0
121 %1 = add i32 %0, %0 ; yields {i32}:%1
122 %result = add i32 %1, %1
124 This last way of multiplying ``%X`` by 8 illustrates several important
125 lexical features of LLVM:
127 #. Comments are delimited with a '``;``' and go until the end of line.
128 #. Unnamed temporaries are created when the result of a computation is
129 not assigned to a named value.
130 #. Unnamed temporaries are numbered sequentially (using a per-function
131 incrementing counter, starting with 0).
133 It also shows a convention that we follow in this document. When
134 demonstrating instructions, we will follow an instruction with a comment
135 that defines the type and name of value produced.
143 LLVM programs are composed of ``Module``'s, each of which is a
144 translation unit of the input programs. Each module consists of
145 functions, global variables, and symbol table entries. Modules may be
146 combined together with the LLVM linker, which merges function (and
147 global variable) definitions, resolves forward declarations, and merges
148 symbol table entries. Here is an example of the "hello world" module:
152 ; Declare the string constant as a global constant.
153 @.str = private unnamed_addr constant [13 x i8] c"hello world\0A\00"
155 ; External declaration of the puts function
156 declare i32 @puts(i8* nocapture) nounwind
158 ; Definition of main function
159 define i32 @main() { ; i32()*
160 ; Convert [13 x i8]* to i8 *...
161 %cast210 = getelementptr [13 x i8]* @.str, i64 0, i64 0
163 ; Call puts function to write out the string to stdout.
164 call i32 @puts(i8* %cast210)
169 !1 = metadata !{i32 42}
172 This example is made up of a :ref:`global variable <globalvars>` named
173 "``.str``", an external declaration of the "``puts``" function, a
174 :ref:`function definition <functionstructure>` for "``main``" and
175 :ref:`named metadata <namedmetadatastructure>` "``foo``".
177 In general, a module is made up of a list of global values (where both
178 functions and global variables are global values). Global values are
179 represented by a pointer to a memory location (in this case, a pointer
180 to an array of char, and a pointer to a function), and have one of the
181 following :ref:`linkage types <linkage>`.
188 All Global Variables and Functions have one of the following types of
192 Global values with "``private``" linkage are only directly
193 accessible by objects in the current module. In particular, linking
194 code into a module with an private global value may cause the
195 private to be renamed as necessary to avoid collisions. Because the
196 symbol is private to the module, all references can be updated. This
197 doesn't show up in any symbol table in the object file.
199 Similar to ``private``, but the symbol is passed through the
200 assembler and evaluated by the linker. Unlike normal strong symbols,
201 they are removed by the linker from the final linked image
202 (executable or dynamic library).
203 ``linker_private_weak``
204 Similar to "``linker_private``", but the symbol is weak. Note that
205 ``linker_private_weak`` symbols are subject to coalescing by the
206 linker. The symbols are removed by the linker from the final linked
207 image (executable or dynamic library).
209 Similar to private, but the value shows as a local symbol
210 (``STB_LOCAL`` in the case of ELF) in the object file. This
211 corresponds to the notion of the '``static``' keyword in C.
212 ``available_externally``
213 Globals with "``available_externally``" linkage are never emitted
214 into the object file corresponding to the LLVM module. They exist to
215 allow inlining and other optimizations to take place given knowledge
216 of the definition of the global, which is known to be somewhere
217 outside the module. Globals with ``available_externally`` linkage
218 are allowed to be discarded at will, and are otherwise the same as
219 ``linkonce_odr``. This linkage type is only allowed on definitions,
222 Globals with "``linkonce``" linkage are merged with other globals of
223 the same name when linkage occurs. This can be used to implement
224 some forms of inline functions, templates, or other code which must
225 be generated in each translation unit that uses it, but where the
226 body may be overridden with a more definitive definition later.
227 Unreferenced ``linkonce`` globals are allowed to be discarded. Note
228 that ``linkonce`` linkage does not actually allow the optimizer to
229 inline the body of this function into callers because it doesn't
230 know if this definition of the function is the definitive definition
231 within the program or whether it will be overridden by a stronger
232 definition. To enable inlining and other optimizations, use
233 "``linkonce_odr``" linkage.
235 "``weak``" linkage has the same merging semantics as ``linkonce``
236 linkage, except that unreferenced globals with ``weak`` linkage may
237 not be discarded. This is used for globals that are declared "weak"
240 "``common``" linkage is most similar to "``weak``" linkage, but they
241 are used for tentative definitions in C, such as "``int X;``" at
242 global scope. Symbols with "``common``" linkage are merged in the
243 same way as ``weak symbols``, and they may not be deleted if
244 unreferenced. ``common`` symbols may not have an explicit section,
245 must have a zero initializer, and may not be marked
246 ':ref:`constant <globalvars>`'. Functions and aliases may not have
249 .. _linkage_appending:
252 "``appending``" linkage may only be applied to global variables of
253 pointer to array type. When two global variables with appending
254 linkage are linked together, the two global arrays are appended
255 together. This is the LLVM, typesafe, equivalent of having the
256 system linker append together "sections" with identical names when
259 The semantics of this linkage follow the ELF object file model: the
260 symbol is weak until linked, if not linked, the symbol becomes null
261 instead of being an undefined reference.
262 ``linkonce_odr``, ``weak_odr``
263 Some languages allow differing globals to be merged, such as two
264 functions with different semantics. Other languages, such as
265 ``C++``, ensure that only equivalent globals are ever merged (the
266 "one definition rule" --- "ODR"). Such languages can use the
267 ``linkonce_odr`` and ``weak_odr`` linkage types to indicate that the
268 global will only be merged with equivalent globals. These linkage
269 types are otherwise the same as their non-``odr`` versions.
270 ``linkonce_odr_auto_hide``
271 Similar to "``linkonce_odr``", but nothing in the translation unit
272 takes the address of this definition. For instance, functions that
273 had an inline definition, but the compiler decided not to inline it.
274 ``linkonce_odr_auto_hide`` may have only ``default`` visibility. The
275 symbols are removed by the linker from the final linked image
276 (executable or dynamic library).
278 If none of the above identifiers are used, the global is externally
279 visible, meaning that it participates in linkage and can be used to
280 resolve external symbol references.
282 The next two types of linkage are targeted for Microsoft Windows
283 platform only. They are designed to support importing (exporting)
284 symbols from (to) DLLs (Dynamic Link Libraries).
287 "``dllimport``" linkage causes the compiler to reference a function
288 or variable via a global pointer to a pointer that is set up by the
289 DLL exporting the symbol. On Microsoft Windows targets, the pointer
290 name is formed by combining ``__imp_`` and the function or variable
293 "``dllexport``" linkage causes the compiler to provide a global
294 pointer to a pointer in a DLL, so that it can be referenced with the
295 ``dllimport`` attribute. On Microsoft Windows targets, the pointer
296 name is formed by combining ``__imp_`` and the function or variable
299 For example, since the "``.LC0``" variable is defined to be internal, if
300 another module defined a "``.LC0``" variable and was linked with this
301 one, one of the two would be renamed, preventing a collision. Since
302 "``main``" and "``puts``" are external (i.e., lacking any linkage
303 declarations), they are accessible outside of the current module.
305 It is illegal for a function *declaration* to have any linkage type
306 other than ``external``, ``dllimport`` or ``extern_weak``.
308 Aliases can have only ``external``, ``internal``, ``weak`` or
309 ``weak_odr`` linkages.
316 LLVM :ref:`functions <functionstructure>`, :ref:`calls <i_call>` and
317 :ref:`invokes <i_invoke>` can all have an optional calling convention
318 specified for the call. The calling convention of any pair of dynamic
319 caller/callee must match, or the behavior of the program is undefined.
320 The following calling conventions are supported by LLVM, and more may be
323 "``ccc``" - The C calling convention
324 This calling convention (the default if no other calling convention
325 is specified) matches the target C calling conventions. This calling
326 convention supports varargs function calls and tolerates some
327 mismatch in the declared prototype and implemented declaration of
328 the function (as does normal C).
329 "``fastcc``" - The fast calling convention
330 This calling convention attempts to make calls as fast as possible
331 (e.g. by passing things in registers). This calling convention
332 allows the target to use whatever tricks it wants to produce fast
333 code for the target, without having to conform to an externally
334 specified ABI (Application Binary Interface). `Tail calls can only
335 be optimized when this, the GHC or the HiPE convention is
336 used. <CodeGenerator.html#id80>`_ This calling convention does not
337 support varargs and requires the prototype of all callees to exactly
338 match the prototype of the function definition.
339 "``coldcc``" - The cold calling convention
340 This calling convention attempts to make code in the caller as
341 efficient as possible under the assumption that the call is not
342 commonly executed. As such, these calls often preserve all registers
343 so that the call does not break any live ranges in the caller side.
344 This calling convention does not support varargs and requires the
345 prototype of all callees to exactly match the prototype of the
347 "``cc 10``" - GHC convention
348 This calling convention has been implemented specifically for use by
349 the `Glasgow Haskell Compiler (GHC) <http://www.haskell.org/ghc>`_.
350 It passes everything in registers, going to extremes to achieve this
351 by disabling callee save registers. This calling convention should
352 not be used lightly but only for specific situations such as an
353 alternative to the *register pinning* performance technique often
354 used when implementing functional programming languages. At the
355 moment only X86 supports this convention and it has the following
358 - On *X86-32* only supports up to 4 bit type parameters. No
359 floating point types are supported.
360 - On *X86-64* only supports up to 10 bit type parameters and 6
361 floating point parameters.
363 This calling convention supports `tail call
364 optimization <CodeGenerator.html#id80>`_ but requires both the
365 caller and callee are using it.
366 "``cc 11``" - The HiPE calling convention
367 This calling convention has been implemented specifically for use by
368 the `High-Performance Erlang
369 (HiPE) <http://www.it.uu.se/research/group/hipe/>`_ compiler, *the*
370 native code compiler of the `Ericsson's Open Source Erlang/OTP
371 system <http://www.erlang.org/download.shtml>`_. It uses more
372 registers for argument passing than the ordinary C calling
373 convention and defines no callee-saved registers. The calling
374 convention properly supports `tail call
375 optimization <CodeGenerator.html#id80>`_ but requires that both the
376 caller and the callee use it. It uses a *register pinning*
377 mechanism, similar to GHC's convention, for keeping frequently
378 accessed runtime components pinned to specific hardware registers.
379 At the moment only X86 supports this convention (both 32 and 64
381 "``cc <n>``" - Numbered convention
382 Any calling convention may be specified by number, allowing
383 target-specific calling conventions to be used. Target specific
384 calling conventions start at 64.
386 More calling conventions can be added/defined on an as-needed basis, to
387 support Pascal conventions or any other well-known target-independent
393 All Global Variables and Functions have one of the following visibility
396 "``default``" - Default style
397 On targets that use the ELF object file format, default visibility
398 means that the declaration is visible to other modules and, in
399 shared libraries, means that the declared entity may be overridden.
400 On Darwin, default visibility means that the declaration is visible
401 to other modules. Default visibility corresponds to "external
402 linkage" in the language.
403 "``hidden``" - Hidden style
404 Two declarations of an object with hidden visibility refer to the
405 same object if they are in the same shared object. Usually, hidden
406 visibility indicates that the symbol will not be placed into the
407 dynamic symbol table, so no other module (executable or shared
408 library) can reference it directly.
409 "``protected``" - Protected style
410 On ELF, protected visibility indicates that the symbol will be
411 placed in the dynamic symbol table, but that references within the
412 defining module will bind to the local symbol. That is, the symbol
413 cannot be overridden by another module.
418 LLVM IR allows you to specify name aliases for certain types. This can
419 make it easier to read the IR and make the IR more condensed
420 (particularly when recursive types are involved). An example of a name
425 %mytype = type { %mytype*, i32 }
427 You may give a name to any :ref:`type <typesystem>` except
428 ":ref:`void <t_void>`". Type name aliases may be used anywhere a type is
429 expected with the syntax "%mytype".
431 Note that type names are aliases for the structural type that they
432 indicate, and that you can therefore specify multiple names for the same
433 type. This often leads to confusing behavior when dumping out a .ll
434 file. Since LLVM IR uses structural typing, the name is not part of the
435 type. When printing out LLVM IR, the printer will pick *one name* to
436 render all types of a particular shape. This means that if you have code
437 where two different source types end up having the same LLVM type, that
438 the dumper will sometimes print the "wrong" or unexpected type. This is
439 an important design point and isn't going to change.
446 Global variables define regions of memory allocated at compilation time
447 instead of run-time. Global variables may optionally be initialized, may
448 have an explicit section to be placed in, and may have an optional
449 explicit alignment specified.
451 A variable may be defined as ``thread_local``, which means that it will
452 not be shared by threads (each thread will have a separated copy of the
453 variable). Not all targets support thread-local variables. Optionally, a
454 TLS model may be specified:
457 For variables that are only used within the current shared library.
459 For variables in modules that will not be loaded dynamically.
461 For variables defined in the executable and only used within it.
463 The models correspond to the ELF TLS models; see `ELF Handling For
464 Thread-Local Storage <http://people.redhat.com/drepper/tls.pdf>`_ for
465 more information on under which circumstances the different models may
466 be used. The target may choose a different TLS model if the specified
467 model is not supported, or if a better choice of model can be made.
469 A variable may be defined as a global ``constant``, which indicates that
470 the contents of the variable will **never** be modified (enabling better
471 optimization, allowing the global data to be placed in the read-only
472 section of an executable, etc). Note that variables that need runtime
473 initialization cannot be marked ``constant`` as there is a store to the
476 LLVM explicitly allows *declarations* of global variables to be marked
477 constant, even if the final definition of the global is not. This
478 capability can be used to enable slightly better optimization of the
479 program, but requires the language definition to guarantee that
480 optimizations based on the 'constantness' are valid for the translation
481 units that do not include the definition.
483 As SSA values, global variables define pointer values that are in scope
484 (i.e. they dominate) all basic blocks in the program. Global variables
485 always define a pointer to their "content" type because they describe a
486 region of memory, and all memory objects in LLVM are accessed through
489 Global variables can be marked with ``unnamed_addr`` which indicates
490 that the address is not significant, only the content. Constants marked
491 like this can be merged with other constants if they have the same
492 initializer. Note that a constant with significant address *can* be
493 merged with a ``unnamed_addr`` constant, the result being a constant
494 whose address is significant.
496 A global variable may be declared to reside in a target-specific
497 numbered address space. For targets that support them, address spaces
498 may affect how optimizations are performed and/or what target
499 instructions are used to access the variable. The default address space
500 is zero. The address space qualifier must precede any other attributes.
502 LLVM allows an explicit section to be specified for globals. If the
503 target supports it, it will emit globals to the section specified.
505 By default, global initializers are optimized by assuming that global
506 variables defined within the module are not modified from their
507 initial values before the start of the global initializer. This is
508 true even for variables potentially accessible from outside the
509 module, including those with external linkage or appearing in
510 ``@llvm.used``. This assumption may be suppressed by marking the
511 variable with ``externally_initialized``.
513 An explicit alignment may be specified for a global, which must be a
514 power of 2. If not present, or if the alignment is set to zero, the
515 alignment of the global is set by the target to whatever it feels
516 convenient. If an explicit alignment is specified, the global is forced
517 to have exactly that alignment. Targets and optimizers are not allowed
518 to over-align the global if the global has an assigned section. In this
519 case, the extra alignment could be observable: for example, code could
520 assume that the globals are densely packed in their section and try to
521 iterate over them as an array, alignment padding would break this
524 For example, the following defines a global in a numbered address space
525 with an initializer, section, and alignment:
529 @G = addrspace(5) constant float 1.0, section "foo", align 4
531 The following example defines a thread-local global with the
532 ``initialexec`` TLS model:
536 @G = thread_local(initialexec) global i32 0, align 4
538 .. _functionstructure:
543 LLVM function definitions consist of the "``define``" keyword, an
544 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
545 style <visibility>`, an optional :ref:`calling convention <callingconv>`,
546 an optional ``unnamed_addr`` attribute, a return type, an optional
547 :ref:`parameter attribute <paramattrs>` for the return type, a function
548 name, a (possibly empty) argument list (each with optional :ref:`parameter
549 attributes <paramattrs>`), optional :ref:`function attributes <fnattrs>`,
550 an optional section, an optional alignment, an optional :ref:`garbage
551 collector name <gc>`, an opening curly brace, a list of basic blocks,
552 and a closing curly brace.
554 LLVM function declarations consist of the "``declare``" keyword, an
555 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
556 style <visibility>`, an optional :ref:`calling convention <callingconv>`,
557 an optional ``unnamed_addr`` attribute, a return type, an optional
558 :ref:`parameter attribute <paramattrs>` for the return type, a function
559 name, a possibly empty list of arguments, an optional alignment, and an
560 optional :ref:`garbage collector name <gc>`.
562 A function definition contains a list of basic blocks, forming the CFG
563 (Control Flow Graph) for the function. Each basic block may optionally
564 start with a label (giving the basic block a symbol table entry),
565 contains a list of instructions, and ends with a
566 :ref:`terminator <terminators>` instruction (such as a branch or function
567 return). If explicit label is not provided, a block is assigned an
568 implicit numbered label, using a next value from the same counter as used
569 for unnamed temporaries (:ref:`see above<identifiers>`). For example, if a
570 function entry block does not have explicit label, it will be assigned
571 label "%0", then first unnamed temporary in that block will be "%1", etc.
573 The first basic block in a function is special in two ways: it is
574 immediately executed on entrance to the function, and it is not allowed
575 to have predecessor basic blocks (i.e. there can not be any branches to
576 the entry block of a function). Because the block can have no
577 predecessors, it also cannot have any :ref:`PHI nodes <i_phi>`.
579 LLVM allows an explicit section to be specified for functions. If the
580 target supports it, it will emit functions to the section specified.
582 An explicit alignment may be specified for a function. If not present,
583 or if the alignment is set to zero, the alignment of the function is set
584 by the target to whatever it feels convenient. If an explicit alignment
585 is specified, the function is forced to have at least that much
586 alignment. All alignments must be a power of 2.
588 If the ``unnamed_addr`` attribute is given, the address is know to not
589 be significant and two identical functions can be merged.
593 define [linkage] [visibility]
595 <ResultType> @<FunctionName> ([argument list])
596 [fn Attrs] [section "name"] [align N]
602 Aliases act as "second name" for the aliasee value (which can be either
603 function, global variable, another alias or bitcast of global value).
604 Aliases may have an optional :ref:`linkage type <linkage>`, and an optional
605 :ref:`visibility style <visibility>`.
609 @<Name> = alias [Linkage] [Visibility] <AliaseeTy> @<Aliasee>
611 .. _namedmetadatastructure:
616 Named metadata is a collection of metadata. :ref:`Metadata
617 nodes <metadata>` (but not metadata strings) are the only valid
618 operands for a named metadata.
622 ; Some unnamed metadata nodes, which are referenced by the named metadata.
623 !0 = metadata !{metadata !"zero"}
624 !1 = metadata !{metadata !"one"}
625 !2 = metadata !{metadata !"two"}
627 !name = !{!0, !1, !2}
634 The return type and each parameter of a function type may have a set of
635 *parameter attributes* associated with them. Parameter attributes are
636 used to communicate additional information about the result or
637 parameters of a function. Parameter attributes are considered to be part
638 of the function, not of the function type, so functions with different
639 parameter attributes can have the same function type.
641 Parameter attributes are simple keywords that follow the type specified.
642 If multiple parameter attributes are needed, they are space separated.
647 declare i32 @printf(i8* noalias nocapture, ...)
648 declare i32 @atoi(i8 zeroext)
649 declare signext i8 @returns_signed_char()
651 Note that any attributes for the function result (``nounwind``,
652 ``readonly``) come immediately after the argument list.
654 Currently, only the following parameter attributes are defined:
657 This indicates to the code generator that the parameter or return
658 value should be zero-extended to the extent required by the target's
659 ABI (which is usually 32-bits, but is 8-bits for a i1 on x86-64) by
660 the caller (for a parameter) or the callee (for a return value).
662 This indicates to the code generator that the parameter or return
663 value should be sign-extended to the extent required by the target's
664 ABI (which is usually 32-bits) by the caller (for a parameter) or
665 the callee (for a return value).
667 This indicates that this parameter or return value should be treated
668 in a special target-dependent fashion during while emitting code for
669 a function call or return (usually, by putting it in a register as
670 opposed to memory, though some targets use it to distinguish between
671 two different kinds of registers). Use of this attribute is
674 This indicates that the pointer parameter should really be passed by
675 value to the function. The attribute implies that a hidden copy of
676 the pointee is made between the caller and the callee, so the callee
677 is unable to modify the value in the caller. This attribute is only
678 valid on LLVM pointer arguments. It is generally used to pass
679 structs and arrays by value, but is also valid on pointers to
680 scalars. The copy is considered to belong to the caller not the
681 callee (for example, ``readonly`` functions should not write to
682 ``byval`` parameters). This is not a valid attribute for return
685 The byval attribute also supports specifying an alignment with the
686 align attribute. It indicates the alignment of the stack slot to
687 form and the known alignment of the pointer specified to the call
688 site. If the alignment is not specified, then the code generator
689 makes a target-specific assumption.
692 This indicates that the pointer parameter specifies the address of a
693 structure that is the return value of the function in the source
694 program. This pointer must be guaranteed by the caller to be valid:
695 loads and stores to the structure may be assumed by the callee
696 not to trap and to be properly aligned. This may only be applied to
697 the first parameter. This is not a valid attribute for return
700 This indicates that pointer values `*based* <pointeraliasing>` on
701 the argument or return value do not alias pointer values which are
702 not *based* on it, ignoring certain "irrelevant" dependencies. For a
703 call to the parent function, dependencies between memory references
704 from before or after the call and from those during the call are
705 "irrelevant" to the ``noalias`` keyword for the arguments and return
706 value used in that call. The caller shares the responsibility with
707 the callee for ensuring that these requirements are met. For further
708 details, please see the discussion of the NoAlias response in `alias
709 analysis <AliasAnalysis.html#MustMayNo>`_.
711 Note that this definition of ``noalias`` is intentionally similar
712 to the definition of ``restrict`` in C99 for function arguments,
713 though it is slightly weaker.
715 For function return values, C99's ``restrict`` is not meaningful,
716 while LLVM's ``noalias`` is.
718 This indicates that the callee does not make any copies of the
719 pointer that outlive the callee itself. This is not a valid
720 attribute for return values.
725 This indicates that the pointer parameter can be excised using the
726 :ref:`trampoline intrinsics <int_trampoline>`. This is not a valid
727 attribute for return values and can only be applied to one parameter.
730 This indicates that the value of the function always returns the value
731 of the parameter as its return value. This is an optimization hint to
732 the code generator when generating the caller, allowing tail call
733 optimization and omission of register saves and restores in some cases;
734 it is not checked or enforced when generating the callee. The parameter
735 and the function return type must be valid operands for the
736 :ref:`bitcast instruction <i_bitcast>`. This is not a valid attribute for
737 return values and can only be applied to one parameter.
741 Garbage Collector Names
742 -----------------------
744 Each function may specify a garbage collector name, which is simply a
749 define void @f() gc "name" { ... }
751 The compiler declares the supported values of *name*. Specifying a
752 collector which will cause the compiler to alter its output in order to
753 support the named garbage collection algorithm.
760 Attribute groups are groups of attributes that are referenced by objects within
761 the IR. They are important for keeping ``.ll`` files readable, because a lot of
762 functions will use the same set of attributes. In the degenerative case of a
763 ``.ll`` file that corresponds to a single ``.c`` file, the single attribute
764 group will capture the important command line flags used to build that file.
766 An attribute group is a module-level object. To use an attribute group, an
767 object references the attribute group's ID (e.g. ``#37``). An object may refer
768 to more than one attribute group. In that situation, the attributes from the
769 different groups are merged.
771 Here is an example of attribute groups for a function that should always be
772 inlined, has a stack alignment of 4, and which shouldn't use SSE instructions:
776 ; Target-independent attributes:
777 attributes #0 = { alwaysinline alignstack=4 }
779 ; Target-dependent attributes:
780 attributes #1 = { "no-sse" }
782 ; Function @f has attributes: alwaysinline, alignstack=4, and "no-sse".
783 define void @f() #0 #1 { ... }
790 Function attributes are set to communicate additional information about
791 a function. Function attributes are considered to be part of the
792 function, not of the function type, so functions with different function
793 attributes can have the same function type.
795 Function attributes are simple keywords that follow the type specified.
796 If multiple attributes are needed, they are space separated. For
801 define void @f() noinline { ... }
802 define void @f() alwaysinline { ... }
803 define void @f() alwaysinline optsize { ... }
804 define void @f() optsize { ... }
807 This attribute indicates that, when emitting the prologue and
808 epilogue, the backend should forcibly align the stack pointer.
809 Specify the desired alignment, which must be a power of two, in
812 This attribute indicates that the inliner should attempt to inline
813 this function into callers whenever possible, ignoring any active
814 inlining size threshold for this caller.
816 This attribute suppresses lazy symbol binding for the function. This
817 may make calls to the function faster, at the cost of extra program
818 startup time if the function is not called during program startup.
820 This attribute indicates that the source code contained a hint that
821 inlining this function is desirable (such as the "inline" keyword in
822 C/C++). It is just a hint; it imposes no requirements on the
825 This attribute disables prologue / epilogue emission for the
826 function. This can have very system-specific consequences.
828 This indicates that the callee function at a call site is not
829 recognized as a built-in function. LLVM will retain the original call
830 and not replace it with equivalent code based on the semantics of the
831 built-in function. This is only valid at call sites, not on function
832 declarations or definitions.
834 This attribute indicates that calls to the function cannot be
835 duplicated. A call to a ``noduplicate`` function may be moved
836 within its parent function, but may not be duplicated within
839 A function containing a ``noduplicate`` call may still
840 be an inlining candidate, provided that the call is not
841 duplicated by inlining. That implies that the function has
842 internal linkage and only has one call site, so the original
843 call is dead after inlining.
845 This attributes disables implicit floating point instructions.
847 This attribute indicates that the inliner should never inline this
848 function in any situation. This attribute may not be used together
849 with the ``alwaysinline`` attribute.
851 This attribute indicates that the code generator should not use a
852 red zone, even if the target-specific ABI normally permits it.
854 This function attribute indicates that the function never returns
855 normally. This produces undefined behavior at runtime if the
856 function ever does dynamically return.
858 This function attribute indicates that the function never returns
859 with an unwind or exceptional control flow. If the function does
860 unwind, its runtime behavior is undefined.
862 This attribute suggests that optimization passes and code generator
863 passes make choices that keep the code size of this function low,
864 and otherwise do optimizations specifically to reduce code size.
866 This attribute indicates that the function computes its result (or
867 decides to unwind an exception) based strictly on its arguments,
868 without dereferencing any pointer arguments or otherwise accessing
869 any mutable state (e.g. memory, control registers, etc) visible to
870 caller functions. It does not write through any pointer arguments
871 (including ``byval`` arguments) and never changes any state visible
872 to callers. This means that it cannot unwind exceptions by calling
873 the ``C++`` exception throwing methods.
875 This attribute indicates that the function does not write through
876 any pointer arguments (including ``byval`` arguments) or otherwise
877 modify any state (e.g. memory, control registers, etc) visible to
878 caller functions. It may dereference pointer arguments and read
879 state that may be set in the caller. A readonly function always
880 returns the same value (or unwinds an exception identically) when
881 called with the same set of arguments and global state. It cannot
882 unwind an exception by calling the ``C++`` exception throwing
885 This attribute indicates that this function can return twice. The C
886 ``setjmp`` is an example of such a function. The compiler disables
887 some optimizations (like tail calls) in the caller of these
890 This attribute indicates that AddressSanitizer checks
891 (dynamic address safety analysis) are enabled for this function.
893 This attribute indicates that MemorySanitizer checks (dynamic detection
894 of accesses to uninitialized memory) are enabled for this function.
896 This attribute indicates that ThreadSanitizer checks
897 (dynamic thread safety analysis) are enabled for this function.
899 This attribute indicates that the function should emit a stack
900 smashing protector. It is in the form of a "canary" --- a random value
901 placed on the stack before the local variables that's checked upon
902 return from the function to see if it has been overwritten. A
903 heuristic is used to determine if a function needs stack protectors
904 or not. The heuristic used will enable protectors for functions with:
906 - Character arrays larger than ``ssp-buffer-size`` (default 8).
907 - Aggregates containing character arrays larger than ``ssp-buffer-size``.
908 - Calls to alloca() with variable sizes or constant sizes greater than
911 If a function that has an ``ssp`` attribute is inlined into a
912 function that doesn't have an ``ssp`` attribute, then the resulting
913 function will have an ``ssp`` attribute.
915 This attribute indicates that the function should *always* emit a
916 stack smashing protector. This overrides the ``ssp`` function
919 If a function that has an ``sspreq`` attribute is inlined into a
920 function that doesn't have an ``sspreq`` attribute or which has an
921 ``ssp`` or ``sspstrong`` attribute, then the resulting function will have
922 an ``sspreq`` attribute.
924 This attribute indicates that the function should emit a stack smashing
925 protector. This attribute causes a strong heuristic to be used when
926 determining if a function needs stack protectors. The strong heuristic
927 will enable protectors for functions with:
929 - Arrays of any size and type
930 - Aggregates containing an array of any size and type.
932 - Local variables that have had their address taken.
934 This overrides the ``ssp`` function attribute.
936 If a function that has an ``sspstrong`` attribute is inlined into a
937 function that doesn't have an ``sspstrong`` attribute, then the
938 resulting function will have an ``sspstrong`` attribute.
940 This attribute indicates that the ABI being targeted requires that
941 an unwind table entry be produce for this function even if we can
942 show that no exceptions passes by it. This is normally the case for
943 the ELF x86-64 abi, but it can be disabled for some compilation
948 Module-Level Inline Assembly
949 ----------------------------
951 Modules may contain "module-level inline asm" blocks, which corresponds
952 to the GCC "file scope inline asm" blocks. These blocks are internally
953 concatenated by LLVM and treated as a single unit, but may be separated
954 in the ``.ll`` file if desired. The syntax is very simple:
958 module asm "inline asm code goes here"
959 module asm "more can go here"
961 The strings can contain any character by escaping non-printable
962 characters. The escape sequence used is simply "\\xx" where "xx" is the
963 two digit hex code for the number.
965 The inline asm code is simply printed to the machine code .s file when
966 assembly code is generated.
971 A module may specify a target specific data layout string that specifies
972 how data is to be laid out in memory. The syntax for the data layout is
977 target datalayout = "layout specification"
979 The *layout specification* consists of a list of specifications
980 separated by the minus sign character ('-'). Each specification starts
981 with a letter and may include other information after the letter to
982 define some aspect of the data layout. The specifications accepted are
986 Specifies that the target lays out data in big-endian form. That is,
987 the bits with the most significance have the lowest address
990 Specifies that the target lays out data in little-endian form. That
991 is, the bits with the least significance have the lowest address
994 Specifies the natural alignment of the stack in bits. Alignment
995 promotion of stack variables is limited to the natural stack
996 alignment to avoid dynamic stack realignment. The stack alignment
997 must be a multiple of 8-bits. If omitted, the natural stack
998 alignment defaults to "unspecified", which does not prevent any
999 alignment promotions.
