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). Note that basic blocks are
132 included in this numbering. For example, if the entry basic block is not
133 given a label name, then it will get number 0.
135 It also shows a convention that we follow in this document. When
136 demonstrating instructions, we will follow an instruction with a comment
137 that defines the type and name of value produced.
145 LLVM programs are composed of ``Module``'s, each of which is a
146 translation unit of the input programs. Each module consists of
147 functions, global variables, and symbol table entries. Modules may be
148 combined together with the LLVM linker, which merges function (and
149 global variable) definitions, resolves forward declarations, and merges
150 symbol table entries. Here is an example of the "hello world" module:
154 ; Declare the string constant as a global constant.
155 @.str = private unnamed_addr constant [13 x i8] c"hello world\0A\00"
157 ; External declaration of the puts function
158 declare i32 @puts(i8* nocapture) nounwind
160 ; Definition of main function
161 define i32 @main() { ; i32()*
162 ; Convert [13 x i8]* to i8 *...
163 %cast210 = getelementptr [13 x i8]* @.str, i64 0, i64 0
165 ; Call puts function to write out the string to stdout.
166 call i32 @puts(i8* %cast210)
171 !1 = metadata !{i32 42}
174 This example is made up of a :ref:`global variable <globalvars>` named
175 "``.str``", an external declaration of the "``puts``" function, a
176 :ref:`function definition <functionstructure>` for "``main``" and
177 :ref:`named metadata <namedmetadatastructure>` "``foo``".
179 In general, a module is made up of a list of global values (where both
180 functions and global variables are global values). Global values are
181 represented by a pointer to a memory location (in this case, a pointer
182 to an array of char, and a pointer to a function), and have one of the
183 following :ref:`linkage types <linkage>`.
190 All Global Variables and Functions have one of the following types of
194 Global values with "``private``" linkage are only directly
195 accessible by objects in the current module. In particular, linking
196 code into a module with an private global value may cause the
197 private to be renamed as necessary to avoid collisions. Because the
198 symbol is private to the module, all references can be updated. This
199 doesn't show up in any symbol table in the object file.
201 Similar to ``private``, but the symbol is passed through the
202 assembler and evaluated by the linker. Unlike normal strong symbols,
203 they are removed by the linker from the final linked image
204 (executable or dynamic library).
205 ``linker_private_weak``
206 Similar to "``linker_private``", but the symbol is weak. Note that
207 ``linker_private_weak`` symbols are subject to coalescing by the
208 linker. The symbols are removed by the linker from the final linked
209 image (executable or dynamic library).
211 Similar to private, but the value shows as a local symbol
212 (``STB_LOCAL`` in the case of ELF) in the object file. This
213 corresponds to the notion of the '``static``' keyword in C.
214 ``available_externally``
215 Globals with "``available_externally``" linkage are never emitted
216 into the object file corresponding to the LLVM module. They exist to
217 allow inlining and other optimizations to take place given knowledge
218 of the definition of the global, which is known to be somewhere
219 outside the module. Globals with ``available_externally`` linkage
220 are allowed to be discarded at will, and are otherwise the same as
221 ``linkonce_odr``. This linkage type is only allowed on definitions,
224 Globals with "``linkonce``" linkage are merged with other globals of
225 the same name when linkage occurs. This can be used to implement
226 some forms of inline functions, templates, or other code which must
227 be generated in each translation unit that uses it, but where the
228 body may be overridden with a more definitive definition later.
229 Unreferenced ``linkonce`` globals are allowed to be discarded. Note
230 that ``linkonce`` linkage does not actually allow the optimizer to
231 inline the body of this function into callers because it doesn't
232 know if this definition of the function is the definitive definition
233 within the program or whether it will be overridden by a stronger
234 definition. To enable inlining and other optimizations, use
235 "``linkonce_odr``" linkage.
237 "``weak``" linkage has the same merging semantics as ``linkonce``
238 linkage, except that unreferenced globals with ``weak`` linkage may
239 not be discarded. This is used for globals that are declared "weak"
242 "``common``" linkage is most similar to "``weak``" linkage, but they
243 are used for tentative definitions in C, such as "``int X;``" at
244 global scope. Symbols with "``common``" linkage are merged in the
245 same way as ``weak symbols``, and they may not be deleted if
246 unreferenced. ``common`` symbols may not have an explicit section,
247 must have a zero initializer, and may not be marked
248 ':ref:`constant <globalvars>`'. Functions and aliases may not have
251 .. _linkage_appending:
254 "``appending``" linkage may only be applied to global variables of
255 pointer to array type. When two global variables with appending
256 linkage are linked together, the two global arrays are appended
257 together. This is the LLVM, typesafe, equivalent of having the
258 system linker append together "sections" with identical names when
261 The semantics of this linkage follow the ELF object file model: the
262 symbol is weak until linked, if not linked, the symbol becomes null
263 instead of being an undefined reference.
264 ``linkonce_odr``, ``weak_odr``
265 Some languages allow differing globals to be merged, such as two
266 functions with different semantics. Other languages, such as
267 ``C++``, ensure that only equivalent globals are ever merged (the
268 "one definition rule" --- "ODR"). Such languages can use the
269 ``linkonce_odr`` and ``weak_odr`` linkage types to indicate that the
270 global will only be merged with equivalent globals. These linkage
271 types are otherwise the same as their non-``odr`` versions.
273 If none of the above identifiers are used, the global is externally
274 visible, meaning that it participates in linkage and can be used to
275 resolve external symbol references.
277 It is illegal for a function *declaration* to have any linkage type
278 other than ``external`` or ``extern_weak``.
285 LLVM :ref:`functions <functionstructure>`, :ref:`calls <i_call>` and
286 :ref:`invokes <i_invoke>` can all have an optional calling convention
287 specified for the call. The calling convention of any pair of dynamic
288 caller/callee must match, or the behavior of the program is undefined.
289 The following calling conventions are supported by LLVM, and more may be
292 "``ccc``" - The C calling convention
293 This calling convention (the default if no other calling convention
294 is specified) matches the target C calling conventions. This calling
295 convention supports varargs function calls and tolerates some
296 mismatch in the declared prototype and implemented declaration of
297 the function (as does normal C).
298 "``fastcc``" - The fast calling convention
299 This calling convention attempts to make calls as fast as possible
300 (e.g. by passing things in registers). This calling convention
301 allows the target to use whatever tricks it wants to produce fast
302 code for the target, without having to conform to an externally
303 specified ABI (Application Binary Interface). `Tail calls can only
304 be optimized when this, the GHC or the HiPE convention is
305 used. <CodeGenerator.html#id80>`_ This calling convention does not
306 support varargs and requires the prototype of all callees to exactly
307 match the prototype of the function definition.
308 "``coldcc``" - The cold calling convention
309 This calling convention attempts to make code in the caller as
310 efficient as possible under the assumption that the call is not
311 commonly executed. As such, these calls often preserve all registers
312 so that the call does not break any live ranges in the caller side.
313 This calling convention does not support varargs and requires the
314 prototype of all callees to exactly match the prototype of the
315 function definition. Furthermore the inliner doesn't consider such function
317 "``cc 10``" - GHC convention
318 This calling convention has been implemented specifically for use by
319 the `Glasgow Haskell Compiler (GHC) <http://www.haskell.org/ghc>`_.
320 It passes everything in registers, going to extremes to achieve this
321 by disabling callee save registers. This calling convention should
322 not be used lightly but only for specific situations such as an
323 alternative to the *register pinning* performance technique often
324 used when implementing functional programming languages. At the
325 moment only X86 supports this convention and it has the following
328 - On *X86-32* only supports up to 4 bit type parameters. No
329 floating point types are supported.
330 - On *X86-64* only supports up to 10 bit type parameters and 6
331 floating point parameters.
333 This calling convention supports `tail call
334 optimization <CodeGenerator.html#id80>`_ but requires both the
335 caller and callee are using it.
336 "``cc 11``" - The HiPE calling convention
337 This calling convention has been implemented specifically for use by
338 the `High-Performance Erlang
339 (HiPE) <http://www.it.uu.se/research/group/hipe/>`_ compiler, *the*
340 native code compiler of the `Ericsson's Open Source Erlang/OTP
341 system <http://www.erlang.org/download.shtml>`_. It uses more
342 registers for argument passing than the ordinary C calling
343 convention and defines no callee-saved registers. The calling
344 convention properly supports `tail call
345 optimization <CodeGenerator.html#id80>`_ but requires that both the
346 caller and the callee use it. It uses a *register pinning*
347 mechanism, similar to GHC's convention, for keeping frequently
348 accessed runtime components pinned to specific hardware registers.
349 At the moment only X86 supports this convention (both 32 and 64
351 "``webkit_jscc``" - WebKit's JavaScript calling convention
352 This calling convention has been implemented for `WebKit FTL JIT
353 <https://trac.webkit.org/wiki/FTLJIT>`_. It passes arguments on the
354 stack right to left (as cdecl does), and returns a value in the
355 platform's customary return register.
356 "``anyregcc``" - Dynamic calling convention for code patching
357 This is a special convention that supports patching an arbitrary code
358 sequence in place of a call site. This convention forces the call
359 arguments into registers but allows them to be dynamcially
360 allocated. This can currently only be used with calls to
361 llvm.experimental.patchpoint because only this intrinsic records
362 the location of its arguments in a side table. See :doc:`StackMaps`.
363 "``preserve_mostcc``" - The `PreserveMost` calling convention
364 This calling convention attempts to make the code in the caller as little
365 intrusive as possible. This calling convention behaves identical to the `C`
366 calling convention on how arguments and return values are passed, but it
367 uses a different set of caller/callee-saved registers. This alleviates the
368 burden of saving and recovering a large register set before and after the
369 call in the caller. If the arguments are passed in callee-saved registers,
370 then they will be preserved by the callee across the call. This doesn't
371 apply for values returned in callee-saved registers.
373 - On X86-64 the callee preserves all general purpose registers, except for
374 R11. R11 can be used as a scratch register. Floating-point registers
375 (XMMs/YMMs) are not preserved and need to be saved by the caller.
377 The idea behind this convention is to support calls to runtime functions
378 that have a hot path and a cold path. The hot path is usually a small piece
379 of code that doesn't many registers. The cold path might need to call out to
380 another function and therefore only needs to preserve the caller-saved
381 registers, which haven't already been saved by the caller. The
382 `PreserveMost` calling convention is very similar to the `cold` calling
383 convention in terms of caller/callee-saved registers, but they are used for
384 different types of function calls. `coldcc` is for function calls that are
385 rarely executed, whereas `preserve_mostcc` function calls are intended to be
386 on the hot path and definitely executed a lot. Furthermore `preserve_mostcc`
387 doesn't prevent the inliner from inlining the function call.
389 This calling convention will be used by a future version of the ObjectiveC
390 runtime and should therefore still be considered experimental at this time.
391 Although this convention was created to optimize certain runtime calls to
392 the ObjectiveC runtime, it is not limited to this runtime and might be used
393 by other runtimes in the future too. The current implementation only
394 supports X86-64, but the intention is to support more architectures in the
396 "``preserve_allcc``" - The `PreserveAll` calling convention
397 This calling convention attempts to make the code in the caller even less
398 intrusive than the `PreserveMost` calling convention. This calling
399 convention also behaves identical to the `C` calling convention on how
400 arguments and return values are passed, but it uses a different set of
401 caller/callee-saved registers. This removes the burden of saving and
402 recovering a large register set before and after the call in the caller. If
403 the arguments are passed in callee-saved registers, then they will be
404 preserved by the callee across the call. This doesn't apply for values
405 returned in callee-saved registers.
407 - On X86-64 the callee preserves all general purpose registers, except for
408 R11. R11 can be used as a scratch register. Furthermore it also preserves
409 all floating-point registers (XMMs/YMMs).
411 The idea behind this convention is to support calls to runtime functions
412 that don't need to call out to any other functions.
414 This calling convention, like the `PreserveMost` calling convention, will be
415 used by a future version of the ObjectiveC runtime and should be considered
416 experimental at this time.
417 "``cc <n>``" - Numbered convention
418 Any calling convention may be specified by number, allowing
419 target-specific calling conventions to be used. Target specific
420 calling conventions start at 64.
422 More calling conventions can be added/defined on an as-needed basis, to
423 support Pascal conventions or any other well-known target-independent
426 .. _visibilitystyles:
431 All Global Variables and Functions have one of the following visibility
434 "``default``" - Default style
435 On targets that use the ELF object file format, default visibility
436 means that the declaration is visible to other modules and, in
437 shared libraries, means that the declared entity may be overridden.
438 On Darwin, default visibility means that the declaration is visible
439 to other modules. Default visibility corresponds to "external
440 linkage" in the language.
441 "``hidden``" - Hidden style
442 Two declarations of an object with hidden visibility refer to the
443 same object if they are in the same shared object. Usually, hidden
444 visibility indicates that the symbol will not be placed into the
445 dynamic symbol table, so no other module (executable or shared
446 library) can reference it directly.
447 "``protected``" - Protected style
448 On ELF, protected visibility indicates that the symbol will be
449 placed in the dynamic symbol table, but that references within the
450 defining module will bind to the local symbol. That is, the symbol
451 cannot be overridden by another module.
458 All Global Variables, Functions and Aliases can have one of the following
462 "``dllimport``" causes the compiler to reference a function or variable via
463 a global pointer to a pointer that is set up by the DLL exporting the
464 symbol. On Microsoft Windows targets, the pointer name is formed by
465 combining ``__imp_`` and the function or variable name.
467 "``dllexport``" causes the compiler to provide a global pointer to a pointer
468 in a DLL, so that it can be referenced with the ``dllimport`` attribute. On
469 Microsoft Windows targets, the pointer name is formed by combining
470 ``__imp_`` and the function or variable name. Since this storage class
471 exists for defining a dll interface, the compiler, assembler and linker know
472 it is externally referenced and must refrain from deleting the symbol.
477 LLVM IR allows you to specify name aliases for certain types. This can
478 make it easier to read the IR and make the IR more condensed
479 (particularly when recursive types are involved). An example of a name
484 %mytype = type { %mytype*, i32 }
486 You may give a name to any :ref:`type <typesystem>` except
487 ":ref:`void <t_void>`". Type name aliases may be used anywhere a type is
488 expected with the syntax "%mytype".
490 Note that type names are aliases for the structural type that they
491 indicate, and that you can therefore specify multiple names for the same
492 type. This often leads to confusing behavior when dumping out a .ll
493 file. Since LLVM IR uses structural typing, the name is not part of the
494 type. When printing out LLVM IR, the printer will pick *one name* to
495 render all types of a particular shape. This means that if you have code
496 where two different source types end up having the same LLVM type, that
497 the dumper will sometimes print the "wrong" or unexpected type. This is
498 an important design point and isn't going to change.
505 Global variables define regions of memory allocated at compilation time
508 Global variables definitions must be initialized, may have an explicit section
509 to be placed in, and may have an optional explicit alignment specified.
511 Global variables in other translation units can also be declared, in which
512 case they don't have an initializer.
514 A variable may be defined as ``thread_local``, which means that it will
515 not be shared by threads (each thread will have a separated copy of the
516 variable). Not all targets support thread-local variables. Optionally, a
517 TLS model may be specified:
520 For variables that are only used within the current shared library.
522 For variables in modules that will not be loaded dynamically.
524 For variables defined in the executable and only used within it.
526 The models correspond to the ELF TLS models; see `ELF Handling For
527 Thread-Local Storage <http://people.redhat.com/drepper/tls.pdf>`_ for
528 more information on under which circumstances the different models may
529 be used. The target may choose a different TLS model if the specified
530 model is not supported, or if a better choice of model can be made.
532 A variable may be defined as a global ``constant``, which indicates that
533 the contents of the variable will **never** be modified (enabling better
534 optimization, allowing the global data to be placed in the read-only
535 section of an executable, etc). Note that variables that need runtime
536 initialization cannot be marked ``constant`` as there is a store to the
539 LLVM explicitly allows *declarations* of global variables to be marked
540 constant, even if the final definition of the global is not. This
541 capability can be used to enable slightly better optimization of the
542 program, but requires the language definition to guarantee that
543 optimizations based on the 'constantness' are valid for the translation
544 units that do not include the definition.
546 As SSA values, global variables define pointer values that are in scope
547 (i.e. they dominate) all basic blocks in the program. Global variables
548 always define a pointer to their "content" type because they describe a
549 region of memory, and all memory objects in LLVM are accessed through
552 Global variables can be marked with ``unnamed_addr`` which indicates
553 that the address is not significant, only the content. Constants marked
554 like this can be merged with other constants if they have the same
555 initializer. Note that a constant with significant address *can* be
556 merged with a ``unnamed_addr`` constant, the result being a constant
557 whose address is significant.
559 A global variable may be declared to reside in a target-specific
560 numbered address space. For targets that support them, address spaces
561 may affect how optimizations are performed and/or what target
562 instructions are used to access the variable. The default address space
563 is zero. The address space qualifier must precede any other attributes.
565 LLVM allows an explicit section to be specified for globals. If the
566 target supports it, it will emit globals to the section specified.
568 By default, global initializers are optimized by assuming that global
569 variables defined within the module are not modified from their
570 initial values before the start of the global initializer. This is
571 true even for variables potentially accessible from outside the
572 module, including those with external linkage or appearing in
573 ``@llvm.used`` or dllexported variables. This assumption may be suppressed
574 by marking the variable with ``externally_initialized``.
