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
390 .. _visibilitystyles:
395 All Global Variables and Functions have one of the following visibility
398 "``default``" - Default style
399 On targets that use the ELF object file format, default visibility
400 means that the declaration is visible to other modules and, in
401 shared libraries, means that the declared entity may be overridden.
402 On Darwin, default visibility means that the declaration is visible
403 to other modules. Default visibility corresponds to "external
404 linkage" in the language.
405 "``hidden``" - Hidden style
406 Two declarations of an object with hidden visibility refer to the
407 same object if they are in the same shared object. Usually, hidden
408 visibility indicates that the symbol will not be placed into the
409 dynamic symbol table, so no other module (executable or shared
410 library) can reference it directly.
411 "``protected``" - Protected style
412 On ELF, protected visibility indicates that the symbol will be
413 placed in the dynamic symbol table, but that references within the
414 defining module will bind to the local symbol. That is, the symbol
415 cannot be overridden by another module.
422 LLVM IR allows you to specify name aliases for certain types. This can
423 make it easier to read the IR and make the IR more condensed
424 (particularly when recursive types are involved). An example of a name
429 %mytype = type { %mytype*, i32 }
431 You may give a name to any :ref:`type <typesystem>` except
432 ":ref:`void <t_void>`". Type name aliases may be used anywhere a type is
433 expected with the syntax "%mytype".
435 Note that type names are aliases for the structural type that they
436 indicate, and that you can therefore specify multiple names for the same
437 type. This often leads to confusing behavior when dumping out a .ll
438 file. Since LLVM IR uses structural typing, the name is not part of the
439 type. When printing out LLVM IR, the printer will pick *one name* to
440 render all types of a particular shape. This means that if you have code
441 where two different source types end up having the same LLVM type, that
442 the dumper will sometimes print the "wrong" or unexpected type. This is
443 an important design point and isn't going to change.
450 Global variables define regions of memory allocated at compilation time
451 instead of run-time. Global variables may optionally be initialized, may
452 have an explicit section to be placed in, and may have an optional
453 explicit alignment specified.
455 A variable may be defined as ``thread_local``, which means that it will
456 not be shared by threads (each thread will have a separated copy of the
457 variable). Not all targets support thread-local variables. Optionally, a
458 TLS model may be specified:
461 For variables that are only used within the current shared library.
463 For variables in modules that will not be loaded dynamically.
465 For variables defined in the executable and only used within it.
467 The models correspond to the ELF TLS models; see `ELF Handling For
468 Thread-Local Storage <http://people.redhat.com/drepper/tls.pdf>`_ for
469 more information on under which circumstances the different models may
470 be used. The target may choose a different TLS model if the specified
471 model is not supported, or if a better choice of model can be made.
473 A variable may be defined as a global ``constant``, which indicates that
474 the contents of the variable will **never** be modified (enabling better
475 optimization, allowing the global data to be placed in the read-only
476 section of an executable, etc). Note that variables that need runtime
477 initialization cannot be marked ``constant`` as there is a store to the
480 LLVM explicitly allows *declarations* of global variables to be marked
481 constant, even if the final definition of the global is not. This
482 capability can be used to enable slightly better optimization of the
483 program, but requires the language definition to guarantee that
484 optimizations based on the 'constantness' are valid for the translation
485 units that do not include the definition.
487 As SSA values, global variables define pointer values that are in scope
488 (i.e. they dominate) all basic blocks in the program. Global variables
489 always define a pointer to their "content" type because they describe a
490 region of memory, and all memory objects in LLVM are accessed through
493 Global variables can be marked with ``unnamed_addr`` which indicates
494 that the address is not significant, only the content. Constants marked
495 like this can be merged with other constants if they have the same
496 initializer. Note that a constant with significant address *can* be
497 merged with a ``unnamed_addr`` constant, the result being a constant
498 whose address is significant.
500 A global variable may be declared to reside in a target-specific
501 numbered address space. For targets that support them, address spaces
502 may affect how optimizations are performed and/or what target
503 instructions are used to access the variable. The default address space
504 is zero. The address space qualifier must precede any other attributes.
506 LLVM allows an explicit section to be specified for globals. If the
507 target supports it, it will emit globals to the section specified.
509 By default, global initializers are optimized by assuming that global
510 variables defined within the module are not modified from their
511 initial values before the start of the global initializer. This is
512 true even for variables potentially accessible from outside the
513 module, including those with external linkage or appearing in
514 ``@llvm.used``. This assumption may be suppressed by marking the
515 variable with ``externally_initialized``.
517 An explicit alignment may be specified for a global, which must be a
518 power of 2. If not present, or if the alignment is set to zero, the
519 alignment of the global is set by the target to whatever it feels
520 convenient. If an explicit alignment is specified, the global is forced
521 to have exactly that alignment. Targets and optimizers are not allowed
522 to over-align the global if the global has an assigned section. In this
523 case, the extra alignment could be observable: for example, code could
524 assume that the globals are densely packed in their section and try to
525 iterate over them as an array, alignment padding would break this
528 For example, the following defines a global in a numbered address space
529 with an initializer, section, and alignment:
533 @G = addrspace(5) constant float 1.0, section "foo", align 4
535 The following example defines a thread-local global with the
536 ``initialexec`` TLS model:
540 @G = thread_local(initialexec) global i32 0, align 4
542 .. _functionstructure:
547 LLVM function definitions consist of the "``define``" keyword, an
548 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
549 style <visibility>`, an optional :ref:`calling convention <callingconv>`,
550 an optional ``unnamed_addr`` attribute, a return type, an optional
551 :ref:`parameter attribute <paramattrs>` for the return type, a function
552 name, a (possibly empty) argument list (each with optional :ref:`parameter
553 attributes <paramattrs>`), optional :ref:`function attributes <fnattrs>`,
554 an optional section, an optional alignment, an optional :ref:`garbage
555 collector name <gc>`, an opening curly brace, a list of basic blocks,
556 and a closing curly brace.
558 LLVM function declarations consist of the "``declare``" keyword, an
559 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
560 style <visibility>`, an optional :ref:`calling convention <callingconv>`,
561 an optional ``unnamed_addr`` attribute, a return type, an optional
562 :ref:`parameter attribute <paramattrs>` for the return type, a function
563 name, a possibly empty list of arguments, an optional alignment, and an
564 optional :ref:`garbage collector name <gc>`.
566 A function definition contains a list of basic blocks, forming the CFG
567 (Control Flow Graph) for the function. Each basic block may optionally
568 start with a label (giving the basic block a symbol table entry),
569 contains a list of instructions, and ends with a
570 :ref:`terminator <terminators>` instruction (such as a branch or function
571 return). If explicit label is not provided, a block is assigned an
572 implicit numbered label, using a next value from the same counter as used
573 for unnamed temporaries (:ref:`see above<identifiers>`). For example, if a
574 function entry block does not have explicit label, it will be assigned
575 label "%0", then first unnamed temporary in that block will be "%1", etc.
577 The first basic block in a function is special in two ways: it is
578 immediately executed on entrance to the function, and it is not allowed
579 to have predecessor basic blocks (i.e. there can not be any branches to
580 the entry block of a function). Because the block can have no
581 predecessors, it also cannot have any :ref:`PHI nodes <i_phi>`.
583 LLVM allows an explicit section to be specified for functions. If the
584 target supports it, it will emit functions to the section specified.
586 An explicit alignment may be specified for a function. If not present,
587 or if the alignment is set to zero, the alignment of the function is set
588 by the target to whatever it feels convenient. If an explicit alignment
589 is specified, the function is forced to have at least that much
590 alignment. All alignments must be a power of 2.
592 If the ``unnamed_addr`` attribute is given, the address is know to not
593 be significant and two identical functions can be merged.
597 define [linkage] [visibility]
599 <ResultType> @<FunctionName> ([argument list])
600 [fn Attrs] [section "name"] [align N]
608 Aliases act as "second name" for the aliasee value (which can be either
609 function, global variable, another alias or bitcast of global value).
610 Aliases may have an optional :ref:`linkage type <linkage>`, and an optional
611 :ref:`visibility style <visibility>`.
615 @<Name> = alias [Linkage] [Visibility] <AliaseeTy> @<Aliasee>
617 .. _namedmetadatastructure:
622 Named metadata is a collection of metadata. :ref:`Metadata
623 nodes <metadata>` (but not metadata strings) are the only valid
624 operands for a named metadata.
628 ; Some unnamed metadata nodes, which are referenced by the named metadata.
629 !0 = metadata !{metadata !"zero"}
630 !1 = metadata !{metadata !"one"}
631 !2 = metadata !{metadata !"two"}
633 !name = !{!0, !1, !2}
640 The return type and each parameter of a function type may have a set of
641 *parameter attributes* associated with them. Parameter attributes are
642 used to communicate additional information about the result or
643 parameters of a function. Parameter attributes are considered to be part
644 of the function, not of the function type, so functions with different
645 parameter attributes can have the same function type.
647 Parameter attributes are simple keywords that follow the type specified.
648 If multiple parameter attributes are needed, they are space separated.
653 declare i32 @printf(i8* noalias nocapture, ...)
654 declare i32 @atoi(i8 zeroext)
655 declare signext i8 @returns_signed_char()
657 Note that any attributes for the function result (``nounwind``,
658 ``readonly``) come immediately after the argument list.
660 Currently, only the following parameter attributes are defined:
663 This indicates to the code generator that the parameter or return
664 value should be zero-extended to the extent required by the target's
665 ABI (which is usually 32-bits, but is 8-bits for a i1 on x86-64) by
666 the caller (for a parameter) or the callee (for a return value).
668 This indicates to the code generator that the parameter or return
669 value should be sign-extended to the extent required by the target's
670 ABI (which is usually 32-bits) by the caller (for a parameter) or
671 the callee (for a return value).
673 This indicates that this parameter or return value should be treated
674 in a special target-dependent fashion during while emitting code for
675 a function call or return (usually, by putting it in a register as
676 opposed to memory, though some targets use it to distinguish between
677 two different kinds of registers). Use of this attribute is
680 This indicates that the pointer parameter should really be passed by
681 value to the function. The attribute implies that a hidden copy of
682 the pointee is made between the caller and the callee, so the callee
683 is unable to modify the value in the caller. This attribute is only
684 valid on LLVM pointer arguments. It is generally used to pass
685 structs and arrays by value, but is also valid on pointers to
686 scalars. The copy is considered to belong to the caller not the
687 callee (for example, ``readonly`` functions should not write to
688 ``byval`` parameters). This is not a valid attribute for return
691 The byval attribute also supports specifying an alignment with the
692 align attribute. It indicates the alignment of the stack slot to
693 form and the known alignment of the pointer specified to the call
694 site. If the alignment is not specified, then the code generator
695 makes a target-specific assumption.
698 This indicates that the pointer parameter specifies the address of a
699 structure that is the return value of the function in the source
700 program. This pointer must be guaranteed by the caller to be valid:
701 loads and stores to the structure may be assumed by the callee
702 not to trap and to be properly aligned. This may only be applied to
703 the first parameter. This is not a valid attribute for return
706 This indicates that pointer values :ref:`based <pointeraliasing>` on
707 the argument or return value do not alias pointer values which are
708 not *based* on it, ignoring certain "irrelevant" dependencies. For a
709 call to the parent function, dependencies between memory references
710 from before or after the call and from those during the call are
711 "irrelevant" to the ``noalias`` keyword for the arguments and return
712 value used in that call. The caller shares the responsibility with
713 the callee for ensuring that these requirements are met. For further
714 details, please see the discussion of the NoAlias response in `alias
715 analysis <AliasAnalysis.html#MustMayNo>`_.
717 Note that this definition of ``noalias`` is intentionally similar
718 to the definition of ``restrict`` in C99 for function arguments,
719 though it is slightly weaker.
721 For function return values, C99's ``restrict`` is not meaningful,
722 while LLVM's ``noalias`` is.
724 This indicates that the callee does not make any copies of the
725 pointer that outlive the callee itself. This is not a valid
726 attribute for return values.
731 This indicates that the pointer parameter can be excised using the
732 :ref:`trampoline intrinsics <int_trampoline>`. This is not a valid
733 attribute for return values and can only be applied to one parameter.
736 This indicates that the value of the function always returns the value
737 of the parameter as its return value. This is an optimization hint to
738 the code generator when generating the caller, allowing tail call
739 optimization and omission of register saves and restores in some cases;
740 it is not checked or enforced when generating the callee. The parameter
741 and the function return type must be valid operands for the
742 :ref:`bitcast instruction <i_bitcast>`. This is not a valid attribute for
743 return values and can only be applied to one parameter.
747 Garbage Collector Names
748 -----------------------
750 Each function may specify a garbage collector name, which is simply a
755 define void @f() gc "name" { ... }
757 The compiler declares the supported values of *name*. Specifying a
758 collector which will cause the compiler to alter its output in order to
759 support the named garbage collection algorithm.
766 Attribute groups are groups of attributes that are referenced by objects within
767 the IR. They are important for keeping ``.ll`` files readable, because a lot of
768 functions will use the same set of attributes. In the degenerative case of a
769 ``.ll`` file that corresponds to a single ``.c`` file, the single attribute
770 group will capture the important command line flags used to build that file.
772 An attribute group is a module-level object. To use an attribute group, an
773 object references the attribute group's ID (e.g. ``#37``). An object may refer
774 to more than one attribute group. In that situation, the attributes from the
775 different groups are merged.
777 Here is an example of attribute groups for a function that should always be
778 inlined, has a stack alignment of 4, and which shouldn't use SSE instructions:
782 ; Target-independent attributes:
783 attributes #0 = { alwaysinline alignstack=4 }
785 ; Target-dependent attributes:
786 attributes #1 = { "no-sse" }
788 ; Function @f has attributes: alwaysinline, alignstack=4, and "no-sse".
789 define void @f() #0 #1 { ... }
796 Function attributes are set to communicate additional information about
797 a function. Function attributes are considered to be part of the
798 function, not of the function type, so functions with different function
799 attributes can have the same function type.
801 Function attributes are simple keywords that follow the type specified.
802 If multiple attributes are needed, they are space separated. For
807 define void @f() noinline { ... }
808 define void @f() alwaysinline { ... }
809 define void @f() alwaysinline optsize { ... }
810 define void @f() optsize { ... }
813 This attribute indicates that, when emitting the prologue and
814 epilogue, the backend should forcibly align the stack pointer.
815 Specify the desired alignment, which must be a power of two, in
818 This attribute indicates that the inliner should attempt to inline
819 this function into callers whenever possible, ignoring any active
820 inlining size threshold for this caller.
822 This attribute indicates that this function is rarely called. When
823 computing edge weights, basic blocks post-dominated by a cold
824 function call are also considered to be cold; and, thus, given low
827 This attribute suppresses lazy symbol binding for the function. This
828 may make calls to the function faster, at the cost of extra program
829 startup time if the function is not called during program startup.
831 This attribute indicates that the source code contained a hint that
832 inlining this function is desirable (such as the "inline" keyword in
833 C/C++). It is just a hint; it imposes no requirements on the
836 This attribute disables prologue / epilogue emission for the
837 function. This can have very system-specific consequences.
839 This indicates that the callee function at a call site is not
840 recognized as a built-in function. LLVM will retain the original call
841 and not replace it with equivalent code based on the semantics of the
842 built-in function. This is only valid at call sites, not on function
843 declarations or definitions.
845 This attribute indicates that calls to the function cannot be
846 duplicated. A call to a ``noduplicate`` function may be moved
847 within its parent function, but may not be duplicated within
850 A function containing a ``noduplicate`` call may still
851 be an inlining candidate, provided that the call is not
852 duplicated by inlining. That implies that the function has
853 internal linkage and only has one call site, so the original
854 call is dead after inlining.
856 This attributes disables implicit floating point instructions.
858 This attribute indicates that the inliner should never inline this
859 function in any situation. This attribute may not be used together
860 with the ``alwaysinline`` attribute.
862 This attribute indicates that the code generator should not use a
863 red zone, even if the target-specific ABI normally permits it.
865 This function attribute indicates that the function never returns
866 normally. This produces undefined behavior at runtime if the
867 function ever does dynamically return.
869 This function attribute indicates that the function never returns
870 with an unwind or exceptional control flow. If the function does
871 unwind, its runtime behavior is undefined.
873 This attribute suggests that optimization passes and code generator
874 passes make choices that keep the code size of this function low,
875 and otherwise do optimizations specifically to reduce code size.
877 This attribute indicates that the function computes its result (or
878 decides to unwind an exception) based strictly on its arguments,
879 without dereferencing any pointer arguments or otherwise accessing
880 any mutable state (e.g. memory, control registers, etc) visible to
881 caller functions. It does not write through any pointer arguments
882 (including ``byval`` arguments) and never changes any state visible
883 to callers. This means that it cannot unwind exceptions by calling
884 the ``C++`` exception throwing methods.
886 This attribute indicates that the function does not write through
887 any pointer arguments (including ``byval`` arguments) or otherwise
888 modify any state (e.g. memory, control registers, etc) visible to
889 caller functions. It may dereference pointer arguments and read
890 state that may be set in the caller. A readonly function always
891 returns the same value (or unwinds an exception identically) when
892 called with the same set of arguments and global state. It cannot
893 unwind an exception by calling the ``C++`` exception throwing
896 This attribute indicates that this function can return twice. The C
897 ``setjmp`` is an example of such a function. The compiler disables
898 some optimizations (like tail calls) in the caller of these
901 This attribute indicates that AddressSanitizer checks
902 (dynamic address safety analysis) are enabled for this function.
904 This attribute indicates that MemorySanitizer checks (dynamic detection
905 of accesses to uninitialized memory) are enabled for this function.
907 This attribute indicates that ThreadSanitizer checks
908 (dynamic thread safety analysis) are enabled for this function.
910 This attribute indicates that the function should emit a stack
911 smashing protector. It is in the form of a "canary" --- a random value
912 placed on the stack before the local variables that's checked upon
913 return from the function to see if it has been overwritten. A
914 heuristic is used to determine if a function needs stack protectors
915 or not. The heuristic used will enable protectors for functions with:
917 - Character arrays larger than ``ssp-buffer-size`` (default 8).
918 - Aggregates containing character arrays larger than ``ssp-buffer-size``.
919 - Calls to alloca() with variable sizes or constant sizes greater than
922 If a function that has an ``ssp`` attribute is inlined into a
923 function that doesn't have an ``ssp`` attribute, then the resulting
924 function will have an ``ssp`` attribute.
926 This attribute indicates that the function should *always* emit a
927 stack smashing protector. This overrides the ``ssp`` function
930 If a function that has an ``sspreq`` attribute is inlined into a
931 function that doesn't have an ``sspreq`` attribute or which has an
932 ``ssp`` or ``sspstrong`` attribute, then the resulting function will have
933 an ``sspreq`` attribute.
935 This attribute indicates that the function should emit a stack smashing
936 protector. This attribute causes a strong heuristic to be used when
937 determining if a function needs stack protectors. The strong heuristic
938 will enable protectors for functions with:
940 - Arrays of any size and type
941 - Aggregates containing an array of any size and type.
943 - Local variables that have had their address taken.
945 This overrides the ``ssp`` function attribute.
947 If a function that has an ``sspstrong`` attribute is inlined into a
948 function that doesn't have an ``sspstrong`` attribute, then the
949 resulting function will have an ``sspstrong`` attribute.
951 This attribute indicates that the ABI being targeted requires that
952 an unwind table entry be produce for this function even if we can
953 show that no exceptions passes by it. This is normally the case for
954 the ELF x86-64 abi, but it can be disabled for some compilation
959 Module-Level Inline Assembly
960 ----------------------------
962 Modules may contain "module-level inline asm" blocks, which corresponds
963 to the GCC "file scope inline asm" blocks. These blocks are internally
964 concatenated by LLVM and treated as a single unit, but may be separated
965 in the ``.ll`` file if desired. The syntax is very simple:
969 module asm "inline asm code goes here"
970 module asm "more can go here"
972 The strings can contain any character by escaping non-printable
973 characters. The escape sequence used is simply "\\xx" where "xx" is the
974 two digit hex code for the number.
976 The inline asm code is simply printed to the machine code .s file when
977 assembly code is generated.
979 .. _langref_datalayout:
984 A module may specify a target specific data layout string that specifies
985 how data is to be laid out in memory. The syntax for the data layout is
990 target datalayout = "layout specification"
992 The *layout specification* consists of a list of specifications
993 separated by the minus sign character ('-'). Each specification starts
994 with a letter and may include other information after the letter to
995 define some aspect of the data layout. The specifications accepted are
999 Specifies that the target lays out data in big-endian form. That is,
1000 the bits with the most significance have the lowest address
1003 Specifies that the target lays out data in little-endian form. That
1004 is, the bits with the least significance have the lowest address
1007 Specifies the natural alignment of the stack in bits. Alignment
1008 promotion of stack variables is limited to the natural stack
1009 alignment to avoid dynamic stack realignment. The stack alignment
1010 must be a multiple of 8-bits. If omitted, the natural stack
1011 alignment defaults to "unspecified", which does not prevent any
1012 alignment promotions.
1013 ``p[n]:<size>:<abi>:<pref>``
1014 This specifies the *size* of a pointer and its ``<abi>`` and
1015 ``<pref>``\erred alignments for address space ``n``. All sizes are in
1016 bits. Specifying the ``<pref>`` alignment is optional. If omitted, the
1017 preceding ``:`` should be omitted too. The address space, ``n`` is
1018 optional, and if not specified, denotes the default address space 0.
1019 The value of ``n`` must be in the range [1,2^23).
1020 ``i<size>:<abi>:<pref>``
1021 This specifies the alignment for an integer type of a given bit
1022 ``<size>``. The value of ``<size>`` must be in the range [1,2^23).
1023 ``v<size>:<abi>:<pref>``
1024 This specifies the alignment for a vector type of a given bit
1026 ``f<size>:<abi>:<pref>``
1027 This specifies the alignment for a floating point type of a given bit
1028 ``<size>``. Only values of ``<size>`` that are supported by the target
1029 will work. 32 (float) and 64 (double) are supported on all targets; 80
1030 or 128 (different flavors of long double) are also supported on some
1032 ``a<size>:<abi>:<pref>``
1033 This specifies the alignment for an aggregate type of a given bit
1035 ``s<size>:<abi>:<pref>``
1036 This specifies the alignment for a stack object of a given bit
1038 ``n<size1>:<size2>:<size3>...``
1039 This specifies a set of native integer widths for the target CPU in
1040 bits. For example, it might contain ``n32`` for 32-bit PowerPC,
1041 ``n32:64`` for PowerPC 64, or ``n8:16:32:64`` for X86-64. Elements of
1042 this set are considered to support most general arithmetic operations
1045 When constructing the data layout for a given target, LLVM starts with a
1046 default set of specifications which are then (possibly) overridden by
1047 the specifications in the ``datalayout`` keyword. The default
1048 specifications are given in this list:
1050 - ``E`` - big endian
1051 - ``p:64:64:64`` - 64-bit pointers with 64-bit alignment
1052 - ``S0`` - natural stack alignment is unspecified
1053 - ``i1:8:8`` - i1 is 8-bit (byte) aligned
1054 - ``i8:8:8`` - i8 is 8-bit (byte) aligned
1055 - ``i16:16:16`` - i16 is 16-bit aligned
1056 - ``i32:32:32`` - i32 is 32-bit aligned
1057 - ``i64:32:64`` - i64 has ABI alignment of 32-bits but preferred
1058 alignment of 64-bits
1059 - ``f16:16:16`` - half is 16-bit aligned
1060 - ``f32:32:32`` - float is 32-bit aligned
1061 - ``f64:64:64`` - double is 64-bit aligned
1062 - ``f128:128:128`` - quad is 128-bit aligned
1063 - ``v64:64:64`` - 64-bit vector is 64-bit aligned
1064 - ``v128:128:128`` - 128-bit vector is 128-bit aligned
1065 - ``a0:0:64`` - aggregates are 64-bit aligned
1067 When LLVM is determining the alignment for a given type, it uses the
1070 #. If the type sought is an exact match for one of the specifications,
1071 that specification is used.
1072 #. If no match is found, and the type sought is an integer type, then
1073 the smallest integer type that is larger than the bitwidth of the
1074 sought type is used. If none of the specifications are larger than
1075 the bitwidth then the largest integer type is used. For example,
1076 given the default specifications above, the i7 type will use the
1077 alignment of i8 (next largest) while both i65 and i256 will use the
1078 alignment of i64 (largest specified).
1079 #. If no match is found, and the type sought is a vector type, then the
1080 largest vector type that is smaller than the sought vector type will
1081 be used as a fall back. This happens because <128 x double> can be
1082 implemented in terms of 64 <2 x double>, for example.
1084 The function of the data layout string may not be what you expect.
1085 Notably, this is not a specification from the frontend of what alignment
1086 the code generator should use.
1088 Instead, if specified, the target data layout is required to match what
1089 the ultimate *code generator* expects. This string is used by the
1090 mid-level optimizers to improve code, and this only works if it matches
1091 what the ultimate code generator uses. If you would like to generate IR
1092 that does not embed this target-specific detail into the IR, then you
1093 don't have to specify the string. This will disable some optimizations
1094 that require precise layout information, but this also prevents those
1095 optimizations from introducing target specificity into the IR.
1097 .. _pointeraliasing:
1099 Pointer Aliasing Rules
1100 ----------------------
1102 Any memory access must be done through a pointer value associated with
1103 an address range of the memory access, otherwise the behavior is
1104 undefined. Pointer values are associated with address ranges according
1105 to the following rules:
1107 - A pointer value is associated with the addresses associated with any
1108 value it is *based* on.
1109 - An address of a global variable is associated with the address range
1110 of the variable's storage.
1111 - The result value of an allocation instruction is associated with the
1112 address range of the allocated storage.
1113 - A null pointer in the default address-space is associated with no
1115 - An integer constant other than zero or a pointer value returned from
1116 a function not defined within LLVM may be associated with address
1117 ranges allocated through mechanisms other than those provided by
1118 LLVM. Such ranges shall not overlap with any ranges of addresses
1119 allocated by mechanisms provided by LLVM.
1121 A pointer value is *based* on another pointer value according to the
1124 - A pointer value formed from a ``getelementptr`` operation is *based*
1125 on the first operand of the ``getelementptr``.
1126 - The result value of a ``bitcast`` is *based* on the operand of the
1128 - A pointer value formed by an ``inttoptr`` is *based* on all pointer
1129 values that contribute (directly or indirectly) to the computation of
1130 the pointer's value.
1131 - The "*based* on" relationship is transitive.
1133 Note that this definition of *"based"* is intentionally similar to the
1134 definition of *"based"* in C99, though it is slightly weaker.
1136 LLVM IR does not associate types with memory. The result type of a
1137 ``load`` merely indicates the size and alignment of the memory from
1138 which to load, as well as the interpretation of the value. The first
1139 operand type of a ``store`` similarly only indicates the size and
1140 alignment of the store.
1142 Consequently, type-based alias analysis, aka TBAA, aka
1143 ``-fstrict-aliasing``, is not applicable to general unadorned LLVM IR.
1144 :ref:`Metadata <metadata>` may be used to encode additional information
1145 which specialized optimization passes may use to implement type-based
1150 Volatile Memory Accesses
1151 ------------------------
1153 Certain memory accesses, such as :ref:`load <i_load>`'s,
1154 :ref:`store <i_store>`'s, and :ref:`llvm.memcpy <int_memcpy>`'s may be
1155 marked ``volatile``. The optimizers must not change the number of
1156 volatile operations or change their order of execution relative to other
1157 volatile operations. The optimizers *may* change the order of volatile
1158 operations relative to non-volatile operations. This is not Java's
1159 "volatile" and has no cross-thread synchronization behavior.
1161 IR-level volatile loads and stores cannot safely be optimized into
1162 llvm.memcpy or llvm.memmove intrinsics even when those intrinsics are
1163 flagged volatile. Likewise, the backend should never split or merge
1164 target-legal volatile load/store instructions.
1166 .. admonition:: Rationale
1168 Platforms may rely on volatile loads and stores of natively supported
1169 data width to be executed as single instruction. For example, in C
1170 this holds for an l-value of volatile primitive type with native
1171 hardware support, but not necessarily for aggregate types. The
1172 frontend upholds these expectations, which are intentionally
1173 unspecified in the IR. The rules above ensure that IR transformation
1174 do not violate the frontend's contract with the language.
1178 Memory Model for Concurrent Operations
1179 --------------------------------------
1181 The LLVM IR does not define any way to start parallel threads of
1182 execution or to register signal handlers. Nonetheless, there are
1183 platform-specific ways to create them, and we define LLVM IR's behavior
1184 in their presence. This model is inspired by the C++0x memory model.
1186 For a more informal introduction to this model, see the :doc:`Atomics`.
1188 We define a *happens-before* partial order as the least partial order
1191 - Is a superset of single-thread program order, and
1192 - When a *synchronizes-with* ``b``, includes an edge from ``a`` to
1193 ``b``. *Synchronizes-with* pairs are introduced by platform-specific
1194 techniques, like pthread locks, thread creation, thread joining,
1195 etc., and by atomic instructions. (See also :ref:`Atomic Memory Ordering
1196 Constraints <ordering>`).
1198 Note that program order does not introduce *happens-before* edges
1199 between a thread and signals executing inside that thread.