1000 ``p[n]:<size>:<abi>:<pref>``
1001 This specifies the *size* of a pointer and its ``<abi>`` and
1002 ``<pref>``\erred alignments for address space ``n``. All sizes are in
1003 bits. Specifying the ``<pref>`` alignment is optional. If omitted, the
1004 preceding ``:`` should be omitted too. The address space, ``n`` is
1005 optional, and if not specified, denotes the default address space 0.
1006 The value of ``n`` must be in the range [1,2^23).
1007 ``i<size>:<abi>:<pref>``
1008 This specifies the alignment for an integer type of a given bit
1009 ``<size>``. The value of ``<size>`` must be in the range [1,2^23).
1010 ``v<size>:<abi>:<pref>``
1011 This specifies the alignment for a vector type of a given bit
1013 ``f<size>:<abi>:<pref>``
1014 This specifies the alignment for a floating point type of a given bit
1015 ``<size>``. Only values of ``<size>`` that are supported by the target
1016 will work. 32 (float) and 64 (double) are supported on all targets; 80
1017 or 128 (different flavors of long double) are also supported on some
1019 ``a<size>:<abi>:<pref>``
1020 This specifies the alignment for an aggregate type of a given bit
1022 ``s<size>:<abi>:<pref>``
1023 This specifies the alignment for a stack object of a given bit
1025 ``n<size1>:<size2>:<size3>...``
1026 This specifies a set of native integer widths for the target CPU in
1027 bits. For example, it might contain ``n32`` for 32-bit PowerPC,
1028 ``n32:64`` for PowerPC 64, or ``n8:16:32:64`` for X86-64. Elements of
1029 this set are considered to support most general arithmetic operations
1032 When constructing the data layout for a given target, LLVM starts with a
1033 default set of specifications which are then (possibly) overridden by
1034 the specifications in the ``datalayout`` keyword. The default
1035 specifications are given in this list:
1037 - ``E`` - big endian
1038 - ``p:64:64:64`` - 64-bit pointers with 64-bit alignment
1039 - ``S0`` - natural stack alignment is unspecified
1040 - ``i1:8:8`` - i1 is 8-bit (byte) aligned
1041 - ``i8:8:8`` - i8 is 8-bit (byte) aligned
1042 - ``i16:16:16`` - i16 is 16-bit aligned
1043 - ``i32:32:32`` - i32 is 32-bit aligned
1044 - ``i64:32:64`` - i64 has ABI alignment of 32-bits but preferred
1045 alignment of 64-bits
1046 - ``f16:16:16`` - half is 16-bit aligned
1047 - ``f32:32:32`` - float is 32-bit aligned
1048 - ``f64:64:64`` - double is 64-bit aligned
1049 - ``f128:128:128`` - quad is 128-bit aligned
1050 - ``v64:64:64`` - 64-bit vector is 64-bit aligned
1051 - ``v128:128:128`` - 128-bit vector is 128-bit aligned
1052 - ``a0:0:64`` - aggregates are 64-bit aligned
1054 When LLVM is determining the alignment for a given type, it uses the
1057 #. If the type sought is an exact match for one of the specifications,
1058 that specification is used.
1059 #. If no match is found, and the type sought is an integer type, then
1060 the smallest integer type that is larger than the bitwidth of the
1061 sought type is used. If none of the specifications are larger than
1062 the bitwidth then the largest integer type is used. For example,
1063 given the default specifications above, the i7 type will use the
1064 alignment of i8 (next largest) while both i65 and i256 will use the
1065 alignment of i64 (largest specified).
1066 #. If no match is found, and the type sought is a vector type, then the
1067 largest vector type that is smaller than the sought vector type will
1068 be used as a fall back. This happens because <128 x double> can be
1069 implemented in terms of 64 <2 x double>, for example.
1071 The function of the data layout string may not be what you expect.
1072 Notably, this is not a specification from the frontend of what alignment
1073 the code generator should use.
1075 Instead, if specified, the target data layout is required to match what
1076 the ultimate *code generator* expects. This string is used by the
1077 mid-level optimizers to improve code, and this only works if it matches
1078 what the ultimate code generator uses. If you would like to generate IR
1079 that does not embed this target-specific detail into the IR, then you
1080 don't have to specify the string. This will disable some optimizations
1081 that require precise layout information, but this also prevents those
1082 optimizations from introducing target specificity into the IR.
1084 .. _pointeraliasing:
1086 Pointer Aliasing Rules
1087 ----------------------
1089 Any memory access must be done through a pointer value associated with
1090 an address range of the memory access, otherwise the behavior is
1091 undefined. Pointer values are associated with address ranges according
1092 to the following rules:
1094 - A pointer value is associated with the addresses associated with any
1095 value it is *based* on.
1096 - An address of a global variable is associated with the address range
1097 of the variable's storage.
1098 - The result value of an allocation instruction is associated with the
1099 address range of the allocated storage.
1100 - A null pointer in the default address-space is associated with no
1102 - An integer constant other than zero or a pointer value returned from
1103 a function not defined within LLVM may be associated with address
1104 ranges allocated through mechanisms other than those provided by
1105 LLVM. Such ranges shall not overlap with any ranges of addresses
1106 allocated by mechanisms provided by LLVM.
1108 A pointer value is *based* on another pointer value according to the
1111 - A pointer value formed from a ``getelementptr`` operation is *based*
1112 on the first operand of the ``getelementptr``.
1113 - The result value of a ``bitcast`` is *based* on the operand of the
1115 - A pointer value formed by an ``inttoptr`` is *based* on all pointer
1116 values that contribute (directly or indirectly) to the computation of
1117 the pointer's value.
1118 - The "*based* on" relationship is transitive.
1120 Note that this definition of *"based"* is intentionally similar to the
1121 definition of *"based"* in C99, though it is slightly weaker.
1123 LLVM IR does not associate types with memory. The result type of a
1124 ``load`` merely indicates the size and alignment of the memory from
1125 which to load, as well as the interpretation of the value. The first
1126 operand type of a ``store`` similarly only indicates the size and
1127 alignment of the store.
1129 Consequently, type-based alias analysis, aka TBAA, aka
1130 ``-fstrict-aliasing``, is not applicable to general unadorned LLVM IR.
1131 :ref:`Metadata <metadata>` may be used to encode additional information
1132 which specialized optimization passes may use to implement type-based
1137 Volatile Memory Accesses
1138 ------------------------
1140 Certain memory accesses, such as :ref:`load <i_load>`'s,
1141 :ref:`store <i_store>`'s, and :ref:`llvm.memcpy <int_memcpy>`'s may be
1142 marked ``volatile``. The optimizers must not change the number of
1143 volatile operations or change their order of execution relative to other
1144 volatile operations. The optimizers *may* change the order of volatile
1145 operations relative to non-volatile operations. This is not Java's
1146 "volatile" and has no cross-thread synchronization behavior.
1148 IR-level volatile loads and stores cannot safely be optimized into
1149 llvm.memcpy or llvm.memmove intrinsics even when those intrinsics are
1150 flagged volatile. Likewise, the backend should never split or merge
1151 target-legal volatile load/store instructions.
1153 .. admonition:: Rationale
1155 Platforms may rely on volatile loads and stores of natively supported
1156 data width to be executed as single instruction. For example, in C
1157 this holds for an l-value of volatile primitive type with native
1158 hardware support, but not necessarily for aggregate types. The
1159 frontend upholds these expectations, which are intentionally
1160 unspecified in the IR. The rules above ensure that IR transformation
1161 do not violate the frontend's contract with the language.
1165 Memory Model for Concurrent Operations
1166 --------------------------------------
1168 The LLVM IR does not define any way to start parallel threads of
1169 execution or to register signal handlers. Nonetheless, there are
1170 platform-specific ways to create them, and we define LLVM IR's behavior
1171 in their presence. This model is inspired by the C++0x memory model.
1173 For a more informal introduction to this model, see the :doc:`Atomics`.
1175 We define a *happens-before* partial order as the least partial order
1178 - Is a superset of single-thread program order, and
1179 - When a *synchronizes-with* ``b``, includes an edge from ``a`` to
1180 ``b``. *Synchronizes-with* pairs are introduced by platform-specific
1181 techniques, like pthread locks, thread creation, thread joining,
1182 etc., and by atomic instructions. (See also :ref:`Atomic Memory Ordering
1183 Constraints <ordering>`).
1185 Note that program order does not introduce *happens-before* edges
1186 between a thread and signals executing inside that thread.
1188 Every (defined) read operation (load instructions, memcpy, atomic
1189 loads/read-modify-writes, etc.) R reads a series of bytes written by
1190 (defined) write operations (store instructions, atomic
1191 stores/read-modify-writes, memcpy, etc.). For the purposes of this
1192 section, initialized globals are considered to have a write of the
1193 initializer which is atomic and happens before any other read or write
1194 of the memory in question. For each byte of a read R, R\ :sub:`byte`
1195 may see any write to the same byte, except:
1197 - If write\ :sub:`1` happens before write\ :sub:`2`, and
1198 write\ :sub:`2` happens before R\ :sub:`byte`, then
1199 R\ :sub:`byte` does not see write\ :sub:`1`.
1200 - If R\ :sub:`byte` happens before write\ :sub:`3`, then
1201 R\ :sub:`byte` does not see write\ :sub:`3`.
1203 Given that definition, R\ :sub:`byte` is defined as follows:
1205 - If R is volatile, the result is target-dependent. (Volatile is
1206 supposed to give guarantees which can support ``sig_atomic_t`` in
1207 C/C++, and may be used for accesses to addresses which do not behave
1208 like normal memory. It does not generally provide cross-thread
1210 - Otherwise, if there is no write to the same byte that happens before
1211 R\ :sub:`byte`, R\ :sub:`byte` returns ``undef`` for that byte.
1212 - Otherwise, if R\ :sub:`byte` may see exactly one write,
1213 R\ :sub:`byte` returns the value written by that write.
1214 - Otherwise, if R is atomic, and all the writes R\ :sub:`byte` may
1215 see are atomic, it chooses one of the values written. See the :ref:`Atomic
1216 Memory Ordering Constraints <ordering>` section for additional
1217 constraints on how the choice is made.
1218 - Otherwise R\ :sub:`byte` returns ``undef``.
1220 R returns the value composed of the series of bytes it read. This
1221 implies that some bytes within the value may be ``undef`` **without**
1222 the entire value being ``undef``. Note that this only defines the
1223 semantics of the operation; it doesn't mean that targets will emit more
1224 than one instruction to read the series of bytes.
1226 Note that in cases where none of the atomic intrinsics are used, this
1227 model places only one restriction on IR transformations on top of what
1228 is required for single-threaded execution: introducing a store to a byte
1229 which might not otherwise be stored is not allowed in general.
1230 (Specifically, in the case where another thread might write to and read
1231 from an address, introducing a store can change a load that may see
1232 exactly one write into a load that may see multiple writes.)
1236 Atomic Memory Ordering Constraints
1237 ----------------------------------
1239 Atomic instructions (:ref:`cmpxchg <i_cmpxchg>`,
1240 :ref:`atomicrmw <i_atomicrmw>`, :ref:`fence <i_fence>`,
1241 :ref:`atomic load <i_load>`, and :ref:`atomic store <i_store>`) take
1242 an ordering parameter that determines which other atomic instructions on
1243 the same address they *synchronize with*. These semantics are borrowed
1244 from Java and C++0x, but are somewhat more colloquial. If these
1245 descriptions aren't precise enough, check those specs (see spec
1246 references in the :doc:`atomics guide <Atomics>`).
1247 :ref:`fence <i_fence>` instructions treat these orderings somewhat
1248 differently since they don't take an address. See that instruction's
1249 documentation for details.
1251 For a simpler introduction to the ordering constraints, see the
1255 The set of values that can be read is governed by the happens-before
1256 partial order. A value cannot be read unless some operation wrote
1257 it. This is intended to provide a guarantee strong enough to model
1258 Java's non-volatile shared variables. This ordering cannot be
1259 specified for read-modify-write operations; it is not strong enough
1260 to make them atomic in any interesting way.
1262 In addition to the guarantees of ``unordered``, there is a single
1263 total order for modifications by ``monotonic`` operations on each
1264 address. All modification orders must be compatible with the
1265 happens-before order. There is no guarantee that the modification
1266 orders can be combined to a global total order for the whole program
1267 (and this often will not be possible). The read in an atomic
1268 read-modify-write operation (:ref:`cmpxchg <i_cmpxchg>` and
1269 :ref:`atomicrmw <i_atomicrmw>`) reads the value in the modification
1270 order immediately before the value it writes. If one atomic read
1271 happens before another atomic read of the same address, the later
1272 read must see the same value or a later value in the address's
1273 modification order. This disallows reordering of ``monotonic`` (or
1274 stronger) operations on the same address. If an address is written
1275 ``monotonic``-ally by one thread, and other threads ``monotonic``-ally
1276 read that address repeatedly, the other threads must eventually see
1277 the write. This corresponds to the C++0x/C1x
1278 ``memory_order_relaxed``.
1280 In addition to the guarantees of ``monotonic``, a
1281 *synchronizes-with* edge may be formed with a ``release`` operation.
1282 This is intended to model C++'s ``memory_order_acquire``.
1284 In addition to the guarantees of ``monotonic``, if this operation
1285 writes a value which is subsequently read by an ``acquire``
1286 operation, it *synchronizes-with* that operation. (This isn't a
1287 complete description; see the C++0x definition of a release
1288 sequence.) This corresponds to the C++0x/C1x
1289 ``memory_order_release``.
1290 ``acq_rel`` (acquire+release)
1291 Acts as both an ``acquire`` and ``release`` operation on its
1292 address. This corresponds to the C++0x/C1x ``memory_order_acq_rel``.
1293 ``seq_cst`` (sequentially consistent)
1294 In addition to the guarantees of ``acq_rel`` (``acquire`` for an
1295 operation which only reads, ``release`` for an operation which only
1296 writes), there is a global total order on all
1297 sequentially-consistent operations on all addresses, which is
1298 consistent with the *happens-before* partial order and with the
1299 modification orders of all the affected addresses. Each
1300 sequentially-consistent read sees the last preceding write to the
1301 same address in this global order. This corresponds to the C++0x/C1x
1302 ``memory_order_seq_cst`` and Java volatile.
1306 If an atomic operation is marked ``singlethread``, it only *synchronizes
1307 with* or participates in modification and seq\_cst total orderings with
1308 other operations running in the same thread (for example, in signal
1316 LLVM IR floating-point binary ops (:ref:`fadd <i_fadd>`,
1317 :ref:`fsub <i_fsub>`, :ref:`fmul <i_fmul>`, :ref:`fdiv <i_fdiv>`,
1318 :ref:`frem <i_frem>`) have the following flags that can set to enable
1319 otherwise unsafe floating point operations
1322 No NaNs - Allow optimizations to assume the arguments and result are not
1323 NaN. Such optimizations are required to retain defined behavior over
1324 NaNs, but the value of the result is undefined.
1327 No Infs - Allow optimizations to assume the arguments and result are not
1328 +/-Inf. Such optimizations are required to retain defined behavior over
1329 +/-Inf, but the value of the result is undefined.
1332 No Signed Zeros - Allow optimizations to treat the sign of a zero
1333 argument or result as insignificant.
1336 Allow Reciprocal - Allow optimizations to use the reciprocal of an
1337 argument rather than perform division.
1340 Fast - Allow algebraically equivalent transformations that may
1341 dramatically change results in floating point (e.g. reassociate). This
1342 flag implies all the others.
1349 The LLVM type system is one of the most important features of the
1350 intermediate representation. Being typed enables a number of
1351 optimizations to be performed on the intermediate representation
1352 directly, without having to do extra analyses on the side before the
1353 transformation. A strong type system makes it easier to read the
1354 generated code and enables novel analyses and transformations that are
1355 not feasible to perform on normal three address code representations.
1357 Type Classifications
1358 --------------------
1360 The types fall into a few useful classifications:
1369 * - :ref:`integer <t_integer>`
1370 - ``i1``, ``i2``, ``i3``, ... ``i8``, ... ``i16``, ... ``i32``, ...
1373 * - :ref:`floating point <t_floating>`
1374 - ``half``, ``float``, ``double``, ``x86_fp80``, ``fp128``,
1382 - :ref:`integer <t_integer>`, :ref:`floating point <t_floating>`,
1383 :ref:`pointer <t_pointer>`, :ref:`vector <t_vector>`,
1384 :ref:`structure <t_struct>`, :ref:`array <t_array>`,
1385 :ref:`label <t_label>`, :ref:`metadata <t_metadata>`.
1387 * - :ref:`primitive <t_primitive>`
1388 - :ref:`label <t_label>`,
1389 :ref:`void <t_void>`,
1390 :ref:`integer <t_integer>`,
1391 :ref:`floating point <t_floating>`,
1392 :ref:`x86mmx <t_x86mmx>`,
1393 :ref:`metadata <t_metadata>`.
1395 * - :ref:`derived <t_derived>`
1396 - :ref:`array <t_array>`,
1397 :ref:`function <t_function>`,
1398 :ref:`pointer <t_pointer>`,
1399 :ref:`structure <t_struct>`,
1400 :ref:`vector <t_vector>`,
1401 :ref:`opaque <t_opaque>`.
1403 The :ref:`first class <t_firstclass>` types are perhaps the most important.
1404 Values of these types are the only ones which can be produced by
1412 The primitive types are the fundamental building blocks of the LLVM
1423 The integer type is a very simple type that simply specifies an
1424 arbitrary bit width for the integer type desired. Any bit width from 1
1425 bit to 2\ :sup:`23`\ -1 (about 8 million) can be specified.
1434 The number of bits the integer will occupy is specified by the ``N``
1440 +----------------+------------------------------------------------+
1441 | ``i1`` | a single-bit integer. |
1442 +----------------+------------------------------------------------+
1443 | ``i32`` | a 32-bit integer. |
1444 +----------------+------------------------------------------------+
1445 | ``i1942652`` | a really big integer of over 1 million bits. |
1446 +----------------+------------------------------------------------+
1450 Floating Point Types
1451 ^^^^^^^^^^^^^^^^^^^^
1460 - 16-bit floating point value
1463 - 32-bit floating point value
1466 - 64-bit floating point value
1469 - 128-bit floating point value (112-bit mantissa)
1472 - 80-bit floating point value (X87)
1475 - 128-bit floating point value (two 64-bits)
1485 The x86mmx type represents a value held in an MMX register on an x86
1486 machine. The operations allowed on it are quite limited: parameters and
1487 return values, load and store, and bitcast. User-specified MMX
1488 instructions are represented as intrinsic or asm calls with arguments
1489 and/or results of this type. There are no arrays, vectors or constants
1507 The void type does not represent any value and has no size.
1524 The label type represents code labels.
1541 The metadata type represents embedded metadata. No derived types may be
1542 created from metadata except for :ref:`function <t_function>` arguments.
1556 The real power in LLVM comes from the derived types in the system. This
1557 is what allows a programmer to represent arrays, functions, pointers,
1558 and other useful types. Each of these types contain one or more element
1559 types which may be a primitive type, or another derived type. For
1560 example, it is possible to have a two dimensional array, using an array
1561 as the element type of another array.
1568 Aggregate Types are a subset of derived types that can contain multiple
1569 member types. :ref:`Arrays <t_array>` and :ref:`structs <t_struct>` are
1570 aggregate types. :ref:`Vectors <t_vector>` are not considered to be
1581 The array type is a very simple derived type that arranges elements
1582 sequentially in memory. The array type requires a size (number of
1583 elements) and an underlying data type.
1590 [<# elements> x <elementtype>]
1592 The number of elements is a constant integer value; ``elementtype`` may
1593 be any type with a size.
1598 +------------------+--------------------------------------+
1599 | ``[40 x i32]`` | Array of 40 32-bit integer values. |
1600 +------------------+--------------------------------------+
1601 | ``[41 x i32]`` | Array of 41 32-bit integer values. |
1602 +------------------+--------------------------------------+
1603 | ``[4 x i8]`` | Array of 4 8-bit integer values. |
1604 +------------------+--------------------------------------+
1606 Here are some examples of multidimensional arrays:
1608 +-----------------------------+----------------------------------------------------------+
1609 | ``[3 x [4 x i32]]`` | 3x4 array of 32-bit integer values. |
1610 +-----------------------------+----------------------------------------------------------+
1611 | ``[12 x [10 x float]]`` | 12x10 array of single precision floating point values. |
1612 +-----------------------------+----------------------------------------------------------+
1613 | ``[2 x [3 x [4 x i16]]]`` | 2x3x4 array of 16-bit integer values. |
1614 +-----------------------------+----------------------------------------------------------+
1616 There is no restriction on indexing beyond the end of the array implied
1617 by a static type (though there are restrictions on indexing beyond the
1618 bounds of an allocated object in some cases). This means that
1619 single-dimension 'variable sized array' addressing can be implemented in
1620 LLVM with a zero length array type. An implementation of 'pascal style
1621 arrays' in LLVM could use the type "``{ i32, [0 x float]}``", for
1632 The function type can be thought of as a function signature. It consists
1633 of a return type and a list of formal parameter types. The return type
1634 of a function type is a first class type or a void type.
1641 <returntype> (<parameter list>)
1643 ...where '``<parameter list>``' is a comma-separated list of type
1644 specifiers. Optionally, the parameter list may include a type ``...``,
1645 which indicates that the function takes a variable number of arguments.
1646 Variable argument functions can access their arguments with the
1647 :ref:`variable argument handling intrinsic <int_varargs>` functions.
1648 '``<returntype>``' is any type except :ref:`label <t_label>`.
1653 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1654 | ``i32 (i32)`` | function taking an ``i32``, returning an ``i32`` |
1655 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1656 | ``float (i16, i32 *) *`` | :ref:`Pointer <t_pointer>` to a function that takes an ``i16`` and a :ref:`pointer <t_pointer>` to ``i32``, returning ``float``. |
1657 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1658 | ``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. |
1659 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1660 | ``{i32, i32} (i32)`` | A function taking an ``i32``, returning a :ref:`structure <t_struct>` containing two ``i32`` values |
1661 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1671 The structure type is used to represent a collection of data members
1672 together in memory. The elements of a structure may be any type that has
1675 Structures in memory are accessed using '``load``' and '``store``' by
1676 getting a pointer to a field with the '``getelementptr``' instruction.
1677 Structures in registers are accessed using the '``extractvalue``' and
1678 '``insertvalue``' instructions.
1680 Structures may optionally be "packed" structures, which indicate that
1681 the alignment of the struct is one byte, and that there is no padding
1682 between the elements. In non-packed structs, padding between field types
1683 is inserted as defined by the DataLayout string in the module, which is
1684 required to match what the underlying code generator expects.
1686 Structures can either be "literal" or "identified". A literal structure
1687 is defined inline with other types (e.g. ``{i32, i32}*``) whereas
1688 identified types are always defined at the top level with a name.
1689 Literal types are uniqued by their contents and can never be recursive
1690 or opaque since there is no way to write one. Identified types can be
1691 recursive, can be opaqued, and are never uniqued.
1698 %T1 = type { <type list> } ; Identified normal struct type
1699 %T2 = type <{ <type list> }> ; Identified packed struct type
1704 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1705 | ``{ i32, i32, i32 }`` | A triple of three ``i32`` values |
1706 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1707 | ``{ 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``. |
1708 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1709 | ``<{ i8, i32 }>`` | A packed struct known to be 5 bytes in size. |
1710 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1714 Opaque Structure Types
1715 ^^^^^^^^^^^^^^^^^^^^^^
1720 Opaque structure types are used to represent named structure types that
1721 do not have a body specified. This corresponds (for example) to the C
1722 notion of a forward declared structure.
1735 +--------------+-------------------+
1736 | ``opaque`` | An opaque type. |
1737 +--------------+-------------------+
1747 The pointer type is used to specify memory locations. Pointers are
1748 commonly used to reference objects in memory.
1750 Pointer types may have an optional address space attribute defining the
1751 numbered address space where the pointed-to object resides. The default
1752 address space is number zero. The semantics of non-zero address spaces
1753 are target-specific.
1755 Note that LLVM does not permit pointers to void (``void*``) nor does it
1756 permit pointers to labels (``label*``). Use ``i8*`` instead.
1768 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1769 | ``[4 x i32]*`` | A :ref:`pointer <t_pointer>` to :ref:`array <t_array>` of four ``i32`` values. |
1770 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1771 | ``i32 (i32*) *`` | A :ref:`pointer <t_pointer>` to a :ref:`function <t_function>` that takes an ``i32*``, returning an ``i32``. |
1772 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1773 | ``i32 addrspace(5)*`` | A :ref:`pointer <t_pointer>` to an ``i32`` value that resides in address space #5. |
1774 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1784 A vector type is a simple derived type that represents a vector of
1785 elements. Vector types are used when multiple primitive data are
1786 operated in parallel using a single instruction (SIMD). A vector type
1787 requires a size (number of elements) and an underlying primitive data
1788 type. Vector types are considered :ref:`first class <t_firstclass>`.
1795 < <# elements> x <elementtype> >
1797 The number of elements is a constant integer value larger than 0;
1798 elementtype may be any integer or floating point type, or a pointer to
1799 these types. Vectors of size zero are not allowed.
1804 +-------------------+--------------------------------------------------+
1805 | ``<4 x i32>`` | Vector of 4 32-bit integer values. |
1806 +-------------------+--------------------------------------------------+
1807 | ``<8 x float>`` | Vector of 8 32-bit floating-point values. |
1808 +-------------------+--------------------------------------------------+
1809 | ``<2 x i64>`` | Vector of 2 64-bit integer values. |
1810 +-------------------+--------------------------------------------------+
1811 | ``<4 x i64*>`` | Vector of 4 pointers to 64-bit integer values. |
1812 +-------------------+--------------------------------------------------+
1817 LLVM has several different basic types of constants. This section
1818 describes them all and their syntax.
1823 **Boolean constants**
1824 The two strings '``true``' and '``false``' are both valid constants
1826 **Integer constants**
1827 Standard integers (such as '4') are constants of the
1828 :ref:`integer <t_integer>` type. Negative numbers may be used with
1830 **Floating point constants**
1831 Floating point constants use standard decimal notation (e.g.
1832 123.421), exponential notation (e.g. 1.23421e+2), or a more precise
1833 hexadecimal notation (see below). The assembler requires the exact
1834 decimal value of a floating-point constant. For example, the
1835 assembler accepts 1.25 but rejects 1.3 because 1.3 is a repeating
1836 decimal in binary. Floating point constants must have a :ref:`floating
1837 point <t_floating>` type.
1838 **Null pointer constants**
1839 The identifier '``null``' is recognized as a null pointer constant
1840 and must be of :ref:`pointer type <t_pointer>`.
1842 The one non-intuitive notation for constants is the hexadecimal form of
1843 floating point constants. For example, the form
1844 '``double 0x432ff973cafa8000``' is equivalent to (but harder to read
1845 than) '``double 4.5e+15``'. The only time hexadecimal floating point
1846 constants are required (and the only time that they are generated by the
1847 disassembler) is when a floating point constant must be emitted but it
1848 cannot be represented as a decimal floating point number in a reasonable
1849 number of digits. For example, NaN's, infinities, and other special
1850 values are represented in their IEEE hexadecimal format so that assembly
1851 and disassembly do not cause any bits to change in the constants.
1853 When using the hexadecimal form, constants of types half, float, and
1854 double are represented using the 16-digit form shown above (which
1855 matches the IEEE754 representation for double); half and float values
1856 must, however, be exactly representable as IEEE 754 half and single
1857 precision, respectively. Hexadecimal format is always used for long
1858 double, and there are three forms of long double. The 80-bit format used
1859 by x86 is represented as ``0xK`` followed by 20 hexadecimal digits. The
1860 128-bit format used by PowerPC (two adjacent doubles) is represented by
1861 ``0xM`` followed by 32 hexadecimal digits. The IEEE 128-bit format is
1862 represented by ``0xL`` followed by 32 hexadecimal digits. Long doubles
1863 will only work if they match the long double format on your target.
1864 The IEEE 16-bit format (half precision) is represented by ``0xH``
1865 followed by 4 hexadecimal digits. All hexadecimal formats are big-endian
1866 (sign bit at the left).
1868 There are no constants of type x86mmx.
1873 Complex constants are a (potentially recursive) combination of simple
1874 constants and smaller complex constants.
1876 **Structure constants**
1877 Structure constants are represented with notation similar to
1878 structure type definitions (a comma separated list of elements,
1879 surrounded by braces (``{}``)). For example:
1880 "``{ i32 4, float 17.0, i32* @G }``", where "``@G``" is declared as
1881 "``@G = external global i32``". Structure constants must have
1882 :ref:`structure type <t_struct>`, and the number and types of elements
1883 must match those specified by the type.
1885 Array constants are represented with notation similar to array type
1886 definitions (a comma separated list of elements, surrounded by
1887 square brackets (``[]``)). For example:
1888 "``[ i32 42, i32 11, i32 74 ]``". Array constants must have
1889 :ref:`array type <t_array>`, and the number and types of elements must
1890 match those specified by the type.
1891 **Vector constants**
1892 Vector constants are represented with notation similar to vector
1893 type definitions (a comma separated list of elements, surrounded by
1894 less-than/greater-than's (``<>``)). For example:
1895 "``< i32 42, i32 11, i32 74, i32 100 >``". Vector constants
1896 must have :ref:`vector type <t_vector>`, and the number and types of
1897 elements must match those specified by the type.
1898 **Zero initialization**
1899 The string '``zeroinitializer``' can be used to zero initialize a
1900 value to zero of *any* type, including scalar and
1901 :ref:`aggregate <t_aggregate>` types. This is often used to avoid
1902 having to print large zero initializers (e.g. for large arrays) and
1903 is always exactly equivalent to using explicit zero initializers.
1905 A metadata node is a structure-like constant with :ref:`metadata
1906 type <t_metadata>`. For example:
1907 "``metadata !{ i32 0, metadata !"test" }``". Unlike other
1908 constants that are meant to be interpreted as part of the
1909 instruction stream, metadata is a place to attach additional
1910 information such as debug info.
1912 Global Variable and Function Addresses
1913 --------------------------------------
1915 The addresses of :ref:`global variables <globalvars>` and
1916 :ref:`functions <functionstructure>` are always implicitly valid
1917 (link-time) constants. These constants are explicitly referenced when
1918 the :ref:`identifier for the global <identifiers>` is used and always have
1919 :ref:`pointer <t_pointer>` type. For example, the following is a legal LLVM
1922 .. code-block:: llvm
1926 @Z = global [2 x i32*] [ i32* @X, i32* @Y ]
1933 The string '``undef``' can be used anywhere a constant is expected, and
1934 indicates that the user of the value may receive an unspecified
1935 bit-pattern. Undefined values may be of any type (other than '``label``'
1936 or '``void``') and be used anywhere a constant is permitted.
1938 Undefined values are useful because they indicate to the compiler that
1939 the program is well defined no matter what value is used. This gives the
1940 compiler more freedom to optimize. Here are some examples of
1941 (potentially surprising) transformations that are valid (in pseudo IR):
1943 .. code-block:: llvm
1953 This is safe because all of the output bits are affected by the undef
1954 bits. Any output bit can have a zero or one depending on the input bits.