576 An explicit alignment may be specified for a global, which must be a
577 power of 2. If not present, or if the alignment is set to zero, the
578 alignment of the global is set by the target to whatever it feels
579 convenient. If an explicit alignment is specified, the global is forced
580 to have exactly that alignment. Targets and optimizers are not allowed
581 to over-align the global if the global has an assigned section. In this
582 case, the extra alignment could be observable: for example, code could
583 assume that the globals are densely packed in their section and try to
584 iterate over them as an array, alignment padding would break this
587 Globals can also have a :ref:`DLL storage class <dllstorageclass>`.
591 [@<GlobalVarName> =] [Linkage] [Visibility] [DLLStorageClass] [ThreadLocal]
592 [AddrSpace] [unnamed_addr] [ExternallyInitialized]
593 <global | constant> <Type>
594 [, section "name"] [, align <Alignment>]
596 For example, the following defines a global in a numbered address space
597 with an initializer, section, and alignment:
601 @G = addrspace(5) constant float 1.0, section "foo", align 4
603 The following example just declares a global variable
607 @G = external global i32
609 The following example defines a thread-local global with the
610 ``initialexec`` TLS model:
614 @G = thread_local(initialexec) global i32 0, align 4
616 .. _functionstructure:
621 LLVM function definitions consist of the "``define``" keyword, an
622 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
623 style <visibility>`, an optional :ref:`DLL storage class <dllstorageclass>`,
624 an optional :ref:`calling convention <callingconv>`,
625 an optional ``unnamed_addr`` attribute, a return type, an optional
626 :ref:`parameter attribute <paramattrs>` for the return type, a function
627 name, a (possibly empty) argument list (each with optional :ref:`parameter
628 attributes <paramattrs>`), optional :ref:`function attributes <fnattrs>`,
629 an optional section, an optional alignment, an optional :ref:`garbage
630 collector name <gc>`, an optional :ref:`prefix <prefixdata>`, an opening
631 curly brace, a list of basic blocks, and a closing curly brace.
633 LLVM function declarations consist of the "``declare``" keyword, an
634 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
635 style <visibility>`, an optional :ref:`DLL storage class <dllstorageclass>`,
636 an optional :ref:`calling convention <callingconv>`,
637 an optional ``unnamed_addr`` attribute, a return type, an optional
638 :ref:`parameter attribute <paramattrs>` for the return type, a function
639 name, a possibly empty list of arguments, an optional alignment, an optional
640 :ref:`garbage collector name <gc>` and an optional :ref:`prefix <prefixdata>`.
642 A function definition contains a list of basic blocks, forming the CFG (Control
643 Flow Graph) for the function. Each basic block may optionally start with a label
644 (giving the basic block a symbol table entry), contains a list of instructions,
645 and ends with a :ref:`terminator <terminators>` instruction (such as a branch or
646 function return). If an explicit label is not provided, a block is assigned an
647 implicit numbered label, using the next value from the same counter as used for
648 unnamed temporaries (:ref:`see above<identifiers>`). For example, if a function
649 entry block does not have an explicit label, it will be assigned label "%0",
650 then the first unnamed temporary in that block will be "%1", etc.
652 The first basic block in a function is special in two ways: it is
653 immediately executed on entrance to the function, and it is not allowed
654 to have predecessor basic blocks (i.e. there can not be any branches to
655 the entry block of a function). Because the block can have no
656 predecessors, it also cannot have any :ref:`PHI nodes <i_phi>`.
658 LLVM allows an explicit section to be specified for functions. If the
659 target supports it, it will emit functions to the section specified.
661 An explicit alignment may be specified for a function. If not present,
662 or if the alignment is set to zero, the alignment of the function is set
663 by the target to whatever it feels convenient. If an explicit alignment
664 is specified, the function is forced to have at least that much
665 alignment. All alignments must be a power of 2.
667 If the ``unnamed_addr`` attribute is given, the address is know to not
668 be significant and two identical functions can be merged.
672 define [linkage] [visibility] [DLLStorageClass]
674 <ResultType> @<FunctionName> ([argument list])
675 [fn Attrs] [section "name"] [align N]
676 [gc] [prefix Constant] { ... }
683 Aliases act as "second name" for the aliasee value (which can be either
684 function, global variable, another alias or bitcast of global value).
685 Aliases may have an optional :ref:`linkage type <linkage>`, an optional
686 :ref:`visibility style <visibility>`, and an optional :ref:`DLL storage class
691 @<Name> = [Visibility] [DLLStorageClass] alias [Linkage] <AliaseeTy> @<Aliasee>
693 The linkage must be one of ``private``, ``linker_private``,
694 ``linker_private_weak``, ``internal``, ``linkonce``, ``weak``,
695 ``linkonce_odr``, ``weak_odr``, ``external``. Note that some system linkers
696 might not correctly handle dropping a weak symbol that is aliased by a non-weak
699 .. _namedmetadatastructure:
704 Named metadata is a collection of metadata. :ref:`Metadata
705 nodes <metadata>` (but not metadata strings) are the only valid
706 operands for a named metadata.
710 ; Some unnamed metadata nodes, which are referenced by the named metadata.
711 !0 = metadata !{metadata !"zero"}
712 !1 = metadata !{metadata !"one"}
713 !2 = metadata !{metadata !"two"}
715 !name = !{!0, !1, !2}
722 The return type and each parameter of a function type may have a set of
723 *parameter attributes* associated with them. Parameter attributes are
724 used to communicate additional information about the result or
725 parameters of a function. Parameter attributes are considered to be part
726 of the function, not of the function type, so functions with different
727 parameter attributes can have the same function type.
729 Parameter attributes are simple keywords that follow the type specified.
730 If multiple parameter attributes are needed, they are space separated.
735 declare i32 @printf(i8* noalias nocapture, ...)
736 declare i32 @atoi(i8 zeroext)
737 declare signext i8 @returns_signed_char()
739 Note that any attributes for the function result (``nounwind``,
740 ``readonly``) come immediately after the argument list.
742 Currently, only the following parameter attributes are defined:
745 This indicates to the code generator that the parameter or return
746 value should be zero-extended to the extent required by the target's
747 ABI (which is usually 32-bits, but is 8-bits for a i1 on x86-64) by
748 the caller (for a parameter) or the callee (for a return value).
750 This indicates to the code generator that the parameter or return
751 value should be sign-extended to the extent required by the target's
752 ABI (which is usually 32-bits) by the caller (for a parameter) or
753 the callee (for a return value).
755 This indicates that this parameter or return value should be treated
756 in a special target-dependent fashion during while emitting code for
757 a function call or return (usually, by putting it in a register as
758 opposed to memory, though some targets use it to distinguish between
759 two different kinds of registers). Use of this attribute is
762 This indicates that the pointer parameter should really be passed by
763 value to the function. The attribute implies that a hidden copy of
764 the pointee is made between the caller and the callee, so the callee
765 is unable to modify the value in the caller. This attribute is only
766 valid on LLVM pointer arguments. It is generally used to pass
767 structs and arrays by value, but is also valid on pointers to
768 scalars. The copy is considered to belong to the caller not the
769 callee (for example, ``readonly`` functions should not write to
770 ``byval`` parameters). This is not a valid attribute for return
773 The byval attribute also supports specifying an alignment with the
774 align attribute. It indicates the alignment of the stack slot to
775 form and the known alignment of the pointer specified to the call
776 site. If the alignment is not specified, then the code generator
777 makes a target-specific assumption.
783 .. Warning:: This feature is unstable and not fully implemented.
785 The ``inalloca`` argument attribute allows the caller to take the
786 address of outgoing stack arguments. An ``inalloca`` argument must
787 be a pointer to stack memory produced by an ``alloca`` instruction.
788 The alloca, or argument allocation, must also be tagged with the
789 inalloca keyword. Only the past argument may have the ``inalloca``
790 attribute, and that argument is guaranteed to be passed in memory.
792 An argument allocation may be used by a call at most once because
793 the call may deallocate it. The ``inalloca`` attribute cannot be
794 used in conjunction with other attributes that affect argument
795 storage, like ``inreg``, ``nest``, ``sret``, or ``byval``.
797 When the call site is reached, the argument allocation must have
798 been the most recent stack allocation that is still live, or the
799 results are undefined. It is possible to allocate additional stack
800 space after an argument allocation and before its call site, but it
801 must be cleared off with :ref:`llvm.stackrestore
804 See :doc:`InAlloca` for more information on how to use this
808 This indicates that the pointer parameter specifies the address of a
809 structure that is the return value of the function in the source
810 program. This pointer must be guaranteed by the caller to be valid:
811 loads and stores to the structure may be assumed by the callee
812 not to trap and to be properly aligned. This may only be applied to
813 the first parameter. This is not a valid attribute for return
816 This indicates that pointer values :ref:`based <pointeraliasing>` on
817 the argument or return value do not alias pointer values which are
818 not *based* on it, ignoring certain "irrelevant" dependencies. For a
819 call to the parent function, dependencies between memory references
820 from before or after the call and from those during the call are
821 "irrelevant" to the ``noalias`` keyword for the arguments and return
822 value used in that call. The caller shares the responsibility with
823 the callee for ensuring that these requirements are met. For further
824 details, please see the discussion of the NoAlias response in `alias
825 analysis <AliasAnalysis.html#MustMayNo>`_.
827 Note that this definition of ``noalias`` is intentionally similar
828 to the definition of ``restrict`` in C99 for function arguments,
829 though it is slightly weaker.
831 For function return values, C99's ``restrict`` is not meaningful,
832 while LLVM's ``noalias`` is.
834 This indicates that the callee does not make any copies of the
835 pointer that outlive the callee itself. This is not a valid
836 attribute for return values.
841 This indicates that the pointer parameter can be excised using the
842 :ref:`trampoline intrinsics <int_trampoline>`. This is not a valid
843 attribute for return values and can only be applied to one parameter.
846 This indicates that the function always returns the argument as its return
847 value. This is an optimization hint to the code generator when generating
848 the caller, allowing tail call optimization and omission of register saves
849 and restores in some cases; it is not checked or enforced when generating
850 the callee. The parameter and the function return type must be valid
851 operands for the :ref:`bitcast instruction <i_bitcast>`. This is not a
852 valid attribute for return values and can only be applied to one parameter.
856 Garbage Collector Names
857 -----------------------
859 Each function may specify a garbage collector name, which is simply a
864 define void @f() gc "name" { ... }
866 The compiler declares the supported values of *name*. Specifying a
867 collector which will cause the compiler to alter its output in order to
868 support the named garbage collection algorithm.
875 Prefix data is data associated with a function which the code generator
876 will emit immediately before the function body. The purpose of this feature
877 is to allow frontends to associate language-specific runtime metadata with
878 specific functions and make it available through the function pointer while
879 still allowing the function pointer to be called. To access the data for a
880 given function, a program may bitcast the function pointer to a pointer to
881 the constant's type. This implies that the IR symbol points to the start
884 To maintain the semantics of ordinary function calls, the prefix data must
885 have a particular format. Specifically, it must begin with a sequence of
886 bytes which decode to a sequence of machine instructions, valid for the
887 module's target, which transfer control to the point immediately succeeding
888 the prefix data, without performing any other visible action. This allows
889 the inliner and other passes to reason about the semantics of the function
890 definition without needing to reason about the prefix data. Obviously this
891 makes the format of the prefix data highly target dependent.
893 Prefix data is laid out as if it were an initializer for a global variable
894 of the prefix data's type. No padding is automatically placed between the
895 prefix data and the function body. If padding is required, it must be part
898 A trivial example of valid prefix data for the x86 architecture is ``i8 144``,
899 which encodes the ``nop`` instruction:
903 define void @f() prefix i8 144 { ... }
905 Generally prefix data can be formed by encoding a relative branch instruction
906 which skips the metadata, as in this example of valid prefix data for the
907 x86_64 architecture, where the first two bytes encode ``jmp .+10``:
911 %0 = type <{ i8, i8, i8* }>
913 define void @f() prefix %0 <{ i8 235, i8 8, i8* @md}> { ... }
915 A function may have prefix data but no body. This has similar semantics
916 to the ``available_externally`` linkage in that the data may be used by the
917 optimizers but will not be emitted in the object file.
924 Attribute groups are groups of attributes that are referenced by objects within
925 the IR. They are important for keeping ``.ll`` files readable, because a lot of
926 functions will use the same set of attributes. In the degenerative case of a
927 ``.ll`` file that corresponds to a single ``.c`` file, the single attribute
928 group will capture the important command line flags used to build that file.
930 An attribute group is a module-level object. To use an attribute group, an
931 object references the attribute group's ID (e.g. ``#37``). An object may refer
932 to more than one attribute group. In that situation, the attributes from the
933 different groups are merged.
935 Here is an example of attribute groups for a function that should always be
936 inlined, has a stack alignment of 4, and which shouldn't use SSE instructions:
940 ; Target-independent attributes:
941 attributes #0 = { alwaysinline alignstack=4 }
943 ; Target-dependent attributes:
944 attributes #1 = { "no-sse" }
946 ; Function @f has attributes: alwaysinline, alignstack=4, and "no-sse".
947 define void @f() #0 #1 { ... }
954 Function attributes are set to communicate additional information about
955 a function. Function attributes are considered to be part of the
956 function, not of the function type, so functions with different function
957 attributes can have the same function type.
959 Function attributes are simple keywords that follow the type specified.
960 If multiple attributes are needed, they are space separated. For
965 define void @f() noinline { ... }
966 define void @f() alwaysinline { ... }
967 define void @f() alwaysinline optsize { ... }
968 define void @f() optsize { ... }
971 This attribute indicates that, when emitting the prologue and
972 epilogue, the backend should forcibly align the stack pointer.
973 Specify the desired alignment, which must be a power of two, in
976 This attribute indicates that the inliner should attempt to inline
977 this function into callers whenever possible, ignoring any active
978 inlining size threshold for this caller.
980 This indicates that the callee function at a call site should be
981 recognized as a built-in function, even though the function's declaration
982 uses the ``nobuiltin`` attribute. This is only valid at call sites for
983 direct calls to functions which are declared with the ``nobuiltin``
986 This attribute indicates that this function is rarely called. When
987 computing edge weights, basic blocks post-dominated by a cold
988 function call are also considered to be cold; and, thus, given low
991 This attribute indicates that the source code contained a hint that
992 inlining this function is desirable (such as the "inline" keyword in
993 C/C++). It is just a hint; it imposes no requirements on the
996 This attribute suggests that optimization passes and code generator
997 passes make choices that keep the code size of this function as small
998 as possible and perform optimizations that may sacrifice runtime
999 performance in order to minimize the size of the generated code.
1001 This attribute disables prologue / epilogue emission for the
1002 function. This can have very system-specific consequences.
1004 This indicates that the callee function at a call site is not recognized as
1005 a built-in function. LLVM will retain the original call and not replace it
1006 with equivalent code based on the semantics of the built-in function, unless
1007 the call site uses the ``builtin`` attribute. This is valid at call sites
1008 and on function declarations and definitions.
1010 This attribute indicates that calls to the function cannot be
1011 duplicated. A call to a ``noduplicate`` function may be moved
1012 within its parent function, but may not be duplicated within
1013 its parent function.
1015 A function containing a ``noduplicate`` call may still
1016 be an inlining candidate, provided that the call is not
1017 duplicated by inlining. That implies that the function has
1018 internal linkage and only has one call site, so the original
1019 call is dead after inlining.
1021 This attributes disables implicit floating point instructions.
1023 This attribute indicates that the inliner should never inline this
1024 function in any situation. This attribute may not be used together
1025 with the ``alwaysinline`` attribute.
1027 This attribute suppresses lazy symbol binding for the function. This
1028 may make calls to the function faster, at the cost of extra program
1029 startup time if the function is not called during program startup.
1031 This attribute indicates that the code generator should not use a
1032 red zone, even if the target-specific ABI normally permits it.
1034 This function attribute indicates that the function never returns
1035 normally. This produces undefined behavior at runtime if the
1036 function ever does dynamically return.
1038 This function attribute indicates that the function never returns
1039 with an unwind or exceptional control flow. If the function does
1040 unwind, its runtime behavior is undefined.
1042 This function attribute indicates that the function is not optimized
1043 by any optimization or code generator passes with the
1044 exception of interprocedural optimization passes.
1045 This attribute cannot be used together with the ``alwaysinline``
1046 attribute; this attribute is also incompatible
1047 with the ``minsize`` attribute and the ``optsize`` attribute.
1049 This attribute requires the ``noinline`` attribute to be specified on
1050 the function as well, so the function is never inlined into any caller.
1051 Only functions with the ``alwaysinline`` attribute are valid
1052 candidates for inlining into the body of this function.
1054 This attribute suggests that optimization passes and code generator
1055 passes make choices that keep the code size of this function low,
1056 and otherwise do optimizations specifically to reduce code size as
1057 long as they do not significantly impact runtime performance.
1059 On a function, this attribute indicates that the function computes its
1060 result (or decides to unwind an exception) based strictly on its arguments,
1061 without dereferencing any pointer arguments or otherwise accessing
1062 any mutable state (e.g. memory, control registers, etc) visible to
1063 caller functions. It does not write through any pointer arguments
1064 (including ``byval`` arguments) and never changes any state visible
1065 to callers. This means that it cannot unwind exceptions by calling
1066 the ``C++`` exception throwing methods.