1201 Every (defined) read operation (load instructions, memcpy, atomic
1202 loads/read-modify-writes, etc.) R reads a series of bytes written by
1203 (defined) write operations (store instructions, atomic
1204 stores/read-modify-writes, memcpy, etc.). For the purposes of this
1205 section, initialized globals are considered to have a write of the
1206 initializer which is atomic and happens before any other read or write
1207 of the memory in question. For each byte of a read R, R\ :sub:`byte`
1208 may see any write to the same byte, except:
1210 - If write\ :sub:`1` happens before write\ :sub:`2`, and
1211 write\ :sub:`2` happens before R\ :sub:`byte`, then
1212 R\ :sub:`byte` does not see write\ :sub:`1`.
1213 - If R\ :sub:`byte` happens before write\ :sub:`3`, then
1214 R\ :sub:`byte` does not see write\ :sub:`3`.
1216 Given that definition, R\ :sub:`byte` is defined as follows:
1218 - If R is volatile, the result is target-dependent. (Volatile is
1219 supposed to give guarantees which can support ``sig_atomic_t`` in
1220 C/C++, and may be used for accesses to addresses which do not behave
1221 like normal memory. It does not generally provide cross-thread
1223 - Otherwise, if there is no write to the same byte that happens before
1224 R\ :sub:`byte`, R\ :sub:`byte` returns ``undef`` for that byte.
1225 - Otherwise, if R\ :sub:`byte` may see exactly one write,
1226 R\ :sub:`byte` returns the value written by that write.
1227 - Otherwise, if R is atomic, and all the writes R\ :sub:`byte` may
1228 see are atomic, it chooses one of the values written. See the :ref:`Atomic
1229 Memory Ordering Constraints <ordering>` section for additional
1230 constraints on how the choice is made.
1231 - Otherwise R\ :sub:`byte` returns ``undef``.
1233 R returns the value composed of the series of bytes it read. This
1234 implies that some bytes within the value may be ``undef`` **without**
1235 the entire value being ``undef``. Note that this only defines the
1236 semantics of the operation; it doesn't mean that targets will emit more
1237 than one instruction to read the series of bytes.
1239 Note that in cases where none of the atomic intrinsics are used, this
1240 model places only one restriction on IR transformations on top of what
1241 is required for single-threaded execution: introducing a store to a byte
1242 which might not otherwise be stored is not allowed in general.
1243 (Specifically, in the case where another thread might write to and read
1244 from an address, introducing a store can change a load that may see
1245 exactly one write into a load that may see multiple writes.)
1249 Atomic Memory Ordering Constraints
1250 ----------------------------------
1252 Atomic instructions (:ref:`cmpxchg <i_cmpxchg>`,
1253 :ref:`atomicrmw <i_atomicrmw>`, :ref:`fence <i_fence>`,
1254 :ref:`atomic load <i_load>`, and :ref:`atomic store <i_store>`) take
1255 an ordering parameter that determines which other atomic instructions on
1256 the same address they *synchronize with*. These semantics are borrowed
1257 from Java and C++0x, but are somewhat more colloquial. If these
1258 descriptions aren't precise enough, check those specs (see spec
1259 references in the :doc:`atomics guide <Atomics>`).
1260 :ref:`fence <i_fence>` instructions treat these orderings somewhat
1261 differently since they don't take an address. See that instruction's
1262 documentation for details.
1264 For a simpler introduction to the ordering constraints, see the
1268 The set of values that can be read is governed by the happens-before
1269 partial order. A value cannot be read unless some operation wrote
1270 it. This is intended to provide a guarantee strong enough to model
1271 Java's non-volatile shared variables. This ordering cannot be
1272 specified for read-modify-write operations; it is not strong enough
1273 to make them atomic in any interesting way.
1275 In addition to the guarantees of ``unordered``, there is a single
1276 total order for modifications by ``monotonic`` operations on each
1277 address. All modification orders must be compatible with the
1278 happens-before order. There is no guarantee that the modification
1279 orders can be combined to a global total order for the whole program
1280 (and this often will not be possible). The read in an atomic
1281 read-modify-write operation (:ref:`cmpxchg <i_cmpxchg>` and
1282 :ref:`atomicrmw <i_atomicrmw>`) reads the value in the modification
1283 order immediately before the value it writes. If one atomic read
1284 happens before another atomic read of the same address, the later
1285 read must see the same value or a later value in the address's
1286 modification order. This disallows reordering of ``monotonic`` (or
1287 stronger) operations on the same address. If an address is written
1288 ``monotonic``-ally by one thread, and other threads ``monotonic``-ally
1289 read that address repeatedly, the other threads must eventually see
1290 the write. This corresponds to the C++0x/C1x
1291 ``memory_order_relaxed``.
1293 In addition to the guarantees of ``monotonic``, a
1294 *synchronizes-with* edge may be formed with a ``release`` operation.
1295 This is intended to model C++'s ``memory_order_acquire``.
1297 In addition to the guarantees of ``monotonic``, if this operation
1298 writes a value which is subsequently read by an ``acquire``
1299 operation, it *synchronizes-with* that operation. (This isn't a
1300 complete description; see the C++0x definition of a release
1301 sequence.) This corresponds to the C++0x/C1x
1302 ``memory_order_release``.
1303 ``acq_rel`` (acquire+release)
1304 Acts as both an ``acquire`` and ``release`` operation on its
1305 address. This corresponds to the C++0x/C1x ``memory_order_acq_rel``.
1306 ``seq_cst`` (sequentially consistent)
1307 In addition to the guarantees of ``acq_rel`` (``acquire`` for an
1308 operation which only reads, ``release`` for an operation which only
1309 writes), there is a global total order on all
1310 sequentially-consistent operations on all addresses, which is
1311 consistent with the *happens-before* partial order and with the
1312 modification orders of all the affected addresses. Each
1313 sequentially-consistent read sees the last preceding write to the
1314 same address in this global order. This corresponds to the C++0x/C1x
1315 ``memory_order_seq_cst`` and Java volatile.
1319 If an atomic operation is marked ``singlethread``, it only *synchronizes
1320 with* or participates in modification and seq\_cst total orderings with
1321 other operations running in the same thread (for example, in signal
1329 LLVM IR floating-point binary ops (:ref:`fadd <i_fadd>`,
1330 :ref:`fsub <i_fsub>`, :ref:`fmul <i_fmul>`, :ref:`fdiv <i_fdiv>`,
1331 :ref:`frem <i_frem>`) have the following flags that can set to enable
1332 otherwise unsafe floating point operations
1335 No NaNs - Allow optimizations to assume the arguments and result are not
1336 NaN. Such optimizations are required to retain defined behavior over
1337 NaNs, but the value of the result is undefined.
1340 No Infs - Allow optimizations to assume the arguments and result are not
1341 +/-Inf. Such optimizations are required to retain defined behavior over
1342 +/-Inf, but the value of the result is undefined.
1345 No Signed Zeros - Allow optimizations to treat the sign of a zero
1346 argument or result as insignificant.
1349 Allow Reciprocal - Allow optimizations to use the reciprocal of an
1350 argument rather than perform division.
1353 Fast - Allow algebraically equivalent transformations that may
1354 dramatically change results in floating point (e.g. reassociate). This
1355 flag implies all the others.
1362 The LLVM type system is one of the most important features of the
1363 intermediate representation. Being typed enables a number of
1364 optimizations to be performed on the intermediate representation
1365 directly, without having to do extra analyses on the side before the
1366 transformation. A strong type system makes it easier to read the
1367 generated code and enables novel analyses and transformations that are
1368 not feasible to perform on normal three address code representations.
1370 Type Classifications
1371 --------------------
1373 The types fall into a few useful classifications:
1382 * - :ref:`integer <t_integer>`
1383 - ``i1``, ``i2``, ``i3``, ... ``i8``, ... ``i16``, ... ``i32``, ...
1386 * - :ref:`floating point <t_floating>`
1387 - ``half``, ``float``, ``double``, ``x86_fp80``, ``fp128``,
1395 - :ref:`integer <t_integer>`, :ref:`floating point <t_floating>`,
1396 :ref:`pointer <t_pointer>`, :ref:`vector <t_vector>`,
1397 :ref:`structure <t_struct>`, :ref:`array <t_array>`,
1398 :ref:`label <t_label>`, :ref:`metadata <t_metadata>`.
1400 * - :ref:`primitive <t_primitive>`
1401 - :ref:`label <t_label>`,
1402 :ref:`void <t_void>`,
1403 :ref:`integer <t_integer>`,
1404 :ref:`floating point <t_floating>`,
1405 :ref:`x86mmx <t_x86mmx>`,
1406 :ref:`metadata <t_metadata>`.
1408 * - :ref:`derived <t_derived>`
1409 - :ref:`array <t_array>`,
1410 :ref:`function <t_function>`,
1411 :ref:`pointer <t_pointer>`,
1412 :ref:`structure <t_struct>`,
1413 :ref:`vector <t_vector>`,
1414 :ref:`opaque <t_opaque>`.
1416 The :ref:`first class <t_firstclass>` types are perhaps the most important.
1417 Values of these types are the only ones which can be produced by
1425 The primitive types are the fundamental building blocks of the LLVM
1436 The integer type is a very simple type that simply specifies an
1437 arbitrary bit width for the integer type desired. Any bit width from 1
1438 bit to 2\ :sup:`23`\ -1 (about 8 million) can be specified.
1447 The number of bits the integer will occupy is specified by the ``N``
1453 +----------------+------------------------------------------------+
1454 | ``i1`` | a single-bit integer. |
1455 +----------------+------------------------------------------------+
1456 | ``i32`` | a 32-bit integer. |
1457 +----------------+------------------------------------------------+
1458 | ``i1942652`` | a really big integer of over 1 million bits. |
1459 +----------------+------------------------------------------------+
1463 Floating Point Types
1464 ^^^^^^^^^^^^^^^^^^^^
1473 - 16-bit floating point value
1476 - 32-bit floating point value
1479 - 64-bit floating point value
1482 - 128-bit floating point value (112-bit mantissa)
1485 - 80-bit floating point value (X87)
1488 - 128-bit floating point value (two 64-bits)
1498 The x86mmx type represents a value held in an MMX register on an x86
1499 machine. The operations allowed on it are quite limited: parameters and
1500 return values, load and store, and bitcast. User-specified MMX
1501 instructions are represented as intrinsic or asm calls with arguments
1502 and/or results of this type. There are no arrays, vectors or constants
1520 The void type does not represent any value and has no size.
1537 The label type represents code labels.
1554 The metadata type represents embedded metadata. No derived types may be
1555 created from metadata except for :ref:`function <t_function>` arguments.
1569 The real power in LLVM comes from the derived types in the system. This
1570 is what allows a programmer to represent arrays, functions, pointers,
1571 and other useful types. Each of these types contain one or more element
1572 types which may be a primitive type, or another derived type. For
1573 example, it is possible to have a two dimensional array, using an array
1574 as the element type of another array.
1581 Aggregate Types are a subset of derived types that can contain multiple
1582 member types. :ref:`Arrays <t_array>` and :ref:`structs <t_struct>` are
1583 aggregate types. :ref:`Vectors <t_vector>` are not considered to be
1594 The array type is a very simple derived type that arranges elements
1595 sequentially in memory. The array type requires a size (number of
1596 elements) and an underlying data type.
1603 [<# elements> x <elementtype>]
1605 The number of elements is a constant integer value; ``elementtype`` may
1606 be any type with a size.
1611 +------------------+--------------------------------------+
1612 | ``[40 x i32]`` | Array of 40 32-bit integer values. |
1613 +------------------+--------------------------------------+
1614 | ``[41 x i32]`` | Array of 41 32-bit integer values. |
1615 +------------------+--------------------------------------+
1616 | ``[4 x i8]`` | Array of 4 8-bit integer values. |
1617 +------------------+--------------------------------------+
1619 Here are some examples of multidimensional arrays:
1621 +-----------------------------+----------------------------------------------------------+
1622 | ``[3 x [4 x i32]]`` | 3x4 array of 32-bit integer values. |
1623 +-----------------------------+----------------------------------------------------------+
1624 | ``[12 x [10 x float]]`` | 12x10 array of single precision floating point values. |
1625 +-----------------------------+----------------------------------------------------------+
1626 | ``[2 x [3 x [4 x i16]]]`` | 2x3x4 array of 16-bit integer values. |
1627 +-----------------------------+----------------------------------------------------------+
1629 There is no restriction on indexing beyond the end of the array implied
1630 by a static type (though there are restrictions on indexing beyond the
1631 bounds of an allocated object in some cases). This means that
1632 single-dimension 'variable sized array' addressing can be implemented in
1633 LLVM with a zero length array type. An implementation of 'pascal style
1634 arrays' in LLVM could use the type "``{ i32, [0 x float]}``", for
1645 The function type can be thought of as a function signature. It consists
1646 of a return type and a list of formal parameter types. The return type
1647 of a function type is a first class type or a void type.
1654 <returntype> (<parameter list>)
1656 ...where '``<parameter list>``' is a comma-separated list of type
1657 specifiers. Optionally, the parameter list may include a type ``...``,
1658 which indicates that the function takes a variable number of arguments.
1659 Variable argument functions can access their arguments with the
1660 :ref:`variable argument handling intrinsic <int_varargs>` functions.
1661 '``<returntype>``' is any type except :ref:`label <t_label>`.
1666 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1667 | ``i32 (i32)`` | function taking an ``i32``, returning an ``i32`` |
1668 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1669 | ``float (i16, i32 *) *`` | :ref:`Pointer <t_pointer>` to a function that takes an ``i16`` and a :ref:`pointer <t_pointer>` to ``i32``, returning ``float``. |
1670 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1671 | ``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. |
1672 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1673 | ``{i32, i32} (i32)`` | A function taking an ``i32``, returning a :ref:`structure <t_struct>` containing two ``i32`` values |
1674 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1684 The structure type is used to represent a collection of data members
1685 together in memory. The elements of a structure may be any type that has
1688 Structures in memory are accessed using '``load``' and '``store``' by
1689 getting a pointer to a field with the '``getelementptr``' instruction.
1690 Structures in registers are accessed using the '``extractvalue``' and
1691 '``insertvalue``' instructions.
1693 Structures may optionally be "packed" structures, which indicate that
1694 the alignment of the struct is one byte, and that there is no padding
1695 between the elements. In non-packed structs, padding between field types
1696 is inserted as defined by the DataLayout string in the module, which is
1697 required to match what the underlying code generator expects.
1699 Structures can either be "literal" or "identified". A literal structure
1700 is defined inline with other types (e.g. ``{i32, i32}*``) whereas
1701 identified types are always defined at the top level with a name.
1702 Literal types are uniqued by their contents and can never be recursive
1703 or opaque since there is no way to write one. Identified types can be
1704 recursive, can be opaqued, and are never uniqued.
1711 %T1 = type { <type list> } ; Identified normal struct type
1712 %T2 = type <{ <type list> }> ; Identified packed struct type
1717 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1718 | ``{ i32, i32, i32 }`` | A triple of three ``i32`` values |
1719 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1720 | ``{ 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``. |
1721 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1722 | ``<{ i8, i32 }>`` | A packed struct known to be 5 bytes in size. |
1723 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1727 Opaque Structure Types
1728 ^^^^^^^^^^^^^^^^^^^^^^
1733 Opaque structure types are used to represent named structure types that
1734 do not have a body specified. This corresponds (for example) to the C
1735 notion of a forward declared structure.
1748 +--------------+-------------------+
1749 | ``opaque`` | An opaque type. |
1750 +--------------+-------------------+
1760 The pointer type is used to specify memory locations. Pointers are
1761 commonly used to reference objects in memory.
1763 Pointer types may have an optional address space attribute defining the
1764 numbered address space where the pointed-to object resides. The default
1765 address space is number zero. The semantics of non-zero address spaces
1766 are target-specific.
1768 Note that LLVM does not permit pointers to void (``void*``) nor does it
1769 permit pointers to labels (``label*``). Use ``i8*`` instead.
1781 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1782 | ``[4 x i32]*`` | A :ref:`pointer <t_pointer>` to :ref:`array <t_array>` of four ``i32`` values. |
1783 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1784 | ``i32 (i32*) *`` | A :ref:`pointer <t_pointer>` to a :ref:`function <t_function>` that takes an ``i32*``, returning an ``i32``. |
1785 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1786 | ``i32 addrspace(5)*`` | A :ref:`pointer <t_pointer>` to an ``i32`` value that resides in address space #5. |
1787 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1797 A vector type is a simple derived type that represents a vector of
1798 elements. Vector types are used when multiple primitive data are
1799 operated in parallel using a single instruction (SIMD). A vector type
1800 requires a size (number of elements) and an underlying primitive data
1801 type. Vector types are considered :ref:`first class <t_firstclass>`.
1808 < <# elements> x <elementtype> >
1810 The number of elements is a constant integer value larger than 0;
1811 elementtype may be any integer or floating point type, or a pointer to
1812 these types. Vectors of size zero are not allowed.
1817 +-------------------+--------------------------------------------------+
1818 | ``<4 x i32>`` | Vector of 4 32-bit integer values. |
1819 +-------------------+--------------------------------------------------+
1820 | ``<8 x float>`` | Vector of 8 32-bit floating-point values. |
1821 +-------------------+--------------------------------------------------+
1822 | ``<2 x i64>`` | Vector of 2 64-bit integer values. |
1823 +-------------------+--------------------------------------------------+
1824 | ``<4 x i64*>`` | Vector of 4 pointers to 64-bit integer values. |
1825 +-------------------+--------------------------------------------------+
1830 LLVM has several different basic types of constants. This section
1831 describes them all and their syntax.
1836 **Boolean constants**
1837 The two strings '``true``' and '``false``' are both valid constants
1839 **Integer constants**
1840 Standard integers (such as '4') are constants of the
1841 :ref:`integer <t_integer>` type. Negative numbers may be used with
1843 **Floating point constants**
1844 Floating point constants use standard decimal notation (e.g.
1845 123.421), exponential notation (e.g. 1.23421e+2), or a more precise
1846 hexadecimal notation (see below). The assembler requires the exact
1847 decimal value of a floating-point constant. For example, the
1848 assembler accepts 1.25 but rejects 1.3 because 1.3 is a repeating
1849 decimal in binary. Floating point constants must have a :ref:`floating
1850 point <t_floating>` type.
1851 **Null pointer constants**
1852 The identifier '``null``' is recognized as a null pointer constant
1853 and must be of :ref:`pointer type <t_pointer>`.
1855 The one non-intuitive notation for constants is the hexadecimal form of
1856 floating point constants. For example, the form
1857 '``double 0x432ff973cafa8000``' is equivalent to (but harder to read
1858 than) '``double 4.5e+15``'. The only time hexadecimal floating point
1859 constants are required (and the only time that they are generated by the
1860 disassembler) is when a floating point constant must be emitted but it
1861 cannot be represented as a decimal floating point number in a reasonable
1862 number of digits. For example, NaN's, infinities, and other special
1863 values are represented in their IEEE hexadecimal format so that assembly
1864 and disassembly do not cause any bits to change in the constants.
1866 When using the hexadecimal form, constants of types half, float, and
1867 double are represented using the 16-digit form shown above (which
1868 matches the IEEE754 representation for double); half and float values
1869 must, however, be exactly representable as IEEE 754 half and single
1870 precision, respectively. Hexadecimal format is always used for long
1871 double, and there are three forms of long double. The 80-bit format used
1872 by x86 is represented as ``0xK`` followed by 20 hexadecimal digits. The
1873 128-bit format used by PowerPC (two adjacent doubles) is represented by
1874 ``0xM`` followed by 32 hexadecimal digits. The IEEE 128-bit format is
1875 represented by ``0xL`` followed by 32 hexadecimal digits. Long doubles
1876 will only work if they match the long double format on your target.
1877 The IEEE 16-bit format (half precision) is represented by ``0xH``
1878 followed by 4 hexadecimal digits. All hexadecimal formats are big-endian
1879 (sign bit at the left).
1881 There are no constants of type x86mmx.
1886 Complex constants are a (potentially recursive) combination of simple
1887 constants and smaller complex constants.
1889 **Structure constants**
1890 Structure constants are represented with notation similar to
1891 structure type definitions (a comma separated list of elements,
1892 surrounded by braces (``{}``)). For example:
1893 "``{ i32 4, float 17.0, i32* @G }``", where "``@G``" is declared as
1894 "``@G = external global i32``". Structure constants must have
1895 :ref:`structure type <t_struct>`, and the number and types of elements
1896 must match those specified by the type.
1898 Array constants are represented with notation similar to array type
1899 definitions (a comma separated list of elements, surrounded by
1900 square brackets (``[]``)). For example:
1901 "``[ i32 42, i32 11, i32 74 ]``". Array constants must have
1902 :ref:`array type <t_array>`, and the number and types of elements must
1903 match those specified by the type.
1904 **Vector constants**
1905 Vector constants are represented with notation similar to vector
1906 type definitions (a comma separated list of elements, surrounded by
1907 less-than/greater-than's (``<>``)). For example:
1908 "``< i32 42, i32 11, i32 74, i32 100 >``". Vector constants
1909 must have :ref:`vector type <t_vector>`, and the number and types of
1910 elements must match those specified by the type.
1911 **Zero initialization**
1912 The string '``zeroinitializer``' can be used to zero initialize a
1913 value to zero of *any* type, including scalar and
1914 :ref:`aggregate <t_aggregate>` types. This is often used to avoid
1915 having to print large zero initializers (e.g. for large arrays) and
1916 is always exactly equivalent to using explicit zero initializers.
1918 A metadata node is a structure-like constant with :ref:`metadata
1919 type <t_metadata>`. For example:
1920 "``metadata !{ i32 0, metadata !"test" }``". Unlike other
1921 constants that are meant to be interpreted as part of the
1922 instruction stream, metadata is a place to attach additional
1923 information such as debug info.
1925 Global Variable and Function Addresses
1926 --------------------------------------
1928 The addresses of :ref:`global variables <globalvars>` and
1929 :ref:`functions <functionstructure>` are always implicitly valid
1930 (link-time) constants. These constants are explicitly referenced when
1931 the :ref:`identifier for the global <identifiers>` is used and always have
1932 :ref:`pointer <t_pointer>` type. For example, the following is a legal LLVM
1935 .. code-block:: llvm
1939 @Z = global [2 x i32*] [ i32* @X, i32* @Y ]
1946 The string '``undef``' can be used anywhere a constant is expected, and
1947 indicates that the user of the value may receive an unspecified
1948 bit-pattern. Undefined values may be of any type (other than '``label``'
1949 or '``void``') and be used anywhere a constant is permitted.
1951 Undefined values are useful because they indicate to the compiler that
1952 the program is well defined no matter what value is used. This gives the
1953 compiler more freedom to optimize. Here are some examples of
1954 (potentially surprising) transformations that are valid (in pseudo IR):
1956 .. code-block:: llvm
1966 This is safe because all of the output bits are affected by the undef
1967 bits. Any output bit can have a zero or one depending on the input bits.
1969 .. code-block:: llvm
1980 These logical operations have bits that are not always affected by the
1981 input. For example, if ``%X`` has a zero bit, then the output of the
1982 '``and``' operation will always be a zero for that bit, no matter what
1983 the corresponding bit from the '``undef``' is. As such, it is unsafe to
1984 optimize or assume that the result of the '``and``' is '``undef``'.
1985 However, it is safe to assume that all bits of the '``undef``' could be
1986 0, and optimize the '``and``' to 0. Likewise, it is safe to assume that
1987 all the bits of the '``undef``' operand to the '``or``' could be set,
1988 allowing the '``or``' to be folded to -1.
1990 .. code-block:: llvm
1992 %A = select undef, %X, %Y
1993 %B = select undef, 42, %Y
1994 %C = select %X, %Y, undef
2004 This set of examples shows that undefined '``select``' (and conditional
2005 branch) conditions can go *either way*, but they have to come from one
2006 of the two operands. In the ``%A`` example, if ``%X`` and ``%Y`` were
2007 both known to have a clear low bit, then ``%A`` would have to have a
2008 cleared low bit. However, in the ``%C`` example, the optimizer is
2009 allowed to assume that the '``undef``' operand could be the same as
2010 ``%Y``, allowing the whole '``select``' to be eliminated.
2012 .. code-block:: llvm
2014 %A = xor undef, undef
2031 This example points out that two '``undef``' operands are not
2032 necessarily the same. This can be surprising to people (and also matches
2033 C semantics) where they assume that "``X^X``" is always zero, even if
2034 ``X`` is undefined. This isn't true for a number of reasons, but the
2035 short answer is that an '``undef``' "variable" can arbitrarily change
2036 its value over its "live range". This is true because the variable
2037 doesn't actually *have a live range*. Instead, the value is logically
2038 read from arbitrary registers that happen to be around when needed, so
2039 the value is not necessarily consistent over time. In fact, ``%A`` and
2040 ``%C`` need to have the same semantics or the core LLVM "replace all
2041 uses with" concept would not hold.
2043 .. code-block:: llvm
2051 These examples show the crucial difference between an *undefined value*
2052 and *undefined behavior*. An undefined value (like '``undef``') is
2053 allowed to have an arbitrary bit-pattern. This means that the ``%A``
2054 operation can be constant folded to '``undef``', because the '``undef``'
2055 could be an SNaN, and ``fdiv`` is not (currently) defined on SNaN's.
2056 However, in the second example, we can make a more aggressive
2057 assumption: because the ``undef`` is allowed to be an arbitrary value,
2058 we are allowed to assume that it could be zero. Since a divide by zero
2059 has *undefined behavior*, we are allowed to assume that the operation
2060 does not execute at all. This allows us to delete the divide and all
2061 code after it. Because the undefined operation "can't happen", the
2062 optimizer can assume that it occurs in dead code.
2064 .. code-block:: llvm
2066 a: store undef -> %X
2067 b: store %X -> undef
2072 These examples reiterate the ``fdiv`` example: a store *of* an undefined
2073 value can be assumed to not have any effect; we can assume that the
2074 value is overwritten with bits that happen to match what was already
2075 there. However, a store *to* an undefined location could clobber
2076 arbitrary memory, therefore, it has undefined behavior.
2083 Poison values are similar to :ref:`undef values <undefvalues>`, however
2084 they also represent the fact that an instruction or constant expression
2085 which cannot evoke side effects has nevertheless detected a condition
2086 which results in undefined behavior.
2088 There is currently no way of representing a poison value in the IR; they
2089 only exist when produced by operations such as :ref:`add <i_add>` with
2092 Poison value behavior is defined in terms of value *dependence*:
2094 - Values other than :ref:`phi <i_phi>` nodes depend on their operands.
2095 - :ref:`Phi <i_phi>` nodes depend on the operand corresponding to
2096 their dynamic predecessor basic block.
2097 - Function arguments depend on the corresponding actual argument values
2098 in the dynamic callers of their functions.
2099 - :ref:`Call <i_call>` instructions depend on the :ref:`ret <i_ret>`
2100 instructions that dynamically transfer control back to them.
2101 - :ref:`Invoke <i_invoke>` instructions depend on the
2102 :ref:`ret <i_ret>`, :ref:`resume <i_resume>`, or exception-throwing
2103 call instructions that dynamically transfer control back to them.
2104 - Non-volatile loads and stores depend on the most recent stores to all
2105 of the referenced memory addresses, following the order in the IR
2106 (including loads and stores implied by intrinsics such as
2107 :ref:`@llvm.memcpy <int_memcpy>`.)
2108 - An instruction with externally visible side effects depends on the
2109 most recent preceding instruction with externally visible side
2110 effects, following the order in the IR. (This includes :ref:`volatile
2111 operations <volatile>`.)
2112 - An instruction *control-depends* on a :ref:`terminator
2113 instruction <terminators>` if the terminator instruction has
2114 multiple successors and the instruction is always executed when
2115 control transfers to one of the successors, and may not be executed
2116 when control is transferred to another.
2117 - Additionally, an instruction also *control-depends* on a terminator
2118 instruction if the set of instructions it otherwise depends on would
2119 be different if the terminator had transferred control to a different
2121 - Dependence is transitive.
2123 Poison Values have the same behavior as :ref:`undef values <undefvalues>`,
2124 with the additional affect that any instruction which has a *dependence*
2125 on a poison value has undefined behavior.
2127 Here are some examples:
2129 .. code-block:: llvm
2132 %poison = sub nuw i32 0, 1 ; Results in a poison value.
2133 %still_poison = and i32 %poison, 0 ; 0, but also poison.
2134 %poison_yet_again = getelementptr i32* @h, i32 %still_poison
2135 store i32 0, i32* %poison_yet_again ; memory at @h[0] is poisoned
2137 store i32 %poison, i32* @g ; Poison value stored to memory.
2138 %poison2 = load i32* @g ; Poison value loaded back from memory.
2140 store volatile i32 %poison, i32* @g ; External observation; undefined behavior.
2142 %narrowaddr = bitcast i32* @g to i16*
2143 %wideaddr = bitcast i32* @g to i64*
2144 %poison3 = load i16* %narrowaddr ; Returns a poison value.
2145 %poison4 = load i64* %wideaddr ; Returns a poison value.
2147 %cmp = icmp slt i32 %poison, 0 ; Returns a poison value.
2148 br i1 %cmp, label %true, label %end ; Branch to either destination.
2151 store volatile i32 0, i32* @g ; This is control-dependent on %cmp, so
2152 ; it has undefined behavior.
2156 %p = phi i32 [ 0, %entry ], [ 1, %true ]
2157 ; Both edges into this PHI are
2158 ; control-dependent on %cmp, so this
2159 ; always results in a poison value.
2161 store volatile i32 0, i32* @g ; This would depend on the store in %true
2162 ; if %cmp is true, or the store in %entry
2163 ; otherwise, so this is undefined behavior.
2165 br i1 %cmp, label %second_true, label %second_end
2166 ; The same branch again, but this time the
2167 ; true block doesn't have side effects.
2174 store volatile i32 0, i32* @g ; This time, the instruction always depends
2175 ; on the store in %end. Also, it is
2176 ; control-equivalent to %end, so this is
2177 ; well-defined (ignoring earlier undefined
2178 ; behavior in this example).
2182 Addresses of Basic Blocks
2183 -------------------------
2185 ``blockaddress(@function, %block)``
2187 The '``blockaddress``' constant computes the address of the specified
2188 basic block in the specified function, and always has an ``i8*`` type.
2189 Taking the address of the entry block is illegal.
2191 This value only has defined behavior when used as an operand to the
2192 ':ref:`indirectbr <i_indirectbr>`' instruction, or for comparisons
2193 against null. Pointer equality tests between labels addresses results in
2194 undefined behavior --- though, again, comparison against null is ok, and
2195 no label is equal to the null pointer. This may be passed around as an
2196 opaque pointer sized value as long as the bits are not inspected. This
2197 allows ``ptrtoint`` and arithmetic to be performed on these values so
2198 long as the original value is reconstituted before the ``indirectbr``
2201 Finally, some targets may provide defined semantics when using the value
2202 as the operand to an inline assembly, but that is target specific.