1956 .. code-block:: llvm
1967 These logical operations have bits that are not always affected by the
1968 input. For example, if ``%X`` has a zero bit, then the output of the
1969 '``and``' operation will always be a zero for that bit, no matter what
1970 the corresponding bit from the '``undef``' is. As such, it is unsafe to
1971 optimize or assume that the result of the '``and``' is '``undef``'.
1972 However, it is safe to assume that all bits of the '``undef``' could be
1973 0, and optimize the '``and``' to 0. Likewise, it is safe to assume that
1974 all the bits of the '``undef``' operand to the '``or``' could be set,
1975 allowing the '``or``' to be folded to -1.
1977 .. code-block:: llvm
1979 %A = select undef, %X, %Y
1980 %B = select undef, 42, %Y
1981 %C = select %X, %Y, undef
1991 This set of examples shows that undefined '``select``' (and conditional
1992 branch) conditions can go *either way*, but they have to come from one
1993 of the two operands. In the ``%A`` example, if ``%X`` and ``%Y`` were
1994 both known to have a clear low bit, then ``%A`` would have to have a
1995 cleared low bit. However, in the ``%C`` example, the optimizer is
1996 allowed to assume that the '``undef``' operand could be the same as
1997 ``%Y``, allowing the whole '``select``' to be eliminated.
1999 .. code-block:: llvm
2001 %A = xor undef, undef
2018 This example points out that two '``undef``' operands are not
2019 necessarily the same. This can be surprising to people (and also matches
2020 C semantics) where they assume that "``X^X``" is always zero, even if
2021 ``X`` is undefined. This isn't true for a number of reasons, but the
2022 short answer is that an '``undef``' "variable" can arbitrarily change
2023 its value over its "live range". This is true because the variable
2024 doesn't actually *have a live range*. Instead, the value is logically
2025 read from arbitrary registers that happen to be around when needed, so
2026 the value is not necessarily consistent over time. In fact, ``%A`` and
2027 ``%C`` need to have the same semantics or the core LLVM "replace all
2028 uses with" concept would not hold.
2030 .. code-block:: llvm
2038 These examples show the crucial difference between an *undefined value*
2039 and *undefined behavior*. An undefined value (like '``undef``') is
2040 allowed to have an arbitrary bit-pattern. This means that the ``%A``
2041 operation can be constant folded to '``undef``', because the '``undef``'
2042 could be an SNaN, and ``fdiv`` is not (currently) defined on SNaN's.
2043 However, in the second example, we can make a more aggressive
2044 assumption: because the ``undef`` is allowed to be an arbitrary value,
2045 we are allowed to assume that it could be zero. Since a divide by zero
2046 has *undefined behavior*, we are allowed to assume that the operation
2047 does not execute at all. This allows us to delete the divide and all
2048 code after it. Because the undefined operation "can't happen", the
2049 optimizer can assume that it occurs in dead code.
2051 .. code-block:: llvm
2053 a: store undef -> %X
2054 b: store %X -> undef
2059 These examples reiterate the ``fdiv`` example: a store *of* an undefined
2060 value can be assumed to not have any effect; we can assume that the
2061 value is overwritten with bits that happen to match what was already
2062 there. However, a store *to* an undefined location could clobber
2063 arbitrary memory, therefore, it has undefined behavior.
2070 Poison values are similar to :ref:`undef values <undefvalues>`, however
2071 they also represent the fact that an instruction or constant expression
2072 which cannot evoke side effects has nevertheless detected a condition
2073 which results in undefined behavior.
2075 There is currently no way of representing a poison value in the IR; they
2076 only exist when produced by operations such as :ref:`add <i_add>` with
2079 Poison value behavior is defined in terms of value *dependence*:
2081 - Values other than :ref:`phi <i_phi>` nodes depend on their operands.
2082 - :ref:`Phi <i_phi>` nodes depend on the operand corresponding to
2083 their dynamic predecessor basic block.
2084 - Function arguments depend on the corresponding actual argument values
2085 in the dynamic callers of their functions.
2086 - :ref:`Call <i_call>` instructions depend on the :ref:`ret <i_ret>`
2087 instructions that dynamically transfer control back to them.
2088 - :ref:`Invoke <i_invoke>` instructions depend on the
2089 :ref:`ret <i_ret>`, :ref:`resume <i_resume>`, or exception-throwing
2090 call instructions that dynamically transfer control back to them.
2091 - Non-volatile loads and stores depend on the most recent stores to all
2092 of the referenced memory addresses, following the order in the IR
2093 (including loads and stores implied by intrinsics such as
2094 :ref:`@llvm.memcpy <int_memcpy>`.)
2095 - An instruction with externally visible side effects depends on the
2096 most recent preceding instruction with externally visible side
2097 effects, following the order in the IR. (This includes :ref:`volatile
2098 operations <volatile>`.)
2099 - An instruction *control-depends* on a :ref:`terminator
2100 instruction <terminators>` if the terminator instruction has
2101 multiple successors and the instruction is always executed when
2102 control transfers to one of the successors, and may not be executed
2103 when control is transferred to another.
2104 - Additionally, an instruction also *control-depends* on a terminator
2105 instruction if the set of instructions it otherwise depends on would
2106 be different if the terminator had transferred control to a different
2108 - Dependence is transitive.
2110 Poison Values have the same behavior as :ref:`undef values <undefvalues>`,
2111 with the additional affect that any instruction which has a *dependence*
2112 on a poison value has undefined behavior.
2114 Here are some examples:
2116 .. code-block:: llvm
2119 %poison = sub nuw i32 0, 1 ; Results in a poison value.
2120 %still_poison = and i32 %poison, 0 ; 0, but also poison.
2121 %poison_yet_again = getelementptr i32* @h, i32 %still_poison
2122 store i32 0, i32* %poison_yet_again ; memory at @h[0] is poisoned
2124 store i32 %poison, i32* @g ; Poison value stored to memory.
2125 %poison2 = load i32* @g ; Poison value loaded back from memory.
2127 store volatile i32 %poison, i32* @g ; External observation; undefined behavior.
2129 %narrowaddr = bitcast i32* @g to i16*
2130 %wideaddr = bitcast i32* @g to i64*
2131 %poison3 = load i16* %narrowaddr ; Returns a poison value.
2132 %poison4 = load i64* %wideaddr ; Returns a poison value.
2134 %cmp = icmp slt i32 %poison, 0 ; Returns a poison value.
2135 br i1 %cmp, label %true, label %end ; Branch to either destination.
2138 store volatile i32 0, i32* @g ; This is control-dependent on %cmp, so
2139 ; it has undefined behavior.
2143 %p = phi i32 [ 0, %entry ], [ 1, %true ]
2144 ; Both edges into this PHI are
2145 ; control-dependent on %cmp, so this
2146 ; always results in a poison value.
2148 store volatile i32 0, i32* @g ; This would depend on the store in %true
2149 ; if %cmp is true, or the store in %entry
2150 ; otherwise, so this is undefined behavior.
2152 br i1 %cmp, label %second_true, label %second_end
2153 ; The same branch again, but this time the
2154 ; true block doesn't have side effects.
2161 store volatile i32 0, i32* @g ; This time, the instruction always depends
2162 ; on the store in %end. Also, it is
2163 ; control-equivalent to %end, so this is
2164 ; well-defined (ignoring earlier undefined
2165 ; behavior in this example).
2169 Addresses of Basic Blocks
2170 -------------------------
2172 ``blockaddress(@function, %block)``
2174 The '``blockaddress``' constant computes the address of the specified
2175 basic block in the specified function, and always has an ``i8*`` type.
2176 Taking the address of the entry block is illegal.
2178 This value only has defined behavior when used as an operand to the
2179 ':ref:`indirectbr <i_indirectbr>`' instruction, or for comparisons
2180 against null. Pointer equality tests between labels addresses results in
2181 undefined behavior --- though, again, comparison against null is ok, and
2182 no label is equal to the null pointer. This may be passed around as an
2183 opaque pointer sized value as long as the bits are not inspected. This
2184 allows ``ptrtoint`` and arithmetic to be performed on these values so
2185 long as the original value is reconstituted before the ``indirectbr``
2188 Finally, some targets may provide defined semantics when using the value
2189 as the operand to an inline assembly, but that is target specific.
2191 Constant Expressions
2192 --------------------
2194 Constant expressions are used to allow expressions involving other
2195 constants to be used as constants. Constant expressions may be of any
2196 :ref:`first class <t_firstclass>` type and may involve any LLVM operation
2197 that does not have side effects (e.g. load and call are not supported).
2198 The following is the syntax for constant expressions:
2200 ``trunc (CST to TYPE)``
2201 Truncate a constant to another type. The bit size of CST must be
2202 larger than the bit size of TYPE. Both types must be integers.
2203 ``zext (CST to TYPE)``
2204 Zero extend a constant to another type. The bit size of CST must be
2205 smaller than the bit size of TYPE. Both types must be integers.
2206 ``sext (CST to TYPE)``
2207 Sign extend a constant to another type. The bit size of CST must be
2208 smaller than the bit size of TYPE. Both types must be integers.
2209 ``fptrunc (CST to TYPE)``
2210 Truncate a floating point constant to another floating point type.
2211 The size of CST must be larger than the size of TYPE. Both types
2212 must be floating point.
2213 ``fpext (CST to TYPE)``
2214 Floating point extend a constant to another type. The size of CST
2215 must be smaller or equal to the size of TYPE. Both types must be
2217 ``fptoui (CST to TYPE)``
2218 Convert a floating point constant to the corresponding unsigned
2219 integer constant. TYPE must be a scalar or vector integer type. CST
2220 must be of scalar or vector floating point type. Both CST and TYPE
2221 must be scalars, or vectors of the same number of elements. If the
2222 value won't fit in the integer type, the results are undefined.
2223 ``fptosi (CST to TYPE)``
2224 Convert a floating point constant to the corresponding signed
2225 integer constant. TYPE must be a scalar or vector integer type. CST
2226 must be of scalar or vector floating point type. Both CST and TYPE
2227 must be scalars, or vectors of the same number of elements. If the
2228 value won't fit in the integer type, the results are undefined.
2229 ``uitofp (CST to TYPE)``
2230 Convert an unsigned integer constant to the corresponding floating
2231 point constant. TYPE must be a scalar or vector floating point type.
2232 CST must be of scalar or vector integer type. Both CST and TYPE must
2233 be scalars, or vectors of the same number of elements. If the value
2234 won't fit in the floating point type, the results are undefined.
2235 ``sitofp (CST to TYPE)``
2236 Convert a signed integer constant to the corresponding floating
2237 point constant. TYPE must be a scalar or vector floating point type.
2238 CST must be of scalar or vector integer type. Both CST and TYPE must
2239 be scalars, or vectors of the same number of elements. If the value
2240 won't fit in the floating point type, the results are undefined.
2241 ``ptrtoint (CST to TYPE)``
2242 Convert a pointer typed constant to the corresponding integer
2243 constant. ``TYPE`` must be an integer type. ``CST`` must be of
2244 pointer type. The ``CST`` value is zero extended, truncated, or
2245 unchanged to make it fit in ``TYPE``.
2246 ``inttoptr (CST to TYPE)``
2247 Convert an integer constant to a pointer constant. TYPE must be a
2248 pointer type. CST must be of integer type. The CST value is zero
2249 extended, truncated, or unchanged to make it fit in a pointer size.
2250 This one is *really* dangerous!
2251 ``bitcast (CST to TYPE)``
2252 Convert a constant, CST, to another TYPE. The constraints of the
2253 operands are the same as those for the :ref:`bitcast
2254 instruction <i_bitcast>`.
2255 ``getelementptr (CSTPTR, IDX0, IDX1, ...)``, ``getelementptr inbounds (CSTPTR, IDX0, IDX1, ...)``
2256 Perform the :ref:`getelementptr operation <i_getelementptr>` on
2257 constants. As with the :ref:`getelementptr <i_getelementptr>`
2258 instruction, the index list may have zero or more indexes, which are
2259 required to make sense for the type of "CSTPTR".
2260 ``select (COND, VAL1, VAL2)``
2261 Perform the :ref:`select operation <i_select>` on constants.
2262 ``icmp COND (VAL1, VAL2)``
2263 Performs the :ref:`icmp operation <i_icmp>` on constants.
2264 ``fcmp COND (VAL1, VAL2)``
2265 Performs the :ref:`fcmp operation <i_fcmp>` on constants.
2266 ``extractelement (VAL, IDX)``
2267 Perform the :ref:`extractelement operation <i_extractelement>` on
2269 ``insertelement (VAL, ELT, IDX)``
2270 Perform the :ref:`insertelement operation <i_insertelement>` on
2272 ``shufflevector (VEC1, VEC2, IDXMASK)``
2273 Perform the :ref:`shufflevector operation <i_shufflevector>` on
2275 ``extractvalue (VAL, IDX0, IDX1, ...)``
2276 Perform the :ref:`extractvalue operation <i_extractvalue>` on
2277 constants. The index list is interpreted in a similar manner as
2278 indices in a ':ref:`getelementptr <i_getelementptr>`' operation. At
2279 least one index value must be specified.
2280 ``insertvalue (VAL, ELT, IDX0, IDX1, ...)``
2281 Perform the :ref:`insertvalue operation <i_insertvalue>` on constants.
2282 The index list is interpreted in a similar manner as indices in a
2283 ':ref:`getelementptr <i_getelementptr>`' operation. At least one index
2284 value must be specified.
2285 ``OPCODE (LHS, RHS)``
2286 Perform the specified operation of the LHS and RHS constants. OPCODE
2287 may be any of the :ref:`binary <binaryops>` or :ref:`bitwise
2288 binary <bitwiseops>` operations. The constraints on operands are
2289 the same as those for the corresponding instruction (e.g. no bitwise
2290 operations on floating point values are allowed).
2295 Inline Assembler Expressions
2296 ----------------------------
2298 LLVM supports inline assembler expressions (as opposed to :ref:`Module-Level
2299 Inline Assembly <moduleasm>`) through the use of a special value. This
2300 value represents the inline assembler as a string (containing the
2301 instructions to emit), a list of operand constraints (stored as a
2302 string), a flag that indicates whether or not the inline asm expression
2303 has side effects, and a flag indicating whether the function containing
2304 the asm needs to align its stack conservatively. An example inline
2305 assembler expression is:
2307 .. code-block:: llvm
2309 i32 (i32) asm "bswap $0", "=r,r"
2311 Inline assembler expressions may **only** be used as the callee operand
2312 of a :ref:`call <i_call>` or an :ref:`invoke <i_invoke>` instruction.
2313 Thus, typically we have:
2315 .. code-block:: llvm
2317 %X = call i32 asm "bswap $0", "=r,r"(i32 %Y)
2319 Inline asms with side effects not visible in the constraint list must be
2320 marked as having side effects. This is done through the use of the
2321 '``sideeffect``' keyword, like so:
2323 .. code-block:: llvm
2325 call void asm sideeffect "eieio", ""()
2327 In some cases inline asms will contain code that will not work unless
2328 the stack is aligned in some way, such as calls or SSE instructions on
2329 x86, yet will not contain code that does that alignment within the asm.
2330 The compiler should make conservative assumptions about what the asm
2331 might contain and should generate its usual stack alignment code in the
2332 prologue if the '``alignstack``' keyword is present:
2334 .. code-block:: llvm
2336 call void asm alignstack "eieio", ""()
2338 Inline asms also support using non-standard assembly dialects. The
2339 assumed dialect is ATT. When the '``inteldialect``' keyword is present,
2340 the inline asm is using the Intel dialect. Currently, ATT and Intel are
2341 the only supported dialects. An example is:
2343 .. code-block:: llvm
2345 call void asm inteldialect "eieio", ""()
2347 If multiple keywords appear the '``sideeffect``' keyword must come
2348 first, the '``alignstack``' keyword second and the '``inteldialect``'
2354 The call instructions that wrap inline asm nodes may have a
2355 "``!srcloc``" MDNode attached to it that contains a list of constant
2356 integers. If present, the code generator will use the integer as the
2357 location cookie value when report errors through the ``LLVMContext``
2358 error reporting mechanisms. This allows a front-end to correlate backend
2359 errors that occur with inline asm back to the source code that produced
2362 .. code-block:: llvm
2364 call void asm sideeffect "something bad", ""(), !srcloc !42
2366 !42 = !{ i32 1234567 }
2368 It is up to the front-end to make sense of the magic numbers it places
2369 in the IR. If the MDNode contains multiple constants, the code generator
2370 will use the one that corresponds to the line of the asm that the error
2375 Metadata Nodes and Metadata Strings
2376 -----------------------------------
2378 LLVM IR allows metadata to be attached to instructions in the program
2379 that can convey extra information about the code to the optimizers and
2380 code generator. One example application of metadata is source-level
2381 debug information. There are two metadata primitives: strings and nodes.
2382 All metadata has the ``metadata`` type and is identified in syntax by a
2383 preceding exclamation point ('``!``').
2385 A metadata string is a string surrounded by double quotes. It can
2386 contain any character by escaping non-printable characters with
2387 "``\xx``" where "``xx``" is the two digit hex code. For example:
2390 Metadata nodes are represented with notation similar to structure
2391 constants (a comma separated list of elements, surrounded by braces and
2392 preceded by an exclamation point). Metadata nodes can have any values as
2393 their operand. For example:
2395 .. code-block:: llvm
2397 !{ metadata !"test\00", i32 10}
2399 A :ref:`named metadata <namedmetadatastructure>` is a collection of
2400 metadata nodes, which can be looked up in the module symbol table. For
2403 .. code-block:: llvm
2405 !foo = metadata !{!4, !3}
2407 Metadata can be used as function arguments. Here ``llvm.dbg.value``
2408 function is using two metadata arguments:
2410 .. code-block:: llvm
2412 call void @llvm.dbg.value(metadata !24, i64 0, metadata !25)
2414 Metadata can be attached with an instruction. Here metadata ``!21`` is
2415 attached to the ``add`` instruction using the ``!dbg`` identifier:
2417 .. code-block:: llvm
2419 %indvar.next = add i64 %indvar, 1, !dbg !21
2421 More information about specific metadata nodes recognized by the
2422 optimizers and code generator is found below.
2427 In LLVM IR, memory does not have types, so LLVM's own type system is not
2428 suitable for doing TBAA. Instead, metadata is added to the IR to
2429 describe a type system of a higher level language. This can be used to
2430 implement typical C/C++ TBAA, but it can also be used to implement
2431 custom alias analysis behavior for other languages.
2433 The current metadata format is very simple. TBAA metadata nodes have up
2434 to three fields, e.g.:
2436 .. code-block:: llvm
2438 !0 = metadata !{ metadata !"an example type tree" }
2439 !1 = metadata !{ metadata !"int", metadata !0 }
2440 !2 = metadata !{ metadata !"float", metadata !0 }
2441 !3 = metadata !{ metadata !"const float", metadata !2, i64 1 }
2443 The first field is an identity field. It can be any value, usually a
2444 metadata string, which uniquely identifies the type. The most important
2445 name in the tree is the name of the root node. Two trees with different
2446 root node names are entirely disjoint, even if they have leaves with
2449 The second field identifies the type's parent node in the tree, or is
2450 null or omitted for a root node. A type is considered to alias all of
2451 its descendants and all of its ancestors in the tree. Also, a type is
2452 considered to alias all types in other trees, so that bitcode produced
2453 from multiple front-ends is handled conservatively.
2455 If the third field is present, it's an integer which if equal to 1
2456 indicates that the type is "constant" (meaning
2457 ``pointsToConstantMemory`` should return true; see `other useful
2458 AliasAnalysis methods <AliasAnalysis.html#OtherItfs>`_).
2460 '``tbaa.struct``' Metadata
2461 ^^^^^^^^^^^^^^^^^^^^^^^^^^
2463 The :ref:`llvm.memcpy <int_memcpy>` is often used to implement
2464 aggregate assignment operations in C and similar languages, however it
2465 is defined to copy a contiguous region of memory, which is more than
2466 strictly necessary for aggregate types which contain holes due to
2467 padding. Also, it doesn't contain any TBAA information about the fields
2470 ``!tbaa.struct`` metadata can describe which memory subregions in a
2471 memcpy are padding and what the TBAA tags of the struct are.
2473 The current metadata format is very simple. ``!tbaa.struct`` metadata
2474 nodes are a list of operands which are in conceptual groups of three.
2475 For each group of three, the first operand gives the byte offset of a
2476 field in bytes, the second gives its size in bytes, and the third gives
2479 .. code-block:: llvm
2481 !4 = metadata !{ i64 0, i64 4, metadata !1, i64 8, i64 4, metadata !2 }
2483 This describes a struct with two fields. The first is at offset 0 bytes
2484 with size 4 bytes, and has tbaa tag !1. The second is at offset 8 bytes
2485 and has size 4 bytes and has tbaa tag !2.
2487 Note that the fields need not be contiguous. In this example, there is a
2488 4 byte gap between the two fields. This gap represents padding which
2489 does not carry useful data and need not be preserved.
2491 '``fpmath``' Metadata
2492 ^^^^^^^^^^^^^^^^^^^^^
2494 ``fpmath`` metadata may be attached to any instruction of floating point
2495 type. It can be used to express the maximum acceptable error in the
2496 result of that instruction, in ULPs, thus potentially allowing the
2497 compiler to use a more efficient but less accurate method of computing
2498 it. ULP is defined as follows:
2500 If ``x`` is a real number that lies between two finite consecutive
2501 floating-point numbers ``a`` and ``b``, without being equal to one
2502 of them, then ``ulp(x) = |b - a|``, otherwise ``ulp(x)`` is the
2503 distance between the two non-equal finite floating-point numbers
2504 nearest ``x``. Moreover, ``ulp(NaN)`` is ``NaN``.
2506 The metadata node shall consist of a single positive floating point
2507 number representing the maximum relative error, for example:
2509 .. code-block:: llvm
2511 !0 = metadata !{ float 2.5 } ; maximum acceptable inaccuracy is 2.5 ULPs
2513 '``range``' Metadata
2514 ^^^^^^^^^^^^^^^^^^^^
2516 ``range`` metadata may be attached only to loads of integer types. It
2517 expresses the possible ranges the loaded value is in. The ranges are
2518 represented with a flattened list of integers. The loaded value is known
2519 to be in the union of the ranges defined by each consecutive pair. Each
2520 pair has the following properties:
2522 - The type must match the type loaded by the instruction.
2523 - The pair ``a,b`` represents the range ``[a,b)``.
2524 - Both ``a`` and ``b`` are constants.
2525 - The range is allowed to wrap.
2526 - The range should not represent the full or empty set. That is,
2529 In addition, the pairs must be in signed order of the lower bound and
2530 they must be non-contiguous.
2534 .. code-block:: llvm
2536 %a = load i8* %x, align 1, !range !0 ; Can only be 0 or 1
2537 %b = load i8* %y, align 1, !range !1 ; Can only be 255 (-1), 0 or 1
2538 %c = load i8* %z, align 1, !range !2 ; Can only be 0, 1, 3, 4 or 5
2539 %d = load i8* %z, align 1, !range !3 ; Can only be -2, -1, 3, 4 or 5
2541 !0 = metadata !{ i8 0, i8 2 }
2542 !1 = metadata !{ i8 255, i8 2 }
2543 !2 = metadata !{ i8 0, i8 2, i8 3, i8 6 }
2544 !3 = metadata !{ i8 -2, i8 0, i8 3, i8 6 }
2549 It is sometimes useful to attach information to loop constructs. Currently,
2550 loop metadata is implemented as metadata attached to the branch instruction
2551 in the loop latch block. This type of metadata refer to a metadata node that is
2552 guaranteed to be separate for each loop. The loop-level metadata is prefixed
2555 The loop identifier metadata is implemented using a metadata that refers to
2556 itself to avoid merging it with any other identifier metadata, e.g.,
2557 during module linkage or function inlining. That is, each loop should refer
2558 to their own identification metadata even if they reside in separate functions.
2559 The following example contains loop identifier metadata for two separate loop
2562 .. code-block:: llvm
2564 !0 = metadata !{ metadata !0 }
2565 !1 = metadata !{ metadata !1 }
2568 '``llvm.loop.parallel``' Metadata
2569 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2571 This loop metadata can be used to communicate that a loop should be considered
2572 a parallel loop. The semantics of parallel loops in this case is the one
2573 with the strongest cross-iteration instruction ordering freedom: the
2574 iterations in the loop can be considered completely independent of each
2575 other (also known as embarrassingly parallel loops).
2577 This metadata can originate from a programming language with parallel loop
2578 constructs. In such a case it is completely the programmer's responsibility
2579 to ensure the instructions from the different iterations of the loop can be
2580 executed in an arbitrary order, in parallel, or intertwined. No loop-carried
2581 dependency checking at all must be expected from the compiler.
2583 In order to fulfill the LLVM requirement for metadata to be safely ignored,
2584 it is important to ensure that a parallel loop is converted to
2585 a sequential loop in case an optimization (agnostic of the parallel loop
2586 semantics) converts the loop back to such. This happens when new memory
2587 accesses that do not fulfill the requirement of free ordering across iterations
2588 are added to the loop. Therefore, this metadata is required, but not
2589 sufficient, to consider the loop at hand a parallel loop. For a loop
2590 to be parallel, all its memory accessing instructions need to be
2591 marked with the ``llvm.mem.parallel_loop_access`` metadata that refer
2592 to the same loop identifier metadata that identify the loop at hand.
2597 Metadata types used to annotate memory accesses with information helpful
2598 for optimizations are prefixed with ``llvm.mem``.
2600 '``llvm.mem.parallel_loop_access``' Metadata
2601 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2603 For a loop to be parallel, in addition to using
2604 the ``llvm.loop.parallel`` metadata to mark the loop latch branch instruction,
2605 also all of the memory accessing instructions in the loop body need to be
2606 marked with the ``llvm.mem.parallel_loop_access`` metadata. If there
2607 is at least one memory accessing instruction not marked with the metadata,
2608 the loop, despite it possibly using the ``llvm.loop.parallel`` metadata,
2609 must be considered a sequential loop. This causes parallel loops to be
2610 converted to sequential loops due to optimization passes that are unaware of
2611 the parallel semantics and that insert new memory instructions to the loop
2614 Example of a loop that is considered parallel due to its correct use of
2615 both ``llvm.loop.parallel`` and ``llvm.mem.parallel_loop_access``
2616 metadata types that refer to the same loop identifier metadata.
2618 .. code-block:: llvm
2622 %0 = load i32* %arrayidx, align 4, !llvm.mem.parallel_loop_access !0
2624 store i32 %0, i32* %arrayidx4, align 4, !llvm.mem.parallel_loop_access !0
2626 br i1 %exitcond, label %for.end, label %for.body, !llvm.loop.parallel !0
2630 !0 = metadata !{ metadata !0 }
2632 It is also possible to have nested parallel loops. In that case the
2633 memory accesses refer to a list of loop identifier metadata nodes instead of
2634 the loop identifier metadata node directly:
2636 .. code-block:: llvm
2643 %0 = load i32* %arrayidx, align 4, !llvm.mem.parallel_loop_access !0
2645 store i32 %0, i32* %arrayidx4, align 4, !llvm.mem.parallel_loop_access !0
2647 br i1 %exitcond, label %inner.for.end, label %inner.for.body, !llvm.loop.parallel !1
2651 %0 = load i32* %arrayidx, align 4, !llvm.mem.parallel_loop_access !0
2653 store i32 %0, i32* %arrayidx4, align 4, !llvm.mem.parallel_loop_access !0
2655 br i1 %exitcond, label %outer.for.end, label %outer.for.body, !llvm.loop.parallel !2
2657 outer.for.end: ; preds = %for.body
2659 !0 = metadata !{ metadata !1, metadata !2 } ; a list of parallel loop identifiers
2660 !1 = metadata !{ metadata !1 } ; an identifier for the inner parallel loop
2661 !2 = metadata !{ metadata !2 } ; an identifier for the outer parallel loop
2664 Module Flags Metadata
2665 =====================
2667 Information about the module as a whole is difficult to convey to LLVM's
2668 subsystems. The LLVM IR isn't sufficient to transmit this information.
2669 The ``llvm.module.flags`` named metadata exists in order to facilitate
2670 this. These flags are in the form of key / value pairs --- much like a
2671 dictionary --- making it easy for any subsystem who cares about a flag to
2674 The ``llvm.module.flags`` metadata contains a list of metadata triplets.
2675 Each triplet has the following form:
2677 - The first element is a *behavior* flag, which specifies the behavior
2678 when two (or more) modules are merged together, and it encounters two
2679 (or more) metadata with the same ID. The supported behaviors are
2681 - The second element is a metadata string that is a unique ID for the
2682 metadata. Each module may only have one flag entry for each unique ID (not
2683 including entries with the **Require** behavior).
2684 - The third element is the value of the flag.
2686 When two (or more) modules are merged together, the resulting
2687 ``llvm.module.flags`` metadata is the union of the modules' flags. That is, for
2688 each unique metadata ID string, there will be exactly one entry in the merged
2689 modules ``llvm.module.flags`` metadata table, and the value for that entry will
2690 be determined by the merge behavior flag, as described below. The only exception
2691 is that entries with the *Require* behavior are always preserved.
2693 The following behaviors are supported:
2704 Emits an error if two values disagree, otherwise the resulting value
2705 is that of the operands.
2709 Emits a warning if two values disagree. The result value will be the
2710 operand for the flag from the first module being linked.
2714 Adds a requirement that another module flag be present and have a
2715 specified value after linking is performed. The value must be a
2716 metadata pair, where the first element of the pair is the ID of the
2717 module flag to be restricted, and the second element of the pair is
2718 the value the module flag should be restricted to. This behavior can
2719 be used to restrict the allowable results (via triggering of an
2720 error) of linking IDs with the **Override** behavior.
2724 Uses the specified value, regardless of the behavior or value of the
2725 other module. If both modules specify **Override**, but the values
2726 differ, an error will be emitted.
2730 Appends the two values, which are required to be metadata nodes.
2734 Appends the two values, which are required to be metadata
2735 nodes. However, duplicate entries in the second list are dropped
2736 during the append operation.
2738 It is an error for a particular unique flag ID to have multiple behaviors,
2739 except in the case of **Require** (which adds restrictions on another metadata
2740 value) or **Override**.
2742 An example of module flags:
2744 .. code-block:: llvm
2746 !0 = metadata !{ i32 1, metadata !"foo", i32 1 }
2747 !1 = metadata !{ i32 4, metadata !"bar", i32 37 }
2748 !2 = metadata !{ i32 2, metadata !"qux", i32 42 }
2749 !3 = metadata !{ i32 3, metadata !"qux",
2751 metadata !"foo", i32 1
2754 !llvm.module.flags = !{ !0, !1, !2, !3 }
2756 - Metadata ``!0`` has the ID ``!"foo"`` and the value '1'. The behavior
2757 if two or more ``!"foo"`` flags are seen is to emit an error if their
2758 values are not equal.
2760 - Metadata ``!1`` has the ID ``!"bar"`` and the value '37'. The
2761 behavior if two or more ``!"bar"`` flags are seen is to use the value
2764 - Metadata ``!2`` has the ID ``!"qux"`` and the value '42'. The
2765 behavior if two or more ``!"qux"`` flags are seen is to emit a
2766 warning if their values are not equal.