1068 On an argument, this attribute indicates that the function does not
1069 dereference that pointer argument, even though it may read or write the
1070 memory that the pointer points to if accessed through other pointers.
1072 On a function, this attribute indicates that the function does not write
1073 through any pointer arguments (including ``byval`` arguments) or otherwise
1074 modify any state (e.g. memory, control registers, etc) visible to
1075 caller functions. It may dereference pointer arguments and read
1076 state that may be set in the caller. A readonly function always
1077 returns the same value (or unwinds an exception identically) when
1078 called with the same set of arguments and global state. It cannot
1079 unwind an exception by calling the ``C++`` exception throwing
1082 On an argument, this attribute indicates that the function does not write
1083 through this pointer argument, even though it may write to the memory that
1084 the pointer points to.
1086 This attribute indicates that this function can return twice. The C
1087 ``setjmp`` is an example of such a function. The compiler disables
1088 some optimizations (like tail calls) in the caller of these
1090 ``sanitize_address``
1091 This attribute indicates that AddressSanitizer checks
1092 (dynamic address safety analysis) are enabled for this function.
1094 This attribute indicates that MemorySanitizer checks (dynamic detection
1095 of accesses to uninitialized memory) are enabled for this function.
1097 This attribute indicates that ThreadSanitizer checks
1098 (dynamic thread safety analysis) are enabled for this function.
1100 This attribute indicates that the function should emit a stack
1101 smashing protector. It is in the form of a "canary" --- a random value
1102 placed on the stack before the local variables that's checked upon
1103 return from the function to see if it has been overwritten. A
1104 heuristic is used to determine if a function needs stack protectors
1105 or not. The heuristic used will enable protectors for functions with:
1107 - Character arrays larger than ``ssp-buffer-size`` (default 8).
1108 - Aggregates containing character arrays larger than ``ssp-buffer-size``.
1109 - Calls to alloca() with variable sizes or constant sizes greater than
1110 ``ssp-buffer-size``.
1112 If a function that has an ``ssp`` attribute is inlined into a
1113 function that doesn't have an ``ssp`` attribute, then the resulting
1114 function will have an ``ssp`` attribute.
1116 This attribute indicates that the function should *always* emit a
1117 stack smashing protector. This overrides the ``ssp`` function
1120 If a function that has an ``sspreq`` attribute is inlined into a
1121 function that doesn't have an ``sspreq`` attribute or which has an
1122 ``ssp`` or ``sspstrong`` attribute, then the resulting function will have
1123 an ``sspreq`` attribute.
1125 This attribute indicates that the function should emit a stack smashing
1126 protector. This attribute causes a strong heuristic to be used when
1127 determining if a function needs stack protectors. The strong heuristic
1128 will enable protectors for functions with:
1130 - Arrays of any size and type
1131 - Aggregates containing an array of any size and type.
1132 - Calls to alloca().
1133 - Local variables that have had their address taken.
1135 This overrides the ``ssp`` function attribute.
1137 If a function that has an ``sspstrong`` attribute is inlined into a
1138 function that doesn't have an ``sspstrong`` attribute, then the
1139 resulting function will have an ``sspstrong`` attribute.
1141 This attribute indicates that the ABI being targeted requires that
1142 an unwind table entry be produce for this function even if we can
1143 show that no exceptions passes by it. This is normally the case for
1144 the ELF x86-64 abi, but it can be disabled for some compilation
1149 Module-Level Inline Assembly
1150 ----------------------------
1152 Modules may contain "module-level inline asm" blocks, which corresponds
1153 to the GCC "file scope inline asm" blocks. These blocks are internally
1154 concatenated by LLVM and treated as a single unit, but may be separated
1155 in the ``.ll`` file if desired. The syntax is very simple:
1157 .. code-block:: llvm
1159 module asm "inline asm code goes here"
1160 module asm "more can go here"
1162 The strings can contain any character by escaping non-printable
1163 characters. The escape sequence used is simply "\\xx" where "xx" is the
1164 two digit hex code for the number.
1166 The inline asm code is simply printed to the machine code .s file when
1167 assembly code is generated.
1169 .. _langref_datalayout:
1174 A module may specify a target specific data layout string that specifies
1175 how data is to be laid out in memory. The syntax for the data layout is
1178 .. code-block:: llvm
1180 target datalayout = "layout specification"
1182 The *layout specification* consists of a list of specifications
1183 separated by the minus sign character ('-'). Each specification starts
1184 with a letter and may include other information after the letter to
1185 define some aspect of the data layout. The specifications accepted are
1189 Specifies that the target lays out data in big-endian form. That is,
1190 the bits with the most significance have the lowest address
1193 Specifies that the target lays out data in little-endian form. That
1194 is, the bits with the least significance have the lowest address
1197 Specifies the natural alignment of the stack in bits. Alignment
1198 promotion of stack variables is limited to the natural stack
1199 alignment to avoid dynamic stack realignment. The stack alignment
1200 must be a multiple of 8-bits. If omitted, the natural stack
1201 alignment defaults to "unspecified", which does not prevent any
1202 alignment promotions.
1203 ``p[n]:<size>:<abi>:<pref>``
1204 This specifies the *size* of a pointer and its ``<abi>`` and
1205 ``<pref>``\erred alignments for address space ``n``. All sizes are in
1206 bits. The address space, ``n`` is optional, and if not specified,
1207 denotes the default address space 0. The value of ``n`` must be
1208 in the range [1,2^23).
1209 ``i<size>:<abi>:<pref>``
1210 This specifies the alignment for an integer type of a given bit
1211 ``<size>``. The value of ``<size>`` must be in the range [1,2^23).
1212 ``v<size>:<abi>:<pref>``
1213 This specifies the alignment for a vector type of a given bit
1215 ``f<size>:<abi>:<pref>``
1216 This specifies the alignment for a floating point type of a given bit
1217 ``<size>``. Only values of ``<size>`` that are supported by the target
1218 will work. 32 (float) and 64 (double) are supported on all targets; 80
1219 or 128 (different flavors of long double) are also supported on some
1222 This specifies the alignment for an object of aggregate type.
1224 If present, specifies that llvm names are mangled in the output. The
1227 * ``e``: ELF mangling: Private symbols get a ``.L`` prefix.
1228 * ``m``: Mips mangling: Private symbols get a ``$`` prefix.
1229 * ``o``: Mach-O mangling: Private symbols get ``L`` prefix. Other
1230 symbols get a ``_`` prefix.
1231 * ``w``: Windows COFF prefix: Similar to Mach-O, but stdcall and fastcall
1232 functions also get a suffix based on the frame size.
1233 ``n<size1>:<size2>:<size3>...``
1234 This specifies a set of native integer widths for the target CPU in
1235 bits. For example, it might contain ``n32`` for 32-bit PowerPC,
1236 ``n32:64`` for PowerPC 64, or ``n8:16:32:64`` for X86-64. Elements of
1237 this set are considered to support most general arithmetic operations
1240 On every specification that takes a ``<abi>:<pref>``, specifying the
1241 ``<pref>`` alignment is optional. If omitted, the preceding ``:``
1242 should be omitted too and ``<pref>`` will be equal to ``<abi>``.
1244 When constructing the data layout for a given target, LLVM starts with a
1245 default set of specifications which are then (possibly) overridden by
1246 the specifications in the ``datalayout`` keyword. The default
1247 specifications are given in this list:
1249 - ``E`` - big endian
1250 - ``p:64:64:64`` - 64-bit pointers with 64-bit alignment.
1251 - ``p[n]:64:64:64`` - Other address spaces are assumed to be the
1252 same as the default address space.
1253 - ``S0`` - natural stack alignment is unspecified
1254 - ``i1:8:8`` - i1 is 8-bit (byte) aligned
1255 - ``i8:8:8`` - i8 is 8-bit (byte) aligned
1256 - ``i16:16:16`` - i16 is 16-bit aligned
1257 - ``i32:32:32`` - i32 is 32-bit aligned
1258 - ``i64:32:64`` - i64 has ABI alignment of 32-bits but preferred
1259 alignment of 64-bits
1260 - ``f16:16:16`` - half is 16-bit aligned
1261 - ``f32:32:32`` - float is 32-bit aligned
1262 - ``f64:64:64`` - double is 64-bit aligned
1263 - ``f128:128:128`` - quad is 128-bit aligned
1264 - ``v64:64:64`` - 64-bit vector is 64-bit aligned
1265 - ``v128:128:128`` - 128-bit vector is 128-bit aligned
1266 - ``a:0:64`` - aggregates are 64-bit aligned
1268 When LLVM is determining the alignment for a given type, it uses the
1271 #. If the type sought is an exact match for one of the specifications,
1272 that specification is used.
1273 #. If no match is found, and the type sought is an integer type, then
1274 the smallest integer type that is larger than the bitwidth of the
1275 sought type is used. If none of the specifications are larger than
1276 the bitwidth then the largest integer type is used. For example,
1277 given the default specifications above, the i7 type will use the
1278 alignment of i8 (next largest) while both i65 and i256 will use the
1279 alignment of i64 (largest specified).
1280 #. If no match is found, and the type sought is a vector type, then the
1281 largest vector type that is smaller than the sought vector type will
1282 be used as a fall back. This happens because <128 x double> can be
1283 implemented in terms of 64 <2 x double>, for example.
1285 The function of the data layout string may not be what you expect.
1286 Notably, this is not a specification from the frontend of what alignment
1287 the code generator should use.
1289 Instead, if specified, the target data layout is required to match what
1290 the ultimate *code generator* expects. This string is used by the
1291 mid-level optimizers to improve code, and this only works if it matches
1292 what the ultimate code generator uses. If you would like to generate IR
1293 that does not embed this target-specific detail into the IR, then you
1294 don't have to specify the string. This will disable some optimizations
1295 that require precise layout information, but this also prevents those
1296 optimizations from introducing target specificity into the IR.
1303 A module may specify a target triple string that describes the target
1304 host. The syntax for the target triple is simply:
1306 .. code-block:: llvm
1308 target triple = "x86_64-apple-macosx10.7.0"
1310 The *target triple* string consists of a series of identifiers delimited
1311 by the minus sign character ('-'). The canonical forms are:
1315 ARCHITECTURE-VENDOR-OPERATING_SYSTEM
1316 ARCHITECTURE-VENDOR-OPERATING_SYSTEM-ENVIRONMENT
1318 This information is passed along to the backend so that it generates
1319 code for the proper architecture. It's possible to override this on the
1320 command line with the ``-mtriple`` command line option.
1322 .. _pointeraliasing:
1324 Pointer Aliasing Rules
1325 ----------------------
1327 Any memory access must be done through a pointer value associated with
1328 an address range of the memory access, otherwise the behavior is
1329 undefined. Pointer values are associated with address ranges according
1330 to the following rules:
1332 - A pointer value is associated with the addresses associated with any
1333 value it is *based* on.
1334 - An address of a global variable is associated with the address range
1335 of the variable's storage.
1336 - The result value of an allocation instruction is associated with the
1337 address range of the allocated storage.
1338 - A null pointer in the default address-space is associated with no
1340 - An integer constant other than zero or a pointer value returned from
1341 a function not defined within LLVM may be associated with address
1342 ranges allocated through mechanisms other than those provided by
1343 LLVM. Such ranges shall not overlap with any ranges of addresses
1344 allocated by mechanisms provided by LLVM.
1346 A pointer value is *based* on another pointer value according to the
1349 - A pointer value formed from a ``getelementptr`` operation is *based*
1350 on the first operand of the ``getelementptr``.
1351 - The result value of a ``bitcast`` is *based* on the operand of the
1353 - A pointer value formed by an ``inttoptr`` is *based* on all pointer
1354 values that contribute (directly or indirectly) to the computation of
1355 the pointer's value.
1356 - The "*based* on" relationship is transitive.
1358 Note that this definition of *"based"* is intentionally similar to the
1359 definition of *"based"* in C99, though it is slightly weaker.
1361 LLVM IR does not associate types with memory. The result type of a
1362 ``load`` merely indicates the size and alignment of the memory from
1363 which to load, as well as the interpretation of the value. The first
1364 operand type of a ``store`` similarly only indicates the size and
1365 alignment of the store.
1367 Consequently, type-based alias analysis, aka TBAA, aka
1368 ``-fstrict-aliasing``, is not applicable to general unadorned LLVM IR.
1369 :ref:`Metadata <metadata>` may be used to encode additional information
1370 which specialized optimization passes may use to implement type-based
1375 Volatile Memory Accesses
1376 ------------------------
1378 Certain memory accesses, such as :ref:`load <i_load>`'s,
1379 :ref:`store <i_store>`'s, and :ref:`llvm.memcpy <int_memcpy>`'s may be
1380 marked ``volatile``. The optimizers must not change the number of
1381 volatile operations or change their order of execution relative to other
1382 volatile operations. The optimizers *may* change the order of volatile
1383 operations relative to non-volatile operations. This is not Java's
1384 "volatile" and has no cross-thread synchronization behavior.
1386 IR-level volatile loads and stores cannot safely be optimized into
1387 llvm.memcpy or llvm.memmove intrinsics even when those intrinsics are
1388 flagged volatile. Likewise, the backend should never split or merge
1389 target-legal volatile load/store instructions.
1391 .. admonition:: Rationale
1393 Platforms may rely on volatile loads and stores of natively supported
1394 data width to be executed as single instruction. For example, in C
1395 this holds for an l-value of volatile primitive type with native
1396 hardware support, but not necessarily for aggregate types. The
1397 frontend upholds these expectations, which are intentionally
1398 unspecified in the IR. The rules above ensure that IR transformation
1399 do not violate the frontend's contract with the language.
1403 Memory Model for Concurrent Operations
1404 --------------------------------------
1406 The LLVM IR does not define any way to start parallel threads of
1407 execution or to register signal handlers. Nonetheless, there are
1408 platform-specific ways to create them, and we define LLVM IR's behavior
1409 in their presence. This model is inspired by the C++0x memory model.
1411 For a more informal introduction to this model, see the :doc:`Atomics`.
1413 We define a *happens-before* partial order as the least partial order
1416 - Is a superset of single-thread program order, and
1417 - When a *synchronizes-with* ``b``, includes an edge from ``a`` to
1418 ``b``. *Synchronizes-with* pairs are introduced by platform-specific
1419 techniques, like pthread locks, thread creation, thread joining,
1420 etc., and by atomic instructions. (See also :ref:`Atomic Memory Ordering
1421 Constraints <ordering>`).
1423 Note that program order does not introduce *happens-before* edges
1424 between a thread and signals executing inside that thread.
1426 Every (defined) read operation (load instructions, memcpy, atomic
1427 loads/read-modify-writes, etc.) R reads a series of bytes written by
1428 (defined) write operations (store instructions, atomic
1429 stores/read-modify-writes, memcpy, etc.). For the purposes of this
1430 section, initialized globals are considered to have a write of the
1431 initializer which is atomic and happens before any other read or write
1432 of the memory in question. For each byte of a read R, R\ :sub:`byte`
1433 may see any write to the same byte, except:
1435 - If write\ :sub:`1` happens before write\ :sub:`2`, and
1436 write\ :sub:`2` happens before R\ :sub:`byte`, then
1437 R\ :sub:`byte` does not see write\ :sub:`1`.
1438 - If R\ :sub:`byte` happens before write\ :sub:`3`, then
1439 R\ :sub:`byte` does not see write\ :sub:`3`.
1441 Given that definition, R\ :sub:`byte` is defined as follows:
1443 - If R is volatile, the result is target-dependent. (Volatile is
1444 supposed to give guarantees which can support ``sig_atomic_t`` in
1445 C/C++, and may be used for accesses to addresses which do not behave
1446 like normal memory. It does not generally provide cross-thread
1448 - Otherwise, if there is no write to the same byte that happens before
1449 R\ :sub:`byte`, R\ :sub:`byte` returns ``undef`` for that byte.
1450 - Otherwise, if R\ :sub:`byte` may see exactly one write,
1451 R\ :sub:`byte` returns the value written by that write.
1452 - Otherwise, if R is atomic, and all the writes R\ :sub:`byte` may
1453 see are atomic, it chooses one of the values written. See the :ref:`Atomic
1454 Memory Ordering Constraints <ordering>` section for additional
1455 constraints on how the choice is made.
1456 - Otherwise R\ :sub:`byte` returns ``undef``.
1458 R returns the value composed of the series of bytes it read. This
1459 implies that some bytes within the value may be ``undef`` **without**
1460 the entire value being ``undef``. Note that this only defines the
1461 semantics of the operation; it doesn't mean that targets will emit more
1462 than one instruction to read the series of bytes.
1464 Note that in cases where none of the atomic intrinsics are used, this
1465 model places only one restriction on IR transformations on top of what
1466 is required for single-threaded execution: introducing a store to a byte
1467 which might not otherwise be stored is not allowed in general.
1468 (Specifically, in the case where another thread might write to and read
1469 from an address, introducing a store can change a load that may see
1470 exactly one write into a load that may see multiple writes.)
1474 Atomic Memory Ordering Constraints
1475 ----------------------------------
1477 Atomic instructions (:ref:`cmpxchg <i_cmpxchg>`,
1478 :ref:`atomicrmw <i_atomicrmw>`, :ref:`fence <i_fence>`,
1479 :ref:`atomic load <i_load>`, and :ref:`atomic store <i_store>`) take
1480 an ordering parameter that determines which other atomic instructions on
1481 the same address they *synchronize with*. These semantics are borrowed
1482 from Java and C++0x, but are somewhat more colloquial. If these
1483 descriptions aren't precise enough, check those specs (see spec
1484 references in the :doc:`atomics guide <Atomics>`).