2204 Constant Expressions
2205 --------------------
2207 Constant expressions are used to allow expressions involving other
2208 constants to be used as constants. Constant expressions may be of any
2209 :ref:`first class <t_firstclass>` type and may involve any LLVM operation
2210 that does not have side effects (e.g. load and call are not supported).
2211 The following is the syntax for constant expressions:
2213 ``trunc (CST to TYPE)``
2214 Truncate a constant to another type. The bit size of CST must be
2215 larger than the bit size of TYPE. Both types must be integers.
2216 ``zext (CST to TYPE)``
2217 Zero extend a constant to another type. The bit size of CST must be
2218 smaller than the bit size of TYPE. Both types must be integers.
2219 ``sext (CST to TYPE)``
2220 Sign extend a constant to another type. The bit size of CST must be
2221 smaller than the bit size of TYPE. Both types must be integers.
2222 ``fptrunc (CST to TYPE)``
2223 Truncate a floating point constant to another floating point type.
2224 The size of CST must be larger than the size of TYPE. Both types
2225 must be floating point.
2226 ``fpext (CST to TYPE)``
2227 Floating point extend a constant to another type. The size of CST
2228 must be smaller or equal to the size of TYPE. Both types must be
2230 ``fptoui (CST to TYPE)``
2231 Convert a floating point constant to the corresponding unsigned
2232 integer constant. TYPE must be a scalar or vector integer type. CST
2233 must be of scalar or vector floating point type. Both CST and TYPE
2234 must be scalars, or vectors of the same number of elements. If the
2235 value won't fit in the integer type, the results are undefined.
2236 ``fptosi (CST to TYPE)``
2237 Convert a floating point constant to the corresponding signed
2238 integer constant. TYPE must be a scalar or vector integer type. CST
2239 must be of scalar or vector floating point type. Both CST and TYPE
2240 must be scalars, or vectors of the same number of elements. If the
2241 value won't fit in the integer type, the results are undefined.
2242 ``uitofp (CST to TYPE)``
2243 Convert an unsigned integer constant to the corresponding floating
2244 point constant. TYPE must be a scalar or vector floating point type.
2245 CST must be of scalar or vector integer type. Both CST and TYPE must
2246 be scalars, or vectors of the same number of elements. If the value
2247 won't fit in the floating point type, the results are undefined.
2248 ``sitofp (CST to TYPE)``
2249 Convert a signed integer constant to the corresponding floating
2250 point constant. TYPE must be a scalar or vector floating point type.
2251 CST must be of scalar or vector integer type. Both CST and TYPE must
2252 be scalars, or vectors of the same number of elements. If the value
2253 won't fit in the floating point type, the results are undefined.
2254 ``ptrtoint (CST to TYPE)``
2255 Convert a pointer typed constant to the corresponding integer
2256 constant. ``TYPE`` must be an integer type. ``CST`` must be of
2257 pointer type. The ``CST`` value is zero extended, truncated, or
2258 unchanged to make it fit in ``TYPE``.
2259 ``inttoptr (CST to TYPE)``
2260 Convert an integer constant to a pointer constant. TYPE must be a
2261 pointer type. CST must be of integer type. The CST value is zero
2262 extended, truncated, or unchanged to make it fit in a pointer size.
2263 This one is *really* dangerous!
2264 ``bitcast (CST to TYPE)``
2265 Convert a constant, CST, to another TYPE. The constraints of the
2266 operands are the same as those for the :ref:`bitcast
2267 instruction <i_bitcast>`.
2268 ``getelementptr (CSTPTR, IDX0, IDX1, ...)``, ``getelementptr inbounds (CSTPTR, IDX0, IDX1, ...)``
2269 Perform the :ref:`getelementptr operation <i_getelementptr>` on
2270 constants. As with the :ref:`getelementptr <i_getelementptr>`
2271 instruction, the index list may have zero or more indexes, which are
2272 required to make sense for the type of "CSTPTR".
2273 ``select (COND, VAL1, VAL2)``
2274 Perform the :ref:`select operation <i_select>` on constants.
2275 ``icmp COND (VAL1, VAL2)``
2276 Performs the :ref:`icmp operation <i_icmp>` on constants.
2277 ``fcmp COND (VAL1, VAL2)``
2278 Performs the :ref:`fcmp operation <i_fcmp>` on constants.
2279 ``extractelement (VAL, IDX)``
2280 Perform the :ref:`extractelement operation <i_extractelement>` on
2282 ``insertelement (VAL, ELT, IDX)``
2283 Perform the :ref:`insertelement operation <i_insertelement>` on
2285 ``shufflevector (VEC1, VEC2, IDXMASK)``
2286 Perform the :ref:`shufflevector operation <i_shufflevector>` on
2288 ``extractvalue (VAL, IDX0, IDX1, ...)``
2289 Perform the :ref:`extractvalue operation <i_extractvalue>` on
2290 constants. The index list is interpreted in a similar manner as
2291 indices in a ':ref:`getelementptr <i_getelementptr>`' operation. At
2292 least one index value must be specified.
2293 ``insertvalue (VAL, ELT, IDX0, IDX1, ...)``
2294 Perform the :ref:`insertvalue operation <i_insertvalue>` on constants.
2295 The index list is interpreted in a similar manner as indices in a
2296 ':ref:`getelementptr <i_getelementptr>`' operation. At least one index
2297 value must be specified.
2298 ``OPCODE (LHS, RHS)``
2299 Perform the specified operation of the LHS and RHS constants. OPCODE
2300 may be any of the :ref:`binary <binaryops>` or :ref:`bitwise
2301 binary <bitwiseops>` operations. The constraints on operands are
2302 the same as those for the corresponding instruction (e.g. no bitwise
2303 operations on floating point values are allowed).
2308 Inline Assembler Expressions
2309 ----------------------------
2311 LLVM supports inline assembler expressions (as opposed to :ref:`Module-Level
2312 Inline Assembly <moduleasm>`) through the use of a special value. This
2313 value represents the inline assembler as a string (containing the
2314 instructions to emit), a list of operand constraints (stored as a
2315 string), a flag that indicates whether or not the inline asm expression
2316 has side effects, and a flag indicating whether the function containing
2317 the asm needs to align its stack conservatively. An example inline
2318 assembler expression is:
2320 .. code-block:: llvm
2322 i32 (i32) asm "bswap $0", "=r,r"
2324 Inline assembler expressions may **only** be used as the callee operand
2325 of a :ref:`call <i_call>` or an :ref:`invoke <i_invoke>` instruction.
2326 Thus, typically we have:
2328 .. code-block:: llvm
2330 %X = call i32 asm "bswap $0", "=r,r"(i32 %Y)
2332 Inline asms with side effects not visible in the constraint list must be
2333 marked as having side effects. This is done through the use of the
2334 '``sideeffect``' keyword, like so:
2336 .. code-block:: llvm
2338 call void asm sideeffect "eieio", ""()
2340 In some cases inline asms will contain code that will not work unless
2341 the stack is aligned in some way, such as calls or SSE instructions on
2342 x86, yet will not contain code that does that alignment within the asm.
2343 The compiler should make conservative assumptions about what the asm
2344 might contain and should generate its usual stack alignment code in the
2345 prologue if the '``alignstack``' keyword is present:
2347 .. code-block:: llvm
2349 call void asm alignstack "eieio", ""()
2351 Inline asms also support using non-standard assembly dialects. The
2352 assumed dialect is ATT. When the '``inteldialect``' keyword is present,
2353 the inline asm is using the Intel dialect. Currently, ATT and Intel are
2354 the only supported dialects. An example is:
2356 .. code-block:: llvm
2358 call void asm inteldialect "eieio", ""()
2360 If multiple keywords appear the '``sideeffect``' keyword must come
2361 first, the '``alignstack``' keyword second and the '``inteldialect``'
2367 The call instructions that wrap inline asm nodes may have a
2368 "``!srcloc``" MDNode attached to it that contains a list of constant
2369 integers. If present, the code generator will use the integer as the
2370 location cookie value when report errors through the ``LLVMContext``
2371 error reporting mechanisms. This allows a front-end to correlate backend
2372 errors that occur with inline asm back to the source code that produced
2375 .. code-block:: llvm
2377 call void asm sideeffect "something bad", ""(), !srcloc !42
2379 !42 = !{ i32 1234567 }
2381 It is up to the front-end to make sense of the magic numbers it places
2382 in the IR. If the MDNode contains multiple constants, the code generator
2383 will use the one that corresponds to the line of the asm that the error
2388 Metadata Nodes and Metadata Strings
2389 -----------------------------------
2391 LLVM IR allows metadata to be attached to instructions in the program
2392 that can convey extra information about the code to the optimizers and
2393 code generator. One example application of metadata is source-level
2394 debug information. There are two metadata primitives: strings and nodes.
2395 All metadata has the ``metadata`` type and is identified in syntax by a
2396 preceding exclamation point ('``!``').
2398 A metadata string is a string surrounded by double quotes. It can
2399 contain any character by escaping non-printable characters with
2400 "``\xx``" where "``xx``" is the two digit hex code. For example:
2403 Metadata nodes are represented with notation similar to structure
2404 constants (a comma separated list of elements, surrounded by braces and
2405 preceded by an exclamation point). Metadata nodes can have any values as
2406 their operand. For example:
2408 .. code-block:: llvm
2410 !{ metadata !"test\00", i32 10}
2412 A :ref:`named metadata <namedmetadatastructure>` is a collection of
2413 metadata nodes, which can be looked up in the module symbol table. For
2416 .. code-block:: llvm
2418 !foo = metadata !{!4, !3}
2420 Metadata can be used as function arguments. Here ``llvm.dbg.value``
2421 function is using two metadata arguments:
2423 .. code-block:: llvm
2425 call void @llvm.dbg.value(metadata !24, i64 0, metadata !25)
2427 Metadata can be attached with an instruction. Here metadata ``!21`` is
2428 attached to the ``add`` instruction using the ``!dbg`` identifier:
2430 .. code-block:: llvm
2432 %indvar.next = add i64 %indvar, 1, !dbg !21
2434 More information about specific metadata nodes recognized by the
2435 optimizers and code generator is found below.
2440 In LLVM IR, memory does not have types, so LLVM's own type system is not
2441 suitable for doing TBAA. Instead, metadata is added to the IR to
2442 describe a type system of a higher level language. This can be used to
2443 implement typical C/C++ TBAA, but it can also be used to implement
2444 custom alias analysis behavior for other languages.
2446 The current metadata format is very simple. TBAA metadata nodes have up
2447 to three fields, e.g.:
2449 .. code-block:: llvm
2451 !0 = metadata !{ metadata !"an example type tree" }
2452 !1 = metadata !{ metadata !"int", metadata !0 }
2453 !2 = metadata !{ metadata !"float", metadata !0 }
2454 !3 = metadata !{ metadata !"const float", metadata !2, i64 1 }
2456 The first field is an identity field. It can be any value, usually a
2457 metadata string, which uniquely identifies the type. The most important
2458 name in the tree is the name of the root node. Two trees with different
2459 root node names are entirely disjoint, even if they have leaves with
2462 The second field identifies the type's parent node in the tree, or is
2463 null or omitted for a root node. A type is considered to alias all of
2464 its descendants and all of its ancestors in the tree. Also, a type is
2465 considered to alias all types in other trees, so that bitcode produced
2466 from multiple front-ends is handled conservatively.
2468 If the third field is present, it's an integer which if equal to 1
2469 indicates that the type is "constant" (meaning
2470 ``pointsToConstantMemory`` should return true; see `other useful
2471 AliasAnalysis methods <AliasAnalysis.html#OtherItfs>`_).
2473 '``tbaa.struct``' Metadata
2474 ^^^^^^^^^^^^^^^^^^^^^^^^^^
2476 The :ref:`llvm.memcpy <int_memcpy>` is often used to implement
2477 aggregate assignment operations in C and similar languages, however it
2478 is defined to copy a contiguous region of memory, which is more than
2479 strictly necessary for aggregate types which contain holes due to
2480 padding. Also, it doesn't contain any TBAA information about the fields
2483 ``!tbaa.struct`` metadata can describe which memory subregions in a
2484 memcpy are padding and what the TBAA tags of the struct are.
2486 The current metadata format is very simple. ``!tbaa.struct`` metadata
2487 nodes are a list of operands which are in conceptual groups of three.
2488 For each group of three, the first operand gives the byte offset of a
2489 field in bytes, the second gives its size in bytes, and the third gives
2492 .. code-block:: llvm
2494 !4 = metadata !{ i64 0, i64 4, metadata !1, i64 8, i64 4, metadata !2 }
2496 This describes a struct with two fields. The first is at offset 0 bytes
2497 with size 4 bytes, and has tbaa tag !1. The second is at offset 8 bytes
2498 and has size 4 bytes and has tbaa tag !2.
2500 Note that the fields need not be contiguous. In this example, there is a
2501 4 byte gap between the two fields. This gap represents padding which
2502 does not carry useful data and need not be preserved.
2504 '``fpmath``' Metadata
2505 ^^^^^^^^^^^^^^^^^^^^^
2507 ``fpmath`` metadata may be attached to any instruction of floating point
2508 type. It can be used to express the maximum acceptable error in the
2509 result of that instruction, in ULPs, thus potentially allowing the
2510 compiler to use a more efficient but less accurate method of computing
2511 it. ULP is defined as follows:
2513 If ``x`` is a real number that lies between two finite consecutive
2514 floating-point numbers ``a`` and ``b``, without being equal to one
2515 of them, then ``ulp(x) = |b - a|``, otherwise ``ulp(x)`` is the
2516 distance between the two non-equal finite floating-point numbers
2517 nearest ``x``. Moreover, ``ulp(NaN)`` is ``NaN``.
2519 The metadata node shall consist of a single positive floating point
2520 number representing the maximum relative error, for example:
2522 .. code-block:: llvm
2524 !0 = metadata !{ float 2.5 } ; maximum acceptable inaccuracy is 2.5 ULPs
2526 '``range``' Metadata
2527 ^^^^^^^^^^^^^^^^^^^^
2529 ``range`` metadata may be attached only to loads of integer types. It
2530 expresses the possible ranges the loaded value is in. The ranges are
2531 represented with a flattened list of integers. The loaded value is known
2532 to be in the union of the ranges defined by each consecutive pair. Each
2533 pair has the following properties:
2535 - The type must match the type loaded by the instruction.
2536 - The pair ``a,b`` represents the range ``[a,b)``.
2537 - Both ``a`` and ``b`` are constants.
2538 - The range is allowed to wrap.
2539 - The range should not represent the full or empty set. That is,
2542 In addition, the pairs must be in signed order of the lower bound and
2543 they must be non-contiguous.
2547 .. code-block:: llvm
2549 %a = load i8* %x, align 1, !range !0 ; Can only be 0 or 1
2550 %b = load i8* %y, align 1, !range !1 ; Can only be 255 (-1), 0 or 1
2551 %c = load i8* %z, align 1, !range !2 ; Can only be 0, 1, 3, 4 or 5
2552 %d = load i8* %z, align 1, !range !3 ; Can only be -2, -1, 3, 4 or 5
2554 !0 = metadata !{ i8 0, i8 2 }
2555 !1 = metadata !{ i8 255, i8 2 }
2556 !2 = metadata !{ i8 0, i8 2, i8 3, i8 6 }
2557 !3 = metadata !{ i8 -2, i8 0, i8 3, i8 6 }
2562 It is sometimes useful to attach information to loop constructs. Currently,
2563 loop metadata is implemented as metadata attached to the branch instruction
2564 in the loop latch block. This type of metadata refer to a metadata node that is
2565 guaranteed to be separate for each loop. The loop identifier metadata is
2566 specified with the name ``llvm.loop``.
2568 The loop identifier metadata is implemented using a metadata that refers to
2569 itself to avoid merging it with any other identifier metadata, e.g.,
2570 during module linkage or function inlining. That is, each loop should refer
2571 to their own identification metadata even if they reside in separate functions.
2572 The following example contains loop identifier metadata for two separate loop
2575 .. code-block:: llvm
2577 !0 = metadata !{ metadata !0 }
2578 !1 = metadata !{ metadata !1 }
2580 The loop identifier metadata can be used to specify additional per-loop
2581 metadata. Any operands after the first operand can be treated as user-defined
2582 metadata. For example the ``llvm.vectorizer.unroll`` metadata is understood
2583 by the loop vectorizer to indicate how many times to unroll the loop:
2585 .. code-block:: llvm
2587 br i1 %exitcond, label %._crit_edge, label %.lr.ph, !llvm.loop !0
2589 !0 = metadata !{ metadata !0, metadata !1 }
2590 !1 = metadata !{ metadata !"llvm.vectorizer.unroll", i32 2 }
2595 Metadata types used to annotate memory accesses with information helpful
2596 for optimizations are prefixed with ``llvm.mem``.
2598 '``llvm.mem.parallel_loop_access``' Metadata
2599 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2601 For a loop to be parallel, in addition to using
2602 the ``llvm.loop`` metadata to mark the loop latch branch instruction,
2603 also all of the memory accessing instructions in the loop body need to be
2604 marked with the ``llvm.mem.parallel_loop_access`` metadata. If there
2605 is at least one memory accessing instruction not marked with the metadata,
2606 the loop must be considered a sequential loop. This causes parallel loops to be
2607 converted to sequential loops due to optimization passes that are unaware of
2608 the parallel semantics and that insert new memory instructions to the loop
2611 Example of a loop that is considered parallel due to its correct use of
2612 both ``llvm.loop`` and ``llvm.mem.parallel_loop_access``
2613 metadata types that refer to the same loop identifier metadata.
2615 .. code-block:: llvm
2619 %0 = load i32* %arrayidx, align 4, !llvm.mem.parallel_loop_access !0
2621 store i32 %0, i32* %arrayidx4, align 4, !llvm.mem.parallel_loop_access !0
2623 br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !0
2627 !0 = metadata !{ metadata !0 }
2629 It is also possible to have nested parallel loops. In that case the
2630 memory accesses refer to a list of loop identifier metadata nodes instead of
2631 the loop identifier metadata node directly:
2633 .. code-block:: llvm
2640 %0 = load i32* %arrayidx, align 4, !llvm.mem.parallel_loop_access !0
2642 store i32 %0, i32* %arrayidx4, align 4, !llvm.mem.parallel_loop_access !0
2644 br i1 %exitcond, label %inner.for.end, label %inner.for.body, !llvm.loop !1
2648 %0 = load i32* %arrayidx, align 4, !llvm.mem.parallel_loop_access !0
2650 store i32 %0, i32* %arrayidx4, align 4, !llvm.mem.parallel_loop_access !0
2652 br i1 %exitcond, label %outer.for.end, label %outer.for.body, !llvm.loop !2
2654 outer.for.end: ; preds = %for.body
2656 !0 = metadata !{ metadata !1, metadata !2 } ; a list of loop identifiers
2657 !1 = metadata !{ metadata !1 } ; an identifier for the inner loop
2658 !2 = metadata !{ metadata !2 } ; an identifier for the outer loop
2660 '``llvm.vectorizer``'
2661 ^^^^^^^^^^^^^^^^^^^^^
2663 Metadata prefixed with ``llvm.vectorizer`` is used to control per-loop
2664 vectorization parameters such as vectorization factor and unroll factor.
2666 ``llvm.vectorizer`` metadata should be used in conjunction with ``llvm.loop``
2667 loop identification metadata.
2669 '``llvm.vectorizer.unroll``' Metadata
2670 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2672 This metadata instructs the loop vectorizer to unroll the specified
2673 loop exactly ``N`` times.
2675 The first operand is the string ``llvm.vectorizer.unroll`` and the second
2676 operand is an integer specifying the unroll factor. For example:
2678 .. code-block:: llvm
2680 !0 = metadata !{ metadata !"llvm.vectorizer.unroll", i32 4 }
2682 Note that setting ``llvm.vectorizer.unroll`` to 1 disables unrolling of the
2685 If ``llvm.vectorizer.unroll`` is set to 0 then the amount of unrolling will be
2686 determined automatically.
2688 '``llvm.vectorizer.width``' Metadata
2689 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2691 This metadata sets the target width of the vectorizer to ``N``. Without
2692 this metadata, the vectorizer will choose a width automatically.
2693 Regardless of this metadata, the vectorizer will only vectorize loops if
2694 it believes it is valid to do so.
2696 The first operand is the string ``llvm.vectorizer.width`` and the second
2697 operand is an integer specifying the width. For example:
2699 .. code-block:: llvm
2701 !0 = metadata !{ metadata !"llvm.vectorizer.width", i32 4 }
2703 Note that setting ``llvm.vectorizer.width`` to 1 disables vectorization of the
2706 If ``llvm.vectorizer.width`` is set to 0 then the width will be determined
2709 Module Flags Metadata
2710 =====================
2712 Information about the module as a whole is difficult to convey to LLVM's
2713 subsystems. The LLVM IR isn't sufficient to transmit this information.
2714 The ``llvm.module.flags`` named metadata exists in order to facilitate
2715 this. These flags are in the form of key / value pairs --- much like a
2716 dictionary --- making it easy for any subsystem who cares about a flag to
2719 The ``llvm.module.flags`` metadata contains a list of metadata triplets.
2720 Each triplet has the following form:
2722 - The first element is a *behavior* flag, which specifies the behavior
2723 when two (or more) modules are merged together, and it encounters two
2724 (or more) metadata with the same ID. The supported behaviors are
2726 - The second element is a metadata string that is a unique ID for the
2727 metadata. Each module may only have one flag entry for each unique ID (not
2728 including entries with the **Require** behavior).
2729 - The third element is the value of the flag.
2731 When two (or more) modules are merged together, the resulting
2732 ``llvm.module.flags`` metadata is the union of the modules' flags. That is, for
2733 each unique metadata ID string, there will be exactly one entry in the merged
2734 modules ``llvm.module.flags`` metadata table, and the value for that entry will
2735 be determined by the merge behavior flag, as described below. The only exception
2736 is that entries with the *Require* behavior are always preserved.
2738 The following behaviors are supported:
2749 Emits an error if two values disagree, otherwise the resulting value
2750 is that of the operands.
2754 Emits a warning if two values disagree. The result value will be the
2755 operand for the flag from the first module being linked.
2759 Adds a requirement that another module flag be present and have a
2760 specified value after linking is performed. The value must be a
2761 metadata pair, where the first element of the pair is the ID of the
2762 module flag to be restricted, and the second element of the pair is
2763 the value the module flag should be restricted to. This behavior can
2764 be used to restrict the allowable results (via triggering of an
2765 error) of linking IDs with the **Override** behavior.
2769 Uses the specified value, regardless of the behavior or value of the
2770 other module. If both modules specify **Override**, but the values
2771 differ, an error will be emitted.
2775 Appends the two values, which are required to be metadata nodes.
2779 Appends the two values, which are required to be metadata
2780 nodes. However, duplicate entries in the second list are dropped
2781 during the append operation.
2783 It is an error for a particular unique flag ID to have multiple behaviors,
2784 except in the case of **Require** (which adds restrictions on another metadata
2785 value) or **Override**.
2787 An example of module flags:
2789 .. code-block:: llvm
2791 !0 = metadata !{ i32 1, metadata !"foo", i32 1 }
2792 !1 = metadata !{ i32 4, metadata !"bar", i32 37 }
2793 !2 = metadata !{ i32 2, metadata !"qux", i32 42 }
2794 !3 = metadata !{ i32 3, metadata !"qux",
2796 metadata !"foo", i32 1
2799 !llvm.module.flags = !{ !0, !1, !2, !3 }
2801 - Metadata ``!0`` has the ID ``!"foo"`` and the value '1'. The behavior
2802 if two or more ``!"foo"`` flags are seen is to emit an error if their
2803 values are not equal.
2805 - Metadata ``!1`` has the ID ``!"bar"`` and the value '37'. The
2806 behavior if two or more ``!"bar"`` flags are seen is to use the value
2809 - Metadata ``!2`` has the ID ``!"qux"`` and the value '42'. The
2810 behavior if two or more ``!"qux"`` flags are seen is to emit a
2811 warning if their values are not equal.
2813 - Metadata ``!3`` has the ID ``!"qux"`` and the value:
2817 metadata !{ metadata !"foo", i32 1 }
2819 The behavior is to emit an error if the ``llvm.module.flags`` does not
2820 contain a flag with the ID ``!"foo"`` that has the value '1' after linking is
2823 Objective-C Garbage Collection Module Flags Metadata
2824 ----------------------------------------------------
2826 On the Mach-O platform, Objective-C stores metadata about garbage
2827 collection in a special section called "image info". The metadata
2828 consists of a version number and a bitmask specifying what types of
2829 garbage collection are supported (if any) by the file. If two or more
2830 modules are linked together their garbage collection metadata needs to
2831 be merged rather than appended together.
2833 The Objective-C garbage collection module flags metadata consists of the
2834 following key-value pairs:
2843 * - ``Objective-C Version``
2844 - **[Required]** --- The Objective-C ABI version. Valid values are 1 and 2.
2846 * - ``Objective-C Image Info Version``
2847 - **[Required]** --- The version of the image info section. Currently
2850 * - ``Objective-C Image Info Section``
2851 - **[Required]** --- The section to place the metadata. Valid values are
2852 ``"__OBJC, __image_info, regular"`` for Objective-C ABI version 1, and
2853 ``"__DATA,__objc_imageinfo, regular, no_dead_strip"`` for
2854 Objective-C ABI version 2.
2856 * - ``Objective-C Garbage Collection``
2857 - **[Required]** --- Specifies whether garbage collection is supported or
2858 not. Valid values are 0, for no garbage collection, and 2, for garbage
2859 collection supported.
2861 * - ``Objective-C GC Only``
2862 - **[Optional]** --- Specifies that only garbage collection is supported.
2863 If present, its value must be 6. This flag requires that the
2864 ``Objective-C Garbage Collection`` flag have the value 2.
2866 Some important flag interactions:
2868 - If a module with ``Objective-C Garbage Collection`` set to 0 is
2869 merged with a module with ``Objective-C Garbage Collection`` set to
2870 2, then the resulting module has the
2871 ``Objective-C Garbage Collection`` flag set to 0.
2872 - A module with ``Objective-C Garbage Collection`` set to 0 cannot be
2873 merged with a module with ``Objective-C GC Only`` set to 6.
2875 Automatic Linker Flags Module Flags Metadata
2876 --------------------------------------------
2878 Some targets support embedding flags to the linker inside individual object
2879 files. Typically this is used in conjunction with language extensions which
2880 allow source files to explicitly declare the libraries they depend on, and have
2881 these automatically be transmitted to the linker via object files.
2883 These flags are encoded in the IR using metadata in the module flags section,
2884 using the ``Linker Options`` key. The merge behavior for this flag is required
2885 to be ``AppendUnique``, and the value for the key is expected to be a metadata
2886 node which should be a list of other metadata nodes, each of which should be a
2887 list of metadata strings defining linker options.
2889 For example, the following metadata section specifies two separate sets of
2890 linker options, presumably to link against ``libz`` and the ``Cocoa``
2893 !0 = metadata !{ i32 6, metadata !"Linker Options",
2895 metadata !{ metadata !"-lz" },
2896 metadata !{ metadata !"-framework", metadata !"Cocoa" } } }
2897 !llvm.module.flags = !{ !0 }
2899 The metadata encoding as lists of lists of options, as opposed to a collapsed
2900 list of options, is chosen so that the IR encoding can use multiple option
2901 strings to specify e.g., a single library, while still having that specifier be
2902 preserved as an atomic element that can be recognized by a target specific
2903 assembly writer or object file emitter.
2905 Each individual option is required to be either a valid option for the target's
2906 linker, or an option that is reserved by the target specific assembly writer or
2907 object file emitter. No other aspect of these options is defined by the IR.
2909 Intrinsic Global Variables
2910 ==========================
2912 LLVM has a number of "magic" global variables that contain data that
2913 affect code generation or other IR semantics. These are documented here.
2914 All globals of this sort should have a section specified as
2915 "``llvm.metadata``". This section and all globals that start with
2916 "``llvm.``" are reserved for use by LLVM.
2918 The '``llvm.used``' Global Variable
2919 -----------------------------------
2921 The ``@llvm.used`` global is an array which has
2922 :ref:`appending linkage <linkage_appending>`. This array contains a list of
2923 pointers to global variables, functions and aliases which may optionally have a
2924 pointer cast formed of bitcast or getelementptr. For example, a legal
2927 .. code-block:: llvm
2932 @llvm.used = appending global [2 x i8*] [
2934 i8* bitcast (i32* @Y to i8*)
2935 ], section "llvm.metadata"
2937 If a symbol appears in the ``@llvm.used`` list, then the compiler, assembler,
2938 and linker are required to treat the symbol as if there is a reference to the
2939 symbol that it cannot see. For example, if a variable has internal linkage and
2940 no references other than that from the ``@llvm.used`` list, it cannot be
2941 deleted. This is commonly used to represent references from inline asms and
2942 other things the compiler cannot "see", and corresponds to
2943 "``attribute((used))``" in GNU C.
2945 On some targets, the code generator must emit a directive to the
2946 assembler or object file to prevent the assembler and linker from
2947 molesting the symbol.
2949 The '``llvm.compiler.used``' Global Variable
2950 --------------------------------------------
2952 The ``@llvm.compiler.used`` directive is the same as the ``@llvm.used``
2953 directive, except that it only prevents the compiler from touching the
2954 symbol. On targets that support it, this allows an intelligent linker to
2955 optimize references to the symbol without being impeded as it would be
2958 This is a rare construct that should only be used in rare circumstances,
2959 and should not be exposed to source languages.