2768 - Metadata ``!3`` has the ID ``!"qux"`` and the value:
2772 metadata !{ metadata !"foo", i32 1 }
2774 The behavior is to emit an error if the ``llvm.module.flags`` does not
2775 contain a flag with the ID ``!"foo"`` that has the value '1' after linking is
2778 Objective-C Garbage Collection Module Flags Metadata
2779 ----------------------------------------------------
2781 On the Mach-O platform, Objective-C stores metadata about garbage
2782 collection in a special section called "image info". The metadata
2783 consists of a version number and a bitmask specifying what types of
2784 garbage collection are supported (if any) by the file. If two or more
2785 modules are linked together their garbage collection metadata needs to
2786 be merged rather than appended together.
2788 The Objective-C garbage collection module flags metadata consists of the
2789 following key-value pairs:
2798 * - ``Objective-C Version``
2799 - **[Required]** --- The Objective-C ABI version. Valid values are 1 and 2.
2801 * - ``Objective-C Image Info Version``
2802 - **[Required]** --- The version of the image info section. Currently
2805 * - ``Objective-C Image Info Section``
2806 - **[Required]** --- The section to place the metadata. Valid values are
2807 ``"__OBJC, __image_info, regular"`` for Objective-C ABI version 1, and
2808 ``"__DATA,__objc_imageinfo, regular, no_dead_strip"`` for
2809 Objective-C ABI version 2.
2811 * - ``Objective-C Garbage Collection``
2812 - **[Required]** --- Specifies whether garbage collection is supported or
2813 not. Valid values are 0, for no garbage collection, and 2, for garbage
2814 collection supported.
2816 * - ``Objective-C GC Only``
2817 - **[Optional]** --- Specifies that only garbage collection is supported.
2818 If present, its value must be 6. This flag requires that the
2819 ``Objective-C Garbage Collection`` flag have the value 2.
2821 Some important flag interactions:
2823 - If a module with ``Objective-C Garbage Collection`` set to 0 is
2824 merged with a module with ``Objective-C Garbage Collection`` set to
2825 2, then the resulting module has the
2826 ``Objective-C Garbage Collection`` flag set to 0.
2827 - A module with ``Objective-C Garbage Collection`` set to 0 cannot be
2828 merged with a module with ``Objective-C GC Only`` set to 6.
2830 Automatic Linker Flags Module Flags Metadata
2831 --------------------------------------------
2833 Some targets support embedding flags to the linker inside individual object
2834 files. Typically this is used in conjunction with language extensions which
2835 allow source files to explicitly declare the libraries they depend on, and have
2836 these automatically be transmitted to the linker via object files.
2838 These flags are encoded in the IR using metadata in the module flags section,
2839 using the ``Linker Options`` key. The merge behavior for this flag is required
2840 to be ``AppendUnique``, and the value for the key is expected to be a metadata
2841 node which should be a list of other metadata nodes, each of which should be a
2842 list of metadata strings defining linker options.
2844 For example, the following metadata section specifies two separate sets of
2845 linker options, presumably to link against ``libz`` and the ``Cocoa``
2848 !0 = metadata !{ i32 6, metadata !"Linker Options",
2850 metadata !{ metadata !"-lz" },
2851 metadata !{ metadata !"-framework", metadata !"Cocoa" } } }
2852 !llvm.module.flags = !{ !0 }
2854 The metadata encoding as lists of lists of options, as opposed to a collapsed
2855 list of options, is chosen so that the IR encoding can use multiple option
2856 strings to specify e.g., a single library, while still having that specifier be
2857 preserved as an atomic element that can be recognized by a target specific
2858 assembly writer or object file emitter.
2860 Each individual option is required to be either a valid option for the target's
2861 linker, or an option that is reserved by the target specific assembly writer or
2862 object file emitter. No other aspect of these options is defined by the IR.
2864 Intrinsic Global Variables
2865 ==========================
2867 LLVM has a number of "magic" global variables that contain data that
2868 affect code generation or other IR semantics. These are documented here.
2869 All globals of this sort should have a section specified as
2870 "``llvm.metadata``". This section and all globals that start with
2871 "``llvm.``" are reserved for use by LLVM.
2873 The '``llvm.used``' Global Variable
2874 -----------------------------------
2876 The ``@llvm.used`` global is an array which has
2877 :ref:`appending linkage <linkage_appending>`. This array contains a list of
2878 pointers to global variables, functions and aliases which may optionally have a
2879 pointer cast formed of bitcast or getelementptr. For example, a legal
2882 .. code-block:: llvm
2887 @llvm.used = appending global [2 x i8*] [
2889 i8* bitcast (i32* @Y to i8*)
2890 ], section "llvm.metadata"
2892 If a symbol appears in the ``@llvm.used`` list, then the compiler, assembler,
2893 and linker are required to treat the symbol as if there is a reference to the
2894 symbol that it cannot see. For example, if a variable has internal linkage and
2895 no references other than that from the ``@llvm.used`` list, it cannot be
2896 deleted. This is commonly used to represent references from inline asms and
2897 other things the compiler cannot "see", and corresponds to
2898 "``attribute((used))``" in GNU C.
2900 On some targets, the code generator must emit a directive to the
2901 assembler or object file to prevent the assembler and linker from
2902 molesting the symbol.
2904 The '``llvm.compiler.used``' Global Variable
2905 --------------------------------------------
2907 The ``@llvm.compiler.used`` directive is the same as the ``@llvm.used``
2908 directive, except that it only prevents the compiler from touching the
2909 symbol. On targets that support it, this allows an intelligent linker to
2910 optimize references to the symbol without being impeded as it would be
2913 This is a rare construct that should only be used in rare circumstances,
2914 and should not be exposed to source languages.
2916 The '``llvm.global_ctors``' Global Variable
2917 -------------------------------------------
2919 .. code-block:: llvm
2921 %0 = type { i32, void ()* }
2922 @llvm.global_ctors = appending global [1 x %0] [%0 { i32 65535, void ()* @ctor }]
2924 The ``@llvm.global_ctors`` array contains a list of constructor
2925 functions and associated priorities. The functions referenced by this
2926 array will be called in ascending order of priority (i.e. lowest first)
2927 when the module is loaded. The order of functions with the same priority
2930 The '``llvm.global_dtors``' Global Variable
2931 -------------------------------------------
2933 .. code-block:: llvm
2935 %0 = type { i32, void ()* }
2936 @llvm.global_dtors = appending global [1 x %0] [%0 { i32 65535, void ()* @dtor }]
2938 The ``@llvm.global_dtors`` array contains a list of destructor functions
2939 and associated priorities. The functions referenced by this array will
2940 be called in descending order of priority (i.e. highest first) when the
2941 module is loaded. The order of functions with the same priority is not
2944 Instruction Reference
2945 =====================
2947 The LLVM instruction set consists of several different classifications
2948 of instructions: :ref:`terminator instructions <terminators>`, :ref:`binary
2949 instructions <binaryops>`, :ref:`bitwise binary
2950 instructions <bitwiseops>`, :ref:`memory instructions <memoryops>`, and
2951 :ref:`other instructions <otherops>`.
2955 Terminator Instructions
2956 -----------------------
2958 As mentioned :ref:`previously <functionstructure>`, every basic block in a
2959 program ends with a "Terminator" instruction, which indicates which
2960 block should be executed after the current block is finished. These
2961 terminator instructions typically yield a '``void``' value: they produce
2962 control flow, not values (the one exception being the
2963 ':ref:`invoke <i_invoke>`' instruction).
2965 The terminator instructions are: ':ref:`ret <i_ret>`',
2966 ':ref:`br <i_br>`', ':ref:`switch <i_switch>`',
2967 ':ref:`indirectbr <i_indirectbr>`', ':ref:`invoke <i_invoke>`',
2968 ':ref:`resume <i_resume>`', and ':ref:`unreachable <i_unreachable>`'.
2972 '``ret``' Instruction
2973 ^^^^^^^^^^^^^^^^^^^^^
2980 ret <type> <value> ; Return a value from a non-void function
2981 ret void ; Return from void function
2986 The '``ret``' instruction is used to return control flow (and optionally
2987 a value) from a function back to the caller.
2989 There are two forms of the '``ret``' instruction: one that returns a
2990 value and then causes control flow, and one that just causes control
2996 The '``ret``' instruction optionally accepts a single argument, the
2997 return value. The type of the return value must be a ':ref:`first
2998 class <t_firstclass>`' type.
3000 A function is not :ref:`well formed <wellformed>` if it it has a non-void
3001 return type and contains a '``ret``' instruction with no return value or
3002 a return value with a type that does not match its type, or if it has a
3003 void return type and contains a '``ret``' instruction with a return
3009 When the '``ret``' instruction is executed, control flow returns back to
3010 the calling function's context. If the caller is a
3011 ":ref:`call <i_call>`" instruction, execution continues at the
3012 instruction after the call. If the caller was an
3013 ":ref:`invoke <i_invoke>`" instruction, execution continues at the
3014 beginning of the "normal" destination block. If the instruction returns
3015 a value, that value shall set the call or invoke instruction's return
3021 .. code-block:: llvm
3023 ret i32 5 ; Return an integer value of 5
3024 ret void ; Return from a void function
3025 ret { i32, i8 } { i32 4, i8 2 } ; Return a struct of values 4 and 2
3029 '``br``' Instruction
3030 ^^^^^^^^^^^^^^^^^^^^
3037 br i1 <cond>, label <iftrue>, label <iffalse>
3038 br label <dest> ; Unconditional branch
3043 The '``br``' instruction is used to cause control flow to transfer to a
3044 different basic block in the current function. There are two forms of
3045 this instruction, corresponding to a conditional branch and an
3046 unconditional branch.
3051 The conditional branch form of the '``br``' instruction takes a single
3052 '``i1``' value and two '``label``' values. The unconditional form of the
3053 '``br``' instruction takes a single '``label``' value as a target.
3058 Upon execution of a conditional '``br``' instruction, the '``i1``'
3059 argument is evaluated. If the value is ``true``, control flows to the
3060 '``iftrue``' ``label`` argument. If "cond" is ``false``, control flows
3061 to the '``iffalse``' ``label`` argument.
3066 .. code-block:: llvm
3069 %cond = icmp eq i32 %a, %b
3070 br i1 %cond, label %IfEqual, label %IfUnequal
3078 '``switch``' Instruction
3079 ^^^^^^^^^^^^^^^^^^^^^^^^
3086 switch <intty> <value>, label <defaultdest> [ <intty> <val>, label <dest> ... ]
3091 The '``switch``' instruction is used to transfer control flow to one of
3092 several different places. It is a generalization of the '``br``'
3093 instruction, allowing a branch to occur to one of many possible
3099 The '``switch``' instruction uses three parameters: an integer
3100 comparison value '``value``', a default '``label``' destination, and an
3101 array of pairs of comparison value constants and '``label``'s. The table
3102 is not allowed to contain duplicate constant entries.
3107 The ``switch`` instruction specifies a table of values and destinations.
3108 When the '``switch``' instruction is executed, this table is searched
3109 for the given value. If the value is found, control flow is transferred
3110 to the corresponding destination; otherwise, control flow is transferred
3111 to the default destination.
3116 Depending on properties of the target machine and the particular
3117 ``switch`` instruction, this instruction may be code generated in
3118 different ways. For example, it could be generated as a series of
3119 chained conditional branches or with a lookup table.
3124 .. code-block:: llvm
3126 ; Emulate a conditional br instruction
3127 %Val = zext i1 %value to i32
3128 switch i32 %Val, label %truedest [ i32 0, label %falsedest ]
3130 ; Emulate an unconditional br instruction
3131 switch i32 0, label %dest [ ]
3133 ; Implement a jump table:
3134 switch i32 %val, label %otherwise [ i32 0, label %onzero
3136 i32 2, label %ontwo ]
3140 '``indirectbr``' Instruction
3141 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3148 indirectbr <somety>* <address>, [ label <dest1>, label <dest2>, ... ]
3153 The '``indirectbr``' instruction implements an indirect branch to a
3154 label within the current function, whose address is specified by
3155 "``address``". Address must be derived from a
3156 :ref:`blockaddress <blockaddress>` constant.
3161 The '``address``' argument is the address of the label to jump to. The
3162 rest of the arguments indicate the full set of possible destinations
3163 that the address may point to. Blocks are allowed to occur multiple
3164 times in the destination list, though this isn't particularly useful.
3166 This destination list is required so that dataflow analysis has an
3167 accurate understanding of the CFG.
3172 Control transfers to the block specified in the address argument. All
3173 possible destination blocks must be listed in the label list, otherwise
3174 this instruction has undefined behavior. This implies that jumps to
3175 labels defined in other functions have undefined behavior as well.
3180 This is typically implemented with a jump through a register.
3185 .. code-block:: llvm
3187 indirectbr i8* %Addr, [ label %bb1, label %bb2, label %bb3 ]
3191 '``invoke``' Instruction
3192 ^^^^^^^^^^^^^^^^^^^^^^^^
3199 <result> = invoke [cconv] [ret attrs] <ptr to function ty> <function ptr val>(<function args>) [fn attrs]
3200 to label <normal label> unwind label <exception label>
3205 The '``invoke``' instruction causes control to transfer to a specified
3206 function, with the possibility of control flow transfer to either the
3207 '``normal``' label or the '``exception``' label. If the callee function
3208 returns with the "``ret``" instruction, control flow will return to the
3209 "normal" label. If the callee (or any indirect callees) returns via the
3210 ":ref:`resume <i_resume>`" instruction or other exception handling
3211 mechanism, control is interrupted and continued at the dynamically
3212 nearest "exception" label.
3214 The '``exception``' label is a `landing
3215 pad <ExceptionHandling.html#overview>`_ for the exception. As such,
3216 '``exception``' label is required to have the
3217 ":ref:`landingpad <i_landingpad>`" instruction, which contains the
3218 information about the behavior of the program after unwinding happens,
3219 as its first non-PHI instruction. The restrictions on the
3220 "``landingpad``" instruction's tightly couples it to the "``invoke``"
3221 instruction, so that the important information contained within the
3222 "``landingpad``" instruction can't be lost through normal code motion.
3227 This instruction requires several arguments:
3229 #. The optional "cconv" marker indicates which :ref:`calling
3230 convention <callingconv>` the call should use. If none is
3231 specified, the call defaults to using C calling conventions.
3232 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
3233 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
3235 #. '``ptr to function ty``': shall be the signature of the pointer to
3236 function value being invoked. In most cases, this is a direct
3237 function invocation, but indirect ``invoke``'s are just as possible,
3238 branching off an arbitrary pointer to function value.
3239 #. '``function ptr val``': An LLVM value containing a pointer to a
3240 function to be invoked.
3241 #. '``function args``': argument list whose types match the function
3242 signature argument types and parameter attributes. All arguments must
3243 be of :ref:`first class <t_firstclass>` type. If the function signature
3244 indicates the function accepts a variable number of arguments, the
3245 extra arguments can be specified.
3246 #. '``normal label``': the label reached when the called function
3247 executes a '``ret``' instruction.
3248 #. '``exception label``': the label reached when a callee returns via
3249 the :ref:`resume <i_resume>` instruction or other exception handling
3251 #. The optional :ref:`function attributes <fnattrs>` list. Only
3252 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
3253 attributes are valid here.
3258 This instruction is designed to operate as a standard '``call``'
3259 instruction in most regards. The primary difference is that it
3260 establishes an association with a label, which is used by the runtime
3261 library to unwind the stack.
3263 This instruction is used in languages with destructors to ensure that
3264 proper cleanup is performed in the case of either a ``longjmp`` or a
3265 thrown exception. Additionally, this is important for implementation of
3266 '``catch``' clauses in high-level languages that support them.
3268 For the purposes of the SSA form, the definition of the value returned
3269 by the '``invoke``' instruction is deemed to occur on the edge from the
3270 current block to the "normal" label. If the callee unwinds then no
3271 return value is available.
3276 .. code-block:: llvm
3278 %retval = invoke i32 @Test(i32 15) to label %Continue
3279 unwind label %TestCleanup ; {i32}:retval set
3280 %retval = invoke coldcc i32 %Testfnptr(i32 15) to label %Continue
3281 unwind label %TestCleanup ; {i32}:retval set
3285 '``resume``' Instruction
3286 ^^^^^^^^^^^^^^^^^^^^^^^^
3293 resume <type> <value>
3298 The '``resume``' instruction is a terminator instruction that has no
3304 The '``resume``' instruction requires one argument, which must have the
3305 same type as the result of any '``landingpad``' instruction in the same
3311 The '``resume``' instruction resumes propagation of an existing
3312 (in-flight) exception whose unwinding was interrupted with a
3313 :ref:`landingpad <i_landingpad>` instruction.
3318 .. code-block:: llvm
3320 resume { i8*, i32 } %exn
3324 '``unreachable``' Instruction
3325 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3337 The '``unreachable``' instruction has no defined semantics. This
3338 instruction is used to inform the optimizer that a particular portion of
3339 the code is not reachable. This can be used to indicate that the code
3340 after a no-return function cannot be reached, and other facts.
3345 The '``unreachable``' instruction has no defined semantics.
3352 Binary operators are used to do most of the computation in a program.
3353 They require two operands of the same type, execute an operation on
3354 them, and produce a single value. The operands might represent multiple
3355 data, as is the case with the :ref:`vector <t_vector>` data type. The
3356 result value has the same type as its operands.
3358 There are several different binary operators:
3362 '``add``' Instruction
3363 ^^^^^^^^^^^^^^^^^^^^^
3370 <result> = add <ty> <op1>, <op2> ; yields {ty}:result
3371 <result> = add nuw <ty> <op1>, <op2> ; yields {ty}:result
3372 <result> = add nsw <ty> <op1>, <op2> ; yields {ty}:result
3373 <result> = add nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3378 The '``add``' instruction returns the sum of its two operands.
3383 The two arguments to the '``add``' instruction must be
3384 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3385 arguments must have identical types.
3390 The value produced is the integer sum of the two operands.
3392 If the sum has unsigned overflow, the result returned is the
3393 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
3396 Because LLVM integers use a two's complement representation, this
3397 instruction is appropriate for both signed and unsigned integers.
3399 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3400 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3401 result value of the ``add`` is a :ref:`poison value <poisonvalues>` if
3402 unsigned and/or signed overflow, respectively, occurs.
3407 .. code-block:: llvm
3409 <result> = add i32 4, %var ; yields {i32}:result = 4 + %var
3413 '``fadd``' Instruction
3414 ^^^^^^^^^^^^^^^^^^^^^^
3421 <result> = fadd [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3426 The '``fadd``' instruction returns the sum of its two operands.
3431 The two arguments to the '``fadd``' instruction must be :ref:`floating
3432 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3433 Both arguments must have identical types.
3438 The value produced is the floating point sum of the two operands. This
3439 instruction can also take any number of :ref:`fast-math flags <fastmath>`,
3440 which are optimization hints to enable otherwise unsafe floating point
3446 .. code-block:: llvm
3448 <result> = fadd float 4.0, %var ; yields {float}:result = 4.0 + %var
3450 '``sub``' Instruction
3451 ^^^^^^^^^^^^^^^^^^^^^
3458 <result> = sub <ty> <op1>, <op2> ; yields {ty}:result
3459 <result> = sub nuw <ty> <op1>, <op2> ; yields {ty}:result
3460 <result> = sub nsw <ty> <op1>, <op2> ; yields {ty}:result
3461 <result> = sub nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3466 The '``sub``' instruction returns the difference of its two operands.
3468 Note that the '``sub``' instruction is used to represent the '``neg``'
3469 instruction present in most other intermediate representations.
3474 The two arguments to the '``sub``' instruction must be
3475 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3476 arguments must have identical types.
3481 The value produced is the integer difference of the two operands.
3483 If the difference has unsigned overflow, the result returned is the
3484 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
3487 Because LLVM integers use a two's complement representation, this
3488 instruction is appropriate for both signed and unsigned integers.
3490 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3491 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3492 result value of the ``sub`` is a :ref:`poison value <poisonvalues>` if
3493 unsigned and/or signed overflow, respectively, occurs.
3498 .. code-block:: llvm
3500 <result> = sub i32 4, %var ; yields {i32}:result = 4 - %var
3501 <result> = sub i32 0, %val ; yields {i32}:result = -%var
3505 '``fsub``' Instruction
3506 ^^^^^^^^^^^^^^^^^^^^^^
3513 <result> = fsub [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3518 The '``fsub``' instruction returns the difference of its two operands.
3520 Note that the '``fsub``' instruction is used to represent the '``fneg``'
3521 instruction present in most other intermediate representations.
3526 The two arguments to the '``fsub``' instruction must be :ref:`floating
3527 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3528 Both arguments must have identical types.
3533 The value produced is the floating point difference of the two operands.
3534 This instruction can also take any number of :ref:`fast-math
3535 flags <fastmath>`, which are optimization hints to enable otherwise
3536 unsafe floating point optimizations:
3541 .. code-block:: llvm
3543 <result> = fsub float 4.0, %var ; yields {float}:result = 4.0 - %var
3544 <result> = fsub float -0.0, %val ; yields {float}:result = -%var
3546 '``mul``' Instruction
3547 ^^^^^^^^^^^^^^^^^^^^^
3554 <result> = mul <ty> <op1>, <op2> ; yields {ty}:result
3555 <result> = mul nuw <ty> <op1>, <op2> ; yields {ty}:result
3556 <result> = mul nsw <ty> <op1>, <op2> ; yields {ty}:result
3557 <result> = mul nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3562 The '``mul``' instruction returns the product of its two operands.
3567 The two arguments to the '``mul``' instruction must be
3568 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3569 arguments must have identical types.
3574 The value produced is the integer product of the two operands.
3576 If the result of the multiplication has unsigned overflow, the result
3577 returned is the mathematical result modulo 2\ :sup:`n`\ , where n is the
3578 bit width of the result.
3580 Because LLVM integers use a two's complement representation, and the
3581 result is the same width as the operands, this instruction returns the
3582 correct result for both signed and unsigned integers. If a full product
3583 (e.g. ``i32`` * ``i32`` -> ``i64``) is needed, the operands should be
3584 sign-extended or zero-extended as appropriate to the width of the full
3587 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3588 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3589 result value of the ``mul`` is a :ref:`poison value <poisonvalues>` if
3590 unsigned and/or signed overflow, respectively, occurs.
3595 .. code-block:: llvm
3597 <result> = mul i32 4, %var ; yields {i32}:result = 4 * %var
3601 '``fmul``' Instruction
3602 ^^^^^^^^^^^^^^^^^^^^^^
3609 <result> = fmul [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3614 The '``fmul``' instruction returns the product of its two operands.
3619 The two arguments to the '``fmul``' instruction must be :ref:`floating
3620 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3621 Both arguments must have identical types.
3626 The value produced is the floating point product of the two operands.
3627 This instruction can also take any number of :ref:`fast-math
3628 flags <fastmath>`, which are optimization hints to enable otherwise
3629 unsafe floating point optimizations:
3634 .. code-block:: llvm
3636 <result> = fmul float 4.0, %var ; yields {float}:result = 4.0 * %var
3638 '``udiv``' Instruction
3639 ^^^^^^^^^^^^^^^^^^^^^^
3646 <result> = udiv <ty> <op1>, <op2> ; yields {ty}:result
3647 <result> = udiv exact <ty> <op1>, <op2> ; yields {ty}:result
3652 The '``udiv``' instruction returns the quotient of its two operands.
3657 The two arguments to the '``udiv``' instruction must be
3658 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3659 arguments must have identical types.
3664 The value produced is the unsigned integer quotient of the two operands.
3666 Note that unsigned integer division and signed integer division are
3667 distinct operations; for signed integer division, use '``sdiv``'.
3669 Division by zero leads to undefined behavior.
3671 If the ``exact`` keyword is present, the result value of the ``udiv`` is
3672 a :ref:`poison value <poisonvalues>` if %op1 is not a multiple of %op2 (as
3673 such, "((a udiv exact b) mul b) == a").
3678 .. code-block:: llvm
3680 <result> = udiv i32 4, %var ; yields {i32}:result = 4 / %var
3682 '``sdiv``' Instruction
3683 ^^^^^^^^^^^^^^^^^^^^^^
3690 <result> = sdiv <ty> <op1>, <op2> ; yields {ty}:result
3691 <result> = sdiv exact <ty> <op1>, <op2> ; yields {ty}:result
3696 The '``sdiv``' instruction returns the quotient of its two operands.
3701 The two arguments to the '``sdiv``' instruction must be
3702 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3703 arguments must have identical types.
3708 The value produced is the signed integer quotient of the two operands
3709 rounded towards zero.
3711 Note that signed integer division and unsigned integer division are
3712 distinct operations; for unsigned integer division, use '``udiv``'.
3714 Division by zero leads to undefined behavior. Overflow also leads to
3715 undefined behavior; this is a rare case, but can occur, for example, by
3716 doing a 32-bit division of -2147483648 by -1.
3718 If the ``exact`` keyword is present, the result value of the ``sdiv`` is
3719 a :ref:`poison value <poisonvalues>` if the result would be rounded.
3724 .. code-block:: llvm
3726 <result> = sdiv i32 4, %var ; yields {i32}:result = 4 / %var
3730 '``fdiv``' Instruction
3731 ^^^^^^^^^^^^^^^^^^^^^^
3738 <result> = fdiv [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3743 The '``fdiv``' instruction returns the quotient of its two operands.
3748 The two arguments to the '``fdiv``' instruction must be :ref:`floating
3749 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3750 Both arguments must have identical types.
3755 The value produced is the floating point quotient of the two operands.
3756 This instruction can also take any number of :ref:`fast-math
3757 flags <fastmath>`, which are optimization hints to enable otherwise
3758 unsafe floating point optimizations:
3763 .. code-block:: llvm
3765 <result> = fdiv float 4.0, %var ; yields {float}:result = 4.0 / %var
3767 '``urem``' Instruction
3768 ^^^^^^^^^^^^^^^^^^^^^^
3775 <result> = urem <ty> <op1>, <op2> ; yields {ty}:result
3780 The '``urem``' instruction returns the remainder from the unsigned
3781 division of its two arguments.
3786 The two arguments to the '``urem``' instruction must be
3787 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3788 arguments must have identical types.
3793 This instruction returns the unsigned integer *remainder* of a division.
3794 This instruction always performs an unsigned division to get the
3797 Note that unsigned integer remainder and signed integer remainder are
3798 distinct operations; for signed integer remainder, use '``srem``'.
3800 Taking the remainder of a division by zero leads to undefined behavior.
3805 .. code-block:: llvm
3807 <result> = urem i32 4, %var ; yields {i32}:result = 4 % %var
3809 '``srem``' Instruction
3810 ^^^^^^^^^^^^^^^^^^^^^^
3817 <result> = srem <ty> <op1>, <op2> ; yields {ty}:result
3822 The '``srem``' instruction returns the remainder from the signed
3823 division of its two operands. This instruction can also take
3824 :ref:`vector <t_vector>` versions of the values in which case the elements
3830 The two arguments to the '``srem``' instruction must be
3831 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3832 arguments must have identical types.
3837 This instruction returns the *remainder* of a division (where the result
3838 is either zero or has the same sign as the dividend, ``op1``), not the
3839 *modulo* operator (where the result is either zero or has the same sign
3840 as the divisor, ``op2``) of a value. For more information about the
3841 difference, see `The Math
3842 Forum <http://mathforum.org/dr.math/problems/anne.4.28.99.html>`_. For a
3843 table of how this is implemented in various languages, please see
3845 operation <http://en.wikipedia.org/wiki/Modulo_operation>`_.
3847 Note that signed integer remainder and unsigned integer remainder are
3848 distinct operations; for unsigned integer remainder, use '``urem``'.
3850 Taking the remainder of a division by zero leads to undefined behavior.
3851 Overflow also leads to undefined behavior; this is a rare case, but can
3852 occur, for example, by taking the remainder of a 32-bit division of
3853 -2147483648 by -1. (The remainder doesn't actually overflow, but this
3854 rule lets srem be implemented using instructions that return both the
3855 result of the division and the remainder.)
3860 .. code-block:: llvm
3862 <result> = srem i32 4, %var ; yields {i32}:result = 4 % %var
3866 '``frem``' Instruction
3867 ^^^^^^^^^^^^^^^^^^^^^^
3874 <result> = frem [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3879 The '``frem``' instruction returns the remainder from the division of
3885 The two arguments to the '``frem``' instruction must be :ref:`floating
3886 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3887 Both arguments must have identical types.
3892 This instruction returns the *remainder* of a division. The remainder
3893 has the same sign as the dividend. This instruction can also take any
3894 number of :ref:`fast-math flags <fastmath>`, which are optimization hints
3895 to enable otherwise unsafe floating point optimizations:
3900 .. code-block:: llvm
3902 <result> = frem float 4.0, %var ; yields {float}:result = 4.0 % %var
3906 Bitwise Binary Operations
3907 -------------------------
3909 Bitwise binary operators are used to do various forms of bit-twiddling
3910 in a program. They are generally very efficient instructions and can
3911 commonly be strength reduced from other instructions. They require two
3912 operands of the same type, execute an operation on them, and produce a
3913 single value. The resulting value is the same type as its operands.
3915 '``shl``' Instruction
3916 ^^^^^^^^^^^^^^^^^^^^^
3923 <result> = shl <ty> <op1>, <op2> ; yields {ty}:result
3924 <result> = shl nuw <ty> <op1>, <op2> ; yields {ty}:result
3925 <result> = shl nsw <ty> <op1>, <op2> ; yields {ty}:result
3926 <result> = shl nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3931 The '``shl``' instruction returns the first operand shifted to the left
3932 a specified number of bits.
3937 Both arguments to the '``shl``' instruction must be the same
3938 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
3939 '``op2``' is treated as an unsigned value.
3944 The value produced is ``op1`` \* 2\ :sup:`op2` mod 2\ :sup:`n`,
3945 where ``n`` is the width of the result. If ``op2`` is (statically or
3946 dynamically) negative or equal to or larger than the number of bits in
3947 ``op1``, the result is undefined. If the arguments are vectors, each
3948 vector element of ``op1`` is shifted by the corresponding shift amount
3951 If the ``nuw`` keyword is present, then the shift produces a :ref:`poison
3952 value <poisonvalues>` if it shifts out any non-zero bits. If the
3953 ``nsw`` keyword is present, then the shift produces a :ref:`poison
3954 value <poisonvalues>` if it shifts out any bits that disagree with the
3955 resultant sign bit. As such, NUW/NSW have the same semantics as they
3956 would if the shift were expressed as a mul instruction with the same
3957 nsw/nuw bits in (mul %op1, (shl 1, %op2)).
3962 .. code-block:: llvm
3964 <result> = shl i32 4, %var ; yields {i32}: 4 << %var
3965 <result> = shl i32 4, 2 ; yields {i32}: 16
3966 <result> = shl i32 1, 10 ; yields {i32}: 1024
3967 <result> = shl i32 1, 32 ; undefined
3968 <result> = shl <2 x i32> < i32 1, i32 1>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 2, i32 4>
3970 '``lshr``' Instruction
3971 ^^^^^^^^^^^^^^^^^^^^^^
3978 <result> = lshr <ty> <op1>, <op2> ; yields {ty}:result
3979 <result> = lshr exact <ty> <op1>, <op2> ; yields {ty}:result
3984 The '``lshr``' instruction (logical shift right) returns the first
3985 operand shifted to the right a specified number of bits with zero fill.