1485 :ref:`fence <i_fence>` instructions treat these orderings somewhat
1486 differently since they don't take an address. See that instruction's
1487 documentation for details.
1489 For a simpler introduction to the ordering constraints, see the
1493 The set of values that can be read is governed by the happens-before
1494 partial order. A value cannot be read unless some operation wrote
1495 it. This is intended to provide a guarantee strong enough to model
1496 Java's non-volatile shared variables. This ordering cannot be
1497 specified for read-modify-write operations; it is not strong enough
1498 to make them atomic in any interesting way.
1500 In addition to the guarantees of ``unordered``, there is a single
1501 total order for modifications by ``monotonic`` operations on each
1502 address. All modification orders must be compatible with the
1503 happens-before order. There is no guarantee that the modification
1504 orders can be combined to a global total order for the whole program
1505 (and this often will not be possible). The read in an atomic
1506 read-modify-write operation (:ref:`cmpxchg <i_cmpxchg>` and
1507 :ref:`atomicrmw <i_atomicrmw>`) reads the value in the modification
1508 order immediately before the value it writes. If one atomic read
1509 happens before another atomic read of the same address, the later
1510 read must see the same value or a later value in the address's
1511 modification order. This disallows reordering of ``monotonic`` (or
1512 stronger) operations on the same address. If an address is written
1513 ``monotonic``-ally by one thread, and other threads ``monotonic``-ally
1514 read that address repeatedly, the other threads must eventually see
1515 the write. This corresponds to the C++0x/C1x
1516 ``memory_order_relaxed``.
1518 In addition to the guarantees of ``monotonic``, a
1519 *synchronizes-with* edge may be formed with a ``release`` operation.
1520 This is intended to model C++'s ``memory_order_acquire``.
1522 In addition to the guarantees of ``monotonic``, if this operation
1523 writes a value which is subsequently read by an ``acquire``
1524 operation, it *synchronizes-with* that operation. (This isn't a
1525 complete description; see the C++0x definition of a release
1526 sequence.) This corresponds to the C++0x/C1x
1527 ``memory_order_release``.
1528 ``acq_rel`` (acquire+release)
1529 Acts as both an ``acquire`` and ``release`` operation on its
1530 address. This corresponds to the C++0x/C1x ``memory_order_acq_rel``.
1531 ``seq_cst`` (sequentially consistent)
1532 In addition to the guarantees of ``acq_rel`` (``acquire`` for an
1533 operation which only reads, ``release`` for an operation which only
1534 writes), there is a global total order on all
1535 sequentially-consistent operations on all addresses, which is
1536 consistent with the *happens-before* partial order and with the
1537 modification orders of all the affected addresses. Each
1538 sequentially-consistent read sees the last preceding write to the
1539 same address in this global order. This corresponds to the C++0x/C1x
1540 ``memory_order_seq_cst`` and Java volatile.
1544 If an atomic operation is marked ``singlethread``, it only *synchronizes
1545 with* or participates in modification and seq\_cst total orderings with
1546 other operations running in the same thread (for example, in signal
1554 LLVM IR floating-point binary ops (:ref:`fadd <i_fadd>`,
1555 :ref:`fsub <i_fsub>`, :ref:`fmul <i_fmul>`, :ref:`fdiv <i_fdiv>`,
1556 :ref:`frem <i_frem>`) have the following flags that can set to enable
1557 otherwise unsafe floating point operations
1560 No NaNs - Allow optimizations to assume the arguments and result are not
1561 NaN. Such optimizations are required to retain defined behavior over
1562 NaNs, but the value of the result is undefined.
1565 No Infs - Allow optimizations to assume the arguments and result are not
1566 +/-Inf. Such optimizations are required to retain defined behavior over
1567 +/-Inf, but the value of the result is undefined.
1570 No Signed Zeros - Allow optimizations to treat the sign of a zero
1571 argument or result as insignificant.
1574 Allow Reciprocal - Allow optimizations to use the reciprocal of an
1575 argument rather than perform division.
1578 Fast - Allow algebraically equivalent transformations that may
1579 dramatically change results in floating point (e.g. reassociate). This
1580 flag implies all the others.
1587 The LLVM type system is one of the most important features of the
1588 intermediate representation. Being typed enables a number of
1589 optimizations to be performed on the intermediate representation
1590 directly, without having to do extra analyses on the side before the
1591 transformation. A strong type system makes it easier to read the
1592 generated code and enables novel analyses and transformations that are
1593 not feasible to perform on normal three address code representations.
1603 The void type does not represent any value and has no size.
1621 The function type can be thought of as a function signature. It consists of a
1622 return type and a list of formal parameter types. The return type of a function
1623 type is a void type or first class type --- except for :ref:`label <t_label>`
1624 and :ref:`metadata <t_metadata>` types.
1630 <returntype> (<parameter list>)
1632 ...where '``<parameter list>``' is a comma-separated list of type
1633 specifiers. Optionally, the parameter list may include a type ``...``, which
1634 indicates that the function takes a variable number of arguments. Variable
1635 argument functions can access their arguments with the :ref:`variable argument
1636 handling intrinsic <int_varargs>` functions. '``<returntype>``' is any type
1637 except :ref:`label <t_label>` and :ref:`metadata <t_metadata>`.
1641 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1642 | ``i32 (i32)`` | function taking an ``i32``, returning an ``i32`` |
1643 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1644 | ``float (i16, i32 *) *`` | :ref:`Pointer <t_pointer>` to a function that takes an ``i16`` and a :ref:`pointer <t_pointer>` to ``i32``, returning ``float``. |
1645 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1646 | ``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. |
1647 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1648 | ``{i32, i32} (i32)`` | A function taking an ``i32``, returning a :ref:`structure <t_struct>` containing two ``i32`` values |
1649 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1656 The :ref:`first class <t_firstclass>` types are perhaps the most important.
1657 Values of these types are the only ones which can be produced by
1665 These are the types that are valid in registers from CodeGen's perspective.
1674 The integer type is a very simple type that simply specifies an
1675 arbitrary bit width for the integer type desired. Any bit width from 1
1676 bit to 2\ :sup:`23`\ -1 (about 8 million) can be specified.
1684 The number of bits the integer will occupy is specified by the ``N``
1690 +----------------+------------------------------------------------+
1691 | ``i1`` | a single-bit integer. |
1692 +----------------+------------------------------------------------+
1693 | ``i32`` | a 32-bit integer. |
1694 +----------------+------------------------------------------------+
1695 | ``i1942652`` | a really big integer of over 1 million bits. |
1696 +----------------+------------------------------------------------+
1700 Floating Point Types
1701 """"""""""""""""""""
1710 - 16-bit floating point value
1713 - 32-bit floating point value
1716 - 64-bit floating point value
1719 - 128-bit floating point value (112-bit mantissa)
1722 - 80-bit floating point value (X87)
1725 - 128-bit floating point value (two 64-bits)
1734 The x86mmx type represents a value held in an MMX register on an x86
1735 machine. The operations allowed on it are quite limited: parameters and
1736 return values, load and store, and bitcast. User-specified MMX
1737 instructions are represented as intrinsic or asm calls with arguments
1738 and/or results of this type. There are no arrays, vectors or constants
1755 The pointer type is used to specify memory locations. Pointers are
1756 commonly used to reference objects in memory.
1758 Pointer types may have an optional address space attribute defining the
1759 numbered address space where the pointed-to object resides. The default
1760 address space is number zero. The semantics of non-zero address spaces
1761 are target-specific.
1763 Note that LLVM does not permit pointers to void (``void*``) nor does it
1764 permit pointers to labels (``label*``). Use ``i8*`` instead.
1774 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1775 | ``[4 x i32]*`` | A :ref:`pointer <t_pointer>` to :ref:`array <t_array>` of four ``i32`` values. |
1776 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1777 | ``i32 (i32*) *`` | A :ref:`pointer <t_pointer>` to a :ref:`function <t_function>` that takes an ``i32*``, returning an ``i32``. |
1778 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1779 | ``i32 addrspace(5)*`` | A :ref:`pointer <t_pointer>` to an ``i32`` value that resides in address space #5. |
1780 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1789 A vector type is a simple derived type that represents a vector of
1790 elements. Vector types are used when multiple primitive data are
1791 operated in parallel using a single instruction (SIMD). A vector type
1792 requires a size (number of elements) and an underlying primitive data
1793 type. Vector types are considered :ref:`first class <t_firstclass>`.
1799 < <# elements> x <elementtype> >
1801 The number of elements is a constant integer value larger than 0;
1802 elementtype may be any integer or floating point type, or a pointer to
1803 these types. Vectors of size zero are not allowed.
1807 +-------------------+--------------------------------------------------+
1808 | ``<4 x i32>`` | Vector of 4 32-bit integer values. |
1809 +-------------------+--------------------------------------------------+
1810 | ``<8 x float>`` | Vector of 8 32-bit floating-point values. |
1811 +-------------------+--------------------------------------------------+
1812 | ``<2 x i64>`` | Vector of 2 64-bit integer values. |
1813 +-------------------+--------------------------------------------------+
1814 | ``<4 x i64*>`` | Vector of 4 pointers to 64-bit integer values. |
1815 +-------------------+--------------------------------------------------+
1824 The label type represents code labels.
1839 The metadata type represents embedded metadata. No derived types may be
1840 created from metadata except for :ref:`function <t_function>` arguments.
1853 Aggregate Types are a subset of derived types that can contain multiple
1854 member types. :ref:`Arrays <t_array>` and :ref:`structs <t_struct>` are
1855 aggregate types. :ref:`Vectors <t_vector>` are not considered to be
1865 The array type is a very simple derived type that arranges elements
1866 sequentially in memory. The array type requires a size (number of
1867 elements) and an underlying data type.
1873 [<# elements> x <elementtype>]
1875 The number of elements is a constant integer value; ``elementtype`` may
1876 be any type with a size.
1880 +------------------+--------------------------------------+
1881 | ``[40 x i32]`` | Array of 40 32-bit integer values. |
1882 +------------------+--------------------------------------+
1883 | ``[41 x i32]`` | Array of 41 32-bit integer values. |
1884 +------------------+--------------------------------------+
1885 | ``[4 x i8]`` | Array of 4 8-bit integer values. |
1886 +------------------+--------------------------------------+
1888 Here are some examples of multidimensional arrays:
1890 +-----------------------------+----------------------------------------------------------+
1891 | ``[3 x [4 x i32]]`` | 3x4 array of 32-bit integer values. |
1892 +-----------------------------+----------------------------------------------------------+
1893 | ``[12 x [10 x float]]`` | 12x10 array of single precision floating point values. |
1894 +-----------------------------+----------------------------------------------------------+
1895 | ``[2 x [3 x [4 x i16]]]`` | 2x3x4 array of 16-bit integer values. |
1896 +-----------------------------+----------------------------------------------------------+
1898 There is no restriction on indexing beyond the end of the array implied
1899 by a static type (though there are restrictions on indexing beyond the
1900 bounds of an allocated object in some cases). This means that
1901 single-dimension 'variable sized array' addressing can be implemented in
1902 LLVM with a zero length array type. An implementation of 'pascal style
1903 arrays' in LLVM could use the type "``{ i32, [0 x float]}``", for
1913 The structure type is used to represent a collection of data members
1914 together in memory. The elements of a structure may be any type that has
1917 Structures in memory are accessed using '``load``' and '``store``' by
1918 getting a pointer to a field with the '``getelementptr``' instruction.
1919 Structures in registers are accessed using the '``extractvalue``' and
1920 '``insertvalue``' instructions.
1922 Structures may optionally be "packed" structures, which indicate that
1923 the alignment of the struct is one byte, and that there is no padding
1924 between the elements. In non-packed structs, padding between field types
1925 is inserted as defined by the DataLayout string in the module, which is
1926 required to match what the underlying code generator expects.
1928 Structures can either be "literal" or "identified". A literal structure
1929 is defined inline with other types (e.g. ``{i32, i32}*``) whereas
1930 identified types are always defined at the top level with a name.
1931 Literal types are uniqued by their contents and can never be recursive
1932 or opaque since there is no way to write one. Identified types can be
1933 recursive, can be opaqued, and are never uniqued.
1939 %T1 = type { <type list> } ; Identified normal struct type
1940 %T2 = type <{ <type list> }> ; Identified packed struct type
1944 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1945 | ``{ i32, i32, i32 }`` | A triple of three ``i32`` values |
1946 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1947 | ``{ 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``. |
1948 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1949 | ``<{ i8, i32 }>`` | A packed struct known to be 5 bytes in size. |
1950 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1954 Opaque Structure Types
1955 """"""""""""""""""""""
1959 Opaque structure types are used to represent named structure types that
1960 do not have a body specified. This corresponds (for example) to the C
1961 notion of a forward declared structure.
1972 +--------------+-------------------+
1973 | ``opaque`` | An opaque type. |
1974 +--------------+-------------------+
1979 LLVM has several different basic types of constants. This section
1980 describes them all and their syntax.
1985 **Boolean constants**
1986 The two strings '``true``' and '``false``' are both valid constants
1988 **Integer constants**
1989 Standard integers (such as '4') are constants of the
1990 :ref:`integer <t_integer>` type. Negative numbers may be used with
1992 **Floating point constants**
1993 Floating point constants use standard decimal notation (e.g.
1994 123.421), exponential notation (e.g. 1.23421e+2), or a more precise
1995 hexadecimal notation (see below). The assembler requires the exact
1996 decimal value of a floating-point constant. For example, the
1997 assembler accepts 1.25 but rejects 1.3 because 1.3 is a repeating
1998 decimal in binary. Floating point constants must have a :ref:`floating
1999 point <t_floating>` type.
2000 **Null pointer constants**
2001 The identifier '``null``' is recognized as a null pointer constant
2002 and must be of :ref:`pointer type <t_pointer>`.
2004 The one non-intuitive notation for constants is the hexadecimal form of
2005 floating point constants. For example, the form
2006 '``double 0x432ff973cafa8000``' is equivalent to (but harder to read
2007 than) '``double 4.5e+15``'. The only time hexadecimal floating point
2008 constants are required (and the only time that they are generated by the
2009 disassembler) is when a floating point constant must be emitted but it
2010 cannot be represented as a decimal floating point number in a reasonable
2011 number of digits. For example, NaN's, infinities, and other special
2012 values are represented in their IEEE hexadecimal format so that assembly
2013 and disassembly do not cause any bits to change in the constants.
2015 When using the hexadecimal form, constants of types half, float, and
2016 double are represented using the 16-digit form shown above (which
2017 matches the IEEE754 representation for double); half and float values
2018 must, however, be exactly representable as IEEE 754 half and single
2019 precision, respectively. Hexadecimal format is always used for long
2020 double, and there are three forms of long double. The 80-bit format used
2021 by x86 is represented as ``0xK`` followed by 20 hexadecimal digits. The
2022 128-bit format used by PowerPC (two adjacent doubles) is represented by
2023 ``0xM`` followed by 32 hexadecimal digits. The IEEE 128-bit format is
2024 represented by ``0xL`` followed by 32 hexadecimal digits. Long doubles
2025 will only work if they match the long double format on your target.
2026 The IEEE 16-bit format (half precision) is represented by ``0xH``
2027 followed by 4 hexadecimal digits. All hexadecimal formats are big-endian
2028 (sign bit at the left).
2030 There are no constants of type x86mmx.
2032 .. _complexconstants:
2037 Complex constants are a (potentially recursive) combination of simple
2038 constants and smaller complex constants.
2040 **Structure constants**
2041 Structure constants are represented with notation similar to
2042 structure type definitions (a comma separated list of elements,
2043 surrounded by braces (``{}``)). For example:
2044 "``{ i32 4, float 17.0, i32* @G }``", where "``@G``" is declared as
2045 "``@G = external global i32``". Structure constants must have
2046 :ref:`structure type <t_struct>`, and the number and types of elements
2047 must match those specified by the type.
2049 Array constants are represented with notation similar to array type
2050 definitions (a comma separated list of elements, surrounded by
2051 square brackets (``[]``)). For example:
2052 "``[ i32 42, i32 11, i32 74 ]``". Array constants must have
2053 :ref:`array type <t_array>`, and the number and types of elements must
2054 match those specified by the type.
2055 **Vector constants**
2056 Vector constants are represented with notation similar to vector
2057 type definitions (a comma separated list of elements, surrounded by
2058 less-than/greater-than's (``<>``)). For example:
2059 "``< i32 42, i32 11, i32 74, i32 100 >``". Vector constants
2060 must have :ref:`vector type <t_vector>`, and the number and types of
2061 elements must match those specified by the type.
2062 **Zero initialization**
2063 The string '``zeroinitializer``' can be used to zero initialize a
2064 value to zero of *any* type, including scalar and
2065 :ref:`aggregate <t_aggregate>` types. This is often used to avoid
2066 having to print large zero initializers (e.g. for large arrays) and
2067 is always exactly equivalent to using explicit zero initializers.
2069 A metadata node is a structure-like constant with :ref:`metadata
2070 type <t_metadata>`. For example:
2071 "``metadata !{ i32 0, metadata !"test" }``". Unlike other
2072 constants that are meant to be interpreted as part of the
2073 instruction stream, metadata is a place to attach additional
2074 information such as debug info.