2961 The '``llvm.global_ctors``' Global Variable
2962 -------------------------------------------
2964 .. code-block:: llvm
2966 %0 = type { i32, void ()* }
2967 @llvm.global_ctors = appending global [1 x %0] [%0 { i32 65535, void ()* @ctor }]
2969 The ``@llvm.global_ctors`` array contains a list of constructor
2970 functions and associated priorities. The functions referenced by this
2971 array will be called in ascending order of priority (i.e. lowest first)
2972 when the module is loaded. The order of functions with the same priority
2975 The '``llvm.global_dtors``' Global Variable
2976 -------------------------------------------
2978 .. code-block:: llvm
2980 %0 = type { i32, void ()* }
2981 @llvm.global_dtors = appending global [1 x %0] [%0 { i32 65535, void ()* @dtor }]
2983 The ``@llvm.global_dtors`` array contains a list of destructor functions
2984 and associated priorities. The functions referenced by this array will
2985 be called in descending order of priority (i.e. highest first) when the
2986 module is loaded. The order of functions with the same priority is not
2989 Instruction Reference
2990 =====================
2992 The LLVM instruction set consists of several different classifications
2993 of instructions: :ref:`terminator instructions <terminators>`, :ref:`binary
2994 instructions <binaryops>`, :ref:`bitwise binary
2995 instructions <bitwiseops>`, :ref:`memory instructions <memoryops>`, and
2996 :ref:`other instructions <otherops>`.
3000 Terminator Instructions
3001 -----------------------
3003 As mentioned :ref:`previously <functionstructure>`, every basic block in a
3004 program ends with a "Terminator" instruction, which indicates which
3005 block should be executed after the current block is finished. These
3006 terminator instructions typically yield a '``void``' value: they produce
3007 control flow, not values (the one exception being the
3008 ':ref:`invoke <i_invoke>`' instruction).
3010 The terminator instructions are: ':ref:`ret <i_ret>`',
3011 ':ref:`br <i_br>`', ':ref:`switch <i_switch>`',
3012 ':ref:`indirectbr <i_indirectbr>`', ':ref:`invoke <i_invoke>`',
3013 ':ref:`resume <i_resume>`', and ':ref:`unreachable <i_unreachable>`'.
3017 '``ret``' Instruction
3018 ^^^^^^^^^^^^^^^^^^^^^
3025 ret <type> <value> ; Return a value from a non-void function
3026 ret void ; Return from void function
3031 The '``ret``' instruction is used to return control flow (and optionally
3032 a value) from a function back to the caller.
3034 There are two forms of the '``ret``' instruction: one that returns a
3035 value and then causes control flow, and one that just causes control
3041 The '``ret``' instruction optionally accepts a single argument, the
3042 return value. The type of the return value must be a ':ref:`first
3043 class <t_firstclass>`' type.
3045 A function is not :ref:`well formed <wellformed>` if it it has a non-void
3046 return type and contains a '``ret``' instruction with no return value or
3047 a return value with a type that does not match its type, or if it has a
3048 void return type and contains a '``ret``' instruction with a return
3054 When the '``ret``' instruction is executed, control flow returns back to
3055 the calling function's context. If the caller is a
3056 ":ref:`call <i_call>`" instruction, execution continues at the
3057 instruction after the call. If the caller was an
3058 ":ref:`invoke <i_invoke>`" instruction, execution continues at the
3059 beginning of the "normal" destination block. If the instruction returns
3060 a value, that value shall set the call or invoke instruction's return
3066 .. code-block:: llvm
3068 ret i32 5 ; Return an integer value of 5
3069 ret void ; Return from a void function
3070 ret { i32, i8 } { i32 4, i8 2 } ; Return a struct of values 4 and 2
3074 '``br``' Instruction
3075 ^^^^^^^^^^^^^^^^^^^^
3082 br i1 <cond>, label <iftrue>, label <iffalse>
3083 br label <dest> ; Unconditional branch
3088 The '``br``' instruction is used to cause control flow to transfer to a
3089 different basic block in the current function. There are two forms of
3090 this instruction, corresponding to a conditional branch and an
3091 unconditional branch.
3096 The conditional branch form of the '``br``' instruction takes a single
3097 '``i1``' value and two '``label``' values. The unconditional form of the
3098 '``br``' instruction takes a single '``label``' value as a target.
3103 Upon execution of a conditional '``br``' instruction, the '``i1``'
3104 argument is evaluated. If the value is ``true``, control flows to the
3105 '``iftrue``' ``label`` argument. If "cond" is ``false``, control flows
3106 to the '``iffalse``' ``label`` argument.
3111 .. code-block:: llvm
3114 %cond = icmp eq i32 %a, %b
3115 br i1 %cond, label %IfEqual, label %IfUnequal
3123 '``switch``' Instruction
3124 ^^^^^^^^^^^^^^^^^^^^^^^^
3131 switch <intty> <value>, label <defaultdest> [ <intty> <val>, label <dest> ... ]
3136 The '``switch``' instruction is used to transfer control flow to one of
3137 several different places. It is a generalization of the '``br``'
3138 instruction, allowing a branch to occur to one of many possible
3144 The '``switch``' instruction uses three parameters: an integer
3145 comparison value '``value``', a default '``label``' destination, and an
3146 array of pairs of comparison value constants and '``label``'s. The table
3147 is not allowed to contain duplicate constant entries.
3152 The ``switch`` instruction specifies a table of values and destinations.
3153 When the '``switch``' instruction is executed, this table is searched
3154 for the given value. If the value is found, control flow is transferred
3155 to the corresponding destination; otherwise, control flow is transferred
3156 to the default destination.
3161 Depending on properties of the target machine and the particular
3162 ``switch`` instruction, this instruction may be code generated in
3163 different ways. For example, it could be generated as a series of
3164 chained conditional branches or with a lookup table.
3169 .. code-block:: llvm
3171 ; Emulate a conditional br instruction
3172 %Val = zext i1 %value to i32
3173 switch i32 %Val, label %truedest [ i32 0, label %falsedest ]
3175 ; Emulate an unconditional br instruction
3176 switch i32 0, label %dest [ ]
3178 ; Implement a jump table:
3179 switch i32 %val, label %otherwise [ i32 0, label %onzero
3181 i32 2, label %ontwo ]
3185 '``indirectbr``' Instruction
3186 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3193 indirectbr <somety>* <address>, [ label <dest1>, label <dest2>, ... ]
3198 The '``indirectbr``' instruction implements an indirect branch to a
3199 label within the current function, whose address is specified by
3200 "``address``". Address must be derived from a
3201 :ref:`blockaddress <blockaddress>` constant.
3206 The '``address``' argument is the address of the label to jump to. The
3207 rest of the arguments indicate the full set of possible destinations
3208 that the address may point to. Blocks are allowed to occur multiple
3209 times in the destination list, though this isn't particularly useful.
3211 This destination list is required so that dataflow analysis has an
3212 accurate understanding of the CFG.
3217 Control transfers to the block specified in the address argument. All
3218 possible destination blocks must be listed in the label list, otherwise
3219 this instruction has undefined behavior. This implies that jumps to
3220 labels defined in other functions have undefined behavior as well.
3225 This is typically implemented with a jump through a register.
3230 .. code-block:: llvm
3232 indirectbr i8* %Addr, [ label %bb1, label %bb2, label %bb3 ]
3236 '``invoke``' Instruction
3237 ^^^^^^^^^^^^^^^^^^^^^^^^
3244 <result> = invoke [cconv] [ret attrs] <ptr to function ty> <function ptr val>(<function args>) [fn attrs]
3245 to label <normal label> unwind label <exception label>
3250 The '``invoke``' instruction causes control to transfer to a specified
3251 function, with the possibility of control flow transfer to either the
3252 '``normal``' label or the '``exception``' label. If the callee function
3253 returns with the "``ret``" instruction, control flow will return to the
3254 "normal" label. If the callee (or any indirect callees) returns via the
3255 ":ref:`resume <i_resume>`" instruction or other exception handling
3256 mechanism, control is interrupted and continued at the dynamically
3257 nearest "exception" label.
3259 The '``exception``' label is a `landing
3260 pad <ExceptionHandling.html#overview>`_ for the exception. As such,
3261 '``exception``' label is required to have the
3262 ":ref:`landingpad <i_landingpad>`" instruction, which contains the
3263 information about the behavior of the program after unwinding happens,
3264 as its first non-PHI instruction. The restrictions on the
3265 "``landingpad``" instruction's tightly couples it to the "``invoke``"
3266 instruction, so that the important information contained within the
3267 "``landingpad``" instruction can't be lost through normal code motion.
3272 This instruction requires several arguments:
3274 #. The optional "cconv" marker indicates which :ref:`calling
3275 convention <callingconv>` the call should use. If none is
3276 specified, the call defaults to using C calling conventions.
3277 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
3278 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
3280 #. '``ptr to function ty``': shall be the signature of the pointer to
3281 function value being invoked. In most cases, this is a direct
3282 function invocation, but indirect ``invoke``'s are just as possible,
3283 branching off an arbitrary pointer to function value.
3284 #. '``function ptr val``': An LLVM value containing a pointer to a
3285 function to be invoked.
3286 #. '``function args``': argument list whose types match the function
3287 signature argument types and parameter attributes. All arguments must
3288 be of :ref:`first class <t_firstclass>` type. If the function signature
3289 indicates the function accepts a variable number of arguments, the
3290 extra arguments can be specified.
3291 #. '``normal label``': the label reached when the called function
3292 executes a '``ret``' instruction.
3293 #. '``exception label``': the label reached when a callee returns via
3294 the :ref:`resume <i_resume>` instruction or other exception handling
3296 #. The optional :ref:`function attributes <fnattrs>` list. Only
3297 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
3298 attributes are valid here.
3303 This instruction is designed to operate as a standard '``call``'
3304 instruction in most regards. The primary difference is that it
3305 establishes an association with a label, which is used by the runtime
3306 library to unwind the stack.
3308 This instruction is used in languages with destructors to ensure that
3309 proper cleanup is performed in the case of either a ``longjmp`` or a
3310 thrown exception. Additionally, this is important for implementation of
3311 '``catch``' clauses in high-level languages that support them.
3313 For the purposes of the SSA form, the definition of the value returned
3314 by the '``invoke``' instruction is deemed to occur on the edge from the
3315 current block to the "normal" label. If the callee unwinds then no
3316 return value is available.
3321 .. code-block:: llvm
3323 %retval = invoke i32 @Test(i32 15) to label %Continue
3324 unwind label %TestCleanup ; {i32}:retval set
3325 %retval = invoke coldcc i32 %Testfnptr(i32 15) to label %Continue
3326 unwind label %TestCleanup ; {i32}:retval set
3330 '``resume``' Instruction
3331 ^^^^^^^^^^^^^^^^^^^^^^^^
3338 resume <type> <value>
3343 The '``resume``' instruction is a terminator instruction that has no
3349 The '``resume``' instruction requires one argument, which must have the
3350 same type as the result of any '``landingpad``' instruction in the same
3356 The '``resume``' instruction resumes propagation of an existing
3357 (in-flight) exception whose unwinding was interrupted with a
3358 :ref:`landingpad <i_landingpad>` instruction.
3363 .. code-block:: llvm
3365 resume { i8*, i32 } %exn
3369 '``unreachable``' Instruction
3370 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3382 The '``unreachable``' instruction has no defined semantics. This
3383 instruction is used to inform the optimizer that a particular portion of
3384 the code is not reachable. This can be used to indicate that the code
3385 after a no-return function cannot be reached, and other facts.
3390 The '``unreachable``' instruction has no defined semantics.
3397 Binary operators are used to do most of the computation in a program.
3398 They require two operands of the same type, execute an operation on
3399 them, and produce a single value. The operands might represent multiple
3400 data, as is the case with the :ref:`vector <t_vector>` data type. The
3401 result value has the same type as its operands.
3403 There are several different binary operators:
3407 '``add``' Instruction
3408 ^^^^^^^^^^^^^^^^^^^^^
3415 <result> = add <ty> <op1>, <op2> ; yields {ty}:result
3416 <result> = add nuw <ty> <op1>, <op2> ; yields {ty}:result
3417 <result> = add nsw <ty> <op1>, <op2> ; yields {ty}:result
3418 <result> = add nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3423 The '``add``' instruction returns the sum of its two operands.
3428 The two arguments to the '``add``' instruction must be
3429 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3430 arguments must have identical types.
3435 The value produced is the integer sum of the two operands.
3437 If the sum has unsigned overflow, the result returned is the
3438 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
3441 Because LLVM integers use a two's complement representation, this
3442 instruction is appropriate for both signed and unsigned integers.
3444 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3445 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3446 result value of the ``add`` is a :ref:`poison value <poisonvalues>` if
3447 unsigned and/or signed overflow, respectively, occurs.
3452 .. code-block:: llvm
3454 <result> = add i32 4, %var ; yields {i32}:result = 4 + %var
3458 '``fadd``' Instruction
3459 ^^^^^^^^^^^^^^^^^^^^^^
3466 <result> = fadd [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3471 The '``fadd``' instruction returns the sum of its two operands.
3476 The two arguments to the '``fadd``' instruction must be :ref:`floating
3477 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3478 Both arguments must have identical types.
3483 The value produced is the floating point sum of the two operands. This
3484 instruction can also take any number of :ref:`fast-math flags <fastmath>`,
3485 which are optimization hints to enable otherwise unsafe floating point
3491 .. code-block:: llvm
3493 <result> = fadd float 4.0, %var ; yields {float}:result = 4.0 + %var
3495 '``sub``' Instruction
3496 ^^^^^^^^^^^^^^^^^^^^^
3503 <result> = sub <ty> <op1>, <op2> ; yields {ty}:result
3504 <result> = sub nuw <ty> <op1>, <op2> ; yields {ty}:result
3505 <result> = sub nsw <ty> <op1>, <op2> ; yields {ty}:result
3506 <result> = sub nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3511 The '``sub``' instruction returns the difference of its two operands.
3513 Note that the '``sub``' instruction is used to represent the '``neg``'
3514 instruction present in most other intermediate representations.
3519 The two arguments to the '``sub``' instruction must be
3520 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3521 arguments must have identical types.
3526 The value produced is the integer difference of the two operands.
3528 If the difference has unsigned overflow, the result returned is the
3529 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
3532 Because LLVM integers use a two's complement representation, this
3533 instruction is appropriate for both signed and unsigned integers.
3535 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3536 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3537 result value of the ``sub`` is a :ref:`poison value <poisonvalues>` if
3538 unsigned and/or signed overflow, respectively, occurs.
3543 .. code-block:: llvm
3545 <result> = sub i32 4, %var ; yields {i32}:result = 4 - %var
3546 <result> = sub i32 0, %val ; yields {i32}:result = -%var
3550 '``fsub``' Instruction
3551 ^^^^^^^^^^^^^^^^^^^^^^
3558 <result> = fsub [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3563 The '``fsub``' instruction returns the difference of its two operands.
3565 Note that the '``fsub``' instruction is used to represent the '``fneg``'
3566 instruction present in most other intermediate representations.
3571 The two arguments to the '``fsub``' instruction must be :ref:`floating
3572 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3573 Both arguments must have identical types.
3578 The value produced is the floating point difference of the two operands.
3579 This instruction can also take any number of :ref:`fast-math
3580 flags <fastmath>`, which are optimization hints to enable otherwise
3581 unsafe floating point optimizations:
3586 .. code-block:: llvm
3588 <result> = fsub float 4.0, %var ; yields {float}:result = 4.0 - %var
3589 <result> = fsub float -0.0, %val ; yields {float}:result = -%var
3591 '``mul``' Instruction
3592 ^^^^^^^^^^^^^^^^^^^^^
3599 <result> = mul <ty> <op1>, <op2> ; yields {ty}:result
3600 <result> = mul nuw <ty> <op1>, <op2> ; yields {ty}:result
3601 <result> = mul nsw <ty> <op1>, <op2> ; yields {ty}:result
3602 <result> = mul nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3607 The '``mul``' instruction returns the product of its two operands.
3612 The two arguments to the '``mul``' instruction must be
3613 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3614 arguments must have identical types.
3619 The value produced is the integer product of the two operands.
3621 If the result of the multiplication has unsigned overflow, the result
3622 returned is the mathematical result modulo 2\ :sup:`n`\ , where n is the
3623 bit width of the result.
3625 Because LLVM integers use a two's complement representation, and the
3626 result is the same width as the operands, this instruction returns the
3627 correct result for both signed and unsigned integers. If a full product
3628 (e.g. ``i32`` * ``i32`` -> ``i64``) is needed, the operands should be
3629 sign-extended or zero-extended as appropriate to the width of the full
3632 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3633 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3634 result value of the ``mul`` is a :ref:`poison value <poisonvalues>` if
3635 unsigned and/or signed overflow, respectively, occurs.
3640 .. code-block:: llvm
3642 <result> = mul i32 4, %var ; yields {i32}:result = 4 * %var
3646 '``fmul``' Instruction
3647 ^^^^^^^^^^^^^^^^^^^^^^
3654 <result> = fmul [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3659 The '``fmul``' instruction returns the product of its two operands.
3664 The two arguments to the '``fmul``' instruction must be :ref:`floating
3665 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3666 Both arguments must have identical types.
3671 The value produced is the floating point product of the two operands.
3672 This instruction can also take any number of :ref:`fast-math
3673 flags <fastmath>`, which are optimization hints to enable otherwise
3674 unsafe floating point optimizations:
3679 .. code-block:: llvm
3681 <result> = fmul float 4.0, %var ; yields {float}:result = 4.0 * %var
3683 '``udiv``' Instruction
3684 ^^^^^^^^^^^^^^^^^^^^^^
3691 <result> = udiv <ty> <op1>, <op2> ; yields {ty}:result
3692 <result> = udiv exact <ty> <op1>, <op2> ; yields {ty}:result
3697 The '``udiv``' instruction returns the quotient of its two operands.
3702 The two arguments to the '``udiv``' instruction must be
3703 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3704 arguments must have identical types.
3709 The value produced is the unsigned integer quotient of the two operands.
3711 Note that unsigned integer division and signed integer division are
3712 distinct operations; for signed integer division, use '``sdiv``'.
3714 Division by zero leads to undefined behavior.
3716 If the ``exact`` keyword is present, the result value of the ``udiv`` is
3717 a :ref:`poison value <poisonvalues>` if %op1 is not a multiple of %op2 (as
3718 such, "((a udiv exact b) mul b) == a").
3723 .. code-block:: llvm
3725 <result> = udiv i32 4, %var ; yields {i32}:result = 4 / %var
3727 '``sdiv``' Instruction
3728 ^^^^^^^^^^^^^^^^^^^^^^
3735 <result> = sdiv <ty> <op1>, <op2> ; yields {ty}:result
3736 <result> = sdiv exact <ty> <op1>, <op2> ; yields {ty}:result
3741 The '``sdiv``' instruction returns the quotient of its two operands.
3746 The two arguments to the '``sdiv``' instruction must be
3747 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3748 arguments must have identical types.
3753 The value produced is the signed integer quotient of the two operands
3754 rounded towards zero.
3756 Note that signed integer division and unsigned integer division are
3757 distinct operations; for unsigned integer division, use '``udiv``'.
3759 Division by zero leads to undefined behavior. Overflow also leads to
3760 undefined behavior; this is a rare case, but can occur, for example, by
3761 doing a 32-bit division of -2147483648 by -1.
3763 If the ``exact`` keyword is present, the result value of the ``sdiv`` is
3764 a :ref:`poison value <poisonvalues>` if the result would be rounded.
3769 .. code-block:: llvm
3771 <result> = sdiv i32 4, %var ; yields {i32}:result = 4 / %var
3775 '``fdiv``' Instruction
3776 ^^^^^^^^^^^^^^^^^^^^^^
3783 <result> = fdiv [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3788 The '``fdiv``' instruction returns the quotient of its two operands.
3793 The two arguments to the '``fdiv``' instruction must be :ref:`floating
3794 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3795 Both arguments must have identical types.
3800 The value produced is the floating point quotient of the two operands.
3801 This instruction can also take any number of :ref:`fast-math
3802 flags <fastmath>`, which are optimization hints to enable otherwise
3803 unsafe floating point optimizations:
3808 .. code-block:: llvm
3810 <result> = fdiv float 4.0, %var ; yields {float}:result = 4.0 / %var
3812 '``urem``' Instruction
3813 ^^^^^^^^^^^^^^^^^^^^^^
3820 <result> = urem <ty> <op1>, <op2> ; yields {ty}:result
3825 The '``urem``' instruction returns the remainder from the unsigned
3826 division of its two arguments.
3831 The two arguments to the '``urem``' instruction must be
3832 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3833 arguments must have identical types.
3838 This instruction returns the unsigned integer *remainder* of a division.
3839 This instruction always performs an unsigned division to get the
3842 Note that unsigned integer remainder and signed integer remainder are
3843 distinct operations; for signed integer remainder, use '``srem``'.
3845 Taking the remainder of a division by zero leads to undefined behavior.
3850 .. code-block:: llvm
3852 <result> = urem i32 4, %var ; yields {i32}:result = 4 % %var
3854 '``srem``' Instruction
3855 ^^^^^^^^^^^^^^^^^^^^^^
3862 <result> = srem <ty> <op1>, <op2> ; yields {ty}:result
3867 The '``srem``' instruction returns the remainder from the signed
3868 division of its two operands. This instruction can also take
3869 :ref:`vector <t_vector>` versions of the values in which case the elements
3875 The two arguments to the '``srem``' instruction must be
3876 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3877 arguments must have identical types.
3882 This instruction returns the *remainder* of a division (where the result
3883 is either zero or has the same sign as the dividend, ``op1``), not the
3884 *modulo* operator (where the result is either zero or has the same sign
3885 as the divisor, ``op2``) of a value. For more information about the
3886 difference, see `The Math
3887 Forum <http://mathforum.org/dr.math/problems/anne.4.28.99.html>`_. For a
3888 table of how this is implemented in various languages, please see
3890 operation <http://en.wikipedia.org/wiki/Modulo_operation>`_.
3892 Note that signed integer remainder and unsigned integer remainder are
3893 distinct operations; for unsigned integer remainder, use '``urem``'.
3895 Taking the remainder of a division by zero leads to undefined behavior.
3896 Overflow also leads to undefined behavior; this is a rare case, but can
3897 occur, for example, by taking the remainder of a 32-bit division of
3898 -2147483648 by -1. (The remainder doesn't actually overflow, but this
3899 rule lets srem be implemented using instructions that return both the
3900 result of the division and the remainder.)
3905 .. code-block:: llvm
3907 <result> = srem i32 4, %var ; yields {i32}:result = 4 % %var
3911 '``frem``' Instruction
3912 ^^^^^^^^^^^^^^^^^^^^^^
3919 <result> = frem [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3924 The '``frem``' instruction returns the remainder from the division of
3930 The two arguments to the '``frem``' instruction must be :ref:`floating
3931 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3932 Both arguments must have identical types.
3937 This instruction returns the *remainder* of a division. The remainder
3938 has the same sign as the dividend. This instruction can also take any
3939 number of :ref:`fast-math flags <fastmath>`, which are optimization hints
3940 to enable otherwise unsafe floating point optimizations:
3945 .. code-block:: llvm
3947 <result> = frem float 4.0, %var ; yields {float}:result = 4.0 % %var
3951 Bitwise Binary Operations
3952 -------------------------
3954 Bitwise binary operators are used to do various forms of bit-twiddling
3955 in a program. They are generally very efficient instructions and can
3956 commonly be strength reduced from other instructions. They require two
3957 operands of the same type, execute an operation on them, and produce a
3958 single value. The resulting value is the same type as its operands.
3960 '``shl``' Instruction
3961 ^^^^^^^^^^^^^^^^^^^^^
3968 <result> = shl <ty> <op1>, <op2> ; yields {ty}:result
3969 <result> = shl nuw <ty> <op1>, <op2> ; yields {ty}:result
3970 <result> = shl nsw <ty> <op1>, <op2> ; yields {ty}:result
3971 <result> = shl nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3976 The '``shl``' instruction returns the first operand shifted to the left
3977 a specified number of bits.
3982 Both arguments to the '``shl``' instruction must be the same
3983 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
3984 '``op2``' is treated as an unsigned value.
3989 The value produced is ``op1`` \* 2\ :sup:`op2` mod 2\ :sup:`n`,
3990 where ``n`` is the width of the result. If ``op2`` is (statically or
3991 dynamically) negative or equal to or larger than the number of bits in
3992 ``op1``, the result is undefined. If the arguments are vectors, each
3993 vector element of ``op1`` is shifted by the corresponding shift amount
3996 If the ``nuw`` keyword is present, then the shift produces a :ref:`poison
3997 value <poisonvalues>` if it shifts out any non-zero bits. If the
3998 ``nsw`` keyword is present, then the shift produces a :ref:`poison
3999 value <poisonvalues>` if it shifts out any bits that disagree with the
4000 resultant sign bit. As such, NUW/NSW have the same semantics as they
4001 would if the shift were expressed as a mul instruction with the same
4002 nsw/nuw bits in (mul %op1, (shl 1, %op2)).
4007 .. code-block:: llvm
4009 <result> = shl i32 4, %var ; yields {i32}: 4 << %var
4010 <result> = shl i32 4, 2 ; yields {i32}: 16
4011 <result> = shl i32 1, 10 ; yields {i32}: 1024
4012 <result> = shl i32 1, 32 ; undefined
4013 <result> = shl <2 x i32> < i32 1, i32 1>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 2, i32 4>
4015 '``lshr``' Instruction
4016 ^^^^^^^^^^^^^^^^^^^^^^
4023 <result> = lshr <ty> <op1>, <op2> ; yields {ty}:result
4024 <result> = lshr exact <ty> <op1>, <op2> ; yields {ty}:result
4029 The '``lshr``' instruction (logical shift right) returns the first
4030 operand shifted to the right a specified number of bits with zero fill.
4035 Both arguments to the '``lshr``' instruction must be the same
4036 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4037 '``op2``' is treated as an unsigned value.
4042 This instruction always performs a logical shift right operation. The
4043 most significant bits of the result will be filled with zero bits after
4044 the shift. If ``op2`` is (statically or dynamically) equal to or larger
4045 than the number of bits in ``op1``, the result is undefined. If the
4046 arguments are vectors, each vector element of ``op1`` is shifted by the
4047 corresponding shift amount in ``op2``.
4049 If the ``exact`` keyword is present, the result value of the ``lshr`` is
4050 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
4056 .. code-block:: llvm
4058 <result> = lshr i32 4, 1 ; yields {i32}:result = 2
4059 <result> = lshr i32 4, 2 ; yields {i32}:result = 1
4060 <result> = lshr i8 4, 3 ; yields {i8}:result = 0
4061 <result> = lshr i8 -2, 1 ; yields {i8}:result = 0x7F
4062 <result> = lshr i32 1, 32 ; undefined
4063 <result> = lshr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 0x7FFFFFFF, i32 1>
4065 '``ashr``' Instruction
4066 ^^^^^^^^^^^^^^^^^^^^^^
4073 <result> = ashr <ty> <op1>, <op2> ; yields {ty}:result
4074 <result> = ashr exact <ty> <op1>, <op2> ; yields {ty}:result
4079 The '``ashr``' instruction (arithmetic shift right) returns the first
4080 operand shifted to the right a specified number of bits with sign
4086 Both arguments to the '``ashr``' instruction must be the same
4087 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4088 '``op2``' is treated as an unsigned value.
4093 This instruction always performs an arithmetic shift right operation,
4094 The most significant bits of the result will be filled with the sign bit
4095 of ``op1``. If ``op2`` is (statically or dynamically) equal to or larger
4096 than the number of bits in ``op1``, the result is undefined. If the
4097 arguments are vectors, each vector element of ``op1`` is shifted by the
4098 corresponding shift amount in ``op2``.
4100 If the ``exact`` keyword is present, the result value of the ``ashr`` is
4101 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
4107 .. code-block:: llvm
4109 <result> = ashr i32 4, 1 ; yields {i32}:result = 2
4110 <result> = ashr i32 4, 2 ; yields {i32}:result = 1
4111 <result> = ashr i8 4, 3 ; yields {i8}:result = 0
4112 <result> = ashr i8 -2, 1 ; yields {i8}:result = -1
4113 <result> = ashr i32 1, 32 ; undefined
4114 <result> = ashr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 3> ; yields: result=<2 x i32> < i32 -1, i32 0>
4116 '``and``' Instruction
4117 ^^^^^^^^^^^^^^^^^^^^^
4124 <result> = and <ty> <op1>, <op2> ; yields {ty}:result
4129 The '``and``' instruction returns the bitwise logical and of its two
4135 The two arguments to the '``and``' instruction must be
4136 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4137 arguments must have identical types.
4142 The truth table used for the '``and``' instruction is:
4159 .. code-block:: llvm
4161 <result> = and i32 4, %var ; yields {i32}:result = 4 & %var
4162 <result> = and i32 15, 40 ; yields {i32}:result = 8
4163 <result> = and i32 4, 8 ; yields {i32}:result = 0
4165 '``or``' Instruction
4166 ^^^^^^^^^^^^^^^^^^^^
4173 <result> = or <ty> <op1>, <op2> ; yields {ty}:result
4178 The '``or``' instruction returns the bitwise logical inclusive or of its
4184 The two arguments to the '``or``' instruction must be
4185 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4186 arguments must have identical types.
4191 The truth table used for the '``or``' instruction is:
4210 <result> = or i32 4, %var ; yields {i32}:result = 4 | %var
4211 <result> = or i32 15, 40 ; yields {i32}:result = 47
4212 <result> = or i32 4, 8 ; yields {i32}:result = 12
4214 '``xor``' Instruction
4215 ^^^^^^^^^^^^^^^^^^^^^
4222 <result> = xor <ty> <op1>, <op2> ; yields {ty}:result
4227 The '``xor``' instruction returns the bitwise logical exclusive or of
4228 its two operands. The ``xor`` is used to implement the "one's
4229 complement" operation, which is the "~" operator in C.
4234 The two arguments to the '``xor``' instruction must be
4235 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4236 arguments must have identical types.