3990 Both arguments to the '``lshr``' instruction must be the same
3991 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
3992 '``op2``' is treated as an unsigned value.
3997 This instruction always performs a logical shift right operation. The
3998 most significant bits of the result will be filled with zero bits after
3999 the shift. If ``op2`` is (statically or dynamically) equal to or larger
4000 than the number of bits in ``op1``, the result is undefined. If the
4001 arguments are vectors, each vector element of ``op1`` is shifted by the
4002 corresponding shift amount in ``op2``.
4004 If the ``exact`` keyword is present, the result value of the ``lshr`` is
4005 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
4011 .. code-block:: llvm
4013 <result> = lshr i32 4, 1 ; yields {i32}:result = 2
4014 <result> = lshr i32 4, 2 ; yields {i32}:result = 1
4015 <result> = lshr i8 4, 3 ; yields {i8}:result = 0
4016 <result> = lshr i8 -2, 1 ; yields {i8}:result = 0x7F
4017 <result> = lshr i32 1, 32 ; undefined
4018 <result> = lshr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 0x7FFFFFFF, i32 1>
4020 '``ashr``' Instruction
4021 ^^^^^^^^^^^^^^^^^^^^^^
4028 <result> = ashr <ty> <op1>, <op2> ; yields {ty}:result
4029 <result> = ashr exact <ty> <op1>, <op2> ; yields {ty}:result
4034 The '``ashr``' instruction (arithmetic shift right) returns the first
4035 operand shifted to the right a specified number of bits with sign
4041 Both arguments to the '``ashr``' instruction must be the same
4042 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4043 '``op2``' is treated as an unsigned value.
4048 This instruction always performs an arithmetic shift right operation,
4049 The most significant bits of the result will be filled with the sign bit
4050 of ``op1``. If ``op2`` is (statically or dynamically) equal to or larger
4051 than the number of bits in ``op1``, the result is undefined. If the
4052 arguments are vectors, each vector element of ``op1`` is shifted by the
4053 corresponding shift amount in ``op2``.
4055 If the ``exact`` keyword is present, the result value of the ``ashr`` is
4056 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
4062 .. code-block:: llvm
4064 <result> = ashr i32 4, 1 ; yields {i32}:result = 2
4065 <result> = ashr i32 4, 2 ; yields {i32}:result = 1
4066 <result> = ashr i8 4, 3 ; yields {i8}:result = 0
4067 <result> = ashr i8 -2, 1 ; yields {i8}:result = -1
4068 <result> = ashr i32 1, 32 ; undefined
4069 <result> = ashr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 3> ; yields: result=<2 x i32> < i32 -1, i32 0>
4071 '``and``' Instruction
4072 ^^^^^^^^^^^^^^^^^^^^^
4079 <result> = and <ty> <op1>, <op2> ; yields {ty}:result
4084 The '``and``' instruction returns the bitwise logical and of its two
4090 The two arguments to the '``and``' instruction must be
4091 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4092 arguments must have identical types.
4097 The truth table used for the '``and``' instruction is:
4114 .. code-block:: llvm
4116 <result> = and i32 4, %var ; yields {i32}:result = 4 & %var
4117 <result> = and i32 15, 40 ; yields {i32}:result = 8
4118 <result> = and i32 4, 8 ; yields {i32}:result = 0
4120 '``or``' Instruction
4121 ^^^^^^^^^^^^^^^^^^^^
4128 <result> = or <ty> <op1>, <op2> ; yields {ty}:result
4133 The '``or``' instruction returns the bitwise logical inclusive or of its
4139 The two arguments to the '``or``' instruction must be
4140 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4141 arguments must have identical types.
4146 The truth table used for the '``or``' instruction is:
4165 <result> = or i32 4, %var ; yields {i32}:result = 4 | %var
4166 <result> = or i32 15, 40 ; yields {i32}:result = 47
4167 <result> = or i32 4, 8 ; yields {i32}:result = 12
4169 '``xor``' Instruction
4170 ^^^^^^^^^^^^^^^^^^^^^
4177 <result> = xor <ty> <op1>, <op2> ; yields {ty}:result
4182 The '``xor``' instruction returns the bitwise logical exclusive or of
4183 its two operands. The ``xor`` is used to implement the "one's
4184 complement" operation, which is the "~" operator in C.
4189 The two arguments to the '``xor``' instruction must be
4190 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4191 arguments must have identical types.
4196 The truth table used for the '``xor``' instruction is:
4213 .. code-block:: llvm
4215 <result> = xor i32 4, %var ; yields {i32}:result = 4 ^ %var
4216 <result> = xor i32 15, 40 ; yields {i32}:result = 39
4217 <result> = xor i32 4, 8 ; yields {i32}:result = 12
4218 <result> = xor i32 %V, -1 ; yields {i32}:result = ~%V
4223 LLVM supports several instructions to represent vector operations in a
4224 target-independent manner. These instructions cover the element-access
4225 and vector-specific operations needed to process vectors effectively.
4226 While LLVM does directly support these vector operations, many
4227 sophisticated algorithms will want to use target-specific intrinsics to
4228 take full advantage of a specific target.
4230 .. _i_extractelement:
4232 '``extractelement``' Instruction
4233 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4240 <result> = extractelement <n x <ty>> <val>, i32 <idx> ; yields <ty>
4245 The '``extractelement``' instruction extracts a single scalar element
4246 from a vector at a specified index.
4251 The first operand of an '``extractelement``' instruction is a value of
4252 :ref:`vector <t_vector>` type. The second operand is an index indicating
4253 the position from which to extract the element. The index may be a
4259 The result is a scalar of the same type as the element type of ``val``.
4260 Its value is the value at position ``idx`` of ``val``. If ``idx``
4261 exceeds the length of ``val``, the results are undefined.
4266 .. code-block:: llvm
4268 <result> = extractelement <4 x i32> %vec, i32 0 ; yields i32
4270 .. _i_insertelement:
4272 '``insertelement``' Instruction
4273 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4280 <result> = insertelement <n x <ty>> <val>, <ty> <elt>, i32 <idx> ; yields <n x <ty>>
4285 The '``insertelement``' instruction inserts a scalar element into a
4286 vector at a specified index.
4291 The first operand of an '``insertelement``' instruction is a value of
4292 :ref:`vector <t_vector>` type. The second operand is a scalar value whose
4293 type must equal the element type of the first operand. The third operand
4294 is an index indicating the position at which to insert the value. The
4295 index may be a variable.
4300 The result is a vector of the same type as ``val``. Its element values
4301 are those of ``val`` except at position ``idx``, where it gets the value
4302 ``elt``. If ``idx`` exceeds the length of ``val``, the results are
4308 .. code-block:: llvm
4310 <result> = insertelement <4 x i32> %vec, i32 1, i32 0 ; yields <4 x i32>
4312 .. _i_shufflevector:
4314 '``shufflevector``' Instruction
4315 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4322 <result> = shufflevector <n x <ty>> <v1>, <n x <ty>> <v2>, <m x i32> <mask> ; yields <m x <ty>>
4327 The '``shufflevector``' instruction constructs a permutation of elements
4328 from two input vectors, returning a vector with the same element type as
4329 the input and length that is the same as the shuffle mask.
4334 The first two operands of a '``shufflevector``' instruction are vectors
4335 with the same type. The third argument is a shuffle mask whose element
4336 type is always 'i32'. The result of the instruction is a vector whose
4337 length is the same as the shuffle mask and whose element type is the
4338 same as the element type of the first two operands.
4340 The shuffle mask operand is required to be a constant vector with either
4341 constant integer or undef values.
4346 The elements of the two input vectors are numbered from left to right
4347 across both of the vectors. The shuffle mask operand specifies, for each
4348 element of the result vector, which element of the two input vectors the
4349 result element gets. The element selector may be undef (meaning "don't
4350 care") and the second operand may be undef if performing a shuffle from
4356 .. code-block:: llvm
4358 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
4359 <4 x i32> <i32 0, i32 4, i32 1, i32 5> ; yields <4 x i32>
4360 <result> = shufflevector <4 x i32> %v1, <4 x i32> undef,
4361 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32> - Identity shuffle.
4362 <result> = shufflevector <8 x i32> %v1, <8 x i32> undef,
4363 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32>
4364 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
4365 <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7 > ; yields <8 x i32>
4367 Aggregate Operations
4368 --------------------
4370 LLVM supports several instructions for working with
4371 :ref:`aggregate <t_aggregate>` values.
4375 '``extractvalue``' Instruction
4376 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4383 <result> = extractvalue <aggregate type> <val>, <idx>{, <idx>}*
4388 The '``extractvalue``' instruction extracts the value of a member field
4389 from an :ref:`aggregate <t_aggregate>` value.
4394 The first operand of an '``extractvalue``' instruction is a value of
4395 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The operands are
4396 constant indices to specify which value to extract in a similar manner
4397 as indices in a '``getelementptr``' instruction.
4399 The major differences to ``getelementptr`` indexing are:
4401 - Since the value being indexed is not a pointer, the first index is
4402 omitted and assumed to be zero.
4403 - At least one index must be specified.
4404 - Not only struct indices but also array indices must be in bounds.
4409 The result is the value at the position in the aggregate specified by
4415 .. code-block:: llvm
4417 <result> = extractvalue {i32, float} %agg, 0 ; yields i32
4421 '``insertvalue``' Instruction
4422 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4429 <result> = insertvalue <aggregate type> <val>, <ty> <elt>, <idx>{, <idx>}* ; yields <aggregate type>
4434 The '``insertvalue``' instruction inserts a value into a member field in
4435 an :ref:`aggregate <t_aggregate>` value.
4440 The first operand of an '``insertvalue``' instruction is a value of
4441 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The second operand is
4442 a first-class value to insert. The following operands are constant
4443 indices indicating the position at which to insert the value in a
4444 similar manner as indices in a '``extractvalue``' instruction. The value
4445 to insert must have the same type as the value identified by the
4451 The result is an aggregate of the same type as ``val``. Its value is
4452 that of ``val`` except that the value at the position specified by the
4453 indices is that of ``elt``.
4458 .. code-block:: llvm
4460 %agg1 = insertvalue {i32, float} undef, i32 1, 0 ; yields {i32 1, float undef}
4461 %agg2 = insertvalue {i32, float} %agg1, float %val, 1 ; yields {i32 1, float %val}
4462 %agg3 = insertvalue {i32, {float}} %agg1, float %val, 1, 0 ; yields {i32 1, float %val}
4466 Memory Access and Addressing Operations
4467 ---------------------------------------
4469 A key design point of an SSA-based representation is how it represents
4470 memory. In LLVM, no memory locations are in SSA form, which makes things
4471 very simple. This section describes how to read, write, and allocate
4476 '``alloca``' Instruction
4477 ^^^^^^^^^^^^^^^^^^^^^^^^
4484 <result> = alloca <type>[, <ty> <NumElements>][, align <alignment>] ; yields {type*}:result
4489 The '``alloca``' instruction allocates memory on the stack frame of the
4490 currently executing function, to be automatically released when this
4491 function returns to its caller. The object is always allocated in the
4492 generic address space (address space zero).
4497 The '``alloca``' instruction allocates ``sizeof(<type>)*NumElements``
4498 bytes of memory on the runtime stack, returning a pointer of the
4499 appropriate type to the program. If "NumElements" is specified, it is
4500 the number of elements allocated, otherwise "NumElements" is defaulted
4501 to be one. If a constant alignment is specified, the value result of the
4502 allocation is guaranteed to be aligned to at least that boundary. If not
4503 specified, or if zero, the target can choose to align the allocation on
4504 any convenient boundary compatible with the type.
4506 '``type``' may be any sized type.
4511 Memory is allocated; a pointer is returned. The operation is undefined
4512 if there is insufficient stack space for the allocation. '``alloca``'d
4513 memory is automatically released when the function returns. The
4514 '``alloca``' instruction is commonly used to represent automatic
4515 variables that must have an address available. When the function returns
4516 (either with the ``ret`` or ``resume`` instructions), the memory is
4517 reclaimed. Allocating zero bytes is legal, but the result is undefined.
4518 The order in which memory is allocated (ie., which way the stack grows)
4524 .. code-block:: llvm
4526 %ptr = alloca i32 ; yields {i32*}:ptr
4527 %ptr = alloca i32, i32 4 ; yields {i32*}:ptr
4528 %ptr = alloca i32, i32 4, align 1024 ; yields {i32*}:ptr
4529 %ptr = alloca i32, align 1024 ; yields {i32*}:ptr
4533 '``load``' Instruction
4534 ^^^^^^^^^^^^^^^^^^^^^^
4541 <result> = load [volatile] <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>][, !invariant.load !<index>]
4542 <result> = load atomic [volatile] <ty>* <pointer> [singlethread] <ordering>, align <alignment>
4543 !<index> = !{ i32 1 }
4548 The '``load``' instruction is used to read from memory.
4553 The argument to the ``load`` instruction specifies the memory address
4554 from which to load. The pointer must point to a :ref:`first
4555 class <t_firstclass>` type. If the ``load`` is marked as ``volatile``,
4556 then the optimizer is not allowed to modify the number or order of
4557 execution of this ``load`` with other :ref:`volatile
4558 operations <volatile>`.
4560 If the ``load`` is marked as ``atomic``, it takes an extra
4561 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
4562 ``release`` and ``acq_rel`` orderings are not valid on ``load``
4563 instructions. Atomic loads produce :ref:`defined <memmodel>` results
4564 when they may see multiple atomic stores. The type of the pointee must
4565 be an integer type whose bit width is a power of two greater than or
4566 equal to eight and less than or equal to a target-specific size limit.
4567 ``align`` must be explicitly specified on atomic loads, and the load has
4568 undefined behavior if the alignment is not set to a value which is at
4569 least the size in bytes of the pointee. ``!nontemporal`` does not have
4570 any defined semantics for atomic loads.
4572 The optional constant ``align`` argument specifies the alignment of the
4573 operation (that is, the alignment of the memory address). A value of 0
4574 or an omitted ``align`` argument means that the operation has the ABI
4575 alignment for the target. It is the responsibility of the code emitter
4576 to ensure that the alignment information is correct. Overestimating the
4577 alignment results in undefined behavior. Underestimating the alignment
4578 may produce less efficient code. An alignment of 1 is always safe.
4580 The optional ``!nontemporal`` metadata must reference a single
4581 metatadata name ``<index>`` corresponding to a metadata node with one
4582 ``i32`` entry of value 1. The existence of the ``!nontemporal``
4583 metatadata on the instruction tells the optimizer and code generator
4584 that this load is not expected to be reused in the cache. The code
4585 generator may select special instructions to save cache bandwidth, such
4586 as the ``MOVNT`` instruction on x86.
4588 The optional ``!invariant.load`` metadata must reference a single
4589 metatadata name ``<index>`` corresponding to a metadata node with no
4590 entries. The existence of the ``!invariant.load`` metatadata on the
4591 instruction tells the optimizer and code generator that this load
4592 address points to memory which does not change value during program
4593 execution. The optimizer may then move this load around, for example, by
4594 hoisting it out of loops using loop invariant code motion.
4599 The location of memory pointed to is loaded. If the value being loaded
4600 is of scalar type then the number of bytes read does not exceed the
4601 minimum number of bytes needed to hold all bits of the type. For
4602 example, loading an ``i24`` reads at most three bytes. When loading a
4603 value of a type like ``i20`` with a size that is not an integral number
4604 of bytes, the result is undefined if the value was not originally
4605 written using a store of the same type.
4610 .. code-block:: llvm
4612 %ptr = alloca i32 ; yields {i32*}:ptr
4613 store i32 3, i32* %ptr ; yields {void}
4614 %val = load i32* %ptr ; yields {i32}:val = i32 3
4618 '``store``' Instruction
4619 ^^^^^^^^^^^^^^^^^^^^^^^
4626 store [volatile] <ty> <value>, <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>] ; yields {void}
4627 store atomic [volatile] <ty> <value>, <ty>* <pointer> [singlethread] <ordering>, align <alignment> ; yields {void}
4632 The '``store``' instruction is used to write to memory.
4637 There are two arguments to the ``store`` instruction: a value to store
4638 and an address at which to store it. The type of the ``<pointer>``
4639 operand must be a pointer to the :ref:`first class <t_firstclass>` type of
4640 the ``<value>`` operand. If the ``store`` is marked as ``volatile``,
4641 then the optimizer is not allowed to modify the number or order of
4642 execution of this ``store`` with other :ref:`volatile
4643 operations <volatile>`.
4645 If the ``store`` is marked as ``atomic``, it takes an extra
4646 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
4647 ``acquire`` and ``acq_rel`` orderings aren't valid on ``store``
4648 instructions. Atomic loads produce :ref:`defined <memmodel>` results
4649 when they may see multiple atomic stores. The type of the pointee must
4650 be an integer type whose bit width is a power of two greater than or
4651 equal to eight and less than or equal to a target-specific size limit.
4652 ``align`` must be explicitly specified on atomic stores, and the store
4653 has undefined behavior if the alignment is not set to a value which is
4654 at least the size in bytes of the pointee. ``!nontemporal`` does not
4655 have any defined semantics for atomic stores.
4657 The optional constant ``align`` argument specifies the alignment of the
4658 operation (that is, the alignment of the memory address). A value of 0
4659 or an omitted ``align`` argument means that the operation has the ABI
4660 alignment for the target. It is the responsibility of the code emitter
4661 to ensure that the alignment information is correct. Overestimating the
4662 alignment results in undefined behavior. Underestimating the
4663 alignment may produce less efficient code. An alignment of 1 is always
4666 The optional ``!nontemporal`` metadata must reference a single metatadata
4667 name ``<index>`` corresponding to a metadata node with one ``i32`` entry of
4668 value 1. The existence of the ``!nontemporal`` metatadata on the instruction
4669 tells the optimizer and code generator that this load is not expected to
4670 be reused in the cache. The code generator may select special
4671 instructions to save cache bandwidth, such as the MOVNT instruction on
4677 The contents of memory are updated to contain ``<value>`` at the
4678 location specified by the ``<pointer>`` operand. If ``<value>`` is
4679 of scalar type then the number of bytes written does not exceed the
4680 minimum number of bytes needed to hold all bits of the type. For
4681 example, storing an ``i24`` writes at most three bytes. When writing a
4682 value of a type like ``i20`` with a size that is not an integral number
4683 of bytes, it is unspecified what happens to the extra bits that do not
4684 belong to the type, but they will typically be overwritten.
4689 .. code-block:: llvm
4691 %ptr = alloca i32 ; yields {i32*}:ptr
4692 store i32 3, i32* %ptr ; yields {void}
4693 %val = load i32* %ptr ; yields {i32}:val = i32 3
4697 '``fence``' Instruction
4698 ^^^^^^^^^^^^^^^^^^^^^^^
4705 fence [singlethread] <ordering> ; yields {void}
4710 The '``fence``' instruction is used to introduce happens-before edges
4716 '``fence``' instructions take an :ref:`ordering <ordering>` argument which
4717 defines what *synchronizes-with* edges they add. They can only be given
4718 ``acquire``, ``release``, ``acq_rel``, and ``seq_cst`` orderings.
4723 A fence A which has (at least) ``release`` ordering semantics
4724 *synchronizes with* a fence B with (at least) ``acquire`` ordering
4725 semantics if and only if there exist atomic operations X and Y, both
4726 operating on some atomic object M, such that A is sequenced before X, X
4727 modifies M (either directly or through some side effect of a sequence
4728 headed by X), Y is sequenced before B, and Y observes M. This provides a
4729 *happens-before* dependency between A and B. Rather than an explicit
4730 ``fence``, one (but not both) of the atomic operations X or Y might
4731 provide a ``release`` or ``acquire`` (resp.) ordering constraint and
4732 still *synchronize-with* the explicit ``fence`` and establish the
4733 *happens-before* edge.
4735 A ``fence`` which has ``seq_cst`` ordering, in addition to having both
4736 ``acquire`` and ``release`` semantics specified above, participates in
4737 the global program order of other ``seq_cst`` operations and/or fences.
4739 The optional ":ref:`singlethread <singlethread>`" argument specifies
4740 that the fence only synchronizes with other fences in the same thread.
4741 (This is useful for interacting with signal handlers.)
4746 .. code-block:: llvm
4748 fence acquire ; yields {void}
4749 fence singlethread seq_cst ; yields {void}
4753 '``cmpxchg``' Instruction
4754 ^^^^^^^^^^^^^^^^^^^^^^^^^
4761 cmpxchg [volatile] <ty>* <pointer>, <ty> <cmp>, <ty> <new> [singlethread] <ordering> ; yields {ty}
4766 The '``cmpxchg``' instruction is used to atomically modify memory. It
4767 loads a value in memory and compares it to a given value. If they are
4768 equal, it stores a new value into the memory.
4773 There are three arguments to the '``cmpxchg``' instruction: an address
4774 to operate on, a value to compare to the value currently be at that
4775 address, and a new value to place at that address if the compared values
4776 are equal. The type of '<cmp>' must be an integer type whose bit width
4777 is a power of two greater than or equal to eight and less than or equal
4778 to a target-specific size limit. '<cmp>' and '<new>' must have the same
4779 type, and the type of '<pointer>' must be a pointer to that type. If the
4780 ``cmpxchg`` is marked as ``volatile``, then the optimizer is not allowed
4781 to modify the number or order of execution of this ``cmpxchg`` with
4782 other :ref:`volatile operations <volatile>`.
4784 The :ref:`ordering <ordering>` argument specifies how this ``cmpxchg``
4785 synchronizes with other atomic operations.
4787 The optional "``singlethread``" argument declares that the ``cmpxchg``
4788 is only atomic with respect to code (usually signal handlers) running in
4789 the same thread as the ``cmpxchg``. Otherwise the cmpxchg is atomic with
4790 respect to all other code in the system.
4792 The pointer passed into cmpxchg must have alignment greater than or
4793 equal to the size in memory of the operand.
4798 The contents of memory at the location specified by the '``<pointer>``'
4799 operand is read and compared to '``<cmp>``'; if the read value is the
4800 equal, '``<new>``' is written. The original value at the location is
4803 A successful ``cmpxchg`` is a read-modify-write instruction for the purpose
4804 of identifying release sequences. A failed ``cmpxchg`` is equivalent to an
4805 atomic load with an ordering parameter determined by dropping any
4806 ``release`` part of the ``cmpxchg``'s ordering.
4811 .. code-block:: llvm
4814 %orig = atomic load i32* %ptr unordered ; yields {i32}
4818 %cmp = phi i32 [ %orig, %entry ], [%old, %loop]
4819 %squared = mul i32 %cmp, %cmp
4820 %old = cmpxchg i32* %ptr, i32 %cmp, i32 %squared ; yields {i32}
4821 %success = icmp eq i32 %cmp, %old
4822 br i1 %success, label %done, label %loop
4829 '``atomicrmw``' Instruction
4830 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
4837 atomicrmw [volatile] <operation> <ty>* <pointer>, <ty> <value> [singlethread] <ordering> ; yields {ty}
4842 The '``atomicrmw``' instruction is used to atomically modify memory.
4847 There are three arguments to the '``atomicrmw``' instruction: an
4848 operation to apply, an address whose value to modify, an argument to the
4849 operation. The operation must be one of the following keywords:
4863 The type of '<value>' must be an integer type whose bit width is a power
4864 of two greater than or equal to eight and less than or equal to a
4865 target-specific size limit. The type of the '``<pointer>``' operand must
4866 be a pointer to that type. If the ``atomicrmw`` is marked as
4867 ``volatile``, then the optimizer is not allowed to modify the number or
4868 order of execution of this ``atomicrmw`` with other :ref:`volatile
4869 operations <volatile>`.
4874 The contents of memory at the location specified by the '``<pointer>``'
4875 operand are atomically read, modified, and written back. The original
4876 value at the location is returned. The modification is specified by the
4879 - xchg: ``*ptr = val``
4880 - add: ``*ptr = *ptr + val``
4881 - sub: ``*ptr = *ptr - val``
4882 - and: ``*ptr = *ptr & val``
4883 - nand: ``*ptr = ~(*ptr & val)``
4884 - or: ``*ptr = *ptr | val``
4885 - xor: ``*ptr = *ptr ^ val``
4886 - max: ``*ptr = *ptr > val ? *ptr : val`` (using a signed comparison)
4887 - min: ``*ptr = *ptr < val ? *ptr : val`` (using a signed comparison)
4888 - umax: ``*ptr = *ptr > val ? *ptr : val`` (using an unsigned
4890 - umin: ``*ptr = *ptr < val ? *ptr : val`` (using an unsigned
4896 .. code-block:: llvm
4898 %old = atomicrmw add i32* %ptr, i32 1 acquire ; yields {i32}
4900 .. _i_getelementptr:
4902 '``getelementptr``' Instruction
4903 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4910 <result> = getelementptr <pty>* <ptrval>{, <ty> <idx>}*
4911 <result> = getelementptr inbounds <pty>* <ptrval>{, <ty> <idx>}*
4912 <result> = getelementptr <ptr vector> ptrval, <vector index type> idx
4917 The '``getelementptr``' instruction is used to get the address of a
4918 subelement of an :ref:`aggregate <t_aggregate>` data structure. It performs
4919 address calculation only and does not access memory.
4924 The first argument is always a pointer or a vector of pointers, and
4925 forms the basis of the calculation. The remaining arguments are indices
4926 that indicate which of the elements of the aggregate object are indexed.
4927 The interpretation of each index is dependent on the type being indexed
4928 into. The first index always indexes the pointer value given as the
4929 first argument, the second index indexes a value of the type pointed to
4930 (not necessarily the value directly pointed to, since the first index
4931 can be non-zero), etc. The first type indexed into must be a pointer
4932 value, subsequent types can be arrays, vectors, and structs. Note that
4933 subsequent types being indexed into can never be pointers, since that
4934 would require loading the pointer before continuing calculation.
4936 The type of each index argument depends on the type it is indexing into.
4937 When indexing into a (optionally packed) structure, only ``i32`` integer
4938 **constants** are allowed (when using a vector of indices they must all
4939 be the **same** ``i32`` integer constant). When indexing into an array,
4940 pointer or vector, integers of any width are allowed, and they are not
4941 required to be constant. These integers are treated as signed values
4944 For example, let's consider a C code fragment and how it gets compiled
4960 int *foo(struct ST *s) {
4961 return &s[1].Z.B[5][13];
4964 The LLVM code generated by Clang is:
4966 .. code-block:: llvm
4968 %struct.RT = type { i8, [10 x [20 x i32]], i8 }
4969 %struct.ST = type { i32, double, %struct.RT }
4971 define i32* @foo(%struct.ST* %s) nounwind uwtable readnone optsize ssp {
4973 %arrayidx = getelementptr inbounds %struct.ST* %s, i64 1, i32 2, i32 1, i64 5, i64 13
4980 In the example above, the first index is indexing into the
4981 '``%struct.ST*``' type, which is a pointer, yielding a '``%struct.ST``'
4982 = '``{ i32, double, %struct.RT }``' type, a structure. The second index
4983 indexes into the third element of the structure, yielding a
4984 '``%struct.RT``' = '``{ i8 , [10 x [20 x i32]], i8 }``' type, another
4985 structure. The third index indexes into the second element of the
4986 structure, yielding a '``[10 x [20 x i32]]``' type, an array. The two
4987 dimensions of the array are subscripted into, yielding an '``i32``'
4988 type. The '``getelementptr``' instruction returns a pointer to this
4989 element, thus computing a value of '``i32*``' type.
4991 Note that it is perfectly legal to index partially through a structure,
4992 returning a pointer to an inner element. Because of this, the LLVM code
4993 for the given testcase is equivalent to:
4995 .. code-block:: llvm
4997 define i32* @foo(%struct.ST* %s) {
4998 %t1 = getelementptr %struct.ST* %s, i32 1 ; yields %struct.ST*:%t1
4999 %t2 = getelementptr %struct.ST* %t1, i32 0, i32 2 ; yields %struct.RT*:%t2
5000 %t3 = getelementptr %struct.RT* %t2, i32 0, i32 1 ; yields [10 x [20 x i32]]*:%t3
5001 %t4 = getelementptr [10 x [20 x i32]]* %t3, i32 0, i32 5 ; yields [20 x i32]*:%t4
5002 %t5 = getelementptr [20 x i32]* %t4, i32 0, i32 13 ; yields i32*:%t5
5006 If the ``inbounds`` keyword is present, the result value of the
5007 ``getelementptr`` is a :ref:`poison value <poisonvalues>` if the base
5008 pointer is not an *in bounds* address of an allocated object, or if any
5009 of the addresses that would be formed by successive addition of the
5010 offsets implied by the indices to the base address with infinitely
5011 precise signed arithmetic are not an *in bounds* address of that
5012 allocated object. The *in bounds* addresses for an allocated object are
5013 all the addresses that point into the object, plus the address one byte
5014 past the end. In cases where the base is a vector of pointers the
5015 ``inbounds`` keyword applies to each of the computations element-wise.
5017 If the ``inbounds`` keyword is not present, the offsets are added to the
5018 base address with silently-wrapping two's complement arithmetic. If the
5019 offsets have a different width from the pointer, they are sign-extended
5020 or truncated to the width of the pointer. The result value of the
5021 ``getelementptr`` may be outside the object pointed to by the base
5022 pointer. The result value may not necessarily be used to access memory
5023 though, even if it happens to point into allocated storage. See the
5024 :ref:`Pointer Aliasing Rules <pointeraliasing>` section for more
5027 The getelementptr instruction is often confusing. For some more insight
5028 into how it works, see :doc:`the getelementptr FAQ <GetElementPtr>`.
5033 .. code-block:: llvm
5035 ; yields [12 x i8]*:aptr
5036 %aptr = getelementptr {i32, [12 x i8]}* %saptr, i64 0, i32 1
5038 %vptr = getelementptr {i32, <2 x i8>}* %svptr, i64 0, i32 1, i32 1
5040 %eptr = getelementptr [12 x i8]* %aptr, i64 0, i32 1
5042 %iptr = getelementptr [10 x i32]* @arr, i16 0, i16 0
5044 In cases where the pointer argument is a vector of pointers, each index
5045 must be a vector with the same number of elements. For example:
5047 .. code-block:: llvm
5049 %A = getelementptr <4 x i8*> %ptrs, <4 x i64> %offsets,
5051 Conversion Operations
5052 ---------------------
5054 The instructions in this category are the conversion instructions
5055 (casting) which all take a single operand and a type. They perform
5056 various bit conversions on the operand.
5058 '``trunc .. to``' Instruction
5059 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5066 <result> = trunc <ty> <value> to <ty2> ; yields ty2
5071 The '``trunc``' instruction truncates its operand to the type ``ty2``.
5076 The '``trunc``' instruction takes a value to trunc, and a type to trunc
5077 it to. Both types must be of :ref:`integer <t_integer>` types, or vectors
5078 of the same number of integers. The bit size of the ``value`` must be
5079 larger than the bit size of the destination type, ``ty2``. Equal sized
5080 types are not allowed.