2076 Global Variable and Function Addresses
2077 --------------------------------------
2079 The addresses of :ref:`global variables <globalvars>` and
2080 :ref:`functions <functionstructure>` are always implicitly valid
2081 (link-time) constants. These constants are explicitly referenced when
2082 the :ref:`identifier for the global <identifiers>` is used and always have
2083 :ref:`pointer <t_pointer>` type. For example, the following is a legal LLVM
2086 .. code-block:: llvm
2090 @Z = global [2 x i32*] [ i32* @X, i32* @Y ]
2097 The string '``undef``' can be used anywhere a constant is expected, and
2098 indicates that the user of the value may receive an unspecified
2099 bit-pattern. Undefined values may be of any type (other than '``label``'
2100 or '``void``') and be used anywhere a constant is permitted.
2102 Undefined values are useful because they indicate to the compiler that
2103 the program is well defined no matter what value is used. This gives the
2104 compiler more freedom to optimize. Here are some examples of
2105 (potentially surprising) transformations that are valid (in pseudo IR):
2107 .. code-block:: llvm
2117 This is safe because all of the output bits are affected by the undef
2118 bits. Any output bit can have a zero or one depending on the input bits.
2120 .. code-block:: llvm
2131 These logical operations have bits that are not always affected by the
2132 input. For example, if ``%X`` has a zero bit, then the output of the
2133 '``and``' operation will always be a zero for that bit, no matter what
2134 the corresponding bit from the '``undef``' is. As such, it is unsafe to
2135 optimize or assume that the result of the '``and``' is '``undef``'.
2136 However, it is safe to assume that all bits of the '``undef``' could be
2137 0, and optimize the '``and``' to 0. Likewise, it is safe to assume that
2138 all the bits of the '``undef``' operand to the '``or``' could be set,
2139 allowing the '``or``' to be folded to -1.
2141 .. code-block:: llvm
2143 %A = select undef, %X, %Y
2144 %B = select undef, 42, %Y
2145 %C = select %X, %Y, undef
2155 This set of examples shows that undefined '``select``' (and conditional
2156 branch) conditions can go *either way*, but they have to come from one
2157 of the two operands. In the ``%A`` example, if ``%X`` and ``%Y`` were
2158 both known to have a clear low bit, then ``%A`` would have to have a
2159 cleared low bit. However, in the ``%C`` example, the optimizer is
2160 allowed to assume that the '``undef``' operand could be the same as
2161 ``%Y``, allowing the whole '``select``' to be eliminated.
2163 .. code-block:: llvm
2165 %A = xor undef, undef
2182 This example points out that two '``undef``' operands are not
2183 necessarily the same. This can be surprising to people (and also matches
2184 C semantics) where they assume that "``X^X``" is always zero, even if
2185 ``X`` is undefined. This isn't true for a number of reasons, but the
2186 short answer is that an '``undef``' "variable" can arbitrarily change
2187 its value over its "live range". This is true because the variable
2188 doesn't actually *have a live range*. Instead, the value is logically
2189 read from arbitrary registers that happen to be around when needed, so
2190 the value is not necessarily consistent over time. In fact, ``%A`` and
2191 ``%C`` need to have the same semantics or the core LLVM "replace all
2192 uses with" concept would not hold.
2194 .. code-block:: llvm
2202 These examples show the crucial difference between an *undefined value*
2203 and *undefined behavior*. An undefined value (like '``undef``') is
2204 allowed to have an arbitrary bit-pattern. This means that the ``%A``
2205 operation can be constant folded to '``undef``', because the '``undef``'
2206 could be an SNaN, and ``fdiv`` is not (currently) defined on SNaN's.
2207 However, in the second example, we can make a more aggressive
2208 assumption: because the ``undef`` is allowed to be an arbitrary value,
2209 we are allowed to assume that it could be zero. Since a divide by zero
2210 has *undefined behavior*, we are allowed to assume that the operation
2211 does not execute at all. This allows us to delete the divide and all
2212 code after it. Because the undefined operation "can't happen", the
2213 optimizer can assume that it occurs in dead code.
2215 .. code-block:: llvm
2217 a: store undef -> %X
2218 b: store %X -> undef
2223 These examples reiterate the ``fdiv`` example: a store *of* an undefined
2224 value can be assumed to not have any effect; we can assume that the
2225 value is overwritten with bits that happen to match what was already
2226 there. However, a store *to* an undefined location could clobber
2227 arbitrary memory, therefore, it has undefined behavior.
2234 Poison values are similar to :ref:`undef values <undefvalues>`, however
2235 they also represent the fact that an instruction or constant expression
2236 which cannot evoke side effects has nevertheless detected a condition
2237 which results in undefined behavior.
2239 There is currently no way of representing a poison value in the IR; they
2240 only exist when produced by operations such as :ref:`add <i_add>` with
2243 Poison value behavior is defined in terms of value *dependence*:
2245 - Values other than :ref:`phi <i_phi>` nodes depend on their operands.
2246 - :ref:`Phi <i_phi>` nodes depend on the operand corresponding to
2247 their dynamic predecessor basic block.
2248 - Function arguments depend on the corresponding actual argument values
2249 in the dynamic callers of their functions.
2250 - :ref:`Call <i_call>` instructions depend on the :ref:`ret <i_ret>`
2251 instructions that dynamically transfer control back to them.
2252 - :ref:`Invoke <i_invoke>` instructions depend on the
2253 :ref:`ret <i_ret>`, :ref:`resume <i_resume>`, or exception-throwing
2254 call instructions that dynamically transfer control back to them.
2255 - Non-volatile loads and stores depend on the most recent stores to all
2256 of the referenced memory addresses, following the order in the IR
2257 (including loads and stores implied by intrinsics such as
2258 :ref:`@llvm.memcpy <int_memcpy>`.)
2259 - An instruction with externally visible side effects depends on the
2260 most recent preceding instruction with externally visible side
2261 effects, following the order in the IR. (This includes :ref:`volatile
2262 operations <volatile>`.)
2263 - An instruction *control-depends* on a :ref:`terminator
2264 instruction <terminators>` if the terminator instruction has
2265 multiple successors and the instruction is always executed when
2266 control transfers to one of the successors, and may not be executed
2267 when control is transferred to another.
2268 - Additionally, an instruction also *control-depends* on a terminator
2269 instruction if the set of instructions it otherwise depends on would
2270 be different if the terminator had transferred control to a different
2272 - Dependence is transitive.
2274 Poison Values have the same behavior as :ref:`undef values <undefvalues>`,
2275 with the additional affect that any instruction which has a *dependence*
2276 on a poison value has undefined behavior.
2278 Here are some examples:
2280 .. code-block:: llvm
2283 %poison = sub nuw i32 0, 1 ; Results in a poison value.
2284 %still_poison = and i32 %poison, 0 ; 0, but also poison.
2285 %poison_yet_again = getelementptr i32* @h, i32 %still_poison
2286 store i32 0, i32* %poison_yet_again ; memory at @h[0] is poisoned
2288 store i32 %poison, i32* @g ; Poison value stored to memory.
2289 %poison2 = load i32* @g ; Poison value loaded back from memory.
2291 store volatile i32 %poison, i32* @g ; External observation; undefined behavior.
2293 %narrowaddr = bitcast i32* @g to i16*
2294 %wideaddr = bitcast i32* @g to i64*
2295 %poison3 = load i16* %narrowaddr ; Returns a poison value.
2296 %poison4 = load i64* %wideaddr ; Returns a poison value.
2298 %cmp = icmp slt i32 %poison, 0 ; Returns a poison value.
2299 br i1 %cmp, label %true, label %end ; Branch to either destination.
2302 store volatile i32 0, i32* @g ; This is control-dependent on %cmp, so
2303 ; it has undefined behavior.
2307 %p = phi i32 [ 0, %entry ], [ 1, %true ]
2308 ; Both edges into this PHI are
2309 ; control-dependent on %cmp, so this
2310 ; always results in a poison value.
2312 store volatile i32 0, i32* @g ; This would depend on the store in %true
2313 ; if %cmp is true, or the store in %entry
2314 ; otherwise, so this is undefined behavior.
2316 br i1 %cmp, label %second_true, label %second_end
2317 ; The same branch again, but this time the
2318 ; true block doesn't have side effects.
2325 store volatile i32 0, i32* @g ; This time, the instruction always depends
2326 ; on the store in %end. Also, it is
2327 ; control-equivalent to %end, so this is
2328 ; well-defined (ignoring earlier undefined
2329 ; behavior in this example).
2333 Addresses of Basic Blocks
2334 -------------------------
2336 ``blockaddress(@function, %block)``
2338 The '``blockaddress``' constant computes the address of the specified
2339 basic block in the specified function, and always has an ``i8*`` type.
2340 Taking the address of the entry block is illegal.
2342 This value only has defined behavior when used as an operand to the
2343 ':ref:`indirectbr <i_indirectbr>`' instruction, or for comparisons
2344 against null. Pointer equality tests between labels addresses results in
2345 undefined behavior --- though, again, comparison against null is ok, and
2346 no label is equal to the null pointer. This may be passed around as an
2347 opaque pointer sized value as long as the bits are not inspected. This
2348 allows ``ptrtoint`` and arithmetic to be performed on these values so
2349 long as the original value is reconstituted before the ``indirectbr``
2352 Finally, some targets may provide defined semantics when using the value
2353 as the operand to an inline assembly, but that is target specific.
2357 Constant Expressions
2358 --------------------
2360 Constant expressions are used to allow expressions involving other
2361 constants to be used as constants. Constant expressions may be of any
2362 :ref:`first class <t_firstclass>` type and may involve any LLVM operation
2363 that does not have side effects (e.g. load and call are not supported).
2364 The following is the syntax for constant expressions:
2366 ``trunc (CST to TYPE)``
2367 Truncate a constant to another type. The bit size of CST must be
2368 larger than the bit size of TYPE. Both types must be integers.
2369 ``zext (CST to TYPE)``
2370 Zero extend a constant to another type. The bit size of CST must be
2371 smaller than the bit size of TYPE. Both types must be integers.
2372 ``sext (CST to TYPE)``
2373 Sign extend a constant to another type. The bit size of CST must be
2374 smaller than the bit size of TYPE. Both types must be integers.
2375 ``fptrunc (CST to TYPE)``
2376 Truncate a floating point constant to another floating point type.
2377 The size of CST must be larger than the size of TYPE. Both types
2378 must be floating point.
2379 ``fpext (CST to TYPE)``
2380 Floating point extend a constant to another type. The size of CST
2381 must be smaller or equal to the size of TYPE. Both types must be
2383 ``fptoui (CST to TYPE)``
2384 Convert a floating point constant to the corresponding unsigned
2385 integer constant. TYPE must be a scalar or vector integer type. CST
2386 must be of scalar or vector floating point type. Both CST and TYPE
2387 must be scalars, or vectors of the same number of elements. If the
2388 value won't fit in the integer type, the results are undefined.
2389 ``fptosi (CST to TYPE)``
2390 Convert a floating point constant to the corresponding signed
2391 integer constant. TYPE must be a scalar or vector integer type. CST
2392 must be of scalar or vector floating point type. Both CST and TYPE
2393 must be scalars, or vectors of the same number of elements. If the
2394 value won't fit in the integer type, the results are undefined.
2395 ``uitofp (CST to TYPE)``
2396 Convert an unsigned integer constant to the corresponding floating
2397 point constant. TYPE must be a scalar or vector floating point type.
2398 CST must be of scalar or vector integer type. Both CST and TYPE must
2399 be scalars, or vectors of the same number of elements. If the value
2400 won't fit in the floating point type, the results are undefined.
2401 ``sitofp (CST to TYPE)``
2402 Convert a signed integer constant to the corresponding floating
2403 point constant. TYPE must be a scalar or vector floating point type.
2404 CST must be of scalar or vector integer type. Both CST and TYPE must
2405 be scalars, or vectors of the same number of elements. If the value
2406 won't fit in the floating point type, the results are undefined.
2407 ``ptrtoint (CST to TYPE)``
2408 Convert a pointer typed constant to the corresponding integer
2409 constant. ``TYPE`` must be an integer type. ``CST`` must be of
2410 pointer type. The ``CST`` value is zero extended, truncated, or
2411 unchanged to make it fit in ``TYPE``.
2412 ``inttoptr (CST to TYPE)``
2413 Convert an integer constant to a pointer constant. TYPE must be a
2414 pointer type. CST must be of integer type. The CST value is zero
2415 extended, truncated, or unchanged to make it fit in a pointer size.
2416 This one is *really* dangerous!
2417 ``bitcast (CST to TYPE)``
2418 Convert a constant, CST, to another TYPE. The constraints of the
2419 operands are the same as those for the :ref:`bitcast
2420 instruction <i_bitcast>`.
2421 ``addrspacecast (CST to TYPE)``
2422 Convert a constant pointer or constant vector of pointer, CST, to another
2423 TYPE in a different address space. The constraints of the operands are the
2424 same as those for the :ref:`addrspacecast instruction <i_addrspacecast>`.
2425 ``getelementptr (CSTPTR, IDX0, IDX1, ...)``, ``getelementptr inbounds (CSTPTR, IDX0, IDX1, ...)``
2426 Perform the :ref:`getelementptr operation <i_getelementptr>` on
2427 constants. As with the :ref:`getelementptr <i_getelementptr>`
2428 instruction, the index list may have zero or more indexes, which are
2429 required to make sense for the type of "CSTPTR".
2430 ``select (COND, VAL1, VAL2)``
2431 Perform the :ref:`select operation <i_select>` on constants.
2432 ``icmp COND (VAL1, VAL2)``
2433 Performs the :ref:`icmp operation <i_icmp>` on constants.
2434 ``fcmp COND (VAL1, VAL2)``
2435 Performs the :ref:`fcmp operation <i_fcmp>` on constants.
2436 ``extractelement (VAL, IDX)``
2437 Perform the :ref:`extractelement operation <i_extractelement>` on
2439 ``insertelement (VAL, ELT, IDX)``
2440 Perform the :ref:`insertelement operation <i_insertelement>` on
2442 ``shufflevector (VEC1, VEC2, IDXMASK)``
2443 Perform the :ref:`shufflevector operation <i_shufflevector>` on
2445 ``extractvalue (VAL, IDX0, IDX1, ...)``
2446 Perform the :ref:`extractvalue operation <i_extractvalue>` on
2447 constants. The index list is interpreted in a similar manner as
2448 indices in a ':ref:`getelementptr <i_getelementptr>`' operation. At
2449 least one index value must be specified.
2450 ``insertvalue (VAL, ELT, IDX0, IDX1, ...)``
2451 Perform the :ref:`insertvalue operation <i_insertvalue>` on constants.
2452 The index list is interpreted in a similar manner as indices in a
2453 ':ref:`getelementptr <i_getelementptr>`' operation. At least one index
2454 value must be specified.
2455 ``OPCODE (LHS, RHS)``
2456 Perform the specified operation of the LHS and RHS constants. OPCODE
2457 may be any of the :ref:`binary <binaryops>` or :ref:`bitwise
2458 binary <bitwiseops>` operations. The constraints on operands are
2459 the same as those for the corresponding instruction (e.g. no bitwise
2460 operations on floating point values are allowed).
2467 Inline Assembler Expressions
2468 ----------------------------
2470 LLVM supports inline assembler expressions (as opposed to :ref:`Module-Level
2471 Inline Assembly <moduleasm>`) through the use of a special value. This
2472 value represents the inline assembler as a string (containing the
2473 instructions to emit), a list of operand constraints (stored as a
2474 string), a flag that indicates whether or not the inline asm expression
2475 has side effects, and a flag indicating whether the function containing
2476 the asm needs to align its stack conservatively. An example inline
2477 assembler expression is:
2479 .. code-block:: llvm
2481 i32 (i32) asm "bswap $0", "=r,r"
2483 Inline assembler expressions may **only** be used as the callee operand
2484 of a :ref:`call <i_call>` or an :ref:`invoke <i_invoke>` instruction.
2485 Thus, typically we have:
2487 .. code-block:: llvm
2489 %X = call i32 asm "bswap $0", "=r,r"(i32 %Y)
2491 Inline asms with side effects not visible in the constraint list must be
2492 marked as having side effects. This is done through the use of the
2493 '``sideeffect``' keyword, like so:
2495 .. code-block:: llvm
2497 call void asm sideeffect "eieio", ""()
2499 In some cases inline asms will contain code that will not work unless
2500 the stack is aligned in some way, such as calls or SSE instructions on
2501 x86, yet will not contain code that does that alignment within the asm.