4241 The truth table used for the '``xor``' instruction is:
4258 .. code-block:: llvm
4260 <result> = xor i32 4, %var ; yields {i32}:result = 4 ^ %var
4261 <result> = xor i32 15, 40 ; yields {i32}:result = 39
4262 <result> = xor i32 4, 8 ; yields {i32}:result = 12
4263 <result> = xor i32 %V, -1 ; yields {i32}:result = ~%V
4268 LLVM supports several instructions to represent vector operations in a
4269 target-independent manner. These instructions cover the element-access
4270 and vector-specific operations needed to process vectors effectively.
4271 While LLVM does directly support these vector operations, many
4272 sophisticated algorithms will want to use target-specific intrinsics to
4273 take full advantage of a specific target.
4275 .. _i_extractelement:
4277 '``extractelement``' Instruction
4278 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4285 <result> = extractelement <n x <ty>> <val>, i32 <idx> ; yields <ty>
4290 The '``extractelement``' instruction extracts a single scalar element
4291 from a vector at a specified index.
4296 The first operand of an '``extractelement``' instruction is a value of
4297 :ref:`vector <t_vector>` type. The second operand is an index indicating
4298 the position from which to extract the element. The index may be a
4304 The result is a scalar of the same type as the element type of ``val``.
4305 Its value is the value at position ``idx`` of ``val``. If ``idx``
4306 exceeds the length of ``val``, the results are undefined.
4311 .. code-block:: llvm
4313 <result> = extractelement <4 x i32> %vec, i32 0 ; yields i32
4315 .. _i_insertelement:
4317 '``insertelement``' Instruction
4318 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4325 <result> = insertelement <n x <ty>> <val>, <ty> <elt>, i32 <idx> ; yields <n x <ty>>
4330 The '``insertelement``' instruction inserts a scalar element into a
4331 vector at a specified index.
4336 The first operand of an '``insertelement``' instruction is a value of
4337 :ref:`vector <t_vector>` type. The second operand is a scalar value whose
4338 type must equal the element type of the first operand. The third operand
4339 is an index indicating the position at which to insert the value. The
4340 index may be a variable.
4345 The result is a vector of the same type as ``val``. Its element values
4346 are those of ``val`` except at position ``idx``, where it gets the value
4347 ``elt``. If ``idx`` exceeds the length of ``val``, the results are
4353 .. code-block:: llvm
4355 <result> = insertelement <4 x i32> %vec, i32 1, i32 0 ; yields <4 x i32>
4357 .. _i_shufflevector:
4359 '``shufflevector``' Instruction
4360 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4367 <result> = shufflevector <n x <ty>> <v1>, <n x <ty>> <v2>, <m x i32> <mask> ; yields <m x <ty>>
4372 The '``shufflevector``' instruction constructs a permutation of elements
4373 from two input vectors, returning a vector with the same element type as
4374 the input and length that is the same as the shuffle mask.
4379 The first two operands of a '``shufflevector``' instruction are vectors
4380 with the same type. The third argument is a shuffle mask whose element
4381 type is always 'i32'. The result of the instruction is a vector whose
4382 length is the same as the shuffle mask and whose element type is the
4383 same as the element type of the first two operands.
4385 The shuffle mask operand is required to be a constant vector with either
4386 constant integer or undef values.
4391 The elements of the two input vectors are numbered from left to right
4392 across both of the vectors. The shuffle mask operand specifies, for each
4393 element of the result vector, which element of the two input vectors the
4394 result element gets. The element selector may be undef (meaning "don't
4395 care") and the second operand may be undef if performing a shuffle from
4401 .. code-block:: llvm
4403 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
4404 <4 x i32> <i32 0, i32 4, i32 1, i32 5> ; yields <4 x i32>
4405 <result> = shufflevector <4 x i32> %v1, <4 x i32> undef,
4406 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32> - Identity shuffle.
4407 <result> = shufflevector <8 x i32> %v1, <8 x i32> undef,
4408 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32>
4409 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
4410 <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7 > ; yields <8 x i32>
4412 Aggregate Operations
4413 --------------------
4415 LLVM supports several instructions for working with
4416 :ref:`aggregate <t_aggregate>` values.
4420 '``extractvalue``' Instruction
4421 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4428 <result> = extractvalue <aggregate type> <val>, <idx>{, <idx>}*
4433 The '``extractvalue``' instruction extracts the value of a member field
4434 from an :ref:`aggregate <t_aggregate>` value.
4439 The first operand of an '``extractvalue``' instruction is a value of
4440 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The operands are
4441 constant indices to specify which value to extract in a similar manner
4442 as indices in a '``getelementptr``' instruction.
4444 The major differences to ``getelementptr`` indexing are:
4446 - Since the value being indexed is not a pointer, the first index is
4447 omitted and assumed to be zero.
4448 - At least one index must be specified.
4449 - Not only struct indices but also array indices must be in bounds.
4454 The result is the value at the position in the aggregate specified by
4460 .. code-block:: llvm
4462 <result> = extractvalue {i32, float} %agg, 0 ; yields i32
4466 '``insertvalue``' Instruction
4467 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4474 <result> = insertvalue <aggregate type> <val>, <ty> <elt>, <idx>{, <idx>}* ; yields <aggregate type>
4479 The '``insertvalue``' instruction inserts a value into a member field in
4480 an :ref:`aggregate <t_aggregate>` value.
4485 The first operand of an '``insertvalue``' instruction is a value of
4486 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The second operand is
4487 a first-class value to insert. The following operands are constant
4488 indices indicating the position at which to insert the value in a
4489 similar manner as indices in a '``extractvalue``' instruction. The value
4490 to insert must have the same type as the value identified by the
4496 The result is an aggregate of the same type as ``val``. Its value is
4497 that of ``val`` except that the value at the position specified by the
4498 indices is that of ``elt``.
4503 .. code-block:: llvm
4505 %agg1 = insertvalue {i32, float} undef, i32 1, 0 ; yields {i32 1, float undef}
4506 %agg2 = insertvalue {i32, float} %agg1, float %val, 1 ; yields {i32 1, float %val}
4507 %agg3 = insertvalue {i32, {float}} %agg1, float %val, 1, 0 ; yields {i32 1, float %val}
4511 Memory Access and Addressing Operations
4512 ---------------------------------------
4514 A key design point of an SSA-based representation is how it represents
4515 memory. In LLVM, no memory locations are in SSA form, which makes things
4516 very simple. This section describes how to read, write, and allocate
4521 '``alloca``' Instruction
4522 ^^^^^^^^^^^^^^^^^^^^^^^^
4529 <result> = alloca <type>[, <ty> <NumElements>][, align <alignment>] ; yields {type*}:result
4534 The '``alloca``' instruction allocates memory on the stack frame of the
4535 currently executing function, to be automatically released when this
4536 function returns to its caller. The object is always allocated in the
4537 generic address space (address space zero).
4542 The '``alloca``' instruction allocates ``sizeof(<type>)*NumElements``
4543 bytes of memory on the runtime stack, returning a pointer of the
4544 appropriate type to the program. If "NumElements" is specified, it is
4545 the number of elements allocated, otherwise "NumElements" is defaulted
4546 to be one. If a constant alignment is specified, the value result of the
4547 allocation is guaranteed to be aligned to at least that boundary. If not
4548 specified, or if zero, the target can choose to align the allocation on
4549 any convenient boundary compatible with the type.
4551 '``type``' may be any sized type.
4556 Memory is allocated; a pointer is returned. The operation is undefined
4557 if there is insufficient stack space for the allocation. '``alloca``'d
4558 memory is automatically released when the function returns. The
4559 '``alloca``' instruction is commonly used to represent automatic
4560 variables that must have an address available. When the function returns
4561 (either with the ``ret`` or ``resume`` instructions), the memory is
4562 reclaimed. Allocating zero bytes is legal, but the result is undefined.
4563 The order in which memory is allocated (ie., which way the stack grows)
4569 .. code-block:: llvm
4571 %ptr = alloca i32 ; yields {i32*}:ptr
4572 %ptr = alloca i32, i32 4 ; yields {i32*}:ptr
4573 %ptr = alloca i32, i32 4, align 1024 ; yields {i32*}:ptr
4574 %ptr = alloca i32, align 1024 ; yields {i32*}:ptr
4578 '``load``' Instruction
4579 ^^^^^^^^^^^^^^^^^^^^^^
4586 <result> = load [volatile] <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>][, !invariant.load !<index>]
4587 <result> = load atomic [volatile] <ty>* <pointer> [singlethread] <ordering>, align <alignment>
4588 !<index> = !{ i32 1 }
4593 The '``load``' instruction is used to read from memory.
4598 The argument to the ``load`` instruction specifies the memory address
4599 from which to load. The pointer must point to a :ref:`first
4600 class <t_firstclass>` type. If the ``load`` is marked as ``volatile``,
4601 then the optimizer is not allowed to modify the number or order of
4602 execution of this ``load`` with other :ref:`volatile
4603 operations <volatile>`.
4605 If the ``load`` is marked as ``atomic``, it takes an extra
4606 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
4607 ``release`` and ``acq_rel`` orderings are not valid on ``load``
4608 instructions. Atomic loads produce :ref:`defined <memmodel>` results
4609 when they may see multiple atomic stores. The type of the pointee must
4610 be an integer type whose bit width is a power of two greater than or
4611 equal to eight and less than or equal to a target-specific size limit.
4612 ``align`` must be explicitly specified on atomic loads, and the load has
4613 undefined behavior if the alignment is not set to a value which is at
4614 least the size in bytes of the pointee. ``!nontemporal`` does not have
4615 any defined semantics for atomic loads.
4617 The optional constant ``align`` argument specifies the alignment of the
4618 operation (that is, the alignment of the memory address). A value of 0
4619 or an omitted ``align`` argument means that the operation has the ABI
4620 alignment for the target. It is the responsibility of the code emitter
4621 to ensure that the alignment information is correct. Overestimating the
4622 alignment results in undefined behavior. Underestimating the alignment
4623 may produce less efficient code. An alignment of 1 is always safe.
4625 The optional ``!nontemporal`` metadata must reference a single
4626 metatadata name ``<index>`` corresponding to a metadata node with one
4627 ``i32`` entry of value 1. The existence of the ``!nontemporal``
4628 metatadata on the instruction tells the optimizer and code generator
4629 that this load is not expected to be reused in the cache. The code
4630 generator may select special instructions to save cache bandwidth, such
4631 as the ``MOVNT`` instruction on x86.
4633 The optional ``!invariant.load`` metadata must reference a single
4634 metatadata name ``<index>`` corresponding to a metadata node with no
4635 entries. The existence of the ``!invariant.load`` metatadata on the
4636 instruction tells the optimizer and code generator that this load
4637 address points to memory which does not change value during program
4638 execution. The optimizer may then move this load around, for example, by
4639 hoisting it out of loops using loop invariant code motion.
4644 The location of memory pointed to is loaded. If the value being loaded
4645 is of scalar type then the number of bytes read does not exceed the
4646 minimum number of bytes needed to hold all bits of the type. For
4647 example, loading an ``i24`` reads at most three bytes. When loading a
4648 value of a type like ``i20`` with a size that is not an integral number
4649 of bytes, the result is undefined if the value was not originally
4650 written using a store of the same type.
4655 .. code-block:: llvm
4657 %ptr = alloca i32 ; yields {i32*}:ptr
4658 store i32 3, i32* %ptr ; yields {void}
4659 %val = load i32* %ptr ; yields {i32}:val = i32 3
4663 '``store``' Instruction
4664 ^^^^^^^^^^^^^^^^^^^^^^^
4671 store [volatile] <ty> <value>, <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>] ; yields {void}
4672 store atomic [volatile] <ty> <value>, <ty>* <pointer> [singlethread] <ordering>, align <alignment> ; yields {void}
4677 The '``store``' instruction is used to write to memory.
4682 There are two arguments to the ``store`` instruction: a value to store
4683 and an address at which to store it. The type of the ``<pointer>``
4684 operand must be a pointer to the :ref:`first class <t_firstclass>` type of
4685 the ``<value>`` operand. If the ``store`` is marked as ``volatile``,
4686 then the optimizer is not allowed to modify the number or order of
4687 execution of this ``store`` with other :ref:`volatile
4688 operations <volatile>`.
4690 If the ``store`` is marked as ``atomic``, it takes an extra
4691 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
4692 ``acquire`` and ``acq_rel`` orderings aren't valid on ``store``
4693 instructions. Atomic loads produce :ref:`defined <memmodel>` results
4694 when they may see multiple atomic stores. The type of the pointee must
4695 be an integer type whose bit width is a power of two greater than or
4696 equal to eight and less than or equal to a target-specific size limit.
4697 ``align`` must be explicitly specified on atomic stores, and the store
4698 has undefined behavior if the alignment is not set to a value which is
4699 at least the size in bytes of the pointee. ``!nontemporal`` does not
4700 have any defined semantics for atomic stores.
4702 The optional constant ``align`` argument specifies the alignment of the
4703 operation (that is, the alignment of the memory address). A value of 0
4704 or an omitted ``align`` argument means that the operation has the ABI
4705 alignment for the target. It is the responsibility of the code emitter
4706 to ensure that the alignment information is correct. Overestimating the
4707 alignment results in undefined behavior. Underestimating the
4708 alignment may produce less efficient code. An alignment of 1 is always
4711 The optional ``!nontemporal`` metadata must reference a single metatadata
4712 name ``<index>`` corresponding to a metadata node with one ``i32`` entry of
4713 value 1. The existence of the ``!nontemporal`` metatadata on the instruction
4714 tells the optimizer and code generator that this load is not expected to
4715 be reused in the cache. The code generator may select special
4716 instructions to save cache bandwidth, such as the MOVNT instruction on
4722 The contents of memory are updated to contain ``<value>`` at the
4723 location specified by the ``<pointer>`` operand. If ``<value>`` is
4724 of scalar type then the number of bytes written does not exceed the
4725 minimum number of bytes needed to hold all bits of the type. For
4726 example, storing an ``i24`` writes at most three bytes. When writing a
4727 value of a type like ``i20`` with a size that is not an integral number
4728 of bytes, it is unspecified what happens to the extra bits that do not
4729 belong to the type, but they will typically be overwritten.
4734 .. code-block:: llvm
4736 %ptr = alloca i32 ; yields {i32*}:ptr
4737 store i32 3, i32* %ptr ; yields {void}
4738 %val = load i32* %ptr ; yields {i32}:val = i32 3
4742 '``fence``' Instruction
4743 ^^^^^^^^^^^^^^^^^^^^^^^
4750 fence [singlethread] <ordering> ; yields {void}
4755 The '``fence``' instruction is used to introduce happens-before edges
4761 '``fence``' instructions take an :ref:`ordering <ordering>` argument which
4762 defines what *synchronizes-with* edges they add. They can only be given
4763 ``acquire``, ``release``, ``acq_rel``, and ``seq_cst`` orderings.
4768 A fence A which has (at least) ``release`` ordering semantics
4769 *synchronizes with* a fence B with (at least) ``acquire`` ordering
4770 semantics if and only if there exist atomic operations X and Y, both
4771 operating on some atomic object M, such that A is sequenced before X, X
4772 modifies M (either directly or through some side effect of a sequence
4773 headed by X), Y is sequenced before B, and Y observes M. This provides a
4774 *happens-before* dependency between A and B. Rather than an explicit
4775 ``fence``, one (but not both) of the atomic operations X or Y might
4776 provide a ``release`` or ``acquire`` (resp.) ordering constraint and
4777 still *synchronize-with* the explicit ``fence`` and establish the
4778 *happens-before* edge.
4780 A ``fence`` which has ``seq_cst`` ordering, in addition to having both
4781 ``acquire`` and ``release`` semantics specified above, participates in
4782 the global program order of other ``seq_cst`` operations and/or fences.
4784 The optional ":ref:`singlethread <singlethread>`" argument specifies
4785 that the fence only synchronizes with other fences in the same thread.
4786 (This is useful for interacting with signal handlers.)
4791 .. code-block:: llvm
4793 fence acquire ; yields {void}
4794 fence singlethread seq_cst ; yields {void}
4798 '``cmpxchg``' Instruction
4799 ^^^^^^^^^^^^^^^^^^^^^^^^^
4806 cmpxchg [volatile] <ty>* <pointer>, <ty> <cmp>, <ty> <new> [singlethread] <ordering> ; yields {ty}
4811 The '``cmpxchg``' instruction is used to atomically modify memory. It
4812 loads a value in memory and compares it to a given value. If they are
4813 equal, it stores a new value into the memory.
4818 There are three arguments to the '``cmpxchg``' instruction: an address
4819 to operate on, a value to compare to the value currently be at that
4820 address, and a new value to place at that address if the compared values
4821 are equal. The type of '<cmp>' must be an integer type whose bit width
4822 is a power of two greater than or equal to eight and less than or equal
4823 to a target-specific size limit. '<cmp>' and '<new>' must have the same
4824 type, and the type of '<pointer>' must be a pointer to that type. If the
4825 ``cmpxchg`` is marked as ``volatile``, then the optimizer is not allowed
4826 to modify the number or order of execution of this ``cmpxchg`` with
4827 other :ref:`volatile operations <volatile>`.
4829 The :ref:`ordering <ordering>` argument specifies how this ``cmpxchg``
4830 synchronizes with other atomic operations.
4832 The optional "``singlethread``" argument declares that the ``cmpxchg``
4833 is only atomic with respect to code (usually signal handlers) running in
4834 the same thread as the ``cmpxchg``. Otherwise the cmpxchg is atomic with
4835 respect to all other code in the system.
4837 The pointer passed into cmpxchg must have alignment greater than or
4838 equal to the size in memory of the operand.
4843 The contents of memory at the location specified by the '``<pointer>``'
4844 operand is read and compared to '``<cmp>``'; if the read value is the
4845 equal, '``<new>``' is written. The original value at the location is
4848 A successful ``cmpxchg`` is a read-modify-write instruction for the purpose
4849 of identifying release sequences. A failed ``cmpxchg`` is equivalent to an
4850 atomic load with an ordering parameter determined by dropping any
4851 ``release`` part of the ``cmpxchg``'s ordering.
4856 .. code-block:: llvm
4859 %orig = atomic load i32* %ptr unordered ; yields {i32}
4863 %cmp = phi i32 [ %orig, %entry ], [%old, %loop]
4864 %squared = mul i32 %cmp, %cmp
4865 %old = cmpxchg i32* %ptr, i32 %cmp, i32 %squared ; yields {i32}
4866 %success = icmp eq i32 %cmp, %old
4867 br i1 %success, label %done, label %loop
4874 '``atomicrmw``' Instruction
4875 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
4882 atomicrmw [volatile] <operation> <ty>* <pointer>, <ty> <value> [singlethread] <ordering> ; yields {ty}
4887 The '``atomicrmw``' instruction is used to atomically modify memory.
4892 There are three arguments to the '``atomicrmw``' instruction: an
4893 operation to apply, an address whose value to modify, an argument to the
4894 operation. The operation must be one of the following keywords:
4908 The type of '<value>' must be an integer type whose bit width is a power
4909 of two greater than or equal to eight and less than or equal to a
4910 target-specific size limit. The type of the '``<pointer>``' operand must
4911 be a pointer to that type. If the ``atomicrmw`` is marked as
4912 ``volatile``, then the optimizer is not allowed to modify the number or
4913 order of execution of this ``atomicrmw`` with other :ref:`volatile
4914 operations <volatile>`.
4919 The contents of memory at the location specified by the '``<pointer>``'
4920 operand are atomically read, modified, and written back. The original
4921 value at the location is returned. The modification is specified by the
4924 - xchg: ``*ptr = val``
4925 - add: ``*ptr = *ptr + val``
4926 - sub: ``*ptr = *ptr - val``
4927 - and: ``*ptr = *ptr & val``
4928 - nand: ``*ptr = ~(*ptr & val)``
4929 - or: ``*ptr = *ptr | val``
4930 - xor: ``*ptr = *ptr ^ val``
4931 - max: ``*ptr = *ptr > val ? *ptr : val`` (using a signed comparison)
4932 - min: ``*ptr = *ptr < val ? *ptr : val`` (using a signed comparison)
4933 - umax: ``*ptr = *ptr > val ? *ptr : val`` (using an unsigned
4935 - umin: ``*ptr = *ptr < val ? *ptr : val`` (using an unsigned
4941 .. code-block:: llvm
4943 %old = atomicrmw add i32* %ptr, i32 1 acquire ; yields {i32}
4945 .. _i_getelementptr:
4947 '``getelementptr``' Instruction
4948 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4955 <result> = getelementptr <pty>* <ptrval>{, <ty> <idx>}*
4956 <result> = getelementptr inbounds <pty>* <ptrval>{, <ty> <idx>}*
4957 <result> = getelementptr <ptr vector> ptrval, <vector index type> idx
4962 The '``getelementptr``' instruction is used to get the address of a
4963 subelement of an :ref:`aggregate <t_aggregate>` data structure. It performs
4964 address calculation only and does not access memory.
4969 The first argument is always a pointer or a vector of pointers, and
4970 forms the basis of the calculation. The remaining arguments are indices
4971 that indicate which of the elements of the aggregate object are indexed.
4972 The interpretation of each index is dependent on the type being indexed
4973 into. The first index always indexes the pointer value given as the
4974 first argument, the second index indexes a value of the type pointed to
4975 (not necessarily the value directly pointed to, since the first index
4976 can be non-zero), etc. The first type indexed into must be a pointer
4977 value, subsequent types can be arrays, vectors, and structs. Note that
4978 subsequent types being indexed into can never be pointers, since that
4979 would require loading the pointer before continuing calculation.
4981 The type of each index argument depends on the type it is indexing into.
4982 When indexing into a (optionally packed) structure, only ``i32`` integer
4983 **constants** are allowed (when using a vector of indices they must all
4984 be the **same** ``i32`` integer constant). When indexing into an array,
4985 pointer or vector, integers of any width are allowed, and they are not
4986 required to be constant. These integers are treated as signed values
4989 For example, let's consider a C code fragment and how it gets compiled
5005 int *foo(struct ST *s) {
5006 return &s[1].Z.B[5][13];
5009 The LLVM code generated by Clang is:
5011 .. code-block:: llvm
5013 %struct.RT = type { i8, [10 x [20 x i32]], i8 }
5014 %struct.ST = type { i32, double, %struct.RT }
5016 define i32* @foo(%struct.ST* %s) nounwind uwtable readnone optsize ssp {
5018 %arrayidx = getelementptr inbounds %struct.ST* %s, i64 1, i32 2, i32 1, i64 5, i64 13
5025 In the example above, the first index is indexing into the
5026 '``%struct.ST*``' type, which is a pointer, yielding a '``%struct.ST``'
5027 = '``{ i32, double, %struct.RT }``' type, a structure. The second index
5028 indexes into the third element of the structure, yielding a
5029 '``%struct.RT``' = '``{ i8 , [10 x [20 x i32]], i8 }``' type, another
5030 structure. The third index indexes into the second element of the
5031 structure, yielding a '``[10 x [20 x i32]]``' type, an array. The two
5032 dimensions of the array are subscripted into, yielding an '``i32``'
5033 type. The '``getelementptr``' instruction returns a pointer to this
5034 element, thus computing a value of '``i32*``' type.
5036 Note that it is perfectly legal to index partially through a structure,
5037 returning a pointer to an inner element. Because of this, the LLVM code
5038 for the given testcase is equivalent to:
5040 .. code-block:: llvm
5042 define i32* @foo(%struct.ST* %s) {
5043 %t1 = getelementptr %struct.ST* %s, i32 1 ; yields %struct.ST*:%t1
5044 %t2 = getelementptr %struct.ST* %t1, i32 0, i32 2 ; yields %struct.RT*:%t2
5045 %t3 = getelementptr %struct.RT* %t2, i32 0, i32 1 ; yields [10 x [20 x i32]]*:%t3
5046 %t4 = getelementptr [10 x [20 x i32]]* %t3, i32 0, i32 5 ; yields [20 x i32]*:%t4
5047 %t5 = getelementptr [20 x i32]* %t4, i32 0, i32 13 ; yields i32*:%t5
5051 If the ``inbounds`` keyword is present, the result value of the
5052 ``getelementptr`` is a :ref:`poison value <poisonvalues>` if the base
5053 pointer is not an *in bounds* address of an allocated object, or if any
5054 of the addresses that would be formed by successive addition of the
5055 offsets implied by the indices to the base address with infinitely
5056 precise signed arithmetic are not an *in bounds* address of that
5057 allocated object. The *in bounds* addresses for an allocated object are
5058 all the addresses that point into the object, plus the address one byte
5059 past the end. In cases where the base is a vector of pointers the
5060 ``inbounds`` keyword applies to each of the computations element-wise.
5062 If the ``inbounds`` keyword is not present, the offsets are added to the
5063 base address with silently-wrapping two's complement arithmetic. If the
5064 offsets have a different width from the pointer, they are sign-extended
5065 or truncated to the width of the pointer. The result value of the
5066 ``getelementptr`` may be outside the object pointed to by the base
5067 pointer. The result value may not necessarily be used to access memory
5068 though, even if it happens to point into allocated storage. See the
5069 :ref:`Pointer Aliasing Rules <pointeraliasing>` section for more
5072 The getelementptr instruction is often confusing. For some more insight
5073 into how it works, see :doc:`the getelementptr FAQ <GetElementPtr>`.
5078 .. code-block:: llvm
5080 ; yields [12 x i8]*:aptr
5081 %aptr = getelementptr {i32, [12 x i8]}* %saptr, i64 0, i32 1
5083 %vptr = getelementptr {i32, <2 x i8>}* %svptr, i64 0, i32 1, i32 1
5085 %eptr = getelementptr [12 x i8]* %aptr, i64 0, i32 1
5087 %iptr = getelementptr [10 x i32]* @arr, i16 0, i16 0
5089 In cases where the pointer argument is a vector of pointers, each index
5090 must be a vector with the same number of elements. For example:
5092 .. code-block:: llvm
5094 %A = getelementptr <4 x i8*> %ptrs, <4 x i64> %offsets,
5096 Conversion Operations
5097 ---------------------
5099 The instructions in this category are the conversion instructions
5100 (casting) which all take a single operand and a type. They perform
5101 various bit conversions on the operand.
5103 '``trunc .. to``' Instruction
5104 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5111 <result> = trunc <ty> <value> to <ty2> ; yields ty2
5116 The '``trunc``' instruction truncates its operand to the type ``ty2``.
5121 The '``trunc``' instruction takes a value to trunc, and a type to trunc
5122 it to. Both types must be of :ref:`integer <t_integer>` types, or vectors
5123 of the same number of integers. The bit size of the ``value`` must be
5124 larger than the bit size of the destination type, ``ty2``. Equal sized
5125 types are not allowed.
5130 The '``trunc``' instruction truncates the high order bits in ``value``
5131 and converts the remaining bits to ``ty2``. Since the source size must
5132 be larger than the destination size, ``trunc`` cannot be a *no-op cast*.
5133 It will always truncate bits.
5138 .. code-block:: llvm
5140 %X = trunc i32 257 to i8 ; yields i8:1
5141 %Y = trunc i32 123 to i1 ; yields i1:true
5142 %Z = trunc i32 122 to i1 ; yields i1:false
5143 %W = trunc <2 x i16> <i16 8, i16 7> to <2 x i8> ; yields <i8 8, i8 7>
5145 '``zext .. to``' Instruction
5146 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5153 <result> = zext <ty> <value> to <ty2> ; yields ty2
5158 The '``zext``' instruction zero extends its operand to type ``ty2``.
5163 The '``zext``' instruction takes a value to cast, and a type to cast it
5164 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
5165 the same number of integers. The bit size of the ``value`` must be
5166 smaller than the bit size of the destination type, ``ty2``.
5171 The ``zext`` fills the high order bits of the ``value`` with zero bits
5172 until it reaches the size of the destination type, ``ty2``.
5174 When zero extending from i1, the result will always be either 0 or 1.
5179 .. code-block:: llvm
5181 %X = zext i32 257 to i64 ; yields i64:257
5182 %Y = zext i1 true to i32 ; yields i32:1
5183 %Z = zext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
5185 '``sext .. to``' Instruction
5186 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5193 <result> = sext <ty> <value> to <ty2> ; yields ty2
5198 The '``sext``' sign extends ``value`` to the type ``ty2``.
5203 The '``sext``' instruction takes a value to cast, and a type to cast it
5204 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
5205 the same number of integers. The bit size of the ``value`` must be
5206 smaller than the bit size of the destination type, ``ty2``.
5211 The '``sext``' instruction performs a sign extension by copying the sign
5212 bit (highest order bit) of the ``value`` until it reaches the bit size
5213 of the type ``ty2``.
5215 When sign extending from i1, the extension always results in -1 or 0.
5220 .. code-block:: llvm
5222 %X = sext i8 -1 to i16 ; yields i16 :65535
5223 %Y = sext i1 true to i32 ; yields i32:-1
5224 %Z = sext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
5226 '``fptrunc .. to``' Instruction
5227 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5234 <result> = fptrunc <ty> <value> to <ty2> ; yields ty2
5239 The '``fptrunc``' instruction truncates ``value`` to type ``ty2``.
5244 The '``fptrunc``' instruction takes a :ref:`floating point <t_floating>`
5245 value to cast and a :ref:`floating point <t_floating>` type to cast it to.
5246 The size of ``value`` must be larger than the size of ``ty2``. This
5247 implies that ``fptrunc`` cannot be used to make a *no-op cast*.
5252 The '``fptrunc``' instruction truncates a ``value`` from a larger
5253 :ref:`floating point <t_floating>` type to a smaller :ref:`floating
5254 point <t_floating>` type. If the value cannot fit within the
5255 destination type, ``ty2``, then the results are undefined.
5260 .. code-block:: llvm
5262 %X = fptrunc double 123.0 to float ; yields float:123.0
5263 %Y = fptrunc double 1.0E+300 to float ; yields undefined
5265 '``fpext .. to``' Instruction
5266 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5273 <result> = fpext <ty> <value> to <ty2> ; yields ty2
5278 The '``fpext``' extends a floating point ``value`` to a larger floating
5284 The '``fpext``' instruction takes a :ref:`floating point <t_floating>`
5285 ``value`` to cast, and a :ref:`floating point <t_floating>` type to cast it
5286 to. The source type must be smaller than the destination type.