5085 The '``trunc``' instruction truncates the high order bits in ``value``
5086 and converts the remaining bits to ``ty2``. Since the source size must
5087 be larger than the destination size, ``trunc`` cannot be a *no-op cast*.
5088 It will always truncate bits.
5093 .. code-block:: llvm
5095 %X = trunc i32 257 to i8 ; yields i8:1
5096 %Y = trunc i32 123 to i1 ; yields i1:true
5097 %Z = trunc i32 122 to i1 ; yields i1:false
5098 %W = trunc <2 x i16> <i16 8, i16 7> to <2 x i8> ; yields <i8 8, i8 7>
5100 '``zext .. to``' Instruction
5101 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5108 <result> = zext <ty> <value> to <ty2> ; yields ty2
5113 The '``zext``' instruction zero extends its operand to type ``ty2``.
5118 The '``zext``' instruction takes a value to cast, and a type to cast it
5119 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
5120 the same number of integers. The bit size of the ``value`` must be
5121 smaller than the bit size of the destination type, ``ty2``.
5126 The ``zext`` fills the high order bits of the ``value`` with zero bits
5127 until it reaches the size of the destination type, ``ty2``.
5129 When zero extending from i1, the result will always be either 0 or 1.
5134 .. code-block:: llvm
5136 %X = zext i32 257 to i64 ; yields i64:257
5137 %Y = zext i1 true to i32 ; yields i32:1
5138 %Z = zext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
5140 '``sext .. to``' Instruction
5141 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5148 <result> = sext <ty> <value> to <ty2> ; yields ty2
5153 The '``sext``' sign extends ``value`` to the type ``ty2``.
5158 The '``sext``' instruction takes a value to cast, and a type to cast it
5159 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
5160 the same number of integers. The bit size of the ``value`` must be
5161 smaller than the bit size of the destination type, ``ty2``.
5166 The '``sext``' instruction performs a sign extension by copying the sign
5167 bit (highest order bit) of the ``value`` until it reaches the bit size
5168 of the type ``ty2``.
5170 When sign extending from i1, the extension always results in -1 or 0.
5175 .. code-block:: llvm
5177 %X = sext i8 -1 to i16 ; yields i16 :65535
5178 %Y = sext i1 true to i32 ; yields i32:-1
5179 %Z = sext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
5181 '``fptrunc .. to``' Instruction
5182 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5189 <result> = fptrunc <ty> <value> to <ty2> ; yields ty2
5194 The '``fptrunc``' instruction truncates ``value`` to type ``ty2``.
5199 The '``fptrunc``' instruction takes a :ref:`floating point <t_floating>`
5200 value to cast and a :ref:`floating point <t_floating>` type to cast it to.
5201 The size of ``value`` must be larger than the size of ``ty2``. This
5202 implies that ``fptrunc`` cannot be used to make a *no-op cast*.
5207 The '``fptrunc``' instruction truncates a ``value`` from a larger
5208 :ref:`floating point <t_floating>` type to a smaller :ref:`floating
5209 point <t_floating>` type. If the value cannot fit within the
5210 destination type, ``ty2``, then the results are undefined.
5215 .. code-block:: llvm
5217 %X = fptrunc double 123.0 to float ; yields float:123.0
5218 %Y = fptrunc double 1.0E+300 to float ; yields undefined
5220 '``fpext .. to``' Instruction
5221 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5228 <result> = fpext <ty> <value> to <ty2> ; yields ty2
5233 The '``fpext``' extends a floating point ``value`` to a larger floating
5239 The '``fpext``' instruction takes a :ref:`floating point <t_floating>`
5240 ``value`` to cast, and a :ref:`floating point <t_floating>` type to cast it
5241 to. The source type must be smaller than the destination type.
5246 The '``fpext``' instruction extends the ``value`` from a smaller
5247 :ref:`floating point <t_floating>` type to a larger :ref:`floating
5248 point <t_floating>` type. The ``fpext`` cannot be used to make a
5249 *no-op cast* because it always changes bits. Use ``bitcast`` to make a
5250 *no-op cast* for a floating point cast.
5255 .. code-block:: llvm
5257 %X = fpext float 3.125 to double ; yields double:3.125000e+00
5258 %Y = fpext double %X to fp128 ; yields fp128:0xL00000000000000004000900000000000
5260 '``fptoui .. to``' Instruction
5261 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5268 <result> = fptoui <ty> <value> to <ty2> ; yields ty2
5273 The '``fptoui``' converts a floating point ``value`` to its unsigned
5274 integer equivalent of type ``ty2``.
5279 The '``fptoui``' instruction takes a value to cast, which must be a
5280 scalar or vector :ref:`floating point <t_floating>` value, and a type to
5281 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
5282 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
5283 type with the same number of elements as ``ty``
5288 The '``fptoui``' instruction converts its :ref:`floating
5289 point <t_floating>` operand into the nearest (rounding towards zero)
5290 unsigned integer value. If the value cannot fit in ``ty2``, the results
5296 .. code-block:: llvm
5298 %X = fptoui double 123.0 to i32 ; yields i32:123
5299 %Y = fptoui float 1.0E+300 to i1 ; yields undefined:1
5300 %Z = fptoui float 1.04E+17 to i8 ; yields undefined:1
5302 '``fptosi .. to``' Instruction
5303 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5310 <result> = fptosi <ty> <value> to <ty2> ; yields ty2
5315 The '``fptosi``' instruction converts :ref:`floating point <t_floating>`
5316 ``value`` to type ``ty2``.
5321 The '``fptosi``' instruction takes a value to cast, which must be a
5322 scalar or vector :ref:`floating point <t_floating>` value, and a type to
5323 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
5324 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
5325 type with the same number of elements as ``ty``
5330 The '``fptosi``' instruction converts its :ref:`floating
5331 point <t_floating>` operand into the nearest (rounding towards zero)
5332 signed integer value. If the value cannot fit in ``ty2``, the results
5338 .. code-block:: llvm
5340 %X = fptosi double -123.0 to i32 ; yields i32:-123
5341 %Y = fptosi float 1.0E-247 to i1 ; yields undefined:1
5342 %Z = fptosi float 1.04E+17 to i8 ; yields undefined:1
5344 '``uitofp .. to``' Instruction
5345 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5352 <result> = uitofp <ty> <value> to <ty2> ; yields ty2
5357 The '``uitofp``' instruction regards ``value`` as an unsigned integer
5358 and converts that value to the ``ty2`` type.
5363 The '``uitofp``' instruction takes a value to cast, which must be a
5364 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
5365 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
5366 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
5367 type with the same number of elements as ``ty``
5372 The '``uitofp``' instruction interprets its operand as an unsigned
5373 integer quantity and converts it to the corresponding floating point
5374 value. If the value cannot fit in the floating point value, the results
5380 .. code-block:: llvm
5382 %X = uitofp i32 257 to float ; yields float:257.0
5383 %Y = uitofp i8 -1 to double ; yields double:255.0
5385 '``sitofp .. to``' Instruction
5386 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5393 <result> = sitofp <ty> <value> to <ty2> ; yields ty2
5398 The '``sitofp``' instruction regards ``value`` as a signed integer and
5399 converts that value to the ``ty2`` type.
5404 The '``sitofp``' instruction takes a value to cast, which must be a
5405 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
5406 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
5407 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
5408 type with the same number of elements as ``ty``
5413 The '``sitofp``' instruction interprets its operand as a signed integer
5414 quantity and converts it to the corresponding floating point value. If
5415 the value cannot fit in the floating point value, the results are
5421 .. code-block:: llvm
5423 %X = sitofp i32 257 to float ; yields float:257.0
5424 %Y = sitofp i8 -1 to double ; yields double:-1.0
5428 '``ptrtoint .. to``' Instruction
5429 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5436 <result> = ptrtoint <ty> <value> to <ty2> ; yields ty2
5441 The '``ptrtoint``' instruction converts the pointer or a vector of
5442 pointers ``value`` to the integer (or vector of integers) type ``ty2``.
5447 The '``ptrtoint``' instruction takes a ``value`` to cast, which must be
5448 a a value of type :ref:`pointer <t_pointer>` or a vector of pointers, and a
5449 type to cast it to ``ty2``, which must be an :ref:`integer <t_integer>` or
5450 a vector of integers type.
5455 The '``ptrtoint``' instruction converts ``value`` to integer type
5456 ``ty2`` by interpreting the pointer value as an integer and either
5457 truncating or zero extending that value to the size of the integer type.
5458 If ``value`` is smaller than ``ty2`` then a zero extension is done. If
5459 ``value`` is larger than ``ty2`` then a truncation is done. If they are
5460 the same size, then nothing is done (*no-op cast*) other than a type
5466 .. code-block:: llvm
5468 %X = ptrtoint i32* %P to i8 ; yields truncation on 32-bit architecture
5469 %Y = ptrtoint i32* %P to i64 ; yields zero extension on 32-bit architecture
5470 %Z = ptrtoint <4 x i32*> %P to <4 x i64>; yields vector zero extension for a vector of addresses on 32-bit architecture
5474 '``inttoptr .. to``' Instruction
5475 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5482 <result> = inttoptr <ty> <value> to <ty2> ; yields ty2
5487 The '``inttoptr``' instruction converts an integer ``value`` to a
5488 pointer type, ``ty2``.
5493 The '``inttoptr``' instruction takes an :ref:`integer <t_integer>` value to
5494 cast, and a type to cast it to, which must be a :ref:`pointer <t_pointer>`
5500 The '``inttoptr``' instruction converts ``value`` to type ``ty2`` by
5501 applying either a zero extension or a truncation depending on the size
5502 of the integer ``value``. If ``value`` is larger than the size of a
5503 pointer then a truncation is done. If ``value`` is smaller than the size
5504 of a pointer then a zero extension is done. If they are the same size,
5505 nothing is done (*no-op cast*).
5510 .. code-block:: llvm
5512 %X = inttoptr i32 255 to i32* ; yields zero extension on 64-bit architecture
5513 %Y = inttoptr i32 255 to i32* ; yields no-op on 32-bit architecture
5514 %Z = inttoptr i64 0 to i32* ; yields truncation on 32-bit architecture
5515 %Z = inttoptr <4 x i32> %G to <4 x i8*>; yields truncation of vector G to four pointers
5519 '``bitcast .. to``' Instruction
5520 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5527 <result> = bitcast <ty> <value> to <ty2> ; yields ty2
5532 The '``bitcast``' instruction converts ``value`` to type ``ty2`` without
5538 The '``bitcast``' instruction takes a value to cast, which must be a
5539 non-aggregate first class value, and a type to cast it to, which must
5540 also be a non-aggregate :ref:`first class <t_firstclass>` type. The bit
5541 sizes of ``value`` and the destination type, ``ty2``, must be identical.
5542 If the source type is a pointer, the destination type must also be a
5543 pointer. This instruction supports bitwise conversion of vectors to
5544 integers and to vectors of other types (as long as they have the same
5550 The '``bitcast``' instruction converts ``value`` to type ``ty2``. It is
5551 always a *no-op cast* because no bits change with this conversion. The
5552 conversion is done as if the ``value`` had been stored to memory and
5553 read back as type ``ty2``. Pointer (or vector of pointers) types may
5554 only be converted to other pointer (or vector of pointers) types with
5555 this instruction. To convert pointers to other types, use the
5556 :ref:`inttoptr <i_inttoptr>` or :ref:`ptrtoint <i_ptrtoint>` instructions
5562 .. code-block:: llvm
5564 %X = bitcast i8 255 to i8 ; yields i8 :-1
5565 %Y = bitcast i32* %x to sint* ; yields sint*:%x
5566 %Z = bitcast <2 x int> %V to i64; ; yields i64: %V
5567 %Z = bitcast <2 x i32*> %V to <2 x i64*> ; yields <2 x i64*>
5574 The instructions in this category are the "miscellaneous" instructions,
5575 which defy better classification.
5579 '``icmp``' Instruction
5580 ^^^^^^^^^^^^^^^^^^^^^^
5587 <result> = icmp <cond> <ty> <op1>, <op2> ; yields {i1} or {<N x i1>}:result
5592 The '``icmp``' instruction returns a boolean value or a vector of
5593 boolean values based on comparison of its two integer, integer vector,
5594 pointer, or pointer vector operands.
5599 The '``icmp``' instruction takes three operands. The first operand is
5600 the condition code indicating the kind of comparison to perform. It is
5601 not a value, just a keyword. The possible condition code are:
5604 #. ``ne``: not equal
5605 #. ``ugt``: unsigned greater than
5606 #. ``uge``: unsigned greater or equal
5607 #. ``ult``: unsigned less than
5608 #. ``ule``: unsigned less or equal
5609 #. ``sgt``: signed greater than
5610 #. ``sge``: signed greater or equal
5611 #. ``slt``: signed less than
5612 #. ``sle``: signed less or equal
5614 The remaining two arguments must be :ref:`integer <t_integer>` or
5615 :ref:`pointer <t_pointer>` or integer :ref:`vector <t_vector>` typed. They
5616 must also be identical types.
5621 The '``icmp``' compares ``op1`` and ``op2`` according to the condition
5622 code given as ``cond``. The comparison performed always yields either an
5623 :ref:`i1 <t_integer>` or vector of ``i1`` result, as follows:
5625 #. ``eq``: yields ``true`` if the operands are equal, ``false``
5626 otherwise. No sign interpretation is necessary or performed.
5627 #. ``ne``: yields ``true`` if the operands are unequal, ``false``
5628 otherwise. No sign interpretation is necessary or performed.
5629 #. ``ugt``: interprets the operands as unsigned values and yields
5630 ``true`` if ``op1`` is greater than ``op2``.
5631 #. ``uge``: interprets the operands as unsigned values and yields
5632 ``true`` if ``op1`` is greater than or equal to ``op2``.
5633 #. ``ult``: interprets the operands as unsigned values and yields
5634 ``true`` if ``op1`` is less than ``op2``.
5635 #. ``ule``: interprets the operands as unsigned values and yields
5636 ``true`` if ``op1`` is less than or equal to ``op2``.
5637 #. ``sgt``: interprets the operands as signed values and yields ``true``
5638 if ``op1`` is greater than ``op2``.
5639 #. ``sge``: interprets the operands as signed values and yields ``true``
5640 if ``op1`` is greater than or equal to ``op2``.
5641 #. ``slt``: interprets the operands as signed values and yields ``true``
5642 if ``op1`` is less than ``op2``.
5643 #. ``sle``: interprets the operands as signed values and yields ``true``
5644 if ``op1`` is less than or equal to ``op2``.
5646 If the operands are :ref:`pointer <t_pointer>` typed, the pointer values
5647 are compared as if they were integers.
5649 If the operands are integer vectors, then they are compared element by
5650 element. The result is an ``i1`` vector with the same number of elements
5651 as the values being compared. Otherwise, the result is an ``i1``.
5656 .. code-block:: llvm
5658 <result> = icmp eq i32 4, 5 ; yields: result=false
5659 <result> = icmp ne float* %X, %X ; yields: result=false
5660 <result> = icmp ult i16 4, 5 ; yields: result=true
5661 <result> = icmp sgt i16 4, 5 ; yields: result=false
5662 <result> = icmp ule i16 -4, 5 ; yields: result=false
5663 <result> = icmp sge i16 4, 5 ; yields: result=false
5665 Note that the code generator does not yet support vector types with the
5666 ``icmp`` instruction.
5670 '``fcmp``' Instruction
5671 ^^^^^^^^^^^^^^^^^^^^^^
5678 <result> = fcmp <cond> <ty> <op1>, <op2> ; yields {i1} or {<N x i1>}:result
5683 The '``fcmp``' instruction returns a boolean value or vector of boolean
5684 values based on comparison of its operands.
5686 If the operands are floating point scalars, then the result type is a
5687 boolean (:ref:`i1 <t_integer>`).
5689 If the operands are floating point vectors, then the result type is a
5690 vector of boolean with the same number of elements as the operands being
5696 The '``fcmp``' instruction takes three operands. The first operand is
5697 the condition code indicating the kind of comparison to perform. It is
5698 not a value, just a keyword. The possible condition code are:
5700 #. ``false``: no comparison, always returns false
5701 #. ``oeq``: ordered and equal
5702 #. ``ogt``: ordered and greater than
5703 #. ``oge``: ordered and greater than or equal
5704 #. ``olt``: ordered and less than
5705 #. ``ole``: ordered and less than or equal
5706 #. ``one``: ordered and not equal
5707 #. ``ord``: ordered (no nans)
5708 #. ``ueq``: unordered or equal
5709 #. ``ugt``: unordered or greater than
5710 #. ``uge``: unordered or greater than or equal
5711 #. ``ult``: unordered or less than
5712 #. ``ule``: unordered or less than or equal
5713 #. ``une``: unordered or not equal
5714 #. ``uno``: unordered (either nans)
5715 #. ``true``: no comparison, always returns true
5717 *Ordered* means that neither operand is a QNAN while *unordered* means
5718 that either operand may be a QNAN.
5720 Each of ``val1`` and ``val2`` arguments must be either a :ref:`floating
5721 point <t_floating>` type or a :ref:`vector <t_vector>` of floating point
5722 type. They must have identical types.
5727 The '``fcmp``' instruction compares ``op1`` and ``op2`` according to the
5728 condition code given as ``cond``. If the operands are vectors, then the
5729 vectors are compared element by element. Each comparison performed
5730 always yields an :ref:`i1 <t_integer>` result, as follows:
5732 #. ``false``: always yields ``false``, regardless of operands.
5733 #. ``oeq``: yields ``true`` if both operands are not a QNAN and ``op1``
5734 is equal to ``op2``.
5735 #. ``ogt``: yields ``true`` if both operands are not a QNAN and ``op1``
5736 is greater than ``op2``.
5737 #. ``oge``: yields ``true`` if both operands are not a QNAN and ``op1``
5738 is greater than or equal to ``op2``.
5739 #. ``olt``: yields ``true`` if both operands are not a QNAN and ``op1``
5740 is less than ``op2``.
5741 #. ``ole``: yields ``true`` if both operands are not a QNAN and ``op1``
5742 is less than or equal to ``op2``.
5743 #. ``one``: yields ``true`` if both operands are not a QNAN and ``op1``
5744 is not equal to ``op2``.
5745 #. ``ord``: yields ``true`` if both operands are not a QNAN.
5746 #. ``ueq``: yields ``true`` if either operand is a QNAN or ``op1`` is
5748 #. ``ugt``: yields ``true`` if either operand is a QNAN or ``op1`` is
5749 greater than ``op2``.
5750 #. ``uge``: yields ``true`` if either operand is a QNAN or ``op1`` is
5751 greater than or equal to ``op2``.
5752 #. ``ult``: yields ``true`` if either operand is a QNAN or ``op1`` is
5754 #. ``ule``: yields ``true`` if either operand is a QNAN or ``op1`` is
5755 less than or equal to ``op2``.
5756 #. ``une``: yields ``true`` if either operand is a QNAN or ``op1`` is
5757 not equal to ``op2``.
5758 #. ``uno``: yields ``true`` if either operand is a QNAN.
5759 #. ``true``: always yields ``true``, regardless of operands.
5764 .. code-block:: llvm
5766 <result> = fcmp oeq float 4.0, 5.0 ; yields: result=false
5767 <result> = fcmp one float 4.0, 5.0 ; yields: result=true
5768 <result> = fcmp olt float 4.0, 5.0 ; yields: result=true
5769 <result> = fcmp ueq double 1.0, 2.0 ; yields: result=false
5771 Note that the code generator does not yet support vector types with the
5772 ``fcmp`` instruction.
5776 '``phi``' Instruction
5777 ^^^^^^^^^^^^^^^^^^^^^
5784 <result> = phi <ty> [ <val0>, <label0>], ...
5789 The '``phi``' instruction is used to implement the φ node in the SSA
5790 graph representing the function.
5795 The type of the incoming values is specified with the first type field.
5796 After this, the '``phi``' instruction takes a list of pairs as
5797 arguments, with one pair for each predecessor basic block of the current
5798 block. Only values of :ref:`first class <t_firstclass>` type may be used as
5799 the value arguments to the PHI node. Only labels may be used as the
5802 There must be no non-phi instructions between the start of a basic block
5803 and the PHI instructions: i.e. PHI instructions must be first in a basic
5806 For the purposes of the SSA form, the use of each incoming value is
5807 deemed to occur on the edge from the corresponding predecessor block to
5808 the current block (but after any definition of an '``invoke``'
5809 instruction's return value on the same edge).
5814 At runtime, the '``phi``' instruction logically takes on the value
5815 specified by the pair corresponding to the predecessor basic block that
5816 executed just prior to the current block.
5821 .. code-block:: llvm
5823 Loop: ; Infinite loop that counts from 0 on up...
5824 %indvar = phi i32 [ 0, %LoopHeader ], [ %nextindvar, %Loop ]
5825 %nextindvar = add i32 %indvar, 1
5830 '``select``' Instruction
5831 ^^^^^^^^^^^^^^^^^^^^^^^^
5838 <result> = select selty <cond>, <ty> <val1>, <ty> <val2> ; yields ty
5840 selty is either i1 or {<N x i1>}
5845 The '``select``' instruction is used to choose one value based on a
5846 condition, without branching.
5851 The '``select``' instruction requires an 'i1' value or a vector of 'i1'
5852 values indicating the condition, and two values of the same :ref:`first
5853 class <t_firstclass>` type. If the val1/val2 are vectors and the
5854 condition is a scalar, then entire vectors are selected, not individual
5860 If the condition is an i1 and it evaluates to 1, the instruction returns
5861 the first value argument; otherwise, it returns the second value
5864 If the condition is a vector of i1, then the value arguments must be
5865 vectors of the same size, and the selection is done element by element.
5870 .. code-block:: llvm
5872 %X = select i1 true, i8 17, i8 42 ; yields i8:17
5876 '``call``' Instruction
5877 ^^^^^^^^^^^^^^^^^^^^^^
5884 <result> = [tail] call [cconv] [ret attrs] <ty> [<fnty>*] <fnptrval>(<function args>) [fn attrs]
5889 The '``call``' instruction represents a simple function call.
5894 This instruction requires several arguments:
5896 #. The optional "tail" marker indicates that the callee function does
5897 not access any allocas or varargs in the caller. Note that calls may
5898 be marked "tail" even if they do not occur before a
5899 :ref:`ret <i_ret>` instruction. If the "tail" marker is present, the
5900 function call is eligible for tail call optimization, but `might not
5901 in fact be optimized into a jump <CodeGenerator.html#tailcallopt>`_.
5902 The code generator may optimize calls marked "tail" with either 1)
5903 automatic `sibling call
5904 optimization <CodeGenerator.html#sibcallopt>`_ when the caller and
5905 callee have matching signatures, or 2) forced tail call optimization
5906 when the following extra requirements are met:
5908 - Caller and callee both have the calling convention ``fastcc``.
5909 - The call is in tail position (ret immediately follows call and ret
5910 uses value of call or is void).
5911 - Option ``-tailcallopt`` is enabled, or
5912 ``llvm::GuaranteedTailCallOpt`` is ``true``.
5913 - `Platform specific constraints are
5914 met. <CodeGenerator.html#tailcallopt>`_
5916 #. The optional "cconv" marker indicates which :ref:`calling
5917 convention <callingconv>` the call should use. If none is
5918 specified, the call defaults to using C calling conventions. The
5919 calling convention of the call must match the calling convention of
5920 the target function, or else the behavior is undefined.
5921 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
5922 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
5924 #. '``ty``': the type of the call instruction itself which is also the
5925 type of the return value. Functions that return no value are marked
5927 #. '``fnty``': shall be the signature of the pointer to function value
5928 being invoked. The argument types must match the types implied by
5929 this signature. This type can be omitted if the function is not
5930 varargs and if the function type does not return a pointer to a
5932 #. '``fnptrval``': An LLVM value containing a pointer to a function to
5933 be invoked. In most cases, this is a direct function invocation, but
5934 indirect ``call``'s are just as possible, calling an arbitrary pointer
5936 #. '``function args``': argument list whose types match the function
5937 signature argument types and parameter attributes. All arguments must
5938 be of :ref:`first class <t_firstclass>` type. If the function signature
5939 indicates the function accepts a variable number of arguments, the
5940 extra arguments can be specified.
5941 #. The optional :ref:`function attributes <fnattrs>` list. Only
5942 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
5943 attributes are valid here.
5948 The '``call``' instruction is used to cause control flow to transfer to
5949 a specified function, with its incoming arguments bound to the specified
5950 values. Upon a '``ret``' instruction in the called function, control
5951 flow continues with the instruction after the function call, and the
5952 return value of the function is bound to the result argument.
5957 .. code-block:: llvm
5959 %retval = call i32 @test(i32 %argc)
5960 call i32 (i8*, ...)* @printf(i8* %msg, i32 12, i8 42) ; yields i32
5961 %X = tail call i32 @foo() ; yields i32
5962 %Y = tail call fastcc i32 @foo() ; yields i32
5963 call void %foo(i8 97 signext)
5965 %struct.A = type { i32, i8 }
5966 %r = call %struct.A @foo() ; yields { 32, i8 }
5967 %gr = extractvalue %struct.A %r, 0 ; yields i32
5968 %gr1 = extractvalue %struct.A %r, 1 ; yields i8
5969 %Z = call void @foo() noreturn ; indicates that %foo never returns normally
5970 %ZZ = call zeroext i32 @bar() ; Return value is %zero extended
5972 llvm treats calls to some functions with names and arguments that match
5973 the standard C99 library as being the C99 library functions, and may
5974 perform optimizations or generate code for them under that assumption.
5975 This is something we'd like to change in the future to provide better
5976 support for freestanding environments and non-C-based languages.
5980 '``va_arg``' Instruction
5981 ^^^^^^^^^^^^^^^^^^^^^^^^
5988 <resultval> = va_arg <va_list*> <arglist>, <argty>
5993 The '``va_arg``' instruction is used to access arguments passed through
5994 the "variable argument" area of a function call. It is used to implement
5995 the ``va_arg`` macro in C.
6000 This instruction takes a ``va_list*`` value and the type of the
6001 argument. It returns a value of the specified argument type and
6002 increments the ``va_list`` to point to the next argument. The actual
6003 type of ``va_list`` is target specific.
6008 The '``va_arg``' instruction loads an argument of the specified type
6009 from the specified ``va_list`` and causes the ``va_list`` to point to
6010 the next argument. For more information, see the variable argument
6011 handling :ref:`Intrinsic Functions <int_varargs>`.
6013 It is legal for this instruction to be called in a function which does
6014 not take a variable number of arguments, for example, the ``vfprintf``
6017 ``va_arg`` is an LLVM instruction instead of an :ref:`intrinsic
6018 function <intrinsics>` because it takes a type as an argument.
6023 See the :ref:`variable argument processing <int_varargs>` section.
6025 Note that the code generator does not yet fully support va\_arg on many
6026 targets. Also, it does not currently support va\_arg with aggregate
6027 types on any target.
6031 '``landingpad``' Instruction
6032 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6039 <resultval> = landingpad <resultty> personality <type> <pers_fn> <clause>+
6040 <resultval> = landingpad <resultty> personality <type> <pers_fn> cleanup <clause>*
6042 <clause> := catch <type> <value>
6043 <clause> := filter <array constant type> <array constant>
6048 The '``landingpad``' instruction is used by `LLVM's exception handling
6049 system <ExceptionHandling.html#overview>`_ to specify that a basic block
6050 is a landing pad --- one where the exception lands, and corresponds to the
6051 code found in the ``catch`` portion of a ``try``/``catch`` sequence. It
6052 defines values supplied by the personality function (``pers_fn``) upon
6053 re-entry to the function. The ``resultval`` has the type ``resultty``.
6058 This instruction takes a ``pers_fn`` value. This is the personality
6059 function associated with the unwinding mechanism. The optional
6060 ``cleanup`` flag indicates that the landing pad block is a cleanup.
6062 A ``clause`` begins with the clause type --- ``catch`` or ``filter`` --- and
6063 contains the global variable representing the "type" that may be caught
6064 or filtered respectively. Unlike the ``catch`` clause, the ``filter``
6065 clause takes an array constant as its argument. Use
6066 "``[0 x i8**] undef``" for a filter which cannot throw. The
6067 '``landingpad``' instruction must contain *at least* one ``clause`` or
6068 the ``cleanup`` flag.
6073 The '``landingpad``' instruction defines the values which are set by the
6074 personality function (``pers_fn``) upon re-entry to the function, and
6075 therefore the "result type" of the ``landingpad`` instruction. As with
6076 calling conventions, how the personality function results are
6077 represented in LLVM IR is target specific.
6079 The clauses are applied in order from top to bottom. If two
6080 ``landingpad`` instructions are merged together through inlining, the
6081 clauses from the calling function are appended to the list of clauses.
6082 When the call stack is being unwound due to an exception being thrown,
6083 the exception is compared against each ``clause`` in turn. If it doesn't
6084 match any of the clauses, and the ``cleanup`` flag is not set, then
6085 unwinding continues further up the call stack.
6087 The ``landingpad`` instruction has several restrictions:
6089 - A landing pad block is a basic block which is the unwind destination
6090 of an '``invoke``' instruction.
6091 - A landing pad block must have a '``landingpad``' instruction as its
6092 first non-PHI instruction.
6093 - There can be only one '``landingpad``' instruction within the landing
6095 - A basic block that is not a landing pad block may not include a
6096 '``landingpad``' instruction.
6097 - All '``landingpad``' instructions in a function must have the same
6098 personality function.
6103 .. code-block:: llvm
6105 ;; A landing pad which can catch an integer.
6106 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6108 ;; A landing pad that is a cleanup.
6109 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6111 ;; A landing pad which can catch an integer and can only throw a double.
6112 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6114 filter [1 x i8**] [@_ZTId]
6121 LLVM supports the notion of an "intrinsic function". These functions
6122 have well known names and semantics and are required to follow certain
6123 restrictions. Overall, these intrinsics represent an extension mechanism
6124 for the LLVM language that does not require changing all of the
6125 transformations in LLVM when adding to the language (or the bitcode
6126 reader/writer, the parser, etc...).
6128 Intrinsic function names must all start with an "``llvm.``" prefix. This
6129 prefix is reserved in LLVM for intrinsic names; thus, function names may
6130 not begin with this prefix. Intrinsic functions must always be external
6131 functions: you cannot define the body of intrinsic functions. Intrinsic
6132 functions may only be used in call or invoke instructions: it is illegal
6133 to take the address of an intrinsic function. Additionally, because
6134 intrinsic functions are part of the LLVM language, it is required if any
6135 are added that they be documented here.
6137 Some intrinsic functions can be overloaded, i.e., the intrinsic
6138 represents a family of functions that perform the same operation but on
6139 different data types. Because LLVM can represent over 8 million
6140 different integer types, overloading is used commonly to allow an
6141 intrinsic function to operate on any integer type. One or more of the
6142 argument types or the result type can be overloaded to accept any
6143 integer type. Argument types may also be defined as exactly matching a
6144 previous argument's type or the result type. This allows an intrinsic
6145 function which accepts multiple arguments, but needs all of them to be
6146 of the same type, to only be overloaded with respect to a single
6147 argument or the result.