2502 The compiler should make conservative assumptions about what the asm
2503 might contain and should generate its usual stack alignment code in the
2504 prologue if the '``alignstack``' keyword is present:
2506 .. code-block:: llvm
2508 call void asm alignstack "eieio", ""()
2510 Inline asms also support using non-standard assembly dialects. The
2511 assumed dialect is ATT. When the '``inteldialect``' keyword is present,
2512 the inline asm is using the Intel dialect. Currently, ATT and Intel are
2513 the only supported dialects. An example is:
2515 .. code-block:: llvm
2517 call void asm inteldialect "eieio", ""()
2519 If multiple keywords appear the '``sideeffect``' keyword must come
2520 first, the '``alignstack``' keyword second and the '``inteldialect``'
2526 The call instructions that wrap inline asm nodes may have a
2527 "``!srcloc``" MDNode attached to it that contains a list of constant
2528 integers. If present, the code generator will use the integer as the
2529 location cookie value when report errors through the ``LLVMContext``
2530 error reporting mechanisms. This allows a front-end to correlate backend
2531 errors that occur with inline asm back to the source code that produced
2534 .. code-block:: llvm
2536 call void asm sideeffect "something bad", ""(), !srcloc !42
2538 !42 = !{ i32 1234567 }
2540 It is up to the front-end to make sense of the magic numbers it places
2541 in the IR. If the MDNode contains multiple constants, the code generator
2542 will use the one that corresponds to the line of the asm that the error
2547 Metadata Nodes and Metadata Strings
2548 -----------------------------------
2550 LLVM IR allows metadata to be attached to instructions in the program
2551 that can convey extra information about the code to the optimizers and
2552 code generator. One example application of metadata is source-level
2553 debug information. There are two metadata primitives: strings and nodes.
2554 All metadata has the ``metadata`` type and is identified in syntax by a
2555 preceding exclamation point ('``!``').
2557 A metadata string is a string surrounded by double quotes. It can
2558 contain any character by escaping non-printable characters with
2559 "``\xx``" where "``xx``" is the two digit hex code. For example:
2562 Metadata nodes are represented with notation similar to structure
2563 constants (a comma separated list of elements, surrounded by braces and
2564 preceded by an exclamation point). Metadata nodes can have any values as
2565 their operand. For example:
2567 .. code-block:: llvm
2569 !{ metadata !"test\00", i32 10}
2571 A :ref:`named metadata <namedmetadatastructure>` is a collection of
2572 metadata nodes, which can be looked up in the module symbol table. For
2575 .. code-block:: llvm
2577 !foo = metadata !{!4, !3}
2579 Metadata can be used as function arguments. Here ``llvm.dbg.value``
2580 function is using two metadata arguments:
2582 .. code-block:: llvm
2584 call void @llvm.dbg.value(metadata !24, i64 0, metadata !25)
2586 Metadata can be attached with an instruction. Here metadata ``!21`` is
2587 attached to the ``add`` instruction using the ``!dbg`` identifier:
2589 .. code-block:: llvm
2591 %indvar.next = add i64 %indvar, 1, !dbg !21
2593 More information about specific metadata nodes recognized by the
2594 optimizers and code generator is found below.
2599 In LLVM IR, memory does not have types, so LLVM's own type system is not
2600 suitable for doing TBAA. Instead, metadata is added to the IR to
2601 describe a type system of a higher level language. This can be used to
2602 implement typical C/C++ TBAA, but it can also be used to implement
2603 custom alias analysis behavior for other languages.
2605 The current metadata format is very simple. TBAA metadata nodes have up
2606 to three fields, e.g.:
2608 .. code-block:: llvm
2610 !0 = metadata !{ metadata !"an example type tree" }
2611 !1 = metadata !{ metadata !"int", metadata !0 }
2612 !2 = metadata !{ metadata !"float", metadata !0 }
2613 !3 = metadata !{ metadata !"const float", metadata !2, i64 1 }
2615 The first field is an identity field. It can be any value, usually a
2616 metadata string, which uniquely identifies the type. The most important
2617 name in the tree is the name of the root node. Two trees with different
2618 root node names are entirely disjoint, even if they have leaves with
2621 The second field identifies the type's parent node in the tree, or is
2622 null or omitted for a root node. A type is considered to alias all of
2623 its descendants and all of its ancestors in the tree. Also, a type is
2624 considered to alias all types in other trees, so that bitcode produced
2625 from multiple front-ends is handled conservatively.
2627 If the third field is present, it's an integer which if equal to 1
2628 indicates that the type is "constant" (meaning
2629 ``pointsToConstantMemory`` should return true; see `other useful
2630 AliasAnalysis methods <AliasAnalysis.html#OtherItfs>`_).
2632 '``tbaa.struct``' Metadata
2633 ^^^^^^^^^^^^^^^^^^^^^^^^^^
2635 The :ref:`llvm.memcpy <int_memcpy>` is often used to implement
2636 aggregate assignment operations in C and similar languages, however it
2637 is defined to copy a contiguous region of memory, which is more than
2638 strictly necessary for aggregate types which contain holes due to
2639 padding. Also, it doesn't contain any TBAA information about the fields
2642 ``!tbaa.struct`` metadata can describe which memory subregions in a
2643 memcpy are padding and what the TBAA tags of the struct are.
2645 The current metadata format is very simple. ``!tbaa.struct`` metadata
2646 nodes are a list of operands which are in conceptual groups of three.
2647 For each group of three, the first operand gives the byte offset of a
2648 field in bytes, the second gives its size in bytes, and the third gives
2651 .. code-block:: llvm
2653 !4 = metadata !{ i64 0, i64 4, metadata !1, i64 8, i64 4, metadata !2 }
2655 This describes a struct with two fields. The first is at offset 0 bytes
2656 with size 4 bytes, and has tbaa tag !1. The second is at offset 8 bytes
2657 and has size 4 bytes and has tbaa tag !2.
2659 Note that the fields need not be contiguous. In this example, there is a
2660 4 byte gap between the two fields. This gap represents padding which
2661 does not carry useful data and need not be preserved.
2663 '``fpmath``' Metadata
2664 ^^^^^^^^^^^^^^^^^^^^^
2666 ``fpmath`` metadata may be attached to any instruction of floating point
2667 type. It can be used to express the maximum acceptable error in the
2668 result of that instruction, in ULPs, thus potentially allowing the
2669 compiler to use a more efficient but less accurate method of computing
2670 it. ULP is defined as follows:
2672 If ``x`` is a real number that lies between two finite consecutive
2673 floating-point numbers ``a`` and ``b``, without being equal to one
2674 of them, then ``ulp(x) = |b - a|``, otherwise ``ulp(x)`` is the
2675 distance between the two non-equal finite floating-point numbers
2676 nearest ``x``. Moreover, ``ulp(NaN)`` is ``NaN``.
2678 The metadata node shall consist of a single positive floating point
2679 number representing the maximum relative error, for example:
2681 .. code-block:: llvm
2683 !0 = metadata !{ float 2.5 } ; maximum acceptable inaccuracy is 2.5 ULPs
2685 '``range``' Metadata
2686 ^^^^^^^^^^^^^^^^^^^^
2688 ``range`` metadata may be attached only to loads of integer types. It
2689 expresses the possible ranges the loaded value is in. The ranges are
2690 represented with a flattened list of integers. The loaded value is known
2691 to be in the union of the ranges defined by each consecutive pair. Each
2692 pair has the following properties:
2694 - The type must match the type loaded by the instruction.
2695 - The pair ``a,b`` represents the range ``[a,b)``.
2696 - Both ``a`` and ``b`` are constants.
2697 - The range is allowed to wrap.
2698 - The range should not represent the full or empty set. That is,
2701 In addition, the pairs must be in signed order of the lower bound and
2702 they must be non-contiguous.
2706 .. code-block:: llvm
2708 %a = load i8* %x, align 1, !range !0 ; Can only be 0 or 1
2709 %b = load i8* %y, align 1, !range !1 ; Can only be 255 (-1), 0 or 1
2710 %c = load i8* %z, align 1, !range !2 ; Can only be 0, 1, 3, 4 or 5
2711 %d = load i8* %z, align 1, !range !3 ; Can only be -2, -1, 3, 4 or 5
2713 !0 = metadata !{ i8 0, i8 2 }
2714 !1 = metadata !{ i8 255, i8 2 }
2715 !2 = metadata !{ i8 0, i8 2, i8 3, i8 6 }
2716 !3 = metadata !{ i8 -2, i8 0, i8 3, i8 6 }
2721 It is sometimes useful to attach information to loop constructs. Currently,
2722 loop metadata is implemented as metadata attached to the branch instruction
2723 in the loop latch block. This type of metadata refer to a metadata node that is
2724 guaranteed to be separate for each loop. The loop identifier metadata is
2725 specified with the name ``llvm.loop``.
2727 The loop identifier metadata is implemented using a metadata that refers to
2728 itself to avoid merging it with any other identifier metadata, e.g.,
2729 during module linkage or function inlining. That is, each loop should refer
2730 to their own identification metadata even if they reside in separate functions.
2731 The following example contains loop identifier metadata for two separate loop
2734 .. code-block:: llvm
2736 !0 = metadata !{ metadata !0 }
2737 !1 = metadata !{ metadata !1 }
2739 The loop identifier metadata can be used to specify additional per-loop
2740 metadata. Any operands after the first operand can be treated as user-defined
2741 metadata. For example the ``llvm.vectorizer.unroll`` metadata is understood
2742 by the loop vectorizer to indicate how many times to unroll the loop:
2744 .. code-block:: llvm
2746 br i1 %exitcond, label %._crit_edge, label %.lr.ph, !llvm.loop !0
2748 !0 = metadata !{ metadata !0, metadata !1 }
2749 !1 = metadata !{ metadata !"llvm.vectorizer.unroll", i32 2 }
2754 Metadata types used to annotate memory accesses with information helpful
2755 for optimizations are prefixed with ``llvm.mem``.
2757 '``llvm.mem.parallel_loop_access``' Metadata
2758 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2760 For a loop to be parallel, in addition to using
2761 the ``llvm.loop`` metadata to mark the loop latch branch instruction,
2762 also all of the memory accessing instructions in the loop body need to be
2763 marked with the ``llvm.mem.parallel_loop_access`` metadata. If there
2764 is at least one memory accessing instruction not marked with the metadata,
2765 the loop must be considered a sequential loop. This causes parallel loops to be
2766 converted to sequential loops due to optimization passes that are unaware of
2767 the parallel semantics and that insert new memory instructions to the loop
2770 Example of a loop that is considered parallel due to its correct use of
2771 both ``llvm.loop`` and ``llvm.mem.parallel_loop_access``
2772 metadata types that refer to the same loop identifier metadata.
2774 .. code-block:: llvm
2778 %0 = load i32* %arrayidx, align 4, !llvm.mem.parallel_loop_access !0
2780 store i32 %0, i32* %arrayidx4, align 4, !llvm.mem.parallel_loop_access !0
2782 br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !0
2786 !0 = metadata !{ metadata !0 }
2788 It is also possible to have nested parallel loops. In that case the
2789 memory accesses refer to a list of loop identifier metadata nodes instead of
2790 the loop identifier metadata node directly:
2792 .. code-block:: llvm
2799 %0 = load i32* %arrayidx, align 4, !llvm.mem.parallel_loop_access !0
2801 store i32 %0, i32* %arrayidx4, align 4, !llvm.mem.parallel_loop_access !0
2803 br i1 %exitcond, label %inner.for.end, label %inner.for.body, !llvm.loop !1
2807 %0 = load i32* %arrayidx, align 4, !llvm.mem.parallel_loop_access !0
2809 store i32 %0, i32* %arrayidx4, align 4, !llvm.mem.parallel_loop_access !0
2811 br i1 %exitcond, label %outer.for.end, label %outer.for.body, !llvm.loop !2
2813 outer.for.end: ; preds = %for.body
2815 !0 = metadata !{ metadata !1, metadata !2 } ; a list of loop identifiers
2816 !1 = metadata !{ metadata !1 } ; an identifier for the inner loop
2817 !2 = metadata !{ metadata !2 } ; an identifier for the outer loop
2819 '``llvm.vectorizer``'
2820 ^^^^^^^^^^^^^^^^^^^^^
2822 Metadata prefixed with ``llvm.vectorizer`` is used to control per-loop
2823 vectorization parameters such as vectorization factor and unroll factor.
2825 ``llvm.vectorizer`` metadata should be used in conjunction with ``llvm.loop``
2826 loop identification metadata.
2828 '``llvm.vectorizer.unroll``' Metadata
2829 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2831 This metadata instructs the loop vectorizer to unroll the specified
2832 loop exactly ``N`` times.
2834 The first operand is the string ``llvm.vectorizer.unroll`` and the second
2835 operand is an integer specifying the unroll factor. For example:
2837 .. code-block:: llvm
2839 !0 = metadata !{ metadata !"llvm.vectorizer.unroll", i32 4 }
2841 Note that setting ``llvm.vectorizer.unroll`` to 1 disables unrolling of the
2844 If ``llvm.vectorizer.unroll`` is set to 0 then the amount of unrolling will be
2845 determined automatically.
2847 '``llvm.vectorizer.width``' Metadata
2848 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2850 This metadata sets the target width of the vectorizer to ``N``. Without
2851 this metadata, the vectorizer will choose a width automatically.
2852 Regardless of this metadata, the vectorizer will only vectorize loops if
2853 it believes it is valid to do so.
2855 The first operand is the string ``llvm.vectorizer.width`` and the second
2856 operand is an integer specifying the width. For example:
2858 .. code-block:: llvm
2860 !0 = metadata !{ metadata !"llvm.vectorizer.width", i32 4 }
2862 Note that setting ``llvm.vectorizer.width`` to 1 disables vectorization of the
2865 If ``llvm.vectorizer.width`` is set to 0 then the width will be determined
2868 Module Flags Metadata
2869 =====================
2871 Information about the module as a whole is difficult to convey to LLVM's
2872 subsystems. The LLVM IR isn't sufficient to transmit this information.
2873 The ``llvm.module.flags`` named metadata exists in order to facilitate
2874 this. These flags are in the form of key / value pairs --- much like a
2875 dictionary --- making it easy for any subsystem who cares about a flag to
2878 The ``llvm.module.flags`` metadata contains a list of metadata triplets.
2879 Each triplet has the following form:
2881 - The first element is a *behavior* flag, which specifies the behavior
2882 when two (or more) modules are merged together, and it encounters two
2883 (or more) metadata with the same ID. The supported behaviors are
2885 - The second element is a metadata string that is a unique ID for the
2886 metadata. Each module may only have one flag entry for each unique ID (not
2887 including entries with the **Require** behavior).
2888 - The third element is the value of the flag.
2890 When two (or more) modules are merged together, the resulting
2891 ``llvm.module.flags`` metadata is the union of the modules' flags. That is, for
2892 each unique metadata ID string, there will be exactly one entry in the merged
2893 modules ``llvm.module.flags`` metadata table, and the value for that entry will
2894 be determined by the merge behavior flag, as described below. The only exception
2895 is that entries with the *Require* behavior are always preserved.
2897 The following behaviors are supported:
2908 Emits an error if two values disagree, otherwise the resulting value
2909 is that of the operands.
2913 Emits a warning if two values disagree. The result value will be the
2914 operand for the flag from the first module being linked.
2918 Adds a requirement that another module flag be present and have a
2919 specified value after linking is performed. The value must be a
2920 metadata pair, where the first element of the pair is the ID of the
2921 module flag to be restricted, and the second element of the pair is
2922 the value the module flag should be restricted to. This behavior can
2923 be used to restrict the allowable results (via triggering of an
2924 error) of linking IDs with the **Override** behavior.
2928 Uses the specified value, regardless of the behavior or value of the
2929 other module. If both modules specify **Override**, but the values
2930 differ, an error will be emitted.
2934 Appends the two values, which are required to be metadata nodes.
2938 Appends the two values, which are required to be metadata
2939 nodes. However, duplicate entries in the second list are dropped
2940 during the append operation.
2942 It is an error for a particular unique flag ID to have multiple behaviors,
2943 except in the case of **Require** (which adds restrictions on another metadata
2944 value) or **Override**.
2946 An example of module flags:
2948 .. code-block:: llvm
2950 !0 = metadata !{ i32 1, metadata !"foo", i32 1 }
2951 !1 = metadata !{ i32 4, metadata !"bar", i32 37 }
2952 !2 = metadata !{ i32 2, metadata !"qux", i32 42 }
2953 !3 = metadata !{ i32 3, metadata !"qux",
2955 metadata !"foo", i32 1
2958 !llvm.module.flags = !{ !0, !1, !2, !3 }
2960 - Metadata ``!0`` has the ID ``!"foo"`` and the value '1'. The behavior
2961 if two or more ``!"foo"`` flags are seen is to emit an error if their
2962 values are not equal.
2964 - Metadata ``!1`` has the ID ``!"bar"`` and the value '37'. The
2965 behavior if two or more ``!"bar"`` flags are seen is to use the value
2968 - Metadata ``!2`` has the ID ``!"qux"`` and the value '42'. The
2969 behavior if two or more ``!"qux"`` flags are seen is to emit a
2970 warning if their values are not equal.
2972 - Metadata ``!3`` has the ID ``!"qux"`` and the value:
2976 metadata !{ metadata !"foo", i32 1 }
2978 The behavior is to emit an error if the ``llvm.module.flags`` does not
2979 contain a flag with the ID ``!"foo"`` that has the value '1' after linking is
2982 Objective-C Garbage Collection Module Flags Metadata
2983 ----------------------------------------------------
2985 On the Mach-O platform, Objective-C stores metadata about garbage
2986 collection in a special section called "image info". The metadata
2987 consists of a version number and a bitmask specifying what types of
2988 garbage collection are supported (if any) by the file. If two or more
2989 modules are linked together their garbage collection metadata needs to
2990 be merged rather than appended together.
2992 The Objective-C garbage collection module flags metadata consists of the
2993 following key-value pairs:
3002 * - ``Objective-C Version``
3003 - **[Required]** --- The Objective-C ABI version. Valid values are 1 and 2.