5291 The '``fpext``' instruction extends the ``value`` from a smaller
5292 :ref:`floating point <t_floating>` type to a larger :ref:`floating
5293 point <t_floating>` type. The ``fpext`` cannot be used to make a
5294 *no-op cast* because it always changes bits. Use ``bitcast`` to make a
5295 *no-op cast* for a floating point cast.
5300 .. code-block:: llvm
5302 %X = fpext float 3.125 to double ; yields double:3.125000e+00
5303 %Y = fpext double %X to fp128 ; yields fp128:0xL00000000000000004000900000000000
5305 '``fptoui .. to``' Instruction
5306 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5313 <result> = fptoui <ty> <value> to <ty2> ; yields ty2
5318 The '``fptoui``' converts a floating point ``value`` to its unsigned
5319 integer equivalent of type ``ty2``.
5324 The '``fptoui``' instruction takes a value to cast, which must be a
5325 scalar or vector :ref:`floating point <t_floating>` value, and a type to
5326 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
5327 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
5328 type with the same number of elements as ``ty``
5333 The '``fptoui``' instruction converts its :ref:`floating
5334 point <t_floating>` operand into the nearest (rounding towards zero)
5335 unsigned integer value. If the value cannot fit in ``ty2``, the results
5341 .. code-block:: llvm
5343 %X = fptoui double 123.0 to i32 ; yields i32:123
5344 %Y = fptoui float 1.0E+300 to i1 ; yields undefined:1
5345 %Z = fptoui float 1.04E+17 to i8 ; yields undefined:1
5347 '``fptosi .. to``' Instruction
5348 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5355 <result> = fptosi <ty> <value> to <ty2> ; yields ty2
5360 The '``fptosi``' instruction converts :ref:`floating point <t_floating>`
5361 ``value`` to type ``ty2``.
5366 The '``fptosi``' instruction takes a value to cast, which must be a
5367 scalar or vector :ref:`floating point <t_floating>` value, and a type to
5368 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
5369 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
5370 type with the same number of elements as ``ty``
5375 The '``fptosi``' instruction converts its :ref:`floating
5376 point <t_floating>` operand into the nearest (rounding towards zero)
5377 signed integer value. If the value cannot fit in ``ty2``, the results
5383 .. code-block:: llvm
5385 %X = fptosi double -123.0 to i32 ; yields i32:-123
5386 %Y = fptosi float 1.0E-247 to i1 ; yields undefined:1
5387 %Z = fptosi float 1.04E+17 to i8 ; yields undefined:1
5389 '``uitofp .. to``' Instruction
5390 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5397 <result> = uitofp <ty> <value> to <ty2> ; yields ty2
5402 The '``uitofp``' instruction regards ``value`` as an unsigned integer
5403 and converts that value to the ``ty2`` type.
5408 The '``uitofp``' instruction takes a value to cast, which must be a
5409 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
5410 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
5411 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
5412 type with the same number of elements as ``ty``
5417 The '``uitofp``' instruction interprets its operand as an unsigned
5418 integer quantity and converts it to the corresponding floating point
5419 value. If the value cannot fit in the floating point value, the results
5425 .. code-block:: llvm
5427 %X = uitofp i32 257 to float ; yields float:257.0
5428 %Y = uitofp i8 -1 to double ; yields double:255.0
5430 '``sitofp .. to``' Instruction
5431 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5438 <result> = sitofp <ty> <value> to <ty2> ; yields ty2
5443 The '``sitofp``' instruction regards ``value`` as a signed integer and
5444 converts that value to the ``ty2`` type.
5449 The '``sitofp``' instruction takes a value to cast, which must be a
5450 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
5451 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
5452 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
5453 type with the same number of elements as ``ty``
5458 The '``sitofp``' instruction interprets its operand as a signed integer
5459 quantity and converts it to the corresponding floating point value. If
5460 the value cannot fit in the floating point value, the results are
5466 .. code-block:: llvm
5468 %X = sitofp i32 257 to float ; yields float:257.0
5469 %Y = sitofp i8 -1 to double ; yields double:-1.0
5473 '``ptrtoint .. to``' Instruction
5474 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5481 <result> = ptrtoint <ty> <value> to <ty2> ; yields ty2
5486 The '``ptrtoint``' instruction converts the pointer or a vector of
5487 pointers ``value`` to the integer (or vector of integers) type ``ty2``.
5492 The '``ptrtoint``' instruction takes a ``value`` to cast, which must be
5493 a a value of type :ref:`pointer <t_pointer>` or a vector of pointers, and a
5494 type to cast it to ``ty2``, which must be an :ref:`integer <t_integer>` or
5495 a vector of integers type.
5500 The '``ptrtoint``' instruction converts ``value`` to integer type
5501 ``ty2`` by interpreting the pointer value as an integer and either
5502 truncating or zero extending that value to the size of the integer type.
5503 If ``value`` is smaller than ``ty2`` then a zero extension is done. If
5504 ``value`` is larger than ``ty2`` then a truncation is done. If they are
5505 the same size, then nothing is done (*no-op cast*) other than a type
5511 .. code-block:: llvm
5513 %X = ptrtoint i32* %P to i8 ; yields truncation on 32-bit architecture
5514 %Y = ptrtoint i32* %P to i64 ; yields zero extension on 32-bit architecture
5515 %Z = ptrtoint <4 x i32*> %P to <4 x i64>; yields vector zero extension for a vector of addresses on 32-bit architecture
5519 '``inttoptr .. to``' Instruction
5520 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5527 <result> = inttoptr <ty> <value> to <ty2> ; yields ty2
5532 The '``inttoptr``' instruction converts an integer ``value`` to a
5533 pointer type, ``ty2``.
5538 The '``inttoptr``' instruction takes an :ref:`integer <t_integer>` value to
5539 cast, and a type to cast it to, which must be a :ref:`pointer <t_pointer>`
5545 The '``inttoptr``' instruction converts ``value`` to type ``ty2`` by
5546 applying either a zero extension or a truncation depending on the size
5547 of the integer ``value``. If ``value`` is larger than the size of a
5548 pointer then a truncation is done. If ``value`` is smaller than the size
5549 of a pointer then a zero extension is done. If they are the same size,
5550 nothing is done (*no-op cast*).
5555 .. code-block:: llvm
5557 %X = inttoptr i32 255 to i32* ; yields zero extension on 64-bit architecture
5558 %Y = inttoptr i32 255 to i32* ; yields no-op on 32-bit architecture
5559 %Z = inttoptr i64 0 to i32* ; yields truncation on 32-bit architecture
5560 %Z = inttoptr <4 x i32> %G to <4 x i8*>; yields truncation of vector G to four pointers
5564 '``bitcast .. to``' Instruction
5565 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5572 <result> = bitcast <ty> <value> to <ty2> ; yields ty2
5577 The '``bitcast``' instruction converts ``value`` to type ``ty2`` without
5583 The '``bitcast``' instruction takes a value to cast, which must be a
5584 non-aggregate first class value, and a type to cast it to, which must
5585 also be a non-aggregate :ref:`first class <t_firstclass>` type. The bit
5586 sizes of ``value`` and the destination type, ``ty2``, must be identical.
5587 If the source type is a pointer, the destination type must also be a
5588 pointer. This instruction supports bitwise conversion of vectors to
5589 integers and to vectors of other types (as long as they have the same
5595 The '``bitcast``' instruction converts ``value`` to type ``ty2``. It is
5596 always a *no-op cast* because no bits change with this conversion. The
5597 conversion is done as if the ``value`` had been stored to memory and
5598 read back as type ``ty2``. Pointer (or vector of pointers) types may
5599 only be converted to other pointer (or vector of pointers) types with
5600 this instruction. To convert pointers to other types, use the
5601 :ref:`inttoptr <i_inttoptr>` or :ref:`ptrtoint <i_ptrtoint>` instructions
5607 .. code-block:: llvm
5609 %X = bitcast i8 255 to i8 ; yields i8 :-1
5610 %Y = bitcast i32* %x to sint* ; yields sint*:%x
5611 %Z = bitcast <2 x int> %V to i64; ; yields i64: %V
5612 %Z = bitcast <2 x i32*> %V to <2 x i64*> ; yields <2 x i64*>
5619 The instructions in this category are the "miscellaneous" instructions,
5620 which defy better classification.
5624 '``icmp``' Instruction
5625 ^^^^^^^^^^^^^^^^^^^^^^
5632 <result> = icmp <cond> <ty> <op1>, <op2> ; yields {i1} or {<N x i1>}:result
5637 The '``icmp``' instruction returns a boolean value or a vector of
5638 boolean values based on comparison of its two integer, integer vector,
5639 pointer, or pointer vector operands.
5644 The '``icmp``' instruction takes three operands. The first operand is
5645 the condition code indicating the kind of comparison to perform. It is
5646 not a value, just a keyword. The possible condition code are:
5649 #. ``ne``: not equal
5650 #. ``ugt``: unsigned greater than
5651 #. ``uge``: unsigned greater or equal
5652 #. ``ult``: unsigned less than
5653 #. ``ule``: unsigned less or equal
5654 #. ``sgt``: signed greater than
5655 #. ``sge``: signed greater or equal
5656 #. ``slt``: signed less than
5657 #. ``sle``: signed less or equal
5659 The remaining two arguments must be :ref:`integer <t_integer>` or
5660 :ref:`pointer <t_pointer>` or integer :ref:`vector <t_vector>` typed. They
5661 must also be identical types.
5666 The '``icmp``' compares ``op1`` and ``op2`` according to the condition
5667 code given as ``cond``. The comparison performed always yields either an
5668 :ref:`i1 <t_integer>` or vector of ``i1`` result, as follows:
5670 #. ``eq``: yields ``true`` if the operands are equal, ``false``
5671 otherwise. No sign interpretation is necessary or performed.
5672 #. ``ne``: yields ``true`` if the operands are unequal, ``false``
5673 otherwise. No sign interpretation is necessary or performed.
5674 #. ``ugt``: interprets the operands as unsigned values and yields
5675 ``true`` if ``op1`` is greater than ``op2``.
5676 #. ``uge``: interprets the operands as unsigned values and yields
5677 ``true`` if ``op1`` is greater than or equal to ``op2``.
5678 #. ``ult``: interprets the operands as unsigned values and yields
5679 ``true`` if ``op1`` is less than ``op2``.
5680 #. ``ule``: interprets the operands as unsigned values and yields
5681 ``true`` if ``op1`` is less than or equal to ``op2``.
5682 #. ``sgt``: interprets the operands as signed values and yields ``true``
5683 if ``op1`` is greater than ``op2``.
5684 #. ``sge``: interprets the operands as signed values and yields ``true``
5685 if ``op1`` is greater than or equal to ``op2``.
5686 #. ``slt``: interprets the operands as signed values and yields ``true``
5687 if ``op1`` is less than ``op2``.
5688 #. ``sle``: interprets the operands as signed values and yields ``true``
5689 if ``op1`` is less than or equal to ``op2``.
5691 If the operands are :ref:`pointer <t_pointer>` typed, the pointer values
5692 are compared as if they were integers.
5694 If the operands are integer vectors, then they are compared element by
5695 element. The result is an ``i1`` vector with the same number of elements
5696 as the values being compared. Otherwise, the result is an ``i1``.
5701 .. code-block:: llvm
5703 <result> = icmp eq i32 4, 5 ; yields: result=false
5704 <result> = icmp ne float* %X, %X ; yields: result=false
5705 <result> = icmp ult i16 4, 5 ; yields: result=true
5706 <result> = icmp sgt i16 4, 5 ; yields: result=false
5707 <result> = icmp ule i16 -4, 5 ; yields: result=false
5708 <result> = icmp sge i16 4, 5 ; yields: result=false
5710 Note that the code generator does not yet support vector types with the
5711 ``icmp`` instruction.
5715 '``fcmp``' Instruction
5716 ^^^^^^^^^^^^^^^^^^^^^^
5723 <result> = fcmp <cond> <ty> <op1>, <op2> ; yields {i1} or {<N x i1>}:result
5728 The '``fcmp``' instruction returns a boolean value or vector of boolean
5729 values based on comparison of its operands.
5731 If the operands are floating point scalars, then the result type is a
5732 boolean (:ref:`i1 <t_integer>`).
5734 If the operands are floating point vectors, then the result type is a
5735 vector of boolean with the same number of elements as the operands being
5741 The '``fcmp``' instruction takes three operands. The first operand is
5742 the condition code indicating the kind of comparison to perform. It is
5743 not a value, just a keyword. The possible condition code are:
5745 #. ``false``: no comparison, always returns false
5746 #. ``oeq``: ordered and equal
5747 #. ``ogt``: ordered and greater than
5748 #. ``oge``: ordered and greater than or equal
5749 #. ``olt``: ordered and less than
5750 #. ``ole``: ordered and less than or equal
5751 #. ``one``: ordered and not equal
5752 #. ``ord``: ordered (no nans)
5753 #. ``ueq``: unordered or equal
5754 #. ``ugt``: unordered or greater than
5755 #. ``uge``: unordered or greater than or equal
5756 #. ``ult``: unordered or less than
5757 #. ``ule``: unordered or less than or equal
5758 #. ``une``: unordered or not equal
5759 #. ``uno``: unordered (either nans)
5760 #. ``true``: no comparison, always returns true
5762 *Ordered* means that neither operand is a QNAN while *unordered* means
5763 that either operand may be a QNAN.
5765 Each of ``val1`` and ``val2`` arguments must be either a :ref:`floating
5766 point <t_floating>` type or a :ref:`vector <t_vector>` of floating point
5767 type. They must have identical types.
5772 The '``fcmp``' instruction compares ``op1`` and ``op2`` according to the
5773 condition code given as ``cond``. If the operands are vectors, then the
5774 vectors are compared element by element. Each comparison performed
5775 always yields an :ref:`i1 <t_integer>` result, as follows:
5777 #. ``false``: always yields ``false``, regardless of operands.
5778 #. ``oeq``: yields ``true`` if both operands are not a QNAN and ``op1``
5779 is equal to ``op2``.
5780 #. ``ogt``: yields ``true`` if both operands are not a QNAN and ``op1``
5781 is greater than ``op2``.
5782 #. ``oge``: yields ``true`` if both operands are not a QNAN and ``op1``
5783 is greater than or equal to ``op2``.
5784 #. ``olt``: yields ``true`` if both operands are not a QNAN and ``op1``
5785 is less than ``op2``.
5786 #. ``ole``: yields ``true`` if both operands are not a QNAN and ``op1``
5787 is less than or equal to ``op2``.
5788 #. ``one``: yields ``true`` if both operands are not a QNAN and ``op1``
5789 is not equal to ``op2``.
5790 #. ``ord``: yields ``true`` if both operands are not a QNAN.
5791 #. ``ueq``: yields ``true`` if either operand is a QNAN or ``op1`` is
5793 #. ``ugt``: yields ``true`` if either operand is a QNAN or ``op1`` is
5794 greater than ``op2``.
5795 #. ``uge``: yields ``true`` if either operand is a QNAN or ``op1`` is
5796 greater than or equal to ``op2``.
5797 #. ``ult``: yields ``true`` if either operand is a QNAN or ``op1`` is
5799 #. ``ule``: yields ``true`` if either operand is a QNAN or ``op1`` is
5800 less than or equal to ``op2``.
5801 #. ``une``: yields ``true`` if either operand is a QNAN or ``op1`` is
5802 not equal to ``op2``.
5803 #. ``uno``: yields ``true`` if either operand is a QNAN.
5804 #. ``true``: always yields ``true``, regardless of operands.
5809 .. code-block:: llvm
5811 <result> = fcmp oeq float 4.0, 5.0 ; yields: result=false
5812 <result> = fcmp one float 4.0, 5.0 ; yields: result=true
5813 <result> = fcmp olt float 4.0, 5.0 ; yields: result=true
5814 <result> = fcmp ueq double 1.0, 2.0 ; yields: result=false
5816 Note that the code generator does not yet support vector types with the
5817 ``fcmp`` instruction.
5821 '``phi``' Instruction
5822 ^^^^^^^^^^^^^^^^^^^^^
5829 <result> = phi <ty> [ <val0>, <label0>], ...
5834 The '``phi``' instruction is used to implement the φ node in the SSA
5835 graph representing the function.
5840 The type of the incoming values is specified with the first type field.
5841 After this, the '``phi``' instruction takes a list of pairs as
5842 arguments, with one pair for each predecessor basic block of the current
5843 block. Only values of :ref:`first class <t_firstclass>` type may be used as
5844 the value arguments to the PHI node. Only labels may be used as the
5847 There must be no non-phi instructions between the start of a basic block
5848 and the PHI instructions: i.e. PHI instructions must be first in a basic
5851 For the purposes of the SSA form, the use of each incoming value is
5852 deemed to occur on the edge from the corresponding predecessor block to
5853 the current block (but after any definition of an '``invoke``'
5854 instruction's return value on the same edge).
5859 At runtime, the '``phi``' instruction logically takes on the value
5860 specified by the pair corresponding to the predecessor basic block that
5861 executed just prior to the current block.
5866 .. code-block:: llvm
5868 Loop: ; Infinite loop that counts from 0 on up...
5869 %indvar = phi i32 [ 0, %LoopHeader ], [ %nextindvar, %Loop ]
5870 %nextindvar = add i32 %indvar, 1
5875 '``select``' Instruction
5876 ^^^^^^^^^^^^^^^^^^^^^^^^
5883 <result> = select selty <cond>, <ty> <val1>, <ty> <val2> ; yields ty
5885 selty is either i1 or {<N x i1>}
5890 The '``select``' instruction is used to choose one value based on a
5891 condition, without branching.
5896 The '``select``' instruction requires an 'i1' value or a vector of 'i1'
5897 values indicating the condition, and two values of the same :ref:`first
5898 class <t_firstclass>` type. If the val1/val2 are vectors and the
5899 condition is a scalar, then entire vectors are selected, not individual
5905 If the condition is an i1 and it evaluates to 1, the instruction returns
5906 the first value argument; otherwise, it returns the second value
5909 If the condition is a vector of i1, then the value arguments must be
5910 vectors of the same size, and the selection is done element by element.
5915 .. code-block:: llvm
5917 %X = select i1 true, i8 17, i8 42 ; yields i8:17
5921 '``call``' Instruction
5922 ^^^^^^^^^^^^^^^^^^^^^^
5929 <result> = [tail] call [cconv] [ret attrs] <ty> [<fnty>*] <fnptrval>(<function args>) [fn attrs]
5934 The '``call``' instruction represents a simple function call.
5939 This instruction requires several arguments:
5941 #. The optional "tail" marker indicates that the callee function does
5942 not access any allocas or varargs in the caller. Note that calls may
5943 be marked "tail" even if they do not occur before a
5944 :ref:`ret <i_ret>` instruction. If the "tail" marker is present, the
5945 function call is eligible for tail call optimization, but `might not
5946 in fact be optimized into a jump <CodeGenerator.html#tailcallopt>`_.
5947 The code generator may optimize calls marked "tail" with either 1)
5948 automatic `sibling call
5949 optimization <CodeGenerator.html#sibcallopt>`_ when the caller and
5950 callee have matching signatures, or 2) forced tail call optimization
5951 when the following extra requirements are met:
5953 - Caller and callee both have the calling convention ``fastcc``.
5954 - The call is in tail position (ret immediately follows call and ret
5955 uses value of call or is void).
5956 - Option ``-tailcallopt`` is enabled, or
5957 ``llvm::GuaranteedTailCallOpt`` is ``true``.
5958 - `Platform specific constraints are
5959 met. <CodeGenerator.html#tailcallopt>`_
5961 #. The optional "cconv" marker indicates which :ref:`calling
5962 convention <callingconv>` the call should use. If none is
5963 specified, the call defaults to using C calling conventions. The
5964 calling convention of the call must match the calling convention of
5965 the target function, or else the behavior is undefined.
5966 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
5967 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
5969 #. '``ty``': the type of the call instruction itself which is also the
5970 type of the return value. Functions that return no value are marked
5972 #. '``fnty``': shall be the signature of the pointer to function value
5973 being invoked. The argument types must match the types implied by
5974 this signature. This type can be omitted if the function is not
5975 varargs and if the function type does not return a pointer to a
5977 #. '``fnptrval``': An LLVM value containing a pointer to a function to
5978 be invoked. In most cases, this is a direct function invocation, but
5979 indirect ``call``'s are just as possible, calling an arbitrary pointer
5981 #. '``function args``': argument list whose types match the function
5982 signature argument types and parameter attributes. All arguments must
5983 be of :ref:`first class <t_firstclass>` type. If the function signature
5984 indicates the function accepts a variable number of arguments, the
5985 extra arguments can be specified.
5986 #. The optional :ref:`function attributes <fnattrs>` list. Only
5987 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
5988 attributes are valid here.
5993 The '``call``' instruction is used to cause control flow to transfer to
5994 a specified function, with its incoming arguments bound to the specified
5995 values. Upon a '``ret``' instruction in the called function, control
5996 flow continues with the instruction after the function call, and the
5997 return value of the function is bound to the result argument.
6002 .. code-block:: llvm
6004 %retval = call i32 @test(i32 %argc)
6005 call i32 (i8*, ...)* @printf(i8* %msg, i32 12, i8 42) ; yields i32
6006 %X = tail call i32 @foo() ; yields i32
6007 %Y = tail call fastcc i32 @foo() ; yields i32
6008 call void %foo(i8 97 signext)
6010 %struct.A = type { i32, i8 }
6011 %r = call %struct.A @foo() ; yields { 32, i8 }
6012 %gr = extractvalue %struct.A %r, 0 ; yields i32
6013 %gr1 = extractvalue %struct.A %r, 1 ; yields i8
6014 %Z = call void @foo() noreturn ; indicates that %foo never returns normally
6015 %ZZ = call zeroext i32 @bar() ; Return value is %zero extended
6017 llvm treats calls to some functions with names and arguments that match
6018 the standard C99 library as being the C99 library functions, and may
6019 perform optimizations or generate code for them under that assumption.
6020 This is something we'd like to change in the future to provide better
6021 support for freestanding environments and non-C-based languages.
6025 '``va_arg``' Instruction
6026 ^^^^^^^^^^^^^^^^^^^^^^^^
6033 <resultval> = va_arg <va_list*> <arglist>, <argty>
6038 The '``va_arg``' instruction is used to access arguments passed through
6039 the "variable argument" area of a function call. It is used to implement
6040 the ``va_arg`` macro in C.
6045 This instruction takes a ``va_list*`` value and the type of the
6046 argument. It returns a value of the specified argument type and
6047 increments the ``va_list`` to point to the next argument. The actual
6048 type of ``va_list`` is target specific.
6053 The '``va_arg``' instruction loads an argument of the specified type
6054 from the specified ``va_list`` and causes the ``va_list`` to point to
6055 the next argument. For more information, see the variable argument
6056 handling :ref:`Intrinsic Functions <int_varargs>`.
6058 It is legal for this instruction to be called in a function which does
6059 not take a variable number of arguments, for example, the ``vfprintf``
6062 ``va_arg`` is an LLVM instruction instead of an :ref:`intrinsic
6063 function <intrinsics>` because it takes a type as an argument.
6068 See the :ref:`variable argument processing <int_varargs>` section.
6070 Note that the code generator does not yet fully support va\_arg on many
6071 targets. Also, it does not currently support va\_arg with aggregate
6072 types on any target.
6076 '``landingpad``' Instruction
6077 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6084 <resultval> = landingpad <resultty> personality <type> <pers_fn> <clause>+
6085 <resultval> = landingpad <resultty> personality <type> <pers_fn> cleanup <clause>*
6087 <clause> := catch <type> <value>
6088 <clause> := filter <array constant type> <array constant>
6093 The '``landingpad``' instruction is used by `LLVM's exception handling
6094 system <ExceptionHandling.html#overview>`_ to specify that a basic block
6095 is a landing pad --- one where the exception lands, and corresponds to the
6096 code found in the ``catch`` portion of a ``try``/``catch`` sequence. It
6097 defines values supplied by the personality function (``pers_fn``) upon
6098 re-entry to the function. The ``resultval`` has the type ``resultty``.
6103 This instruction takes a ``pers_fn`` value. This is the personality
6104 function associated with the unwinding mechanism. The optional
6105 ``cleanup`` flag indicates that the landing pad block is a cleanup.
6107 A ``clause`` begins with the clause type --- ``catch`` or ``filter`` --- and
6108 contains the global variable representing the "type" that may be caught
6109 or filtered respectively. Unlike the ``catch`` clause, the ``filter``
6110 clause takes an array constant as its argument. Use
6111 "``[0 x i8**] undef``" for a filter which cannot throw. The
6112 '``landingpad``' instruction must contain *at least* one ``clause`` or
6113 the ``cleanup`` flag.
6118 The '``landingpad``' instruction defines the values which are set by the
6119 personality function (``pers_fn``) upon re-entry to the function, and
6120 therefore the "result type" of the ``landingpad`` instruction. As with
6121 calling conventions, how the personality function results are
6122 represented in LLVM IR is target specific.
6124 The clauses are applied in order from top to bottom. If two
6125 ``landingpad`` instructions are merged together through inlining, the
6126 clauses from the calling function are appended to the list of clauses.
6127 When the call stack is being unwound due to an exception being thrown,
6128 the exception is compared against each ``clause`` in turn. If it doesn't
6129 match any of the clauses, and the ``cleanup`` flag is not set, then
6130 unwinding continues further up the call stack.
6132 The ``landingpad`` instruction has several restrictions:
6134 - A landing pad block is a basic block which is the unwind destination
6135 of an '``invoke``' instruction.
6136 - A landing pad block must have a '``landingpad``' instruction as its
6137 first non-PHI instruction.
6138 - There can be only one '``landingpad``' instruction within the landing
6140 - A basic block that is not a landing pad block may not include a
6141 '``landingpad``' instruction.
6142 - All '``landingpad``' instructions in a function must have the same
6143 personality function.
6148 .. code-block:: llvm
6150 ;; A landing pad which can catch an integer.
6151 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6153 ;; A landing pad that is a cleanup.
6154 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6156 ;; A landing pad which can catch an integer and can only throw a double.
6157 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6159 filter [1 x i8**] [@_ZTId]
6166 LLVM supports the notion of an "intrinsic function". These functions
6167 have well known names and semantics and are required to follow certain
6168 restrictions. Overall, these intrinsics represent an extension mechanism
6169 for the LLVM language that does not require changing all of the
6170 transformations in LLVM when adding to the language (or the bitcode
6171 reader/writer, the parser, etc...).
6173 Intrinsic function names must all start with an "``llvm.``" prefix. This
6174 prefix is reserved in LLVM for intrinsic names; thus, function names may
6175 not begin with this prefix. Intrinsic functions must always be external
6176 functions: you cannot define the body of intrinsic functions. Intrinsic
6177 functions may only be used in call or invoke instructions: it is illegal
6178 to take the address of an intrinsic function. Additionally, because
6179 intrinsic functions are part of the LLVM language, it is required if any
6180 are added that they be documented here.
6182 Some intrinsic functions can be overloaded, i.e., the intrinsic
6183 represents a family of functions that perform the same operation but on
6184 different data types. Because LLVM can represent over 8 million
6185 different integer types, overloading is used commonly to allow an
6186 intrinsic function to operate on any integer type. One or more of the
6187 argument types or the result type can be overloaded to accept any
6188 integer type. Argument types may also be defined as exactly matching a
6189 previous argument's type or the result type. This allows an intrinsic
6190 function which accepts multiple arguments, but needs all of them to be
6191 of the same type, to only be overloaded with respect to a single
6192 argument or the result.
6194 Overloaded intrinsics will have the names of its overloaded argument
6195 types encoded into its function name, each preceded by a period. Only
6196 those types which are overloaded result in a name suffix. Arguments
6197 whose type is matched against another type do not. For example, the
6198 ``llvm.ctpop`` function can take an integer of any width and returns an
6199 integer of exactly the same integer width. This leads to a family of
6200 functions such as ``i8 @llvm.ctpop.i8(i8 %val)`` and
6201 ``i29 @llvm.ctpop.i29(i29 %val)``. Only one type, the return type, is
6202 overloaded, and only one type suffix is required. Because the argument's
6203 type is matched against the return type, it does not require its own
6206 To learn how to add an intrinsic function, please see the `Extending
6207 LLVM Guide <ExtendingLLVM.html>`_.
6211 Variable Argument Handling Intrinsics
6212 -------------------------------------
6214 Variable argument support is defined in LLVM with the
6215 :ref:`va_arg <i_va_arg>` instruction and these three intrinsic
6216 functions. These functions are related to the similarly named macros
6217 defined in the ``<stdarg.h>`` header file.
6219 All of these functions operate on arguments that use a target-specific
6220 value type "``va_list``". The LLVM assembly language reference manual
6221 does not define what this type is, so all transformations should be
6222 prepared to handle these functions regardless of the type used.
6224 This example shows how the :ref:`va_arg <i_va_arg>` instruction and the
6225 variable argument handling intrinsic functions are used.
6227 .. code-block:: llvm
6229 define i32 @test(i32 %X, ...) {
6230 ; Initialize variable argument processing
6232 %ap2 = bitcast i8** %ap to i8*
6233 call void @llvm.va_start(i8* %ap2)
6235 ; Read a single integer argument
6236 %tmp = va_arg i8** %ap, i32
6238 ; Demonstrate usage of llvm.va_copy and llvm.va_end
6240 %aq2 = bitcast i8** %aq to i8*
6241 call void @llvm.va_copy(i8* %aq2, i8* %ap2)
6242 call void @llvm.va_end(i8* %aq2)
6244 ; Stop processing of arguments.
6245 call void @llvm.va_end(i8* %ap2)
6249 declare void @llvm.va_start(i8*)
6250 declare void @llvm.va_copy(i8*, i8*)
6251 declare void @llvm.va_end(i8*)
6255 '``llvm.va_start``' Intrinsic
6256 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6263 declare void %llvm.va_start(i8* <arglist>)
6268 The '``llvm.va_start``' intrinsic initializes ``*<arglist>`` for
6269 subsequent use by ``va_arg``.