6149 Overloaded intrinsics will have the names of its overloaded argument
6150 types encoded into its function name, each preceded by a period. Only
6151 those types which are overloaded result in a name suffix. Arguments
6152 whose type is matched against another type do not. For example, the
6153 ``llvm.ctpop`` function can take an integer of any width and returns an
6154 integer of exactly the same integer width. This leads to a family of
6155 functions such as ``i8 @llvm.ctpop.i8(i8 %val)`` and
6156 ``i29 @llvm.ctpop.i29(i29 %val)``. Only one type, the return type, is
6157 overloaded, and only one type suffix is required. Because the argument's
6158 type is matched against the return type, it does not require its own
6161 To learn how to add an intrinsic function, please see the `Extending
6162 LLVM Guide <ExtendingLLVM.html>`_.
6166 Variable Argument Handling Intrinsics
6167 -------------------------------------
6169 Variable argument support is defined in LLVM with the
6170 :ref:`va_arg <i_va_arg>` instruction and these three intrinsic
6171 functions. These functions are related to the similarly named macros
6172 defined in the ``<stdarg.h>`` header file.
6174 All of these functions operate on arguments that use a target-specific
6175 value type "``va_list``". The LLVM assembly language reference manual
6176 does not define what this type is, so all transformations should be
6177 prepared to handle these functions regardless of the type used.
6179 This example shows how the :ref:`va_arg <i_va_arg>` instruction and the
6180 variable argument handling intrinsic functions are used.
6182 .. code-block:: llvm
6184 define i32 @test(i32 %X, ...) {
6185 ; Initialize variable argument processing
6187 %ap2 = bitcast i8** %ap to i8*
6188 call void @llvm.va_start(i8* %ap2)
6190 ; Read a single integer argument
6191 %tmp = va_arg i8** %ap, i32
6193 ; Demonstrate usage of llvm.va_copy and llvm.va_end
6195 %aq2 = bitcast i8** %aq to i8*
6196 call void @llvm.va_copy(i8* %aq2, i8* %ap2)
6197 call void @llvm.va_end(i8* %aq2)
6199 ; Stop processing of arguments.
6200 call void @llvm.va_end(i8* %ap2)
6204 declare void @llvm.va_start(i8*)
6205 declare void @llvm.va_copy(i8*, i8*)
6206 declare void @llvm.va_end(i8*)
6210 '``llvm.va_start``' Intrinsic
6211 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6218 declare void %llvm.va_start(i8* <arglist>)
6223 The '``llvm.va_start``' intrinsic initializes ``*<arglist>`` for
6224 subsequent use by ``va_arg``.
6229 The argument is a pointer to a ``va_list`` element to initialize.
6234 The '``llvm.va_start``' intrinsic works just like the ``va_start`` macro
6235 available in C. In a target-dependent way, it initializes the
6236 ``va_list`` element to which the argument points, so that the next call
6237 to ``va_arg`` will produce the first variable argument passed to the
6238 function. Unlike the C ``va_start`` macro, this intrinsic does not need
6239 to know the last argument of the function as the compiler can figure
6242 '``llvm.va_end``' Intrinsic
6243 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6250 declare void @llvm.va_end(i8* <arglist>)
6255 The '``llvm.va_end``' intrinsic destroys ``*<arglist>``, which has been
6256 initialized previously with ``llvm.va_start`` or ``llvm.va_copy``.
6261 The argument is a pointer to a ``va_list`` to destroy.
6266 The '``llvm.va_end``' intrinsic works just like the ``va_end`` macro
6267 available in C. In a target-dependent way, it destroys the ``va_list``
6268 element to which the argument points. Calls to
6269 :ref:`llvm.va_start <int_va_start>` and
6270 :ref:`llvm.va_copy <int_va_copy>` must be matched exactly with calls to
6275 '``llvm.va_copy``' Intrinsic
6276 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6283 declare void @llvm.va_copy(i8* <destarglist>, i8* <srcarglist>)
6288 The '``llvm.va_copy``' intrinsic copies the current argument position
6289 from the source argument list to the destination argument list.
6294 The first argument is a pointer to a ``va_list`` element to initialize.
6295 The second argument is a pointer to a ``va_list`` element to copy from.
6300 The '``llvm.va_copy``' intrinsic works just like the ``va_copy`` macro
6301 available in C. In a target-dependent way, it copies the source
6302 ``va_list`` element into the destination ``va_list`` element. This
6303 intrinsic is necessary because the `` llvm.va_start`` intrinsic may be
6304 arbitrarily complex and require, for example, memory allocation.
6306 Accurate Garbage Collection Intrinsics
6307 --------------------------------------
6309 LLVM support for `Accurate Garbage Collection <GarbageCollection.html>`_
6310 (GC) requires the implementation and generation of these intrinsics.
6311 These intrinsics allow identification of :ref:`GC roots on the
6312 stack <int_gcroot>`, as well as garbage collector implementations that
6313 require :ref:`read <int_gcread>` and :ref:`write <int_gcwrite>` barriers.
6314 Front-ends for type-safe garbage collected languages should generate
6315 these intrinsics to make use of the LLVM garbage collectors. For more
6316 details, see `Accurate Garbage Collection with
6317 LLVM <GarbageCollection.html>`_.
6319 The garbage collection intrinsics only operate on objects in the generic
6320 address space (address space zero).
6324 '``llvm.gcroot``' Intrinsic
6325 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6332 declare void @llvm.gcroot(i8** %ptrloc, i8* %metadata)
6337 The '``llvm.gcroot``' intrinsic declares the existence of a GC root to
6338 the code generator, and allows some metadata to be associated with it.
6343 The first argument specifies the address of a stack object that contains
6344 the root pointer. The second pointer (which must be either a constant or
6345 a global value address) contains the meta-data to be associated with the
6351 At runtime, a call to this intrinsic stores a null pointer into the
6352 "ptrloc" location. At compile-time, the code generator generates
6353 information to allow the runtime to find the pointer at GC safe points.
6354 The '``llvm.gcroot``' intrinsic may only be used in a function which
6355 :ref:`specifies a GC algorithm <gc>`.
6359 '``llvm.gcread``' Intrinsic
6360 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6367 declare i8* @llvm.gcread(i8* %ObjPtr, i8** %Ptr)
6372 The '``llvm.gcread``' intrinsic identifies reads of references from heap
6373 locations, allowing garbage collector implementations that require read
6379 The second argument is the address to read from, which should be an
6380 address allocated from the garbage collector. The first object is a
6381 pointer to the start of the referenced object, if needed by the language
6382 runtime (otherwise null).
6387 The '``llvm.gcread``' intrinsic has the same semantics as a load
6388 instruction, but may be replaced with substantially more complex code by
6389 the garbage collector runtime, as needed. The '``llvm.gcread``'
6390 intrinsic may only be used in a function which :ref:`specifies a GC
6395 '``llvm.gcwrite``' Intrinsic
6396 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6403 declare void @llvm.gcwrite(i8* %P1, i8* %Obj, i8** %P2)
6408 The '``llvm.gcwrite``' intrinsic identifies writes of references to heap
6409 locations, allowing garbage collector implementations that require write
6410 barriers (such as generational or reference counting collectors).
6415 The first argument is the reference to store, the second is the start of
6416 the object to store it to, and the third is the address of the field of
6417 Obj to store to. If the runtime does not require a pointer to the
6418 object, Obj may be null.
6423 The '``llvm.gcwrite``' intrinsic has the same semantics as a store
6424 instruction, but may be replaced with substantially more complex code by
6425 the garbage collector runtime, as needed. The '``llvm.gcwrite``'
6426 intrinsic may only be used in a function which :ref:`specifies a GC
6429 Code Generator Intrinsics
6430 -------------------------
6432 These intrinsics are provided by LLVM to expose special features that
6433 may only be implemented with code generator support.
6435 '``llvm.returnaddress``' Intrinsic
6436 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6443 declare i8 *@llvm.returnaddress(i32 <level>)
6448 The '``llvm.returnaddress``' intrinsic attempts to compute a
6449 target-specific value indicating the return address of the current
6450 function or one of its callers.
6455 The argument to this intrinsic indicates which function to return the
6456 address for. Zero indicates the calling function, one indicates its
6457 caller, etc. The argument is **required** to be a constant integer
6463 The '``llvm.returnaddress``' intrinsic either returns a pointer
6464 indicating the return address of the specified call frame, or zero if it
6465 cannot be identified. The value returned by this intrinsic is likely to
6466 be incorrect or 0 for arguments other than zero, so it should only be
6467 used for debugging purposes.
6469 Note that calling this intrinsic does not prevent function inlining or
6470 other aggressive transformations, so the value returned may not be that
6471 of the obvious source-language caller.
6473 '``llvm.frameaddress``' Intrinsic
6474 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6481 declare i8* @llvm.frameaddress(i32 <level>)
6486 The '``llvm.frameaddress``' intrinsic attempts to return the
6487 target-specific frame pointer value for the specified stack frame.
6492 The argument to this intrinsic indicates which function to return the
6493 frame pointer for. Zero indicates the calling function, one indicates
6494 its caller, etc. The argument is **required** to be a constant integer
6500 The '``llvm.frameaddress``' intrinsic either returns a pointer
6501 indicating the frame address of the specified call frame, or zero if it
6502 cannot be identified. The value returned by this intrinsic is likely to
6503 be incorrect or 0 for arguments other than zero, so it should only be
6504 used for debugging purposes.
6506 Note that calling this intrinsic does not prevent function inlining or
6507 other aggressive transformations, so the value returned may not be that
6508 of the obvious source-language caller.
6512 '``llvm.stacksave``' Intrinsic
6513 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6520 declare i8* @llvm.stacksave()
6525 The '``llvm.stacksave``' intrinsic is used to remember the current state
6526 of the function stack, for use with
6527 :ref:`llvm.stackrestore <int_stackrestore>`. This is useful for
6528 implementing language features like scoped automatic variable sized
6534 This intrinsic returns a opaque pointer value that can be passed to
6535 :ref:`llvm.stackrestore <int_stackrestore>`. When an
6536 ``llvm.stackrestore`` intrinsic is executed with a value saved from
6537 ``llvm.stacksave``, it effectively restores the state of the stack to
6538 the state it was in when the ``llvm.stacksave`` intrinsic executed. In
6539 practice, this pops any :ref:`alloca <i_alloca>` blocks from the stack that
6540 were allocated after the ``llvm.stacksave`` was executed.
6542 .. _int_stackrestore:
6544 '``llvm.stackrestore``' Intrinsic
6545 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6552 declare void @llvm.stackrestore(i8* %ptr)
6557 The '``llvm.stackrestore``' intrinsic is used to restore the state of
6558 the function stack to the state it was in when the corresponding
6559 :ref:`llvm.stacksave <int_stacksave>` intrinsic executed. This is
6560 useful for implementing language features like scoped automatic variable
6561 sized arrays in C99.
6566 See the description for :ref:`llvm.stacksave <int_stacksave>`.
6568 '``llvm.prefetch``' Intrinsic
6569 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6576 declare void @llvm.prefetch(i8* <address>, i32 <rw>, i32 <locality>, i32 <cache type>)
6581 The '``llvm.prefetch``' intrinsic is a hint to the code generator to
6582 insert a prefetch instruction if supported; otherwise, it is a noop.
6583 Prefetches have no effect on the behavior of the program but can change
6584 its performance characteristics.
6589 ``address`` is the address to be prefetched, ``rw`` is the specifier
6590 determining if the fetch should be for a read (0) or write (1), and
6591 ``locality`` is a temporal locality specifier ranging from (0) - no
6592 locality, to (3) - extremely local keep in cache. The ``cache type``
6593 specifies whether the prefetch is performed on the data (1) or
6594 instruction (0) cache. The ``rw``, ``locality`` and ``cache type``
6595 arguments must be constant integers.
6600 This intrinsic does not modify the behavior of the program. In
6601 particular, prefetches cannot trap and do not produce a value. On
6602 targets that support this intrinsic, the prefetch can provide hints to
6603 the processor cache for better performance.
6605 '``llvm.pcmarker``' Intrinsic
6606 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6613 declare void @llvm.pcmarker(i32 <id>)
6618 The '``llvm.pcmarker``' intrinsic is a method to export a Program
6619 Counter (PC) in a region of code to simulators and other tools. The
6620 method is target specific, but it is expected that the marker will use
6621 exported symbols to transmit the PC of the marker. The marker makes no
6622 guarantees that it will remain with any specific instruction after
6623 optimizations. It is possible that the presence of a marker will inhibit
6624 optimizations. The intended use is to be inserted after optimizations to
6625 allow correlations of simulation runs.
6630 ``id`` is a numerical id identifying the marker.
6635 This intrinsic does not modify the behavior of the program. Backends
6636 that do not support this intrinsic may ignore it.
6638 '``llvm.readcyclecounter``' Intrinsic
6639 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6646 declare i64 @llvm.readcyclecounter()
6651 The '``llvm.readcyclecounter``' intrinsic provides access to the cycle
6652 counter register (or similar low latency, high accuracy clocks) on those
6653 targets that support it. On X86, it should map to RDTSC. On Alpha, it
6654 should map to RPCC. As the backing counters overflow quickly (on the
6655 order of 9 seconds on alpha), this should only be used for small
6661 When directly supported, reading the cycle counter should not modify any
6662 memory. Implementations are allowed to either return a application
6663 specific value or a system wide value. On backends without support, this
6664 is lowered to a constant 0.
6666 Note that runtime support may be conditional on the privilege-level code is
6667 running at and the host platform.
6669 Standard C Library Intrinsics
6670 -----------------------------
6672 LLVM provides intrinsics for a few important standard C library
6673 functions. These intrinsics allow source-language front-ends to pass
6674 information about the alignment of the pointer arguments to the code
6675 generator, providing opportunity for more efficient code generation.
6679 '``llvm.memcpy``' Intrinsic
6680 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6685 This is an overloaded intrinsic. You can use ``llvm.memcpy`` on any
6686 integer bit width and for different address spaces. Not all targets
6687 support all bit widths however.
6691 declare void @llvm.memcpy.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
6692 i32 <len>, i32 <align>, i1 <isvolatile>)
6693 declare void @llvm.memcpy.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
6694 i64 <len>, i32 <align>, i1 <isvolatile>)
6699 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
6700 source location to the destination location.
6702 Note that, unlike the standard libc function, the ``llvm.memcpy.*``
6703 intrinsics do not return a value, takes extra alignment/isvolatile
6704 arguments and the pointers can be in specified address spaces.
6709 The first argument is a pointer to the destination, the second is a
6710 pointer to the source. The third argument is an integer argument
6711 specifying the number of bytes to copy, the fourth argument is the
6712 alignment of the source and destination locations, and the fifth is a
6713 boolean indicating a volatile access.
6715 If the call to this intrinsic has an alignment value that is not 0 or 1,
6716 then the caller guarantees that both the source and destination pointers
6717 are aligned to that boundary.
6719 If the ``isvolatile`` parameter is ``true``, the ``llvm.memcpy`` call is
6720 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
6721 very cleanly specified and it is unwise to depend on it.
6726 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
6727 source location to the destination location, which are not allowed to
6728 overlap. It copies "len" bytes of memory over. If the argument is known
6729 to be aligned to some boundary, this can be specified as the fourth
6730 argument, otherwise it should be set to 0 or 1.
6732 '``llvm.memmove``' Intrinsic
6733 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6738 This is an overloaded intrinsic. You can use llvm.memmove on any integer
6739 bit width and for different address space. Not all targets support all
6744 declare void @llvm.memmove.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
6745 i32 <len>, i32 <align>, i1 <isvolatile>)
6746 declare void @llvm.memmove.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
6747 i64 <len>, i32 <align>, i1 <isvolatile>)
6752 The '``llvm.memmove.*``' intrinsics move a block of memory from the
6753 source location to the destination location. It is similar to the
6754 '``llvm.memcpy``' intrinsic but allows the two memory locations to
6757 Note that, unlike the standard libc function, the ``llvm.memmove.*``
6758 intrinsics do not return a value, takes extra alignment/isvolatile
6759 arguments and the pointers can be in specified address spaces.
6764 The first argument is a pointer to the destination, the second is a
6765 pointer to the source. The third argument is an integer argument
6766 specifying the number of bytes to copy, the fourth argument is the
6767 alignment of the source and destination locations, and the fifth is a
6768 boolean indicating a volatile access.
6770 If the call to this intrinsic has an alignment value that is not 0 or 1,
6771 then the caller guarantees that the source and destination pointers are
6772 aligned to that boundary.
6774 If the ``isvolatile`` parameter is ``true``, the ``llvm.memmove`` call
6775 is a :ref:`volatile operation <volatile>`. The detailed access behavior is
6776 not very cleanly specified and it is unwise to depend on it.
6781 The '``llvm.memmove.*``' intrinsics copy a block of memory from the
6782 source location to the destination location, which may overlap. It
6783 copies "len" bytes of memory over. If the argument is known to be
6784 aligned to some boundary, this can be specified as the fourth argument,
6785 otherwise it should be set to 0 or 1.
6787 '``llvm.memset.*``' Intrinsics
6788 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6793 This is an overloaded intrinsic. You can use llvm.memset on any integer
6794 bit width and for different address spaces. However, not all targets
6795 support all bit widths.
6799 declare void @llvm.memset.p0i8.i32(i8* <dest>, i8 <val>,
6800 i32 <len>, i32 <align>, i1 <isvolatile>)
6801 declare void @llvm.memset.p0i8.i64(i8* <dest>, i8 <val>,
6802 i64 <len>, i32 <align>, i1 <isvolatile>)
6807 The '``llvm.memset.*``' intrinsics fill a block of memory with a
6808 particular byte value.
6810 Note that, unlike the standard libc function, the ``llvm.memset``
6811 intrinsic does not return a value and takes extra alignment/volatile
6812 arguments. Also, the destination can be in an arbitrary address space.
6817 The first argument is a pointer to the destination to fill, the second
6818 is the byte value with which to fill it, the third argument is an
6819 integer argument specifying the number of bytes to fill, and the fourth
6820 argument is the known alignment of the destination location.
6822 If the call to this intrinsic has an alignment value that is not 0 or 1,
6823 then the caller guarantees that the destination pointer is aligned to
6826 If the ``isvolatile`` parameter is ``true``, the ``llvm.memset`` call is
6827 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
6828 very cleanly specified and it is unwise to depend on it.
6833 The '``llvm.memset.*``' intrinsics fill "len" bytes of memory starting
6834 at the destination location. If the argument is known to be aligned to
6835 some boundary, this can be specified as the fourth argument, otherwise
6836 it should be set to 0 or 1.
6838 '``llvm.sqrt.*``' Intrinsic
6839 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6844 This is an overloaded intrinsic. You can use ``llvm.sqrt`` on any
6845 floating point or vector of floating point type. Not all targets support
6850 declare float @llvm.sqrt.f32(float %Val)
6851 declare double @llvm.sqrt.f64(double %Val)
6852 declare x86_fp80 @llvm.sqrt.f80(x86_fp80 %Val)
6853 declare fp128 @llvm.sqrt.f128(fp128 %Val)
6854 declare ppc_fp128 @llvm.sqrt.ppcf128(ppc_fp128 %Val)
6859 The '``llvm.sqrt``' intrinsics return the sqrt of the specified operand,
6860 returning the same value as the libm '``sqrt``' functions would. Unlike
6861 ``sqrt`` in libm, however, ``llvm.sqrt`` has undefined behavior for
6862 negative numbers other than -0.0 (which allows for better optimization,
6863 because there is no need to worry about errno being set).
6864 ``llvm.sqrt(-0.0)`` is defined to return -0.0 like IEEE sqrt.
6869 The argument and return value are floating point numbers of the same
6875 This function returns the sqrt of the specified operand if it is a
6876 nonnegative floating point number.
6878 '``llvm.powi.*``' Intrinsic
6879 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6884 This is an overloaded intrinsic. You can use ``llvm.powi`` on any
6885 floating point or vector of floating point type. Not all targets support
6890 declare float @llvm.powi.f32(float %Val, i32 %power)
6891 declare double @llvm.powi.f64(double %Val, i32 %power)
6892 declare x86_fp80 @llvm.powi.f80(x86_fp80 %Val, i32 %power)
6893 declare fp128 @llvm.powi.f128(fp128 %Val, i32 %power)
6894 declare ppc_fp128 @llvm.powi.ppcf128(ppc_fp128 %Val, i32 %power)
6899 The '``llvm.powi.*``' intrinsics return the first operand raised to the
6900 specified (positive or negative) power. The order of evaluation of
6901 multiplications is not defined. When a vector of floating point type is
6902 used, the second argument remains a scalar integer value.
6907 The second argument is an integer power, and the first is a value to
6908 raise to that power.
6913 This function returns the first value raised to the second power with an
6914 unspecified sequence of rounding operations.
6916 '``llvm.sin.*``' Intrinsic
6917 ^^^^^^^^^^^^^^^^^^^^^^^^^^
6922 This is an overloaded intrinsic. You can use ``llvm.sin`` on any
6923 floating point or vector of floating point type. Not all targets support
6928 declare float @llvm.sin.f32(float %Val)
6929 declare double @llvm.sin.f64(double %Val)
6930 declare x86_fp80 @llvm.sin.f80(x86_fp80 %Val)
6931 declare fp128 @llvm.sin.f128(fp128 %Val)
6932 declare ppc_fp128 @llvm.sin.ppcf128(ppc_fp128 %Val)
6937 The '``llvm.sin.*``' intrinsics return the sine of the operand.
6942 The argument and return value are floating point numbers of the same
6948 This function returns the sine of the specified operand, returning the
6949 same values as the libm ``sin`` functions would, and handles error
6950 conditions in the same way.
6952 '``llvm.cos.*``' Intrinsic
6953 ^^^^^^^^^^^^^^^^^^^^^^^^^^
6958 This is an overloaded intrinsic. You can use ``llvm.cos`` on any
6959 floating point or vector of floating point type. Not all targets support
6964 declare float @llvm.cos.f32(float %Val)
6965 declare double @llvm.cos.f64(double %Val)
6966 declare x86_fp80 @llvm.cos.f80(x86_fp80 %Val)
6967 declare fp128 @llvm.cos.f128(fp128 %Val)
6968 declare ppc_fp128 @llvm.cos.ppcf128(ppc_fp128 %Val)
6973 The '``llvm.cos.*``' intrinsics return the cosine of the operand.
6978 The argument and return value are floating point numbers of the same
6984 This function returns the cosine of the specified operand, returning the
6985 same values as the libm ``cos`` functions would, and handles error
6986 conditions in the same way.
6988 '``llvm.pow.*``' Intrinsic
6989 ^^^^^^^^^^^^^^^^^^^^^^^^^^
6994 This is an overloaded intrinsic. You can use ``llvm.pow`` on any
6995 floating point or vector of floating point type. Not all targets support
7000 declare float @llvm.pow.f32(float %Val, float %Power)
7001 declare double @llvm.pow.f64(double %Val, double %Power)
7002 declare x86_fp80 @llvm.pow.f80(x86_fp80 %Val, x86_fp80 %Power)
7003 declare fp128 @llvm.pow.f128(fp128 %Val, fp128 %Power)
7004 declare ppc_fp128 @llvm.pow.ppcf128(ppc_fp128 %Val, ppc_fp128 Power)
7009 The '``llvm.pow.*``' intrinsics return the first operand raised to the
7010 specified (positive or negative) power.
7015 The second argument is a floating point power, and the first is a value
7016 to raise to that power.
7021 This function returns the first value raised to the second power,
7022 returning the same values as the libm ``pow`` functions would, and
7023 handles error conditions in the same way.
7025 '``llvm.exp.*``' Intrinsic
7026 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7031 This is an overloaded intrinsic. You can use ``llvm.exp`` on any
7032 floating point or vector of floating point type. Not all targets support
7037 declare float @llvm.exp.f32(float %Val)
7038 declare double @llvm.exp.f64(double %Val)
7039 declare x86_fp80 @llvm.exp.f80(x86_fp80 %Val)
7040 declare fp128 @llvm.exp.f128(fp128 %Val)
7041 declare ppc_fp128 @llvm.exp.ppcf128(ppc_fp128 %Val)
7046 The '``llvm.exp.*``' intrinsics perform the exp function.
7051 The argument and return value are floating point numbers of the same
7057 This function returns the same values as the libm ``exp`` functions
7058 would, and handles error conditions in the same way.
7060 '``llvm.exp2.*``' Intrinsic
7061 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7066 This is an overloaded intrinsic. You can use ``llvm.exp2`` on any
7067 floating point or vector of floating point type. Not all targets support
7072 declare float @llvm.exp2.f32(float %Val)
7073 declare double @llvm.exp2.f64(double %Val)
7074 declare x86_fp80 @llvm.exp2.f80(x86_fp80 %Val)
7075 declare fp128 @llvm.exp2.f128(fp128 %Val)
7076 declare ppc_fp128 @llvm.exp2.ppcf128(ppc_fp128 %Val)
7081 The '``llvm.exp2.*``' intrinsics perform the exp2 function.
7086 The argument and return value are floating point numbers of the same
7092 This function returns the same values as the libm ``exp2`` functions
7093 would, and handles error conditions in the same way.
7095 '``llvm.log.*``' Intrinsic
7096 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7101 This is an overloaded intrinsic. You can use ``llvm.log`` on any
7102 floating point or vector of floating point type. Not all targets support
7107 declare float @llvm.log.f32(float %Val)
7108 declare double @llvm.log.f64(double %Val)
7109 declare x86_fp80 @llvm.log.f80(x86_fp80 %Val)
7110 declare fp128 @llvm.log.f128(fp128 %Val)
7111 declare ppc_fp128 @llvm.log.ppcf128(ppc_fp128 %Val)
7116 The '``llvm.log.*``' intrinsics perform the log function.
7121 The argument and return value are floating point numbers of the same
7127 This function returns the same values as the libm ``log`` functions
7128 would, and handles error conditions in the same way.
7130 '``llvm.log10.*``' Intrinsic
7131 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7136 This is an overloaded intrinsic. You can use ``llvm.log10`` on any
7137 floating point or vector of floating point type. Not all targets support
7142 declare float @llvm.log10.f32(float %Val)
7143 declare double @llvm.log10.f64(double %Val)
7144 declare x86_fp80 @llvm.log10.f80(x86_fp80 %Val)
7145 declare fp128 @llvm.log10.f128(fp128 %Val)
7146 declare ppc_fp128 @llvm.log10.ppcf128(ppc_fp128 %Val)
7151 The '``llvm.log10.*``' intrinsics perform the log10 function.
7156 The argument and return value are floating point numbers of the same
7162 This function returns the same values as the libm ``log10`` functions
7163 would, and handles error conditions in the same way.
7165 '``llvm.log2.*``' Intrinsic
7166 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7171 This is an overloaded intrinsic. You can use ``llvm.log2`` on any
7172 floating point or vector of floating point type. Not all targets support
7177 declare float @llvm.log2.f32(float %Val)
7178 declare double @llvm.log2.f64(double %Val)
7179 declare x86_fp80 @llvm.log2.f80(x86_fp80 %Val)
7180 declare fp128 @llvm.log2.f128(fp128 %Val)
7181 declare ppc_fp128 @llvm.log2.ppcf128(ppc_fp128 %Val)
7186 The '``llvm.log2.*``' intrinsics perform the log2 function.
7191 The argument and return value are floating point numbers of the same
7197 This function returns the same values as the libm ``log2`` functions
7198 would, and handles error conditions in the same way.
7200 '``llvm.fma.*``' Intrinsic
7201 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7206 This is an overloaded intrinsic. You can use ``llvm.fma`` on any
7207 floating point or vector of floating point type. Not all targets support
7212 declare float @llvm.fma.f32(float %a, float %b, float %c)
7213 declare double @llvm.fma.f64(double %a, double %b, double %c)
7214 declare x86_fp80 @llvm.fma.f80(x86_fp80 %a, x86_fp80 %b, x86_fp80 %c)
7215 declare fp128 @llvm.fma.f128(fp128 %a, fp128 %b, fp128 %c)
7216 declare ppc_fp128 @llvm.fma.ppcf128(ppc_fp128 %a, ppc_fp128 %b, ppc_fp128 %c)
7221 The '``llvm.fma.*``' intrinsics perform the fused multiply-add
7227 The argument and return value are floating point numbers of the same
7233 This function returns the same values as the libm ``fma`` functions
7236 '``llvm.fabs.*``' Intrinsic
7237 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7242 This is an overloaded intrinsic. You can use ``llvm.fabs`` on any
7243 floating point or vector of floating point type. Not all targets support
7248 declare float @llvm.fabs.f32(float %Val)
7249 declare double @llvm.fabs.f64(double %Val)
7250 declare x86_fp80 @llvm.fabs.f80(x86_fp80 %Val)
7251 declare fp128 @llvm.fabs.f128(fp128 %Val)
7252 declare ppc_fp128 @llvm.fabs.ppcf128(ppc_fp128 %Val)
7257 The '``llvm.fabs.*``' intrinsics return the absolute value of the
7263 The argument and return value are floating point numbers of the same
7269 This function returns the same values as the libm ``fabs`` functions
7270 would, and handles error conditions in the same way.
7272 '``llvm.floor.*``' Intrinsic
7273 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7278 This is an overloaded intrinsic. You can use ``llvm.floor`` on any
7279 floating point or vector of floating point type. Not all targets support
7284 declare float @llvm.floor.f32(float %Val)
7285 declare double @llvm.floor.f64(double %Val)
7286 declare x86_fp80 @llvm.floor.f80(x86_fp80 %Val)
7287 declare fp128 @llvm.floor.f128(fp128 %Val)
7288 declare ppc_fp128 @llvm.floor.ppcf128(ppc_fp128 %Val)
7293 The '``llvm.floor.*``' intrinsics return the floor of the operand.
7298 The argument and return value are floating point numbers of the same
7304 This function returns the same values as the libm ``floor`` functions
7305 would, and handles error conditions in the same way.
7307 '``llvm.ceil.*``' Intrinsic
7308 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7313 This is an overloaded intrinsic. You can use ``llvm.ceil`` on any
7314 floating point or vector of floating point type. Not all targets support
7319 declare float @llvm.ceil.f32(float %Val)
7320 declare double @llvm.ceil.f64(double %Val)
7321 declare x86_fp80 @llvm.ceil.f80(x86_fp80 %Val)
7322 declare fp128 @llvm.ceil.f128(fp128 %Val)
7323 declare ppc_fp128 @llvm.ceil.ppcf128(ppc_fp128 %Val)
7328 The '``llvm.ceil.*``' intrinsics return the ceiling of the operand.
7333 The argument and return value are floating point numbers of the same
7339 This function returns the same values as the libm ``ceil`` functions
7340 would, and handles error conditions in the same way.
7342 '``llvm.trunc.*``' Intrinsic
7343 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7348 This is an overloaded intrinsic. You can use ``llvm.trunc`` on any
7349 floating point or vector of floating point type. Not all targets support
7354 declare float @llvm.trunc.f32(float %Val)
7355 declare double @llvm.trunc.f64(double %Val)
7356 declare x86_fp80 @llvm.trunc.f80(x86_fp80 %Val)
7357 declare fp128 @llvm.trunc.f128(fp128 %Val)
7358 declare ppc_fp128 @llvm.trunc.ppcf128(ppc_fp128 %Val)
7363 The '``llvm.trunc.*``' intrinsics returns the operand rounded to the
7364 nearest integer not larger in magnitude than the operand.