3005 * - ``Objective-C Image Info Version``
3006 - **[Required]** --- The version of the image info section. Currently
3009 * - ``Objective-C Image Info Section``
3010 - **[Required]** --- The section to place the metadata. Valid values are
3011 ``"__OBJC, __image_info, regular"`` for Objective-C ABI version 1, and
3012 ``"__DATA,__objc_imageinfo, regular, no_dead_strip"`` for
3013 Objective-C ABI version 2.
3015 * - ``Objective-C Garbage Collection``
3016 - **[Required]** --- Specifies whether garbage collection is supported or
3017 not. Valid values are 0, for no garbage collection, and 2, for garbage
3018 collection supported.
3020 * - ``Objective-C GC Only``
3021 - **[Optional]** --- Specifies that only garbage collection is supported.
3022 If present, its value must be 6. This flag requires that the
3023 ``Objective-C Garbage Collection`` flag have the value 2.
3025 Some important flag interactions:
3027 - If a module with ``Objective-C Garbage Collection`` set to 0 is
3028 merged with a module with ``Objective-C Garbage Collection`` set to
3029 2, then the resulting module has the
3030 ``Objective-C Garbage Collection`` flag set to 0.
3031 - A module with ``Objective-C Garbage Collection`` set to 0 cannot be
3032 merged with a module with ``Objective-C GC Only`` set to 6.
3034 Automatic Linker Flags Module Flags Metadata
3035 --------------------------------------------
3037 Some targets support embedding flags to the linker inside individual object
3038 files. Typically this is used in conjunction with language extensions which
3039 allow source files to explicitly declare the libraries they depend on, and have
3040 these automatically be transmitted to the linker via object files.
3042 These flags are encoded in the IR using metadata in the module flags section,
3043 using the ``Linker Options`` key. The merge behavior for this flag is required
3044 to be ``AppendUnique``, and the value for the key is expected to be a metadata
3045 node which should be a list of other metadata nodes, each of which should be a
3046 list of metadata strings defining linker options.
3048 For example, the following metadata section specifies two separate sets of
3049 linker options, presumably to link against ``libz`` and the ``Cocoa``
3052 !0 = metadata !{ i32 6, metadata !"Linker Options",
3054 metadata !{ metadata !"-lz" },
3055 metadata !{ metadata !"-framework", metadata !"Cocoa" } } }
3056 !llvm.module.flags = !{ !0 }
3058 The metadata encoding as lists of lists of options, as opposed to a collapsed
3059 list of options, is chosen so that the IR encoding can use multiple option
3060 strings to specify e.g., a single library, while still having that specifier be
3061 preserved as an atomic element that can be recognized by a target specific
3062 assembly writer or object file emitter.
3064 Each individual option is required to be either a valid option for the target's
3065 linker, or an option that is reserved by the target specific assembly writer or
3066 object file emitter. No other aspect of these options is defined by the IR.
3068 .. _intrinsicglobalvariables:
3070 Intrinsic Global Variables
3071 ==========================
3073 LLVM has a number of "magic" global variables that contain data that
3074 affect code generation or other IR semantics. These are documented here.
3075 All globals of this sort should have a section specified as
3076 "``llvm.metadata``". This section and all globals that start with
3077 "``llvm.``" are reserved for use by LLVM.
3081 The '``llvm.used``' Global Variable
3082 -----------------------------------
3084 The ``@llvm.used`` global is an array which has
3085 :ref:`appending linkage <linkage_appending>`. This array contains a list of
3086 pointers to named global variables, functions and aliases which may optionally
3087 have a pointer cast formed of bitcast or getelementptr. For example, a legal
3090 .. code-block:: llvm
3095 @llvm.used = appending global [2 x i8*] [
3097 i8* bitcast (i32* @Y to i8*)
3098 ], section "llvm.metadata"
3100 If a symbol appears in the ``@llvm.used`` list, then the compiler, assembler,
3101 and linker are required to treat the symbol as if there is a reference to the
3102 symbol that it cannot see (which is why they have to be named). For example, if
3103 a variable has internal linkage and no references other than that from the
3104 ``@llvm.used`` list, it cannot be deleted. This is commonly used to represent
3105 references from inline asms and other things the compiler cannot "see", and
3106 corresponds to "``attribute((used))``" in GNU C.
3108 On some targets, the code generator must emit a directive to the
3109 assembler or object file to prevent the assembler and linker from
3110 molesting the symbol.
3112 .. _gv_llvmcompilerused:
3114 The '``llvm.compiler.used``' Global Variable
3115 --------------------------------------------
3117 The ``@llvm.compiler.used`` directive is the same as the ``@llvm.used``
3118 directive, except that it only prevents the compiler from touching the
3119 symbol. On targets that support it, this allows an intelligent linker to
3120 optimize references to the symbol without being impeded as it would be
3123 This is a rare construct that should only be used in rare circumstances,
3124 and should not be exposed to source languages.
3126 .. _gv_llvmglobalctors:
3128 The '``llvm.global_ctors``' Global Variable
3129 -------------------------------------------
3131 .. code-block:: llvm
3133 %0 = type { i32, void ()* }
3134 @llvm.global_ctors = appending global [1 x %0] [%0 { i32 65535, void ()* @ctor }]
3136 The ``@llvm.global_ctors`` array contains a list of constructor
3137 functions and associated priorities. The functions referenced by this
3138 array will be called in ascending order of priority (i.e. lowest first)
3139 when the module is loaded. The order of functions with the same priority
3142 .. _llvmglobaldtors:
3144 The '``llvm.global_dtors``' Global Variable
3145 -------------------------------------------
3147 .. code-block:: llvm
3149 %0 = type { i32, void ()* }
3150 @llvm.global_dtors = appending global [1 x %0] [%0 { i32 65535, void ()* @dtor }]
3152 The ``@llvm.global_dtors`` array contains a list of destructor functions
3153 and associated priorities. The functions referenced by this array will
3154 be called in descending order of priority (i.e. highest first) when the
3155 module is loaded. The order of functions with the same priority is not
3158 Instruction Reference
3159 =====================
3161 The LLVM instruction set consists of several different classifications
3162 of instructions: :ref:`terminator instructions <terminators>`, :ref:`binary
3163 instructions <binaryops>`, :ref:`bitwise binary
3164 instructions <bitwiseops>`, :ref:`memory instructions <memoryops>`, and
3165 :ref:`other instructions <otherops>`.
3169 Terminator Instructions
3170 -----------------------
3172 As mentioned :ref:`previously <functionstructure>`, every basic block in a
3173 program ends with a "Terminator" instruction, which indicates which
3174 block should be executed after the current block is finished. These
3175 terminator instructions typically yield a '``void``' value: they produce
3176 control flow, not values (the one exception being the
3177 ':ref:`invoke <i_invoke>`' instruction).
3179 The terminator instructions are: ':ref:`ret <i_ret>`',
3180 ':ref:`br <i_br>`', ':ref:`switch <i_switch>`',
3181 ':ref:`indirectbr <i_indirectbr>`', ':ref:`invoke <i_invoke>`',
3182 ':ref:`resume <i_resume>`', and ':ref:`unreachable <i_unreachable>`'.
3186 '``ret``' Instruction
3187 ^^^^^^^^^^^^^^^^^^^^^
3194 ret <type> <value> ; Return a value from a non-void function
3195 ret void ; Return from void function
3200 The '``ret``' instruction is used to return control flow (and optionally
3201 a value) from a function back to the caller.
3203 There are two forms of the '``ret``' instruction: one that returns a
3204 value and then causes control flow, and one that just causes control
3210 The '``ret``' instruction optionally accepts a single argument, the
3211 return value. The type of the return value must be a ':ref:`first
3212 class <t_firstclass>`' type.
3214 A function is not :ref:`well formed <wellformed>` if it it has a non-void
3215 return type and contains a '``ret``' instruction with no return value or
3216 a return value with a type that does not match its type, or if it has a
3217 void return type and contains a '``ret``' instruction with a return
3223 When the '``ret``' instruction is executed, control flow returns back to
3224 the calling function's context. If the caller is a
3225 ":ref:`call <i_call>`" instruction, execution continues at the
3226 instruction after the call. If the caller was an
3227 ":ref:`invoke <i_invoke>`" instruction, execution continues at the
3228 beginning of the "normal" destination block. If the instruction returns
3229 a value, that value shall set the call or invoke instruction's return
3235 .. code-block:: llvm
3237 ret i32 5 ; Return an integer value of 5
3238 ret void ; Return from a void function
3239 ret { i32, i8 } { i32 4, i8 2 } ; Return a struct of values 4 and 2
3243 '``br``' Instruction
3244 ^^^^^^^^^^^^^^^^^^^^
3251 br i1 <cond>, label <iftrue>, label <iffalse>
3252 br label <dest> ; Unconditional branch
3257 The '``br``' instruction is used to cause control flow to transfer to a
3258 different basic block in the current function. There are two forms of
3259 this instruction, corresponding to a conditional branch and an
3260 unconditional branch.
3265 The conditional branch form of the '``br``' instruction takes a single
3266 '``i1``' value and two '``label``' values. The unconditional form of the
3267 '``br``' instruction takes a single '``label``' value as a target.
3272 Upon execution of a conditional '``br``' instruction, the '``i1``'
3273 argument is evaluated. If the value is ``true``, control flows to the
3274 '``iftrue``' ``label`` argument. If "cond" is ``false``, control flows
3275 to the '``iffalse``' ``label`` argument.
3280 .. code-block:: llvm
3283 %cond = icmp eq i32 %a, %b
3284 br i1 %cond, label %IfEqual, label %IfUnequal
3292 '``switch``' Instruction
3293 ^^^^^^^^^^^^^^^^^^^^^^^^
3300 switch <intty> <value>, label <defaultdest> [ <intty> <val>, label <dest> ... ]
3305 The '``switch``' instruction is used to transfer control flow to one of
3306 several different places. It is a generalization of the '``br``'
3307 instruction, allowing a branch to occur to one of many possible
3313 The '``switch``' instruction uses three parameters: an integer
3314 comparison value '``value``', a default '``label``' destination, and an
3315 array of pairs of comparison value constants and '``label``'s. The table
3316 is not allowed to contain duplicate constant entries.
3321 The ``switch`` instruction specifies a table of values and destinations.
3322 When the '``switch``' instruction is executed, this table is searched
3323 for the given value. If the value is found, control flow is transferred
3324 to the corresponding destination; otherwise, control flow is transferred
3325 to the default destination.
3330 Depending on properties of the target machine and the particular
3331 ``switch`` instruction, this instruction may be code generated in
3332 different ways. For example, it could be generated as a series of
3333 chained conditional branches or with a lookup table.
3338 .. code-block:: llvm
3340 ; Emulate a conditional br instruction
3341 %Val = zext i1 %value to i32
3342 switch i32 %Val, label %truedest [ i32 0, label %falsedest ]
3344 ; Emulate an unconditional br instruction
3345 switch i32 0, label %dest [ ]
3347 ; Implement a jump table:
3348 switch i32 %val, label %otherwise [ i32 0, label %onzero
3350 i32 2, label %ontwo ]
3354 '``indirectbr``' Instruction
3355 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3362 indirectbr <somety>* <address>, [ label <dest1>, label <dest2>, ... ]
3367 The '``indirectbr``' instruction implements an indirect branch to a
3368 label within the current function, whose address is specified by
3369 "``address``". Address must be derived from a
3370 :ref:`blockaddress <blockaddress>` constant.
3375 The '``address``' argument is the address of the label to jump to. The
3376 rest of the arguments indicate the full set of possible destinations
3377 that the address may point to. Blocks are allowed to occur multiple
3378 times in the destination list, though this isn't particularly useful.
3380 This destination list is required so that dataflow analysis has an
3381 accurate understanding of the CFG.
3386 Control transfers to the block specified in the address argument. All
3387 possible destination blocks must be listed in the label list, otherwise
3388 this instruction has undefined behavior. This implies that jumps to
3389 labels defined in other functions have undefined behavior as well.
3394 This is typically implemented with a jump through a register.
3399 .. code-block:: llvm
3401 indirectbr i8* %Addr, [ label %bb1, label %bb2, label %bb3 ]
3405 '``invoke``' Instruction
3406 ^^^^^^^^^^^^^^^^^^^^^^^^
3413 <result> = invoke [cconv] [ret attrs] <ptr to function ty> <function ptr val>(<function args>) [fn attrs]
3414 to label <normal label> unwind label <exception label>
3419 The '``invoke``' instruction causes control to transfer to a specified
3420 function, with the possibility of control flow transfer to either the
3421 '``normal``' label or the '``exception``' label. If the callee function
3422 returns with the "``ret``" instruction, control flow will return to the
3423 "normal" label. If the callee (or any indirect callees) returns via the
3424 ":ref:`resume <i_resume>`" instruction or other exception handling
3425 mechanism, control is interrupted and continued at the dynamically
3426 nearest "exception" label.
3428 The '``exception``' label is a `landing
3429 pad <ExceptionHandling.html#overview>`_ for the exception. As such,
3430 '``exception``' label is required to have the
3431 ":ref:`landingpad <i_landingpad>`" instruction, which contains the
3432 information about the behavior of the program after unwinding happens,
3433 as its first non-PHI instruction. The restrictions on the
3434 "``landingpad``" instruction's tightly couples it to the "``invoke``"
3435 instruction, so that the important information contained within the
3436 "``landingpad``" instruction can't be lost through normal code motion.
3441 This instruction requires several arguments:
3443 #. The optional "cconv" marker indicates which :ref:`calling
3444 convention <callingconv>` the call should use. If none is
3445 specified, the call defaults to using C calling conventions.
3446 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
3447 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
3449 #. '``ptr to function ty``': shall be the signature of the pointer to
3450 function value being invoked. In most cases, this is a direct
3451 function invocation, but indirect ``invoke``'s are just as possible,
3452 branching off an arbitrary pointer to function value.
3453 #. '``function ptr val``': An LLVM value containing a pointer to a
3454 function to be invoked.
3455 #. '``function args``': argument list whose types match the function
3456 signature argument types and parameter attributes. All arguments must
3457 be of :ref:`first class <t_firstclass>` type. If the function signature
3458 indicates the function accepts a variable number of arguments, the
3459 extra arguments can be specified.
3460 #. '``normal label``': the label reached when the called function
3461 executes a '``ret``' instruction.
3462 #. '``exception label``': the label reached when a callee returns via
3463 the :ref:`resume <i_resume>` instruction or other exception handling
3465 #. The optional :ref:`function attributes <fnattrs>` list. Only
3466 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
3467 attributes are valid here.
3472 This instruction is designed to operate as a standard '``call``'
3473 instruction in most regards. The primary difference is that it
3474 establishes an association with a label, which is used by the runtime
3475 library to unwind the stack.
3477 This instruction is used in languages with destructors to ensure that
3478 proper cleanup is performed in the case of either a ``longjmp`` or a
3479 thrown exception. Additionally, this is important for implementation of
3480 '``catch``' clauses in high-level languages that support them.
3482 For the purposes of the SSA form, the definition of the value returned
3483 by the '``invoke``' instruction is deemed to occur on the edge from the
3484 current block to the "normal" label. If the callee unwinds then no
3485 return value is available.
3490 .. code-block:: llvm
3492 %retval = invoke i32 @Test(i32 15) to label %Continue
3493 unwind label %TestCleanup ; {i32}:retval set
3494 %retval = invoke coldcc i32 %Testfnptr(i32 15) to label %Continue
3495 unwind label %TestCleanup ; {i32}:retval set
3499 '``resume``' Instruction
3500 ^^^^^^^^^^^^^^^^^^^^^^^^
3507 resume <type> <value>
3512 The '``resume``' instruction is a terminator instruction that has no
3518 The '``resume``' instruction requires one argument, which must have the
3519 same type as the result of any '``landingpad``' instruction in the same
3525 The '``resume``' instruction resumes propagation of an existing
3526 (in-flight) exception whose unwinding was interrupted with a
3527 :ref:`landingpad <i_landingpad>` instruction.
3532 .. code-block:: llvm
3534 resume { i8*, i32 } %exn
3538 '``unreachable``' Instruction
3539 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3551 The '``unreachable``' instruction has no defined semantics. This
3552 instruction is used to inform the optimizer that a particular portion of
3553 the code is not reachable. This can be used to indicate that the code
3554 after a no-return function cannot be reached, and other facts.
3559 The '``unreachable``' instruction has no defined semantics.
3566 Binary operators are used to do most of the computation in a program.
3567 They require two operands of the same type, execute an operation on
3568 them, and produce a single value. The operands might represent multiple
3569 data, as is the case with the :ref:`vector <t_vector>` data type. The
3570 result value has the same type as its operands.
3572 There are several different binary operators:
3576 '``add``' Instruction
3577 ^^^^^^^^^^^^^^^^^^^^^
3584 <result> = add <ty> <op1>, <op2> ; yields {ty}:result
3585 <result> = add nuw <ty> <op1>, <op2> ; yields {ty}:result
3586 <result> = add nsw <ty> <op1>, <op2> ; yields {ty}:result
3587 <result> = add nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3592 The '``add``' instruction returns the sum of its two operands.
3597 The two arguments to the '``add``' instruction must be
3598 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3599 arguments must have identical types.
3604 The value produced is the integer sum of the two operands.
3606 If the sum has unsigned overflow, the result returned is the
3607 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
3610 Because LLVM integers use a two's complement representation, this
3611 instruction is appropriate for both signed and unsigned integers.