6274 The argument is a pointer to a ``va_list`` element to initialize.
6279 The '``llvm.va_start``' intrinsic works just like the ``va_start`` macro
6280 available in C. In a target-dependent way, it initializes the
6281 ``va_list`` element to which the argument points, so that the next call
6282 to ``va_arg`` will produce the first variable argument passed to the
6283 function. Unlike the C ``va_start`` macro, this intrinsic does not need
6284 to know the last argument of the function as the compiler can figure
6287 '``llvm.va_end``' Intrinsic
6288 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6295 declare void @llvm.va_end(i8* <arglist>)
6300 The '``llvm.va_end``' intrinsic destroys ``*<arglist>``, which has been
6301 initialized previously with ``llvm.va_start`` or ``llvm.va_copy``.
6306 The argument is a pointer to a ``va_list`` to destroy.
6311 The '``llvm.va_end``' intrinsic works just like the ``va_end`` macro
6312 available in C. In a target-dependent way, it destroys the ``va_list``
6313 element to which the argument points. Calls to
6314 :ref:`llvm.va_start <int_va_start>` and
6315 :ref:`llvm.va_copy <int_va_copy>` must be matched exactly with calls to
6320 '``llvm.va_copy``' Intrinsic
6321 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6328 declare void @llvm.va_copy(i8* <destarglist>, i8* <srcarglist>)
6333 The '``llvm.va_copy``' intrinsic copies the current argument position
6334 from the source argument list to the destination argument list.
6339 The first argument is a pointer to a ``va_list`` element to initialize.
6340 The second argument is a pointer to a ``va_list`` element to copy from.
6345 The '``llvm.va_copy``' intrinsic works just like the ``va_copy`` macro
6346 available in C. In a target-dependent way, it copies the source
6347 ``va_list`` element into the destination ``va_list`` element. This
6348 intrinsic is necessary because the `` llvm.va_start`` intrinsic may be
6349 arbitrarily complex and require, for example, memory allocation.
6351 Accurate Garbage Collection Intrinsics
6352 --------------------------------------
6354 LLVM support for `Accurate Garbage Collection <GarbageCollection.html>`_
6355 (GC) requires the implementation and generation of these intrinsics.
6356 These intrinsics allow identification of :ref:`GC roots on the
6357 stack <int_gcroot>`, as well as garbage collector implementations that
6358 require :ref:`read <int_gcread>` and :ref:`write <int_gcwrite>` barriers.
6359 Front-ends for type-safe garbage collected languages should generate
6360 these intrinsics to make use of the LLVM garbage collectors. For more
6361 details, see `Accurate Garbage Collection with
6362 LLVM <GarbageCollection.html>`_.
6364 The garbage collection intrinsics only operate on objects in the generic
6365 address space (address space zero).
6369 '``llvm.gcroot``' Intrinsic
6370 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6377 declare void @llvm.gcroot(i8** %ptrloc, i8* %metadata)
6382 The '``llvm.gcroot``' intrinsic declares the existence of a GC root to
6383 the code generator, and allows some metadata to be associated with it.
6388 The first argument specifies the address of a stack object that contains
6389 the root pointer. The second pointer (which must be either a constant or
6390 a global value address) contains the meta-data to be associated with the
6396 At runtime, a call to this intrinsic stores a null pointer into the
6397 "ptrloc" location. At compile-time, the code generator generates
6398 information to allow the runtime to find the pointer at GC safe points.
6399 The '``llvm.gcroot``' intrinsic may only be used in a function which
6400 :ref:`specifies a GC algorithm <gc>`.
6404 '``llvm.gcread``' Intrinsic
6405 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6412 declare i8* @llvm.gcread(i8* %ObjPtr, i8** %Ptr)
6417 The '``llvm.gcread``' intrinsic identifies reads of references from heap
6418 locations, allowing garbage collector implementations that require read
6424 The second argument is the address to read from, which should be an
6425 address allocated from the garbage collector. The first object is a
6426 pointer to the start of the referenced object, if needed by the language
6427 runtime (otherwise null).
6432 The '``llvm.gcread``' intrinsic has the same semantics as a load
6433 instruction, but may be replaced with substantially more complex code by
6434 the garbage collector runtime, as needed. The '``llvm.gcread``'
6435 intrinsic may only be used in a function which :ref:`specifies a GC
6440 '``llvm.gcwrite``' Intrinsic
6441 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6448 declare void @llvm.gcwrite(i8* %P1, i8* %Obj, i8** %P2)
6453 The '``llvm.gcwrite``' intrinsic identifies writes of references to heap
6454 locations, allowing garbage collector implementations that require write
6455 barriers (such as generational or reference counting collectors).
6460 The first argument is the reference to store, the second is the start of
6461 the object to store it to, and the third is the address of the field of
6462 Obj to store to. If the runtime does not require a pointer to the
6463 object, Obj may be null.
6468 The '``llvm.gcwrite``' intrinsic has the same semantics as a store
6469 instruction, but may be replaced with substantially more complex code by
6470 the garbage collector runtime, as needed. The '``llvm.gcwrite``'
6471 intrinsic may only be used in a function which :ref:`specifies a GC
6474 Code Generator Intrinsics
6475 -------------------------
6477 These intrinsics are provided by LLVM to expose special features that
6478 may only be implemented with code generator support.
6480 '``llvm.returnaddress``' Intrinsic
6481 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6488 declare i8 *@llvm.returnaddress(i32 <level>)
6493 The '``llvm.returnaddress``' intrinsic attempts to compute a
6494 target-specific value indicating the return address of the current
6495 function or one of its callers.
6500 The argument to this intrinsic indicates which function to return the
6501 address for. Zero indicates the calling function, one indicates its
6502 caller, etc. The argument is **required** to be a constant integer
6508 The '``llvm.returnaddress``' intrinsic either returns a pointer
6509 indicating the return address of the specified call frame, or zero if it
6510 cannot be identified. The value returned by this intrinsic is likely to
6511 be incorrect or 0 for arguments other than zero, so it should only be
6512 used for debugging purposes.
6514 Note that calling this intrinsic does not prevent function inlining or
6515 other aggressive transformations, so the value returned may not be that
6516 of the obvious source-language caller.
6518 '``llvm.frameaddress``' Intrinsic
6519 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6526 declare i8* @llvm.frameaddress(i32 <level>)
6531 The '``llvm.frameaddress``' intrinsic attempts to return the
6532 target-specific frame pointer value for the specified stack frame.
6537 The argument to this intrinsic indicates which function to return the
6538 frame pointer for. Zero indicates the calling function, one indicates
6539 its caller, etc. The argument is **required** to be a constant integer
6545 The '``llvm.frameaddress``' intrinsic either returns a pointer
6546 indicating the frame address of the specified call frame, or zero if it
6547 cannot be identified. The value returned by this intrinsic is likely to
6548 be incorrect or 0 for arguments other than zero, so it should only be
6549 used for debugging purposes.
6551 Note that calling this intrinsic does not prevent function inlining or
6552 other aggressive transformations, so the value returned may not be that
6553 of the obvious source-language caller.
6557 '``llvm.stacksave``' Intrinsic
6558 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6565 declare i8* @llvm.stacksave()
6570 The '``llvm.stacksave``' intrinsic is used to remember the current state
6571 of the function stack, for use with
6572 :ref:`llvm.stackrestore <int_stackrestore>`. This is useful for
6573 implementing language features like scoped automatic variable sized
6579 This intrinsic returns a opaque pointer value that can be passed to
6580 :ref:`llvm.stackrestore <int_stackrestore>`. When an
6581 ``llvm.stackrestore`` intrinsic is executed with a value saved from
6582 ``llvm.stacksave``, it effectively restores the state of the stack to
6583 the state it was in when the ``llvm.stacksave`` intrinsic executed. In
6584 practice, this pops any :ref:`alloca <i_alloca>` blocks from the stack that
6585 were allocated after the ``llvm.stacksave`` was executed.
6587 .. _int_stackrestore:
6589 '``llvm.stackrestore``' Intrinsic
6590 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6597 declare void @llvm.stackrestore(i8* %ptr)
6602 The '``llvm.stackrestore``' intrinsic is used to restore the state of
6603 the function stack to the state it was in when the corresponding
6604 :ref:`llvm.stacksave <int_stacksave>` intrinsic executed. This is
6605 useful for implementing language features like scoped automatic variable
6606 sized arrays in C99.
6611 See the description for :ref:`llvm.stacksave <int_stacksave>`.
6613 '``llvm.prefetch``' Intrinsic
6614 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6621 declare void @llvm.prefetch(i8* <address>, i32 <rw>, i32 <locality>, i32 <cache type>)
6626 The '``llvm.prefetch``' intrinsic is a hint to the code generator to
6627 insert a prefetch instruction if supported; otherwise, it is a noop.
6628 Prefetches have no effect on the behavior of the program but can change
6629 its performance characteristics.
6634 ``address`` is the address to be prefetched, ``rw`` is the specifier
6635 determining if the fetch should be for a read (0) or write (1), and
6636 ``locality`` is a temporal locality specifier ranging from (0) - no
6637 locality, to (3) - extremely local keep in cache. The ``cache type``
6638 specifies whether the prefetch is performed on the data (1) or
6639 instruction (0) cache. The ``rw``, ``locality`` and ``cache type``
6640 arguments must be constant integers.
6645 This intrinsic does not modify the behavior of the program. In
6646 particular, prefetches cannot trap and do not produce a value. On
6647 targets that support this intrinsic, the prefetch can provide hints to
6648 the processor cache for better performance.
6650 '``llvm.pcmarker``' Intrinsic
6651 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6658 declare void @llvm.pcmarker(i32 <id>)
6663 The '``llvm.pcmarker``' intrinsic is a method to export a Program
6664 Counter (PC) in a region of code to simulators and other tools. The
6665 method is target specific, but it is expected that the marker will use
6666 exported symbols to transmit the PC of the marker. The marker makes no
6667 guarantees that it will remain with any specific instruction after
6668 optimizations. It is possible that the presence of a marker will inhibit
6669 optimizations. The intended use is to be inserted after optimizations to
6670 allow correlations of simulation runs.
6675 ``id`` is a numerical id identifying the marker.
6680 This intrinsic does not modify the behavior of the program. Backends
6681 that do not support this intrinsic may ignore it.
6683 '``llvm.readcyclecounter``' Intrinsic
6684 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6691 declare i64 @llvm.readcyclecounter()
6696 The '``llvm.readcyclecounter``' intrinsic provides access to the cycle
6697 counter register (or similar low latency, high accuracy clocks) on those
6698 targets that support it. On X86, it should map to RDTSC. On Alpha, it
6699 should map to RPCC. As the backing counters overflow quickly (on the
6700 order of 9 seconds on alpha), this should only be used for small
6706 When directly supported, reading the cycle counter should not modify any
6707 memory. Implementations are allowed to either return a application
6708 specific value or a system wide value. On backends without support, this
6709 is lowered to a constant 0.
6711 Note that runtime support may be conditional on the privilege-level code is
6712 running at and the host platform.
6714 Standard C Library Intrinsics
6715 -----------------------------
6717 LLVM provides intrinsics for a few important standard C library
6718 functions. These intrinsics allow source-language front-ends to pass
6719 information about the alignment of the pointer arguments to the code
6720 generator, providing opportunity for more efficient code generation.
6724 '``llvm.memcpy``' Intrinsic
6725 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6730 This is an overloaded intrinsic. You can use ``llvm.memcpy`` on any
6731 integer bit width and for different address spaces. Not all targets
6732 support all bit widths however.
6736 declare void @llvm.memcpy.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
6737 i32 <len>, i32 <align>, i1 <isvolatile>)
6738 declare void @llvm.memcpy.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
6739 i64 <len>, i32 <align>, i1 <isvolatile>)
6744 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
6745 source location to the destination location.
6747 Note that, unlike the standard libc function, the ``llvm.memcpy.*``
6748 intrinsics do not return a value, takes extra alignment/isvolatile
6749 arguments and the pointers can be in specified address spaces.
6754 The first argument is a pointer to the destination, the second is a
6755 pointer to the source. The third argument is an integer argument
6756 specifying the number of bytes to copy, the fourth argument is the
6757 alignment of the source and destination locations, and the fifth is a
6758 boolean indicating a volatile access.
6760 If the call to this intrinsic has an alignment value that is not 0 or 1,
6761 then the caller guarantees that both the source and destination pointers
6762 are aligned to that boundary.
6764 If the ``isvolatile`` parameter is ``true``, the ``llvm.memcpy`` call is
6765 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
6766 very cleanly specified and it is unwise to depend on it.
6771 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
6772 source location to the destination location, which are not allowed to
6773 overlap. It copies "len" bytes of memory over. If the argument is known
6774 to be aligned to some boundary, this can be specified as the fourth
6775 argument, otherwise it should be set to 0 or 1.
6777 '``llvm.memmove``' Intrinsic
6778 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6783 This is an overloaded intrinsic. You can use llvm.memmove on any integer
6784 bit width and for different address space. Not all targets support all
6789 declare void @llvm.memmove.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
6790 i32 <len>, i32 <align>, i1 <isvolatile>)
6791 declare void @llvm.memmove.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
6792 i64 <len>, i32 <align>, i1 <isvolatile>)
6797 The '``llvm.memmove.*``' intrinsics move a block of memory from the
6798 source location to the destination location. It is similar to the
6799 '``llvm.memcpy``' intrinsic but allows the two memory locations to
6802 Note that, unlike the standard libc function, the ``llvm.memmove.*``
6803 intrinsics do not return a value, takes extra alignment/isvolatile
6804 arguments and the pointers can be in specified address spaces.
6809 The first argument is a pointer to the destination, the second is a
6810 pointer to the source. The third argument is an integer argument
6811 specifying the number of bytes to copy, the fourth argument is the
6812 alignment of the source and destination locations, and the fifth is a
6813 boolean indicating a volatile access.
6815 If the call to this intrinsic has an alignment value that is not 0 or 1,
6816 then the caller guarantees that the source and destination pointers are
6817 aligned to that boundary.
6819 If the ``isvolatile`` parameter is ``true``, the ``llvm.memmove`` call
6820 is a :ref:`volatile operation <volatile>`. The detailed access behavior is
6821 not very cleanly specified and it is unwise to depend on it.
6826 The '``llvm.memmove.*``' intrinsics copy a block of memory from the
6827 source location to the destination location, which may overlap. It
6828 copies "len" bytes of memory over. If the argument is known to be
6829 aligned to some boundary, this can be specified as the fourth argument,
6830 otherwise it should be set to 0 or 1.
6832 '``llvm.memset.*``' Intrinsics
6833 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6838 This is an overloaded intrinsic. You can use llvm.memset on any integer
6839 bit width and for different address spaces. However, not all targets
6840 support all bit widths.
6844 declare void @llvm.memset.p0i8.i32(i8* <dest>, i8 <val>,
6845 i32 <len>, i32 <align>, i1 <isvolatile>)
6846 declare void @llvm.memset.p0i8.i64(i8* <dest>, i8 <val>,
6847 i64 <len>, i32 <align>, i1 <isvolatile>)
6852 The '``llvm.memset.*``' intrinsics fill a block of memory with a
6853 particular byte value.
6855 Note that, unlike the standard libc function, the ``llvm.memset``
6856 intrinsic does not return a value and takes extra alignment/volatile
6857 arguments. Also, the destination can be in an arbitrary address space.
6862 The first argument is a pointer to the destination to fill, the second
6863 is the byte value with which to fill it, the third argument is an
6864 integer argument specifying the number of bytes to fill, and the fourth
6865 argument is the known alignment of the destination location.
6867 If the call to this intrinsic has an alignment value that is not 0 or 1,
6868 then the caller guarantees that the destination pointer is aligned to
6871 If the ``isvolatile`` parameter is ``true``, the ``llvm.memset`` call is
6872 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
6873 very cleanly specified and it is unwise to depend on it.
6878 The '``llvm.memset.*``' intrinsics fill "len" bytes of memory starting
6879 at the destination location. If the argument is known to be aligned to
6880 some boundary, this can be specified as the fourth argument, otherwise
6881 it should be set to 0 or 1.
6883 '``llvm.sqrt.*``' Intrinsic
6884 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6889 This is an overloaded intrinsic. You can use ``llvm.sqrt`` on any
6890 floating point or vector of floating point type. Not all targets support
6895 declare float @llvm.sqrt.f32(float %Val)
6896 declare double @llvm.sqrt.f64(double %Val)
6897 declare x86_fp80 @llvm.sqrt.f80(x86_fp80 %Val)
6898 declare fp128 @llvm.sqrt.f128(fp128 %Val)
6899 declare ppc_fp128 @llvm.sqrt.ppcf128(ppc_fp128 %Val)
6904 The '``llvm.sqrt``' intrinsics return the sqrt of the specified operand,
6905 returning the same value as the libm '``sqrt``' functions would. Unlike
6906 ``sqrt`` in libm, however, ``llvm.sqrt`` has undefined behavior for
6907 negative numbers other than -0.0 (which allows for better optimization,
6908 because there is no need to worry about errno being set).
6909 ``llvm.sqrt(-0.0)`` is defined to return -0.0 like IEEE sqrt.
6914 The argument and return value are floating point numbers of the same
6920 This function returns the sqrt of the specified operand if it is a
6921 nonnegative floating point number.
6923 '``llvm.powi.*``' Intrinsic
6924 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6929 This is an overloaded intrinsic. You can use ``llvm.powi`` on any
6930 floating point or vector of floating point type. Not all targets support
6935 declare float @llvm.powi.f32(float %Val, i32 %power)
6936 declare double @llvm.powi.f64(double %Val, i32 %power)
6937 declare x86_fp80 @llvm.powi.f80(x86_fp80 %Val, i32 %power)
6938 declare fp128 @llvm.powi.f128(fp128 %Val, i32 %power)
6939 declare ppc_fp128 @llvm.powi.ppcf128(ppc_fp128 %Val, i32 %power)
6944 The '``llvm.powi.*``' intrinsics return the first operand raised to the
6945 specified (positive or negative) power. The order of evaluation of
6946 multiplications is not defined. When a vector of floating point type is
6947 used, the second argument remains a scalar integer value.
6952 The second argument is an integer power, and the first is a value to
6953 raise to that power.
6958 This function returns the first value raised to the second power with an
6959 unspecified sequence of rounding operations.
6961 '``llvm.sin.*``' Intrinsic
6962 ^^^^^^^^^^^^^^^^^^^^^^^^^^
6967 This is an overloaded intrinsic. You can use ``llvm.sin`` on any
6968 floating point or vector of floating point type. Not all targets support
6973 declare float @llvm.sin.f32(float %Val)
6974 declare double @llvm.sin.f64(double %Val)
6975 declare x86_fp80 @llvm.sin.f80(x86_fp80 %Val)
6976 declare fp128 @llvm.sin.f128(fp128 %Val)
6977 declare ppc_fp128 @llvm.sin.ppcf128(ppc_fp128 %Val)
6982 The '``llvm.sin.*``' intrinsics return the sine of the operand.
6987 The argument and return value are floating point numbers of the same
6993 This function returns the sine of the specified operand, returning the
6994 same values as the libm ``sin`` functions would, and handles error
6995 conditions in the same way.
6997 '``llvm.cos.*``' Intrinsic
6998 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7003 This is an overloaded intrinsic. You can use ``llvm.cos`` on any
7004 floating point or vector of floating point type. Not all targets support
7009 declare float @llvm.cos.f32(float %Val)
7010 declare double @llvm.cos.f64(double %Val)
7011 declare x86_fp80 @llvm.cos.f80(x86_fp80 %Val)
7012 declare fp128 @llvm.cos.f128(fp128 %Val)
7013 declare ppc_fp128 @llvm.cos.ppcf128(ppc_fp128 %Val)
7018 The '``llvm.cos.*``' intrinsics return the cosine of the operand.
7023 The argument and return value are floating point numbers of the same
7029 This function returns the cosine of the specified operand, returning the
7030 same values as the libm ``cos`` functions would, and handles error
7031 conditions in the same way.
7033 '``llvm.pow.*``' Intrinsic
7034 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7039 This is an overloaded intrinsic. You can use ``llvm.pow`` on any
7040 floating point or vector of floating point type. Not all targets support
7045 declare float @llvm.pow.f32(float %Val, float %Power)
7046 declare double @llvm.pow.f64(double %Val, double %Power)
7047 declare x86_fp80 @llvm.pow.f80(x86_fp80 %Val, x86_fp80 %Power)
7048 declare fp128 @llvm.pow.f128(fp128 %Val, fp128 %Power)
7049 declare ppc_fp128 @llvm.pow.ppcf128(ppc_fp128 %Val, ppc_fp128 Power)
7054 The '``llvm.pow.*``' intrinsics return the first operand raised to the
7055 specified (positive or negative) power.
7060 The second argument is a floating point power, and the first is a value
7061 to raise to that power.
7066 This function returns the first value raised to the second power,
7067 returning the same values as the libm ``pow`` functions would, and
7068 handles error conditions in the same way.
7070 '``llvm.exp.*``' Intrinsic
7071 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7076 This is an overloaded intrinsic. You can use ``llvm.exp`` on any
7077 floating point or vector of floating point type. Not all targets support
7082 declare float @llvm.exp.f32(float %Val)
7083 declare double @llvm.exp.f64(double %Val)
7084 declare x86_fp80 @llvm.exp.f80(x86_fp80 %Val)
7085 declare fp128 @llvm.exp.f128(fp128 %Val)
7086 declare ppc_fp128 @llvm.exp.ppcf128(ppc_fp128 %Val)
7091 The '``llvm.exp.*``' intrinsics perform the exp function.
7096 The argument and return value are floating point numbers of the same
7102 This function returns the same values as the libm ``exp`` functions
7103 would, and handles error conditions in the same way.
7105 '``llvm.exp2.*``' Intrinsic
7106 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7111 This is an overloaded intrinsic. You can use ``llvm.exp2`` on any
7112 floating point or vector of floating point type. Not all targets support
7117 declare float @llvm.exp2.f32(float %Val)
7118 declare double @llvm.exp2.f64(double %Val)
7119 declare x86_fp80 @llvm.exp2.f80(x86_fp80 %Val)
7120 declare fp128 @llvm.exp2.f128(fp128 %Val)
7121 declare ppc_fp128 @llvm.exp2.ppcf128(ppc_fp128 %Val)
7126 The '``llvm.exp2.*``' intrinsics perform the exp2 function.
7131 The argument and return value are floating point numbers of the same
7137 This function returns the same values as the libm ``exp2`` functions
7138 would, and handles error conditions in the same way.
7140 '``llvm.log.*``' Intrinsic
7141 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7146 This is an overloaded intrinsic. You can use ``llvm.log`` on any
7147 floating point or vector of floating point type. Not all targets support
7152 declare float @llvm.log.f32(float %Val)
7153 declare double @llvm.log.f64(double %Val)
7154 declare x86_fp80 @llvm.log.f80(x86_fp80 %Val)
7155 declare fp128 @llvm.log.f128(fp128 %Val)
7156 declare ppc_fp128 @llvm.log.ppcf128(ppc_fp128 %Val)
7161 The '``llvm.log.*``' intrinsics perform the log function.
7166 The argument and return value are floating point numbers of the same
7172 This function returns the same values as the libm ``log`` functions
7173 would, and handles error conditions in the same way.
7175 '``llvm.log10.*``' Intrinsic
7176 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7181 This is an overloaded intrinsic. You can use ``llvm.log10`` on any
7182 floating point or vector of floating point type. Not all targets support
7187 declare float @llvm.log10.f32(float %Val)
7188 declare double @llvm.log10.f64(double %Val)
7189 declare x86_fp80 @llvm.log10.f80(x86_fp80 %Val)
7190 declare fp128 @llvm.log10.f128(fp128 %Val)
7191 declare ppc_fp128 @llvm.log10.ppcf128(ppc_fp128 %Val)
7196 The '``llvm.log10.*``' intrinsics perform the log10 function.
7201 The argument and return value are floating point numbers of the same
7207 This function returns the same values as the libm ``log10`` functions
7208 would, and handles error conditions in the same way.
7210 '``llvm.log2.*``' Intrinsic
7211 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7216 This is an overloaded intrinsic. You can use ``llvm.log2`` on any
7217 floating point or vector of floating point type. Not all targets support
7222 declare float @llvm.log2.f32(float %Val)
7223 declare double @llvm.log2.f64(double %Val)
7224 declare x86_fp80 @llvm.log2.f80(x86_fp80 %Val)
7225 declare fp128 @llvm.log2.f128(fp128 %Val)
7226 declare ppc_fp128 @llvm.log2.ppcf128(ppc_fp128 %Val)
7231 The '``llvm.log2.*``' intrinsics perform the log2 function.
7236 The argument and return value are floating point numbers of the same
7242 This function returns the same values as the libm ``log2`` functions
7243 would, and handles error conditions in the same way.
7245 '``llvm.fma.*``' Intrinsic
7246 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7251 This is an overloaded intrinsic. You can use ``llvm.fma`` on any
7252 floating point or vector of floating point type. Not all targets support
7257 declare float @llvm.fma.f32(float %a, float %b, float %c)
7258 declare double @llvm.fma.f64(double %a, double %b, double %c)
7259 declare x86_fp80 @llvm.fma.f80(x86_fp80 %a, x86_fp80 %b, x86_fp80 %c)
7260 declare fp128 @llvm.fma.f128(fp128 %a, fp128 %b, fp128 %c)
7261 declare ppc_fp128 @llvm.fma.ppcf128(ppc_fp128 %a, ppc_fp128 %b, ppc_fp128 %c)
7266 The '``llvm.fma.*``' intrinsics perform the fused multiply-add
7272 The argument and return value are floating point numbers of the same
7278 This function returns the same values as the libm ``fma`` functions
7281 '``llvm.fabs.*``' Intrinsic
7282 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7287 This is an overloaded intrinsic. You can use ``llvm.fabs`` on any
7288 floating point or vector of floating point type. Not all targets support
7293 declare float @llvm.fabs.f32(float %Val)
7294 declare double @llvm.fabs.f64(double %Val)
7295 declare x86_fp80 @llvm.fabs.f80(x86_fp80 %Val)
7296 declare fp128 @llvm.fabs.f128(fp128 %Val)
7297 declare ppc_fp128 @llvm.fabs.ppcf128(ppc_fp128 %Val)
7302 The '``llvm.fabs.*``' intrinsics return the absolute value of the
7308 The argument and return value are floating point numbers of the same
7314 This function returns the same values as the libm ``fabs`` functions
7315 would, and handles error conditions in the same way.
7317 '``llvm.floor.*``' Intrinsic
7318 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7323 This is an overloaded intrinsic. You can use ``llvm.floor`` on any
7324 floating point or vector of floating point type. Not all targets support
7329 declare float @llvm.floor.f32(float %Val)
7330 declare double @llvm.floor.f64(double %Val)
7331 declare x86_fp80 @llvm.floor.f80(x86_fp80 %Val)
7332 declare fp128 @llvm.floor.f128(fp128 %Val)
7333 declare ppc_fp128 @llvm.floor.ppcf128(ppc_fp128 %Val)
7338 The '``llvm.floor.*``' intrinsics return the floor of the operand.
7343 The argument and return value are floating point numbers of the same
7349 This function returns the same values as the libm ``floor`` functions
7350 would, and handles error conditions in the same way.
7352 '``llvm.ceil.*``' Intrinsic
7353 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7358 This is an overloaded intrinsic. You can use ``llvm.ceil`` on any
7359 floating point or vector of floating point type. Not all targets support
7364 declare float @llvm.ceil.f32(float %Val)
7365 declare double @llvm.ceil.f64(double %Val)
7366 declare x86_fp80 @llvm.ceil.f80(x86_fp80 %Val)
7367 declare fp128 @llvm.ceil.f128(fp128 %Val)
7368 declare ppc_fp128 @llvm.ceil.ppcf128(ppc_fp128 %Val)
7373 The '``llvm.ceil.*``' intrinsics return the ceiling of the operand.
7378 The argument and return value are floating point numbers of the same
7384 This function returns the same values as the libm ``ceil`` functions
7385 would, and handles error conditions in the same way.
7387 '``llvm.trunc.*``' Intrinsic
7388 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7393 This is an overloaded intrinsic. You can use ``llvm.trunc`` on any
7394 floating point or vector of floating point type. Not all targets support
7399 declare float @llvm.trunc.f32(float %Val)
7400 declare double @llvm.trunc.f64(double %Val)
7401 declare x86_fp80 @llvm.trunc.f80(x86_fp80 %Val)
7402 declare fp128 @llvm.trunc.f128(fp128 %Val)
7403 declare ppc_fp128 @llvm.trunc.ppcf128(ppc_fp128 %Val)
7408 The '``llvm.trunc.*``' intrinsics returns the operand rounded to the
7409 nearest integer not larger in magnitude than the operand.
7414 The argument and return value are floating point numbers of the same
7420 This function returns the same values as the libm ``trunc`` functions
7421 would, and handles error conditions in the same way.
7423 '``llvm.rint.*``' Intrinsic
7424 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7429 This is an overloaded intrinsic. You can use ``llvm.rint`` on any
7430 floating point or vector of floating point type. Not all targets support
7435 declare float @llvm.rint.f32(float %Val)
7436 declare double @llvm.rint.f64(double %Val)
7437 declare x86_fp80 @llvm.rint.f80(x86_fp80 %Val)
7438 declare fp128 @llvm.rint.f128(fp128 %Val)
7439 declare ppc_fp128 @llvm.rint.ppcf128(ppc_fp128 %Val)
7444 The '``llvm.rint.*``' intrinsics returns the operand rounded to the
7445 nearest integer. It may raise an inexact floating-point exception if the
7446 operand isn't an integer.
7451 The argument and return value are floating point numbers of the same
7457 This function returns the same values as the libm ``rint`` functions
7458 would, and handles error conditions in the same way.