7369 The argument and return value are floating point numbers of the same
7375 This function returns the same values as the libm ``trunc`` functions
7376 would, and handles error conditions in the same way.
7378 '``llvm.rint.*``' Intrinsic
7379 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7384 This is an overloaded intrinsic. You can use ``llvm.rint`` on any
7385 floating point or vector of floating point type. Not all targets support
7390 declare float @llvm.rint.f32(float %Val)
7391 declare double @llvm.rint.f64(double %Val)
7392 declare x86_fp80 @llvm.rint.f80(x86_fp80 %Val)
7393 declare fp128 @llvm.rint.f128(fp128 %Val)
7394 declare ppc_fp128 @llvm.rint.ppcf128(ppc_fp128 %Val)
7399 The '``llvm.rint.*``' intrinsics returns the operand rounded to the
7400 nearest integer. It may raise an inexact floating-point exception if the
7401 operand isn't an integer.
7406 The argument and return value are floating point numbers of the same
7412 This function returns the same values as the libm ``rint`` functions
7413 would, and handles error conditions in the same way.
7415 '``llvm.nearbyint.*``' Intrinsic
7416 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7421 This is an overloaded intrinsic. You can use ``llvm.nearbyint`` on any
7422 floating point or vector of floating point type. Not all targets support
7427 declare float @llvm.nearbyint.f32(float %Val)
7428 declare double @llvm.nearbyint.f64(double %Val)
7429 declare x86_fp80 @llvm.nearbyint.f80(x86_fp80 %Val)
7430 declare fp128 @llvm.nearbyint.f128(fp128 %Val)
7431 declare ppc_fp128 @llvm.nearbyint.ppcf128(ppc_fp128 %Val)
7436 The '``llvm.nearbyint.*``' intrinsics returns the operand rounded to the
7442 The argument and return value are floating point numbers of the same
7448 This function returns the same values as the libm ``nearbyint``
7449 functions would, and handles error conditions in the same way.
7451 Bit Manipulation Intrinsics
7452 ---------------------------
7454 LLVM provides intrinsics for a few important bit manipulation
7455 operations. These allow efficient code generation for some algorithms.
7457 '``llvm.bswap.*``' Intrinsics
7458 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7463 This is an overloaded intrinsic function. You can use bswap on any
7464 integer type that is an even number of bytes (i.e. BitWidth % 16 == 0).
7468 declare i16 @llvm.bswap.i16(i16 <id>)
7469 declare i32 @llvm.bswap.i32(i32 <id>)
7470 declare i64 @llvm.bswap.i64(i64 <id>)
7475 The '``llvm.bswap``' family of intrinsics is used to byte swap integer
7476 values with an even number of bytes (positive multiple of 16 bits).
7477 These are useful for performing operations on data that is not in the
7478 target's native byte order.
7483 The ``llvm.bswap.i16`` intrinsic returns an i16 value that has the high
7484 and low byte of the input i16 swapped. Similarly, the ``llvm.bswap.i32``
7485 intrinsic returns an i32 value that has the four bytes of the input i32
7486 swapped, so that if the input bytes are numbered 0, 1, 2, 3 then the
7487 returned i32 will have its bytes in 3, 2, 1, 0 order. The
7488 ``llvm.bswap.i48``, ``llvm.bswap.i64`` and other intrinsics extend this
7489 concept to additional even-byte lengths (6 bytes, 8 bytes and more,
7492 '``llvm.ctpop.*``' Intrinsic
7493 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7498 This is an overloaded intrinsic. You can use llvm.ctpop on any integer
7499 bit width, or on any vector with integer elements. Not all targets
7500 support all bit widths or vector types, however.
7504 declare i8 @llvm.ctpop.i8(i8 <src>)
7505 declare i16 @llvm.ctpop.i16(i16 <src>)
7506 declare i32 @llvm.ctpop.i32(i32 <src>)
7507 declare i64 @llvm.ctpop.i64(i64 <src>)
7508 declare i256 @llvm.ctpop.i256(i256 <src>)
7509 declare <2 x i32> @llvm.ctpop.v2i32(<2 x i32> <src>)
7514 The '``llvm.ctpop``' family of intrinsics counts the number of bits set
7520 The only argument is the value to be counted. The argument may be of any
7521 integer type, or a vector with integer elements. The return type must
7522 match the argument type.
7527 The '``llvm.ctpop``' intrinsic counts the 1's in a variable, or within
7528 each element of a vector.
7530 '``llvm.ctlz.*``' Intrinsic
7531 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7536 This is an overloaded intrinsic. You can use ``llvm.ctlz`` on any
7537 integer bit width, or any vector whose elements are integers. Not all
7538 targets support all bit widths or vector types, however.
7542 declare i8 @llvm.ctlz.i8 (i8 <src>, i1 <is_zero_undef>)
7543 declare i16 @llvm.ctlz.i16 (i16 <src>, i1 <is_zero_undef>)
7544 declare i32 @llvm.ctlz.i32 (i32 <src>, i1 <is_zero_undef>)
7545 declare i64 @llvm.ctlz.i64 (i64 <src>, i1 <is_zero_undef>)
7546 declare i256 @llvm.ctlz.i256(i256 <src>, i1 <is_zero_undef>)
7547 declase <2 x i32> @llvm.ctlz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
7552 The '``llvm.ctlz``' family of intrinsic functions counts the number of
7553 leading zeros in a variable.
7558 The first argument is the value to be counted. This argument may be of
7559 any integer type, or a vectory with integer element type. The return
7560 type must match the first argument type.
7562 The second argument must be a constant and is a flag to indicate whether
7563 the intrinsic should ensure that a zero as the first argument produces a
7564 defined result. Historically some architectures did not provide a
7565 defined result for zero values as efficiently, and many algorithms are
7566 now predicated on avoiding zero-value inputs.
7571 The '``llvm.ctlz``' intrinsic counts the leading (most significant)
7572 zeros in a variable, or within each element of the vector. If
7573 ``src == 0`` then the result is the size in bits of the type of ``src``
7574 if ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
7575 ``llvm.ctlz(i32 2) = 30``.
7577 '``llvm.cttz.*``' Intrinsic
7578 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7583 This is an overloaded intrinsic. You can use ``llvm.cttz`` on any
7584 integer bit width, or any vector of integer elements. Not all targets
7585 support all bit widths or vector types, however.
7589 declare i8 @llvm.cttz.i8 (i8 <src>, i1 <is_zero_undef>)
7590 declare i16 @llvm.cttz.i16 (i16 <src>, i1 <is_zero_undef>)
7591 declare i32 @llvm.cttz.i32 (i32 <src>, i1 <is_zero_undef>)
7592 declare i64 @llvm.cttz.i64 (i64 <src>, i1 <is_zero_undef>)
7593 declare i256 @llvm.cttz.i256(i256 <src>, i1 <is_zero_undef>)
7594 declase <2 x i32> @llvm.cttz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
7599 The '``llvm.cttz``' family of intrinsic functions counts the number of
7605 The first argument is the value to be counted. This argument may be of
7606 any integer type, or a vectory with integer element type. The return
7607 type must match the first argument type.
7609 The second argument must be a constant and is a flag to indicate whether
7610 the intrinsic should ensure that a zero as the first argument produces a
7611 defined result. Historically some architectures did not provide a
7612 defined result for zero values as efficiently, and many algorithms are
7613 now predicated on avoiding zero-value inputs.
7618 The '``llvm.cttz``' intrinsic counts the trailing (least significant)
7619 zeros in a variable, or within each element of a vector. If ``src == 0``
7620 then the result is the size in bits of the type of ``src`` if
7621 ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
7622 ``llvm.cttz(2) = 1``.
7624 Arithmetic with Overflow Intrinsics
7625 -----------------------------------
7627 LLVM provides intrinsics for some arithmetic with overflow operations.
7629 '``llvm.sadd.with.overflow.*``' Intrinsics
7630 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7635 This is an overloaded intrinsic. You can use ``llvm.sadd.with.overflow``
7636 on any integer bit width.
7640 declare {i16, i1} @llvm.sadd.with.overflow.i16(i16 %a, i16 %b)
7641 declare {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
7642 declare {i64, i1} @llvm.sadd.with.overflow.i64(i64 %a, i64 %b)
7647 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
7648 a signed addition of the two arguments, and indicate whether an overflow
7649 occurred during the signed summation.
7654 The arguments (%a and %b) and the first element of the result structure
7655 may be of integer types of any bit width, but they must have the same
7656 bit width. The second element of the result structure must be of type
7657 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
7663 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
7664 a signed addition of the two variables. They return a structure --- the
7665 first element of which is the signed summation, and the second element
7666 of which is a bit specifying if the signed summation resulted in an
7672 .. code-block:: llvm
7674 %res = call {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
7675 %sum = extractvalue {i32, i1} %res, 0
7676 %obit = extractvalue {i32, i1} %res, 1
7677 br i1 %obit, label %overflow, label %normal
7679 '``llvm.uadd.with.overflow.*``' Intrinsics
7680 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7685 This is an overloaded intrinsic. You can use ``llvm.uadd.with.overflow``
7686 on any integer bit width.
7690 declare {i16, i1} @llvm.uadd.with.overflow.i16(i16 %a, i16 %b)
7691 declare {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
7692 declare {i64, i1} @llvm.uadd.with.overflow.i64(i64 %a, i64 %b)
7697 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
7698 an unsigned addition of the two arguments, and indicate whether a carry
7699 occurred during the unsigned summation.
7704 The arguments (%a and %b) and the first element of the result structure
7705 may be of integer types of any bit width, but they must have the same
7706 bit width. The second element of the result structure must be of type
7707 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
7713 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
7714 an unsigned addition of the two arguments. They return a structure --- the
7715 first element of which is the sum, and the second element of which is a
7716 bit specifying if the unsigned summation resulted in a carry.
7721 .. code-block:: llvm
7723 %res = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
7724 %sum = extractvalue {i32, i1} %res, 0
7725 %obit = extractvalue {i32, i1} %res, 1
7726 br i1 %obit, label %carry, label %normal
7728 '``llvm.ssub.with.overflow.*``' Intrinsics
7729 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7734 This is an overloaded intrinsic. You can use ``llvm.ssub.with.overflow``
7735 on any integer bit width.
7739 declare {i16, i1} @llvm.ssub.with.overflow.i16(i16 %a, i16 %b)
7740 declare {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
7741 declare {i64, i1} @llvm.ssub.with.overflow.i64(i64 %a, i64 %b)
7746 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
7747 a signed subtraction of the two arguments, and indicate whether an
7748 overflow occurred during the signed subtraction.
7753 The arguments (%a and %b) and the first element of the result structure
7754 may be of integer types of any bit width, but they must have the same
7755 bit width. The second element of the result structure must be of type
7756 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
7762 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
7763 a signed subtraction of the two arguments. They return a structure --- the
7764 first element of which is the subtraction, and the second element of
7765 which is a bit specifying if the signed subtraction resulted in an
7771 .. code-block:: llvm
7773 %res = call {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
7774 %sum = extractvalue {i32, i1} %res, 0
7775 %obit = extractvalue {i32, i1} %res, 1
7776 br i1 %obit, label %overflow, label %normal
7778 '``llvm.usub.with.overflow.*``' Intrinsics
7779 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7784 This is an overloaded intrinsic. You can use ``llvm.usub.with.overflow``
7785 on any integer bit width.
7789 declare {i16, i1} @llvm.usub.with.overflow.i16(i16 %a, i16 %b)
7790 declare {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
7791 declare {i64, i1} @llvm.usub.with.overflow.i64(i64 %a, i64 %b)
7796 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
7797 an unsigned subtraction of the two arguments, and indicate whether an
7798 overflow occurred during the unsigned subtraction.
7803 The arguments (%a and %b) and the first element of the result structure
7804 may be of integer types of any bit width, but they must have the same
7805 bit width. The second element of the result structure must be of type
7806 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
7812 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
7813 an unsigned subtraction of the two arguments. They return a structure ---
7814 the first element of which is the subtraction, and the second element of
7815 which is a bit specifying if the unsigned subtraction resulted in an
7821 .. code-block:: llvm
7823 %res = call {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
7824 %sum = extractvalue {i32, i1} %res, 0
7825 %obit = extractvalue {i32, i1} %res, 1
7826 br i1 %obit, label %overflow, label %normal
7828 '``llvm.smul.with.overflow.*``' Intrinsics
7829 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7834 This is an overloaded intrinsic. You can use ``llvm.smul.with.overflow``
7835 on any integer bit width.
7839 declare {i16, i1} @llvm.smul.with.overflow.i16(i16 %a, i16 %b)
7840 declare {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
7841 declare {i64, i1} @llvm.smul.with.overflow.i64(i64 %a, i64 %b)
7846 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
7847 a signed multiplication of the two arguments, and indicate whether an
7848 overflow occurred during the signed multiplication.
7853 The arguments (%a and %b) and the first element of the result structure
7854 may be of integer types of any bit width, but they must have the same
7855 bit width. The second element of the result structure must be of type
7856 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
7862 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
7863 a signed multiplication of the two arguments. They return a structure ---
7864 the first element of which is the multiplication, and the second element
7865 of which is a bit specifying if the signed multiplication resulted in an
7871 .. code-block:: llvm
7873 %res = call {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
7874 %sum = extractvalue {i32, i1} %res, 0
7875 %obit = extractvalue {i32, i1} %res, 1
7876 br i1 %obit, label %overflow, label %normal
7878 '``llvm.umul.with.overflow.*``' Intrinsics
7879 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7884 This is an overloaded intrinsic. You can use ``llvm.umul.with.overflow``
7885 on any integer bit width.
7889 declare {i16, i1} @llvm.umul.with.overflow.i16(i16 %a, i16 %b)
7890 declare {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
7891 declare {i64, i1} @llvm.umul.with.overflow.i64(i64 %a, i64 %b)
7896 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
7897 a unsigned multiplication of the two arguments, and indicate whether an
7898 overflow occurred during the unsigned multiplication.
7903 The arguments (%a and %b) and the first element of the result structure
7904 may be of integer types of any bit width, but they must have the same
7905 bit width. The second element of the result structure must be of type
7906 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
7912 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
7913 an unsigned multiplication of the two arguments. They return a structure ---
7914 the first element of which is the multiplication, and the second
7915 element of which is a bit specifying if the unsigned multiplication
7916 resulted in an overflow.
7921 .. code-block:: llvm
7923 %res = call {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
7924 %sum = extractvalue {i32, i1} %res, 0
7925 %obit = extractvalue {i32, i1} %res, 1
7926 br i1 %obit, label %overflow, label %normal
7928 Specialised Arithmetic Intrinsics
7929 ---------------------------------
7931 '``llvm.fmuladd.*``' Intrinsic
7932 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7939 declare float @llvm.fmuladd.f32(float %a, float %b, float %c)
7940 declare double @llvm.fmuladd.f64(double %a, double %b, double %c)
7945 The '``llvm.fmuladd.*``' intrinsic functions represent multiply-add
7946 expressions that can be fused if the code generator determines that (a) the
7947 target instruction set has support for a fused operation, and (b) that the
7948 fused operation is more efficient than the equivalent, separate pair of mul
7949 and add instructions.
7954 The '``llvm.fmuladd.*``' intrinsics each take three arguments: two
7955 multiplicands, a and b, and an addend c.
7964 %0 = call float @llvm.fmuladd.f32(%a, %b, %c)
7966 is equivalent to the expression a \* b + c, except that rounding will
7967 not be performed between the multiplication and addition steps if the
7968 code generator fuses the operations. Fusion is not guaranteed, even if
7969 the target platform supports it. If a fused multiply-add is required the
7970 corresponding llvm.fma.\* intrinsic function should be used instead.
7975 .. code-block:: llvm
7977 %r2 = call float @llvm.fmuladd.f32(float %a, float %b, float %c) ; yields {float}:r2 = (a * b) + c
7979 Half Precision Floating Point Intrinsics
7980 ----------------------------------------
7982 For most target platforms, half precision floating point is a
7983 storage-only format. This means that it is a dense encoding (in memory)
7984 but does not support computation in the format.
7986 This means that code must first load the half-precision floating point
7987 value as an i16, then convert it to float with
7988 :ref:`llvm.convert.from.fp16 <int_convert_from_fp16>`. Computation can
7989 then be performed on the float value (including extending to double
7990 etc). To store the value back to memory, it is first converted to float
7991 if needed, then converted to i16 with
7992 :ref:`llvm.convert.to.fp16 <int_convert_to_fp16>`, then storing as an
7995 .. _int_convert_to_fp16:
7997 '``llvm.convert.to.fp16``' Intrinsic
7998 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8005 declare i16 @llvm.convert.to.fp16(f32 %a)
8010 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion
8011 from single precision floating point format to half precision floating
8017 The intrinsic function contains single argument - the value to be
8023 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion
8024 from single precision floating point format to half precision floating
8025 point format. The return value is an ``i16`` which contains the
8031 .. code-block:: llvm
8033 %res = call i16 @llvm.convert.to.fp16(f32 %a)
8034 store i16 %res, i16* @x, align 2
8036 .. _int_convert_from_fp16:
8038 '``llvm.convert.from.fp16``' Intrinsic
8039 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8046 declare f32 @llvm.convert.from.fp16(i16 %a)
8051 The '``llvm.convert.from.fp16``' intrinsic function performs a
8052 conversion from half precision floating point format to single precision
8053 floating point format.
8058 The intrinsic function contains single argument - the value to be
8064 The '``llvm.convert.from.fp16``' intrinsic function performs a
8065 conversion from half single precision floating point format to single
8066 precision floating point format. The input half-float value is
8067 represented by an ``i16`` value.
8072 .. code-block:: llvm
8074 %a = load i16* @x, align 2
8075 %res = call f32 @llvm.convert.from.fp16(i16 %a)
8080 The LLVM debugger intrinsics (which all start with ``llvm.dbg.``
8081 prefix), are described in the `LLVM Source Level
8082 Debugging <SourceLevelDebugging.html#format_common_intrinsics>`_
8085 Exception Handling Intrinsics
8086 -----------------------------
8088 The LLVM exception handling intrinsics (which all start with
8089 ``llvm.eh.`` prefix), are described in the `LLVM Exception
8090 Handling <ExceptionHandling.html#format_common_intrinsics>`_ document.
8094 Trampoline Intrinsics
8095 ---------------------
8097 These intrinsics make it possible to excise one parameter, marked with
8098 the :ref:`nest <nest>` attribute, from a function. The result is a
8099 callable function pointer lacking the nest parameter - the caller does
8100 not need to provide a value for it. Instead, the value to use is stored
8101 in advance in a "trampoline", a block of memory usually allocated on the
8102 stack, which also contains code to splice the nest value into the
8103 argument list. This is used to implement the GCC nested function address
8106 For example, if the function is ``i32 f(i8* nest %c, i32 %x, i32 %y)``
8107 then the resulting function pointer has signature ``i32 (i32, i32)*``.
8108 It can be created as follows:
8110 .. code-block:: llvm
8112 %tramp = alloca [10 x i8], align 4 ; size and alignment only correct for X86
8113 %tramp1 = getelementptr [10 x i8]* %tramp, i32 0, i32 0
8114 call i8* @llvm.init.trampoline(i8* %tramp1, i8* bitcast (i32 (i8*, i32, i32)* @f to i8*), i8* %nval)
8115 %p = call i8* @llvm.adjust.trampoline(i8* %tramp1)
8116 %fp = bitcast i8* %p to i32 (i32, i32)*
8118 The call ``%val = call i32 %fp(i32 %x, i32 %y)`` is then equivalent to
8119 ``%val = call i32 %f(i8* %nval, i32 %x, i32 %y)``.
8123 '``llvm.init.trampoline``' Intrinsic
8124 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8131 declare void @llvm.init.trampoline(i8* <tramp>, i8* <func>, i8* <nval>)
8136 This fills the memory pointed to by ``tramp`` with executable code,
8137 turning it into a trampoline.
8142 The ``llvm.init.trampoline`` intrinsic takes three arguments, all
8143 pointers. The ``tramp`` argument must point to a sufficiently large and
8144 sufficiently aligned block of memory; this memory is written to by the
8145 intrinsic. Note that the size and the alignment are target-specific -
8146 LLVM currently provides no portable way of determining them, so a
8147 front-end that generates this intrinsic needs to have some
8148 target-specific knowledge. The ``func`` argument must hold a function
8149 bitcast to an ``i8*``.
8154 The block of memory pointed to by ``tramp`` is filled with target
8155 dependent code, turning it into a function. Then ``tramp`` needs to be
8156 passed to :ref:`llvm.adjust.trampoline <int_at>` to get a pointer which can
8157 be :ref:`bitcast (to a new function) and called <int_trampoline>`. The new
8158 function's signature is the same as that of ``func`` with any arguments
8159 marked with the ``nest`` attribute removed. At most one such ``nest``
8160 argument is allowed, and it must be of pointer type. Calling the new
8161 function is equivalent to calling ``func`` with the same argument list,
8162 but with ``nval`` used for the missing ``nest`` argument. If, after
8163 calling ``llvm.init.trampoline``, the memory pointed to by ``tramp`` is
8164 modified, then the effect of any later call to the returned function
8165 pointer is undefined.
8169 '``llvm.adjust.trampoline``' Intrinsic
8170 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8177 declare i8* @llvm.adjust.trampoline(i8* <tramp>)
8182 This performs any required machine-specific adjustment to the address of
8183 a trampoline (passed as ``tramp``).
8188 ``tramp`` must point to a block of memory which already has trampoline
8189 code filled in by a previous call to
8190 :ref:`llvm.init.trampoline <int_it>`.
8195 On some architectures the address of the code to be executed needs to be
8196 different to the address where the trampoline is actually stored. This
8197 intrinsic returns the executable address corresponding to ``tramp``
8198 after performing the required machine specific adjustments. The pointer
8199 returned can then be :ref:`bitcast and executed <int_trampoline>`.
8204 This class of intrinsics exists to information about the lifetime of
8205 memory objects and ranges where variables are immutable.
8207 '``llvm.lifetime.start``' Intrinsic
8208 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8215 declare void @llvm.lifetime.start(i64 <size>, i8* nocapture <ptr>)
8220 The '``llvm.lifetime.start``' intrinsic specifies the start of a memory
8226 The first argument is a constant integer representing the size of the
8227 object, or -1 if it is variable sized. The second argument is a pointer
8233 This intrinsic indicates that before this point in the code, the value
8234 of the memory pointed to by ``ptr`` is dead. This means that it is known
8235 to never be used and has an undefined value. A load from the pointer
8236 that precedes this intrinsic can be replaced with ``'undef'``.
8238 '``llvm.lifetime.end``' Intrinsic
8239 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8246 declare void @llvm.lifetime.end(i64 <size>, i8* nocapture <ptr>)
8251 The '``llvm.lifetime.end``' intrinsic specifies the end of a memory
8257 The first argument is a constant integer representing the size of the
8258 object, or -1 if it is variable sized. The second argument is a pointer
8264 This intrinsic indicates that after this point in the code, the value of
8265 the memory pointed to by ``ptr`` is dead. This means that it is known to
8266 never be used and has an undefined value. Any stores into the memory
8267 object following this intrinsic may be removed as dead.
8269 '``llvm.invariant.start``' Intrinsic
8270 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8277 declare {}* @llvm.invariant.start(i64 <size>, i8* nocapture <ptr>)
8282 The '``llvm.invariant.start``' intrinsic specifies that the contents of
8283 a memory object will not change.
8288 The first argument is a constant integer representing the size of the
8289 object, or -1 if it is variable sized. The second argument is a pointer
8295 This intrinsic indicates that until an ``llvm.invariant.end`` that uses
8296 the return value, the referenced memory location is constant and
8299 '``llvm.invariant.end``' Intrinsic
8300 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8307 declare void @llvm.invariant.end({}* <start>, i64 <size>, i8* nocapture <ptr>)
8312 The '``llvm.invariant.end``' intrinsic specifies that the contents of a
8313 memory object are mutable.
8318 The first argument is the matching ``llvm.invariant.start`` intrinsic.
8319 The second argument is a constant integer representing the size of the
8320 object, or -1 if it is variable sized and the third argument is a
8321 pointer to the object.
8326 This intrinsic indicates that the memory is mutable again.
8331 This class of intrinsics is designed to be generic and has no specific
8334 '``llvm.var.annotation``' Intrinsic
8335 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8342 declare void @llvm.var.annotation(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
8347 The '``llvm.var.annotation``' intrinsic.
8352 The first argument is a pointer to a value, the second is a pointer to a
8353 global string, the third is a pointer to a global string which is the
8354 source file name, and the last argument is the line number.
8359 This intrinsic allows annotation of local variables with arbitrary
8360 strings. This can be useful for special purpose optimizations that want
8361 to look for these annotations. These have no other defined use; they are
8362 ignored by code generation and optimization.
8364 '``llvm.ptr.annotation.*``' Intrinsic
8365 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8370 This is an overloaded intrinsic. You can use '``llvm.ptr.annotation``' on a
8371 pointer to an integer of any width. *NOTE* you must specify an address space for
8372 the pointer. The identifier for the default address space is the integer
8377 declare i8* @llvm.ptr.annotation.p<address space>i8(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
8378 declare i16* @llvm.ptr.annotation.p<address space>i16(i16* <val>, i8* <str>, i8* <str>, i32 <int>)
8379 declare i32* @llvm.ptr.annotation.p<address space>i32(i32* <val>, i8* <str>, i8* <str>, i32 <int>)
8380 declare i64* @llvm.ptr.annotation.p<address space>i64(i64* <val>, i8* <str>, i8* <str>, i32 <int>)
8381 declare i256* @llvm.ptr.annotation.p<address space>i256(i256* <val>, i8* <str>, i8* <str>, i32 <int>)
8386 The '``llvm.ptr.annotation``' intrinsic.
8391 The first argument is a pointer to an integer value of arbitrary bitwidth
8392 (result of some expression), the second is a pointer to a global string, the
8393 third is a pointer to a global string which is the source file name, and the
8394 last argument is the line number. It returns the value of the first argument.
8399 This intrinsic allows annotation of a pointer to an integer with arbitrary
8400 strings. This can be useful for special purpose optimizations that want to look
8401 for these annotations. These have no other defined use; they are ignored by code
8402 generation and optimization.
8404 '``llvm.annotation.*``' Intrinsic
8405 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8410 This is an overloaded intrinsic. You can use '``llvm.annotation``' on
8411 any integer bit width.
8415 declare i8 @llvm.annotation.i8(i8 <val>, i8* <str>, i8* <str>, i32 <int>)
8416 declare i16 @llvm.annotation.i16(i16 <val>, i8* <str>, i8* <str>, i32 <int>)
8417 declare i32 @llvm.annotation.i32(i32 <val>, i8* <str>, i8* <str>, i32 <int>)
8418 declare i64 @llvm.annotation.i64(i64 <val>, i8* <str>, i8* <str>, i32 <int>)
8419 declare i256 @llvm.annotation.i256(i256 <val>, i8* <str>, i8* <str>, i32 <int>)
8424 The '``llvm.annotation``' intrinsic.
8429 The first argument is an integer value (result of some expression), the
8430 second is a pointer to a global string, the third is a pointer to a
8431 global string which is the source file name, and the last argument is
8432 the line number. It returns the value of the first argument.
8437 This intrinsic allows annotations to be put on arbitrary expressions
8438 with arbitrary strings. This can be useful for special purpose
8439 optimizations that want to look for these annotations. These have no
8440 other defined use; they are ignored by code generation and optimization.
8442 '``llvm.trap``' Intrinsic
8443 ^^^^^^^^^^^^^^^^^^^^^^^^^
8450 declare void @llvm.trap() noreturn nounwind
8455 The '``llvm.trap``' intrinsic.
8465 This intrinsic is lowered to the target dependent trap instruction. If
8466 the target does not have a trap instruction, this intrinsic will be
8467 lowered to a call of the ``abort()`` function.
8469 '``llvm.debugtrap``' Intrinsic
8470 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8477 declare void @llvm.debugtrap() nounwind
8482 The '``llvm.debugtrap``' intrinsic.
8492 This intrinsic is lowered to code which is intended to cause an
8493 execution trap with the intention of requesting the attention of a
8496 '``llvm.stackprotector``' Intrinsic
8497 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8504 declare void @llvm.stackprotector(i8* <guard>, i8** <slot>)
8509 The ``llvm.stackprotector`` intrinsic takes the ``guard`` and stores it
8510 onto the stack at ``slot``. The stack slot is adjusted to ensure that it
8511 is placed on the stack before local variables.
8516 The ``llvm.stackprotector`` intrinsic requires two pointer arguments.
8517 The first argument is the value loaded from the stack guard
8518 ``@__stack_chk_guard``. The second variable is an ``alloca`` that has
8519 enough space to hold the value of the guard.
8524 This intrinsic causes the prologue/epilogue inserter to force the
8525 position of the ``AllocaInst`` stack slot to be before local variables
8526 on the stack. This is to ensure that if a local variable on the stack is
8527 overwritten, it will destroy the value of the guard. When the function
8528 exits, the guard on the stack is checked against the original guard. If
8529 they are different, then the program aborts by calling the
8530 ``__stack_chk_fail()`` function.
8532 '``llvm.objectsize``' Intrinsic
8533 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8540 declare i32 @llvm.objectsize.i32(i8* <object>, i1 <min>)
8541 declare i64 @llvm.objectsize.i64(i8* <object>, i1 <min>)
8546 The ``llvm.objectsize`` intrinsic is designed to provide information to
8547 the optimizers to determine at compile time whether a) an operation
8548 (like memcpy) will overflow a buffer that corresponds to an object, or
8549 b) that a runtime check for overflow isn't necessary. An object in this
8550 context means an allocation of a specific class, structure, array, or
8556 The ``llvm.objectsize`` intrinsic takes two arguments. The first
8557 argument is a pointer to or into the ``object``. The second argument is
8558 a boolean and determines whether ``llvm.objectsize`` returns 0 (if true)
8559 or -1 (if false) when the object size is unknown. The second argument
8560 only accepts constants.
8565 The ``llvm.objectsize`` intrinsic is lowered to a constant representing
8566 the size of the object concerned. If the size cannot be determined at
8567 compile time, ``llvm.objectsize`` returns ``i32/i64 -1 or 0`` (depending
8568 on the ``min`` argument).
8570 '``llvm.expect``' Intrinsic
8571 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8578 declare i32 @llvm.expect.i32(i32 <val>, i32 <expected_val>)
8579 declare i64 @llvm.expect.i64(i64 <val>, i64 <expected_val>)
8584 The ``llvm.expect`` intrinsic provides information about expected (the
8585 most probable) value of ``val``, which can be used by optimizers.
8590 The ``llvm.expect`` intrinsic takes two arguments. The first argument is
8591 a value. The second argument is an expected value, this needs to be a
8592 constant value, variables are not allowed.
8597 This intrinsic is lowered to the ``val``.
8599 '``llvm.donothing``' Intrinsic
8600 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8607 declare void @llvm.donothing() nounwind readnone
8612 The ``llvm.donothing`` intrinsic doesn't perform any operation. It's the
8613 only intrinsic that can be called with an invoke instruction.
8623 This intrinsic does nothing, and it's removed by optimizers and ignored