7460 '``llvm.nearbyint.*``' Intrinsic
7461 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7466 This is an overloaded intrinsic. You can use ``llvm.nearbyint`` on any
7467 floating point or vector of floating point type. Not all targets support
7472 declare float @llvm.nearbyint.f32(float %Val)
7473 declare double @llvm.nearbyint.f64(double %Val)
7474 declare x86_fp80 @llvm.nearbyint.f80(x86_fp80 %Val)
7475 declare fp128 @llvm.nearbyint.f128(fp128 %Val)
7476 declare ppc_fp128 @llvm.nearbyint.ppcf128(ppc_fp128 %Val)
7481 The '``llvm.nearbyint.*``' intrinsics returns the operand rounded to the
7487 The argument and return value are floating point numbers of the same
7493 This function returns the same values as the libm ``nearbyint``
7494 functions would, and handles error conditions in the same way.
7496 Bit Manipulation Intrinsics
7497 ---------------------------
7499 LLVM provides intrinsics for a few important bit manipulation
7500 operations. These allow efficient code generation for some algorithms.
7502 '``llvm.bswap.*``' Intrinsics
7503 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7508 This is an overloaded intrinsic function. You can use bswap on any
7509 integer type that is an even number of bytes (i.e. BitWidth % 16 == 0).
7513 declare i16 @llvm.bswap.i16(i16 <id>)
7514 declare i32 @llvm.bswap.i32(i32 <id>)
7515 declare i64 @llvm.bswap.i64(i64 <id>)
7520 The '``llvm.bswap``' family of intrinsics is used to byte swap integer
7521 values with an even number of bytes (positive multiple of 16 bits).
7522 These are useful for performing operations on data that is not in the
7523 target's native byte order.
7528 The ``llvm.bswap.i16`` intrinsic returns an i16 value that has the high
7529 and low byte of the input i16 swapped. Similarly, the ``llvm.bswap.i32``
7530 intrinsic returns an i32 value that has the four bytes of the input i32
7531 swapped, so that if the input bytes are numbered 0, 1, 2, 3 then the
7532 returned i32 will have its bytes in 3, 2, 1, 0 order. The
7533 ``llvm.bswap.i48``, ``llvm.bswap.i64`` and other intrinsics extend this
7534 concept to additional even-byte lengths (6 bytes, 8 bytes and more,
7537 '``llvm.ctpop.*``' Intrinsic
7538 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7543 This is an overloaded intrinsic. You can use llvm.ctpop on any integer
7544 bit width, or on any vector with integer elements. Not all targets
7545 support all bit widths or vector types, however.
7549 declare i8 @llvm.ctpop.i8(i8 <src>)
7550 declare i16 @llvm.ctpop.i16(i16 <src>)
7551 declare i32 @llvm.ctpop.i32(i32 <src>)
7552 declare i64 @llvm.ctpop.i64(i64 <src>)
7553 declare i256 @llvm.ctpop.i256(i256 <src>)
7554 declare <2 x i32> @llvm.ctpop.v2i32(<2 x i32> <src>)
7559 The '``llvm.ctpop``' family of intrinsics counts the number of bits set
7565 The only argument is the value to be counted. The argument may be of any
7566 integer type, or a vector with integer elements. The return type must
7567 match the argument type.
7572 The '``llvm.ctpop``' intrinsic counts the 1's in a variable, or within
7573 each element of a vector.
7575 '``llvm.ctlz.*``' Intrinsic
7576 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7581 This is an overloaded intrinsic. You can use ``llvm.ctlz`` on any
7582 integer bit width, or any vector whose elements are integers. Not all
7583 targets support all bit widths or vector types, however.
7587 declare i8 @llvm.ctlz.i8 (i8 <src>, i1 <is_zero_undef>)
7588 declare i16 @llvm.ctlz.i16 (i16 <src>, i1 <is_zero_undef>)
7589 declare i32 @llvm.ctlz.i32 (i32 <src>, i1 <is_zero_undef>)
7590 declare i64 @llvm.ctlz.i64 (i64 <src>, i1 <is_zero_undef>)
7591 declare i256 @llvm.ctlz.i256(i256 <src>, i1 <is_zero_undef>)
7592 declase <2 x i32> @llvm.ctlz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
7597 The '``llvm.ctlz``' family of intrinsic functions counts the number of
7598 leading zeros in a variable.
7603 The first argument is the value to be counted. This argument may be of
7604 any integer type, or a vectory with integer element type. The return
7605 type must match the first argument type.
7607 The second argument must be a constant and is a flag to indicate whether
7608 the intrinsic should ensure that a zero as the first argument produces a
7609 defined result. Historically some architectures did not provide a
7610 defined result for zero values as efficiently, and many algorithms are
7611 now predicated on avoiding zero-value inputs.
7616 The '``llvm.ctlz``' intrinsic counts the leading (most significant)
7617 zeros in a variable, or within each element of the vector. If
7618 ``src == 0`` then the result is the size in bits of the type of ``src``
7619 if ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
7620 ``llvm.ctlz(i32 2) = 30``.
7622 '``llvm.cttz.*``' Intrinsic
7623 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7628 This is an overloaded intrinsic. You can use ``llvm.cttz`` on any
7629 integer bit width, or any vector of integer elements. Not all targets
7630 support all bit widths or vector types, however.
7634 declare i8 @llvm.cttz.i8 (i8 <src>, i1 <is_zero_undef>)
7635 declare i16 @llvm.cttz.i16 (i16 <src>, i1 <is_zero_undef>)
7636 declare i32 @llvm.cttz.i32 (i32 <src>, i1 <is_zero_undef>)
7637 declare i64 @llvm.cttz.i64 (i64 <src>, i1 <is_zero_undef>)
7638 declare i256 @llvm.cttz.i256(i256 <src>, i1 <is_zero_undef>)
7639 declase <2 x i32> @llvm.cttz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
7644 The '``llvm.cttz``' family of intrinsic functions counts the number of
7650 The first argument is the value to be counted. This argument may be of
7651 any integer type, or a vectory with integer element type. The return
7652 type must match the first argument type.
7654 The second argument must be a constant and is a flag to indicate whether
7655 the intrinsic should ensure that a zero as the first argument produces a
7656 defined result. Historically some architectures did not provide a
7657 defined result for zero values as efficiently, and many algorithms are
7658 now predicated on avoiding zero-value inputs.
7663 The '``llvm.cttz``' intrinsic counts the trailing (least significant)
7664 zeros in a variable, or within each element of a vector. If ``src == 0``
7665 then the result is the size in bits of the type of ``src`` if
7666 ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
7667 ``llvm.cttz(2) = 1``.
7669 Arithmetic with Overflow Intrinsics
7670 -----------------------------------
7672 LLVM provides intrinsics for some arithmetic with overflow operations.
7674 '``llvm.sadd.with.overflow.*``' Intrinsics
7675 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7680 This is an overloaded intrinsic. You can use ``llvm.sadd.with.overflow``
7681 on any integer bit width.
7685 declare {i16, i1} @llvm.sadd.with.overflow.i16(i16 %a, i16 %b)
7686 declare {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
7687 declare {i64, i1} @llvm.sadd.with.overflow.i64(i64 %a, i64 %b)
7692 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
7693 a signed addition of the two arguments, and indicate whether an overflow
7694 occurred during the signed summation.
7699 The arguments (%a and %b) and the first element of the result structure
7700 may be of integer types of any bit width, but they must have the same
7701 bit width. The second element of the result structure must be of type
7702 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
7708 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
7709 a signed addition of the two variables. They return a structure --- the
7710 first element of which is the signed summation, and the second element
7711 of which is a bit specifying if the signed summation resulted in an
7717 .. code-block:: llvm
7719 %res = call {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
7720 %sum = extractvalue {i32, i1} %res, 0
7721 %obit = extractvalue {i32, i1} %res, 1
7722 br i1 %obit, label %overflow, label %normal
7724 '``llvm.uadd.with.overflow.*``' Intrinsics
7725 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7730 This is an overloaded intrinsic. You can use ``llvm.uadd.with.overflow``
7731 on any integer bit width.
7735 declare {i16, i1} @llvm.uadd.with.overflow.i16(i16 %a, i16 %b)
7736 declare {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
7737 declare {i64, i1} @llvm.uadd.with.overflow.i64(i64 %a, i64 %b)
7742 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
7743 an unsigned addition of the two arguments, and indicate whether a carry
7744 occurred during the unsigned summation.
7749 The arguments (%a and %b) and the first element of the result structure
7750 may be of integer types of any bit width, but they must have the same
7751 bit width. The second element of the result structure must be of type
7752 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
7758 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
7759 an unsigned addition of the two arguments. They return a structure --- the
7760 first element of which is the sum, and the second element of which is a
7761 bit specifying if the unsigned summation resulted in a carry.
7766 .. code-block:: llvm
7768 %res = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
7769 %sum = extractvalue {i32, i1} %res, 0
7770 %obit = extractvalue {i32, i1} %res, 1
7771 br i1 %obit, label %carry, label %normal
7773 '``llvm.ssub.with.overflow.*``' Intrinsics
7774 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7779 This is an overloaded intrinsic. You can use ``llvm.ssub.with.overflow``
7780 on any integer bit width.
7784 declare {i16, i1} @llvm.ssub.with.overflow.i16(i16 %a, i16 %b)
7785 declare {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
7786 declare {i64, i1} @llvm.ssub.with.overflow.i64(i64 %a, i64 %b)
7791 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
7792 a signed subtraction of the two arguments, and indicate whether an
7793 overflow occurred during the signed subtraction.
7798 The arguments (%a and %b) and the first element of the result structure
7799 may be of integer types of any bit width, but they must have the same
7800 bit width. The second element of the result structure must be of type
7801 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
7807 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
7808 a signed subtraction of the two arguments. They return a structure --- the
7809 first element of which is the subtraction, and the second element of
7810 which is a bit specifying if the signed subtraction resulted in an
7816 .. code-block:: llvm
7818 %res = call {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
7819 %sum = extractvalue {i32, i1} %res, 0
7820 %obit = extractvalue {i32, i1} %res, 1
7821 br i1 %obit, label %overflow, label %normal
7823 '``llvm.usub.with.overflow.*``' Intrinsics
7824 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7829 This is an overloaded intrinsic. You can use ``llvm.usub.with.overflow``
7830 on any integer bit width.
7834 declare {i16, i1} @llvm.usub.with.overflow.i16(i16 %a, i16 %b)
7835 declare {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
7836 declare {i64, i1} @llvm.usub.with.overflow.i64(i64 %a, i64 %b)
7841 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
7842 an unsigned subtraction of the two arguments, and indicate whether an
7843 overflow occurred during the unsigned subtraction.
7848 The arguments (%a and %b) and the first element of the result structure
7849 may be of integer types of any bit width, but they must have the same
7850 bit width. The second element of the result structure must be of type
7851 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
7857 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
7858 an unsigned subtraction of the two arguments. They return a structure ---
7859 the first element of which is the subtraction, and the second element of
7860 which is a bit specifying if the unsigned subtraction resulted in an
7866 .. code-block:: llvm
7868 %res = call {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
7869 %sum = extractvalue {i32, i1} %res, 0
7870 %obit = extractvalue {i32, i1} %res, 1
7871 br i1 %obit, label %overflow, label %normal
7873 '``llvm.smul.with.overflow.*``' Intrinsics
7874 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7879 This is an overloaded intrinsic. You can use ``llvm.smul.with.overflow``
7880 on any integer bit width.
7884 declare {i16, i1} @llvm.smul.with.overflow.i16(i16 %a, i16 %b)
7885 declare {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
7886 declare {i64, i1} @llvm.smul.with.overflow.i64(i64 %a, i64 %b)
7891 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
7892 a signed multiplication of the two arguments, and indicate whether an
7893 overflow occurred during the signed multiplication.
7898 The arguments (%a and %b) and the first element of the result structure
7899 may be of integer types of any bit width, but they must have the same
7900 bit width. The second element of the result structure must be of type
7901 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
7907 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
7908 a signed multiplication of the two arguments. They return a structure ---
7909 the first element of which is the multiplication, and the second element
7910 of which is a bit specifying if the signed multiplication resulted in an
7916 .. code-block:: llvm
7918 %res = call {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
7919 %sum = extractvalue {i32, i1} %res, 0
7920 %obit = extractvalue {i32, i1} %res, 1
7921 br i1 %obit, label %overflow, label %normal
7923 '``llvm.umul.with.overflow.*``' Intrinsics
7924 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7929 This is an overloaded intrinsic. You can use ``llvm.umul.with.overflow``
7930 on any integer bit width.
7934 declare {i16, i1} @llvm.umul.with.overflow.i16(i16 %a, i16 %b)
7935 declare {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
7936 declare {i64, i1} @llvm.umul.with.overflow.i64(i64 %a, i64 %b)
7941 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
7942 a unsigned multiplication of the two arguments, and indicate whether an
7943 overflow occurred during the unsigned multiplication.
7948 The arguments (%a and %b) and the first element of the result structure
7949 may be of integer types of any bit width, but they must have the same
7950 bit width. The second element of the result structure must be of type
7951 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
7957 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
7958 an unsigned multiplication of the two arguments. They return a structure ---
7959 the first element of which is the multiplication, and the second
7960 element of which is a bit specifying if the unsigned multiplication
7961 resulted in an overflow.
7966 .. code-block:: llvm
7968 %res = call {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
7969 %sum = extractvalue {i32, i1} %res, 0
7970 %obit = extractvalue {i32, i1} %res, 1
7971 br i1 %obit, label %overflow, label %normal
7973 Specialised Arithmetic Intrinsics
7974 ---------------------------------
7976 '``llvm.fmuladd.*``' Intrinsic
7977 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7984 declare float @llvm.fmuladd.f32(float %a, float %b, float %c)
7985 declare double @llvm.fmuladd.f64(double %a, double %b, double %c)
7990 The '``llvm.fmuladd.*``' intrinsic functions represent multiply-add
7991 expressions that can be fused if the code generator determines that (a) the
7992 target instruction set has support for a fused operation, and (b) that the
7993 fused operation is more efficient than the equivalent, separate pair of mul
7994 and add instructions.
7999 The '``llvm.fmuladd.*``' intrinsics each take three arguments: two
8000 multiplicands, a and b, and an addend c.
8009 %0 = call float @llvm.fmuladd.f32(%a, %b, %c)
8011 is equivalent to the expression a \* b + c, except that rounding will
8012 not be performed between the multiplication and addition steps if the
8013 code generator fuses the operations. Fusion is not guaranteed, even if
8014 the target platform supports it. If a fused multiply-add is required the
8015 corresponding llvm.fma.\* intrinsic function should be used instead.
8020 .. code-block:: llvm
8022 %r2 = call float @llvm.fmuladd.f32(float %a, float %b, float %c) ; yields {float}:r2 = (a * b) + c
8024 Half Precision Floating Point Intrinsics
8025 ----------------------------------------
8027 For most target platforms, half precision floating point is a
8028 storage-only format. This means that it is a dense encoding (in memory)
8029 but does not support computation in the format.
8031 This means that code must first load the half-precision floating point
8032 value as an i16, then convert it to float with
8033 :ref:`llvm.convert.from.fp16 <int_convert_from_fp16>`. Computation can
8034 then be performed on the float value (including extending to double
8035 etc). To store the value back to memory, it is first converted to float
8036 if needed, then converted to i16 with
8037 :ref:`llvm.convert.to.fp16 <int_convert_to_fp16>`, then storing as an
8040 .. _int_convert_to_fp16:
8042 '``llvm.convert.to.fp16``' Intrinsic
8043 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8050 declare i16 @llvm.convert.to.fp16(f32 %a)
8055 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion
8056 from single precision floating point format to half precision floating
8062 The intrinsic function contains single argument - the value to be
8068 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion
8069 from single precision floating point format to half precision floating
8070 point format. The return value is an ``i16`` which contains the
8076 .. code-block:: llvm
8078 %res = call i16 @llvm.convert.to.fp16(f32 %a)
8079 store i16 %res, i16* @x, align 2
8081 .. _int_convert_from_fp16:
8083 '``llvm.convert.from.fp16``' Intrinsic
8084 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8091 declare f32 @llvm.convert.from.fp16(i16 %a)
8096 The '``llvm.convert.from.fp16``' intrinsic function performs a
8097 conversion from half precision floating point format to single precision
8098 floating point format.
8103 The intrinsic function contains single argument - the value to be
8109 The '``llvm.convert.from.fp16``' intrinsic function performs a
8110 conversion from half single precision floating point format to single
8111 precision floating point format. The input half-float value is
8112 represented by an ``i16`` value.
8117 .. code-block:: llvm
8119 %a = load i16* @x, align 2
8120 %res = call f32 @llvm.convert.from.fp16(i16 %a)
8125 The LLVM debugger intrinsics (which all start with ``llvm.dbg.``
8126 prefix), are described in the `LLVM Source Level
8127 Debugging <SourceLevelDebugging.html#format_common_intrinsics>`_
8130 Exception Handling Intrinsics
8131 -----------------------------
8133 The LLVM exception handling intrinsics (which all start with
8134 ``llvm.eh.`` prefix), are described in the `LLVM Exception
8135 Handling <ExceptionHandling.html#format_common_intrinsics>`_ document.
8139 Trampoline Intrinsics
8140 ---------------------
8142 These intrinsics make it possible to excise one parameter, marked with
8143 the :ref:`nest <nest>` attribute, from a function. The result is a
8144 callable function pointer lacking the nest parameter - the caller does
8145 not need to provide a value for it. Instead, the value to use is stored
8146 in advance in a "trampoline", a block of memory usually allocated on the
8147 stack, which also contains code to splice the nest value into the
8148 argument list. This is used to implement the GCC nested function address
8151 For example, if the function is ``i32 f(i8* nest %c, i32 %x, i32 %y)``
8152 then the resulting function pointer has signature ``i32 (i32, i32)*``.
8153 It can be created as follows:
8155 .. code-block:: llvm
8157 %tramp = alloca [10 x i8], align 4 ; size and alignment only correct for X86
8158 %tramp1 = getelementptr [10 x i8]* %tramp, i32 0, i32 0
8159 call i8* @llvm.init.trampoline(i8* %tramp1, i8* bitcast (i32 (i8*, i32, i32)* @f to i8*), i8* %nval)
8160 %p = call i8* @llvm.adjust.trampoline(i8* %tramp1)
8161 %fp = bitcast i8* %p to i32 (i32, i32)*
8163 The call ``%val = call i32 %fp(i32 %x, i32 %y)`` is then equivalent to
8164 ``%val = call i32 %f(i8* %nval, i32 %x, i32 %y)``.
8168 '``llvm.init.trampoline``' Intrinsic
8169 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8176 declare void @llvm.init.trampoline(i8* <tramp>, i8* <func>, i8* <nval>)
8181 This fills the memory pointed to by ``tramp`` with executable code,
8182 turning it into a trampoline.
8187 The ``llvm.init.trampoline`` intrinsic takes three arguments, all
8188 pointers. The ``tramp`` argument must point to a sufficiently large and
8189 sufficiently aligned block of memory; this memory is written to by the
8190 intrinsic. Note that the size and the alignment are target-specific -
8191 LLVM currently provides no portable way of determining them, so a
8192 front-end that generates this intrinsic needs to have some
8193 target-specific knowledge. The ``func`` argument must hold a function
8194 bitcast to an ``i8*``.
8199 The block of memory pointed to by ``tramp`` is filled with target
8200 dependent code, turning it into a function. Then ``tramp`` needs to be
8201 passed to :ref:`llvm.adjust.trampoline <int_at>` to get a pointer which can
8202 be :ref:`bitcast (to a new function) and called <int_trampoline>`. The new
8203 function's signature is the same as that of ``func`` with any arguments
8204 marked with the ``nest`` attribute removed. At most one such ``nest``
8205 argument is allowed, and it must be of pointer type. Calling the new
8206 function is equivalent to calling ``func`` with the same argument list,
8207 but with ``nval`` used for the missing ``nest`` argument. If, after
8208 calling ``llvm.init.trampoline``, the memory pointed to by ``tramp`` is
8209 modified, then the effect of any later call to the returned function
8210 pointer is undefined.
8214 '``llvm.adjust.trampoline``' Intrinsic
8215 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8222 declare i8* @llvm.adjust.trampoline(i8* <tramp>)
8227 This performs any required machine-specific adjustment to the address of
8228 a trampoline (passed as ``tramp``).
8233 ``tramp`` must point to a block of memory which already has trampoline
8234 code filled in by a previous call to
8235 :ref:`llvm.init.trampoline <int_it>`.
8240 On some architectures the address of the code to be executed needs to be
8241 different to the address where the trampoline is actually stored. This
8242 intrinsic returns the executable address corresponding to ``tramp``
8243 after performing the required machine specific adjustments. The pointer
8244 returned can then be :ref:`bitcast and executed <int_trampoline>`.
8249 This class of intrinsics exists to information about the lifetime of
8250 memory objects and ranges where variables are immutable.
8252 '``llvm.lifetime.start``' Intrinsic
8253 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8260 declare void @llvm.lifetime.start(i64 <size>, i8* nocapture <ptr>)
8265 The '``llvm.lifetime.start``' intrinsic specifies the start of a memory
8271 The first argument is a constant integer representing the size of the
8272 object, or -1 if it is variable sized. The second argument is a pointer
8278 This intrinsic indicates that before this point in the code, the value
8279 of the memory pointed to by ``ptr`` is dead. This means that it is known
8280 to never be used and has an undefined value. A load from the pointer
8281 that precedes this intrinsic can be replaced with ``'undef'``.
8283 '``llvm.lifetime.end``' Intrinsic
8284 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8291 declare void @llvm.lifetime.end(i64 <size>, i8* nocapture <ptr>)
8296 The '``llvm.lifetime.end``' intrinsic specifies the end of a memory
8302 The first argument is a constant integer representing the size of the
8303 object, or -1 if it is variable sized. The second argument is a pointer
8309 This intrinsic indicates that after this point in the code, the value of
8310 the memory pointed to by ``ptr`` is dead. This means that it is known to
8311 never be used and has an undefined value. Any stores into the memory
8312 object following this intrinsic may be removed as dead.
8314 '``llvm.invariant.start``' Intrinsic
8315 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8322 declare {}* @llvm.invariant.start(i64 <size>, i8* nocapture <ptr>)
8327 The '``llvm.invariant.start``' intrinsic specifies that the contents of
8328 a memory object will not change.
8333 The first argument is a constant integer representing the size of the
8334 object, or -1 if it is variable sized. The second argument is a pointer
8340 This intrinsic indicates that until an ``llvm.invariant.end`` that uses
8341 the return value, the referenced memory location is constant and
8344 '``llvm.invariant.end``' Intrinsic
8345 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8352 declare void @llvm.invariant.end({}* <start>, i64 <size>, i8* nocapture <ptr>)
8357 The '``llvm.invariant.end``' intrinsic specifies that the contents of a
8358 memory object are mutable.
8363 The first argument is the matching ``llvm.invariant.start`` intrinsic.
8364 The second argument is a constant integer representing the size of the
8365 object, or -1 if it is variable sized and the third argument is a
8366 pointer to the object.
8371 This intrinsic indicates that the memory is mutable again.
8376 This class of intrinsics is designed to be generic and has no specific
8379 '``llvm.var.annotation``' Intrinsic
8380 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8387 declare void @llvm.var.annotation(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
8392 The '``llvm.var.annotation``' intrinsic.
8397 The first argument is a pointer to a value, the second is a pointer to a
8398 global string, the third is a pointer to a global string which is the
8399 source file name, and the last argument is the line number.
8404 This intrinsic allows annotation of local variables with arbitrary
8405 strings. This can be useful for special purpose optimizations that want
8406 to look for these annotations. These have no other defined use; they are
8407 ignored by code generation and optimization.
8409 '``llvm.ptr.annotation.*``' Intrinsic
8410 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8415 This is an overloaded intrinsic. You can use '``llvm.ptr.annotation``' on a
8416 pointer to an integer of any width. *NOTE* you must specify an address space for
8417 the pointer. The identifier for the default address space is the integer
8422 declare i8* @llvm.ptr.annotation.p<address space>i8(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
8423 declare i16* @llvm.ptr.annotation.p<address space>i16(i16* <val>, i8* <str>, i8* <str>, i32 <int>)
8424 declare i32* @llvm.ptr.annotation.p<address space>i32(i32* <val>, i8* <str>, i8* <str>, i32 <int>)
8425 declare i64* @llvm.ptr.annotation.p<address space>i64(i64* <val>, i8* <str>, i8* <str>, i32 <int>)
8426 declare i256* @llvm.ptr.annotation.p<address space>i256(i256* <val>, i8* <str>, i8* <str>, i32 <int>)
8431 The '``llvm.ptr.annotation``' intrinsic.
8436 The first argument is a pointer to an integer value of arbitrary bitwidth
8437 (result of some expression), the second is a pointer to a global string, the
8438 third is a pointer to a global string which is the source file name, and the
8439 last argument is the line number. It returns the value of the first argument.
8444 This intrinsic allows annotation of a pointer to an integer with arbitrary
8445 strings. This can be useful for special purpose optimizations that want to look
8446 for these annotations. These have no other defined use; they are ignored by code
8447 generation and optimization.
8449 '``llvm.annotation.*``' Intrinsic
8450 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8455 This is an overloaded intrinsic. You can use '``llvm.annotation``' on
8456 any integer bit width.
8460 declare i8 @llvm.annotation.i8(i8 <val>, i8* <str>, i8* <str>, i32 <int>)
8461 declare i16 @llvm.annotation.i16(i16 <val>, i8* <str>, i8* <str>, i32 <int>)
8462 declare i32 @llvm.annotation.i32(i32 <val>, i8* <str>, i8* <str>, i32 <int>)
8463 declare i64 @llvm.annotation.i64(i64 <val>, i8* <str>, i8* <str>, i32 <int>)
8464 declare i256 @llvm.annotation.i256(i256 <val>, i8* <str>, i8* <str>, i32 <int>)
8469 The '``llvm.annotation``' intrinsic.
8474 The first argument is an integer value (result of some expression), the
8475 second is a pointer to a global string, the third is a pointer to a
8476 global string which is the source file name, and the last argument is
8477 the line number. It returns the value of the first argument.
8482 This intrinsic allows annotations to be put on arbitrary expressions
8483 with arbitrary strings. This can be useful for special purpose
8484 optimizations that want to look for these annotations. These have no
8485 other defined use; they are ignored by code generation and optimization.
8487 '``llvm.trap``' Intrinsic
8488 ^^^^^^^^^^^^^^^^^^^^^^^^^
8495 declare void @llvm.trap() noreturn nounwind
8500 The '``llvm.trap``' intrinsic.
8510 This intrinsic is lowered to the target dependent trap instruction. If
8511 the target does not have a trap instruction, this intrinsic will be
8512 lowered to a call of the ``abort()`` function.
8514 '``llvm.debugtrap``' Intrinsic
8515 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8522 declare void @llvm.debugtrap() nounwind
8527 The '``llvm.debugtrap``' intrinsic.
8537 This intrinsic is lowered to code which is intended to cause an
8538 execution trap with the intention of requesting the attention of a
8541 '``llvm.stackprotector``' Intrinsic
8542 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8549 declare void @llvm.stackprotector(i8* <guard>, i8** <slot>)
8554 The ``llvm.stackprotector`` intrinsic takes the ``guard`` and stores it
8555 onto the stack at ``slot``. The stack slot is adjusted to ensure that it
8556 is placed on the stack before local variables.
8561 The ``llvm.stackprotector`` intrinsic requires two pointer arguments.
8562 The first argument is the value loaded from the stack guard
8563 ``@__stack_chk_guard``. The second variable is an ``alloca`` that has
8564 enough space to hold the value of the guard.
8569 This intrinsic causes the prologue/epilogue inserter to force the
8570 position of the ``AllocaInst`` stack slot to be before local variables
8571 on the stack. This is to ensure that if a local variable on the stack is
8572 overwritten, it will destroy the value of the guard. When the function
8573 exits, the guard on the stack is checked against the original guard. If
8574 they are different, then the program aborts by calling the
8575 ``__stack_chk_fail()`` function.
8577 '``llvm.objectsize``' Intrinsic
8578 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8585 declare i32 @llvm.objectsize.i32(i8* <object>, i1 <min>)
8586 declare i64 @llvm.objectsize.i64(i8* <object>, i1 <min>)
8591 The ``llvm.objectsize`` intrinsic is designed to provide information to
8592 the optimizers to determine at compile time whether a) an operation
8593 (like memcpy) will overflow a buffer that corresponds to an object, or
8594 b) that a runtime check for overflow isn't necessary. An object in this
8595 context means an allocation of a specific class, structure, array, or
8601 The ``llvm.objectsize`` intrinsic takes two arguments. The first
8602 argument is a pointer to or into the ``object``. The second argument is
8603 a boolean and determines whether ``llvm.objectsize`` returns 0 (if true)
8604 or -1 (if false) when the object size is unknown. The second argument
8605 only accepts constants.
8610 The ``llvm.objectsize`` intrinsic is lowered to a constant representing
8611 the size of the object concerned. If the size cannot be determined at
8612 compile time, ``llvm.objectsize`` returns ``i32/i64 -1 or 0`` (depending
8613 on the ``min`` argument).
8615 '``llvm.expect``' Intrinsic
8616 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8623 declare i32 @llvm.expect.i32(i32 <val>, i32 <expected_val>)
8624 declare i64 @llvm.expect.i64(i64 <val>, i64 <expected_val>)
8629 The ``llvm.expect`` intrinsic provides information about expected (the
8630 most probable) value of ``val``, which can be used by optimizers.
8635 The ``llvm.expect`` intrinsic takes two arguments. The first argument is
8636 a value. The second argument is an expected value, this needs to be a
8637 constant value, variables are not allowed.
8642 This intrinsic is lowered to the ``val``.
8644 '``llvm.donothing``' Intrinsic
8645 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8652 declare void @llvm.donothing() nounwind readnone
8657 The ``llvm.donothing`` intrinsic doesn't perform any operation. It's the
8658 only intrinsic that can be called with an invoke instruction.
8668 This intrinsic does nothing, and it's removed by optimizers and ignored