3613 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3614 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3615 result value of the ``add`` is a :ref:`poison value <poisonvalues>` if
3616 unsigned and/or signed overflow, respectively, occurs.
3621 .. code-block:: llvm
3623 <result> = add i32 4, %var ; yields {i32}:result = 4 + %var
3627 '``fadd``' Instruction
3628 ^^^^^^^^^^^^^^^^^^^^^^
3635 <result> = fadd [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3640 The '``fadd``' instruction returns the sum of its two operands.
3645 The two arguments to the '``fadd``' instruction must be :ref:`floating
3646 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3647 Both arguments must have identical types.
3652 The value produced is the floating point sum of the two operands. This
3653 instruction can also take any number of :ref:`fast-math flags <fastmath>`,
3654 which are optimization hints to enable otherwise unsafe floating point
3660 .. code-block:: llvm
3662 <result> = fadd float 4.0, %var ; yields {float}:result = 4.0 + %var
3664 '``sub``' Instruction
3665 ^^^^^^^^^^^^^^^^^^^^^
3672 <result> = sub <ty> <op1>, <op2> ; yields {ty}:result
3673 <result> = sub nuw <ty> <op1>, <op2> ; yields {ty}:result
3674 <result> = sub nsw <ty> <op1>, <op2> ; yields {ty}:result
3675 <result> = sub nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3680 The '``sub``' instruction returns the difference of its two operands.
3682 Note that the '``sub``' instruction is used to represent the '``neg``'
3683 instruction present in most other intermediate representations.
3688 The two arguments to the '``sub``' instruction must be
3689 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3690 arguments must have identical types.
3695 The value produced is the integer difference of the two operands.
3697 If the difference has unsigned overflow, the result returned is the
3698 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
3701 Because LLVM integers use a two's complement representation, this
3702 instruction is appropriate for both signed and unsigned integers.
3704 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3705 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3706 result value of the ``sub`` is a :ref:`poison value <poisonvalues>` if
3707 unsigned and/or signed overflow, respectively, occurs.
3712 .. code-block:: llvm
3714 <result> = sub i32 4, %var ; yields {i32}:result = 4 - %var
3715 <result> = sub i32 0, %val ; yields {i32}:result = -%var
3719 '``fsub``' Instruction
3720 ^^^^^^^^^^^^^^^^^^^^^^
3727 <result> = fsub [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3732 The '``fsub``' instruction returns the difference of its two operands.
3734 Note that the '``fsub``' instruction is used to represent the '``fneg``'
3735 instruction present in most other intermediate representations.
3740 The two arguments to the '``fsub``' instruction must be :ref:`floating
3741 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3742 Both arguments must have identical types.
3747 The value produced is the floating point difference of the two operands.
3748 This instruction can also take any number of :ref:`fast-math
3749 flags <fastmath>`, which are optimization hints to enable otherwise
3750 unsafe floating point optimizations:
3755 .. code-block:: llvm
3757 <result> = fsub float 4.0, %var ; yields {float}:result = 4.0 - %var
3758 <result> = fsub float -0.0, %val ; yields {float}:result = -%var
3760 '``mul``' Instruction
3761 ^^^^^^^^^^^^^^^^^^^^^
3768 <result> = mul <ty> <op1>, <op2> ; yields {ty}:result
3769 <result> = mul nuw <ty> <op1>, <op2> ; yields {ty}:result
3770 <result> = mul nsw <ty> <op1>, <op2> ; yields {ty}:result
3771 <result> = mul nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3776 The '``mul``' instruction returns the product of its two operands.
3781 The two arguments to the '``mul``' instruction must be
3782 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3783 arguments must have identical types.
3788 The value produced is the integer product of the two operands.
3790 If the result of the multiplication has unsigned overflow, the result
3791 returned is the mathematical result modulo 2\ :sup:`n`\ , where n is the
3792 bit width of the result.
3794 Because LLVM integers use a two's complement representation, and the
3795 result is the same width as the operands, this instruction returns the
3796 correct result for both signed and unsigned integers. If a full product
3797 (e.g. ``i32`` * ``i32`` -> ``i64``) is needed, the operands should be
3798 sign-extended or zero-extended as appropriate to the width of the full
3801 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3802 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3803 result value of the ``mul`` is a :ref:`poison value <poisonvalues>` if
3804 unsigned and/or signed overflow, respectively, occurs.
3809 .. code-block:: llvm
3811 <result> = mul i32 4, %var ; yields {i32}:result = 4 * %var
3815 '``fmul``' Instruction
3816 ^^^^^^^^^^^^^^^^^^^^^^
3823 <result> = fmul [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3828 The '``fmul``' instruction returns the product of its two operands.
3833 The two arguments to the '``fmul``' instruction must be :ref:`floating
3834 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3835 Both arguments must have identical types.
3840 The value produced is the floating point product of the two operands.
3841 This instruction can also take any number of :ref:`fast-math
3842 flags <fastmath>`, which are optimization hints to enable otherwise
3843 unsafe floating point optimizations:
3848 .. code-block:: llvm
3850 <result> = fmul float 4.0, %var ; yields {float}:result = 4.0 * %var
3852 '``udiv``' Instruction
3853 ^^^^^^^^^^^^^^^^^^^^^^
3860 <result> = udiv <ty> <op1>, <op2> ; yields {ty}:result
3861 <result> = udiv exact <ty> <op1>, <op2> ; yields {ty}:result
3866 The '``udiv``' instruction returns the quotient of its two operands.
3871 The two arguments to the '``udiv``' instruction must be
3872 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3873 arguments must have identical types.
3878 The value produced is the unsigned integer quotient of the two operands.
3880 Note that unsigned integer division and signed integer division are
3881 distinct operations; for signed integer division, use '``sdiv``'.
3883 Division by zero leads to undefined behavior.
3885 If the ``exact`` keyword is present, the result value of the ``udiv`` is
3886 a :ref:`poison value <poisonvalues>` if %op1 is not a multiple of %op2 (as
3887 such, "((a udiv exact b) mul b) == a").
3892 .. code-block:: llvm
3894 <result> = udiv i32 4, %var ; yields {i32}:result = 4 / %var
3896 '``sdiv``' Instruction
3897 ^^^^^^^^^^^^^^^^^^^^^^
3904 <result> = sdiv <ty> <op1>, <op2> ; yields {ty}:result
3905 <result> = sdiv exact <ty> <op1>, <op2> ; yields {ty}:result
3910 The '``sdiv``' instruction returns the quotient of its two operands.
3915 The two arguments to the '``sdiv``' instruction must be
3916 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3917 arguments must have identical types.
3922 The value produced is the signed integer quotient of the two operands
3923 rounded towards zero.
3925 Note that signed integer division and unsigned integer division are
3926 distinct operations; for unsigned integer division, use '``udiv``'.
3928 Division by zero leads to undefined behavior. Overflow also leads to
3929 undefined behavior; this is a rare case, but can occur, for example, by
3930 doing a 32-bit division of -2147483648 by -1.
3932 If the ``exact`` keyword is present, the result value of the ``sdiv`` is
3933 a :ref:`poison value <poisonvalues>` if the result would be rounded.
3938 .. code-block:: llvm
3940 <result> = sdiv i32 4, %var ; yields {i32}:result = 4 / %var
3944 '``fdiv``' Instruction
3945 ^^^^^^^^^^^^^^^^^^^^^^
3952 <result> = fdiv [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3957 The '``fdiv``' instruction returns the quotient of its two operands.
3962 The two arguments to the '``fdiv``' instruction must be :ref:`floating
3963 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3964 Both arguments must have identical types.
3969 The value produced is the floating point quotient of the two operands.
3970 This instruction can also take any number of :ref:`fast-math
3971 flags <fastmath>`, which are optimization hints to enable otherwise
3972 unsafe floating point optimizations:
3977 .. code-block:: llvm
3979 <result> = fdiv float 4.0, %var ; yields {float}:result = 4.0 / %var
3981 '``urem``' Instruction
3982 ^^^^^^^^^^^^^^^^^^^^^^
3989 <result> = urem <ty> <op1>, <op2> ; yields {ty}:result
3994 The '``urem``' instruction returns the remainder from the unsigned
3995 division of its two arguments.
4000 The two arguments to the '``urem``' instruction must be
4001 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4002 arguments must have identical types.
4007 This instruction returns the unsigned integer *remainder* of a division.
4008 This instruction always performs an unsigned division to get the
4011 Note that unsigned integer remainder and signed integer remainder are
4012 distinct operations; for signed integer remainder, use '``srem``'.
4014 Taking the remainder of a division by zero leads to undefined behavior.
4019 .. code-block:: llvm
4021 <result> = urem i32 4, %var ; yields {i32}:result = 4 % %var
4023 '``srem``' Instruction
4024 ^^^^^^^^^^^^^^^^^^^^^^
4031 <result> = srem <ty> <op1>, <op2> ; yields {ty}:result
4036 The '``srem``' instruction returns the remainder from the signed
4037 division of its two operands. This instruction can also take
4038 :ref:`vector <t_vector>` versions of the values in which case the elements
4044 The two arguments to the '``srem``' instruction must be
4045 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4046 arguments must have identical types.
4051 This instruction returns the *remainder* of a division (where the result
4052 is either zero or has the same sign as the dividend, ``op1``), not the
4053 *modulo* operator (where the result is either zero or has the same sign
4054 as the divisor, ``op2``) of a value. For more information about the
4055 difference, see `The Math
4056 Forum <http://mathforum.org/dr.math/problems/anne.4.28.99.html>`_. For a
4057 table of how this is implemented in various languages, please see
4059 operation <http://en.wikipedia.org/wiki/Modulo_operation>`_.
4061 Note that signed integer remainder and unsigned integer remainder are
4062 distinct operations; for unsigned integer remainder, use '``urem``'.
4064 Taking the remainder of a division by zero leads to undefined behavior.
4065 Overflow also leads to undefined behavior; this is a rare case, but can
4066 occur, for example, by taking the remainder of a 32-bit division of
4067 -2147483648 by -1. (The remainder doesn't actually overflow, but this
4068 rule lets srem be implemented using instructions that return both the
4069 result of the division and the remainder.)
4074 .. code-block:: llvm
4076 <result> = srem i32 4, %var ; yields {i32}:result = 4 % %var
4080 '``frem``' Instruction
4081 ^^^^^^^^^^^^^^^^^^^^^^
4088 <result> = frem [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
4093 The '``frem``' instruction returns the remainder from the division of
4099 The two arguments to the '``frem``' instruction must be :ref:`floating
4100 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
4101 Both arguments must have identical types.
4106 This instruction returns the *remainder* of a division. The remainder
4107 has the same sign as the dividend. This instruction can also take any
4108 number of :ref:`fast-math flags <fastmath>`, which are optimization hints
4109 to enable otherwise unsafe floating point optimizations:
4114 .. code-block:: llvm
4116 <result> = frem float 4.0, %var ; yields {float}:result = 4.0 % %var
4120 Bitwise Binary Operations
4121 -------------------------
4123 Bitwise binary operators are used to do various forms of bit-twiddling
4124 in a program. They are generally very efficient instructions and can
4125 commonly be strength reduced from other instructions. They require two
4126 operands of the same type, execute an operation on them, and produce a
4127 single value. The resulting value is the same type as its operands.
4129 '``shl``' Instruction
4130 ^^^^^^^^^^^^^^^^^^^^^
4137 <result> = shl <ty> <op1>, <op2> ; yields {ty}:result
4138 <result> = shl nuw <ty> <op1>, <op2> ; yields {ty}:result
4139 <result> = shl nsw <ty> <op1>, <op2> ; yields {ty}:result
4140 <result> = shl nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
4145 The '``shl``' instruction returns the first operand shifted to the left
4146 a specified number of bits.
4151 Both arguments to the '``shl``' instruction must be the same
4152 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4153 '``op2``' is treated as an unsigned value.
4158 The value produced is ``op1`` \* 2\ :sup:`op2` mod 2\ :sup:`n`,
4159 where ``n`` is the width of the result. If ``op2`` is (statically or
4160 dynamically) negative or equal to or larger than the number of bits in
4161 ``op1``, the result is undefined. If the arguments are vectors, each
4162 vector element of ``op1`` is shifted by the corresponding shift amount
4165 If the ``nuw`` keyword is present, then the shift produces a :ref:`poison
4166 value <poisonvalues>` if it shifts out any non-zero bits. If the
4167 ``nsw`` keyword is present, then the shift produces a :ref:`poison
4168 value <poisonvalues>` if it shifts out any bits that disagree with the
4169 resultant sign bit. As such, NUW/NSW have the same semantics as they
4170 would if the shift were expressed as a mul instruction with the same
4171 nsw/nuw bits in (mul %op1, (shl 1, %op2)).
4176 .. code-block:: llvm
4178 <result> = shl i32 4, %var ; yields {i32}: 4 << %var
4179 <result> = shl i32 4, 2 ; yields {i32}: 16
4180 <result> = shl i32 1, 10 ; yields {i32}: 1024
4181 <result> = shl i32 1, 32 ; undefined
4182 <result> = shl <2 x i32> < i32 1, i32 1>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 2, i32 4>
4184 '``lshr``' Instruction
4185 ^^^^^^^^^^^^^^^^^^^^^^
4192 <result> = lshr <ty> <op1>, <op2> ; yields {ty}:result
4193 <result> = lshr exact <ty> <op1>, <op2> ; yields {ty}:result
4198 The '``lshr``' instruction (logical shift right) returns the first
4199 operand shifted to the right a specified number of bits with zero fill.
4204 Both arguments to the '``lshr``' instruction must be the same
4205 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4206 '``op2``' is treated as an unsigned value.
4211 This instruction always performs a logical shift right operation. The
4212 most significant bits of the result will be filled with zero bits after
4213 the shift. If ``op2`` is (statically or dynamically) equal to or larger
4214 than the number of bits in ``op1``, the result is undefined. If the
4215 arguments are vectors, each vector element of ``op1`` is shifted by the
4216 corresponding shift amount in ``op2``.
4218 If the ``exact`` keyword is present, the result value of the ``lshr`` is
4219 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
4225 .. code-block:: llvm
4227 <result> = lshr i32 4, 1 ; yields {i32}:result = 2
4228 <result> = lshr i32 4, 2 ; yields {i32}:result = 1
4229 <result> = lshr i8 4, 3 ; yields {i8}:result = 0
4230 <result> = lshr i8 -2, 1 ; yields {i8}:result = 0x7F
4231 <result> = lshr i32 1, 32 ; undefined
4232 <result> = lshr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 0x7FFFFFFF, i32 1>
4234 '``ashr``' Instruction
4235 ^^^^^^^^^^^^^^^^^^^^^^
4242 <result> = ashr <ty> <op1>, <op2> ; yields {ty}:result
4243 <result> = ashr exact <ty> <op1>, <op2> ; yields {ty}:result
4248 The '``ashr``' instruction (arithmetic shift right) returns the first
4249 operand shifted to the right a specified number of bits with sign
4255 Both arguments to the '``ashr``' instruction must be the same
4256 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4257 '``op2``' is treated as an unsigned value.
4262 This instruction always performs an arithmetic shift right operation,
4263 The most significant bits of the result will be filled with the sign bit
4264 of ``op1``. If ``op2`` is (statically or dynamically) equal to or larger
4265 than the number of bits in ``op1``, the result is undefined. If the
4266 arguments are vectors, each vector element of ``op1`` is shifted by the
4267 corresponding shift amount in ``op2``.
4269 If the ``exact`` keyword is present, the result value of the ``ashr`` is
4270 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
4276 .. code-block:: llvm
4278 <result> = ashr i32 4, 1 ; yields {i32}:result = 2
4279 <result> = ashr i32 4, 2 ; yields {i32}:result = 1
4280 <result> = ashr i8 4, 3 ; yields {i8}:result = 0
4281 <result> = ashr i8 -2, 1 ; yields {i8}:result = -1
4282 <result> = ashr i32 1, 32 ; undefined
4283 <result> = ashr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 3> ; yields: result=<2 x i32> < i32 -1, i32 0>
4285 '``and``' Instruction
4286 ^^^^^^^^^^^^^^^^^^^^^
4293 <result> = and <ty> <op1>, <op2> ; yields {ty}:result
4298 The '``and``' instruction returns the bitwise logical and of its two
4304 The two arguments to the '``and``' instruction must be
4305 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4306 arguments must have identical types.
4311 The truth table used for the '``and``' instruction is:
4328 .. code-block:: llvm
4330 <result> = and i32 4, %var ; yields {i32}:result = 4 & %var
4331 <result> = and i32 15, 40 ; yields {i32}:result = 8
4332 <result> = and i32 4, 8 ; yields {i32}:result = 0
4334 '``or``' Instruction
4335 ^^^^^^^^^^^^^^^^^^^^
4342 <result> = or <ty> <op1>, <op2> ; yields {ty}:result
4347 The '``or``' instruction returns the bitwise logical inclusive or of its
4353 The two arguments to the '``or``' instruction must be
4354 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4355 arguments must have identical types.
4360 The truth table used for the '``or``' instruction is:
4379 <result> = or i32 4, %var ; yields {i32}:result = 4 | %var
4380 <result> = or i32 15, 40 ; yields {i32}:result = 47
4381 <result> = or i32 4, 8 ; yields {i32}:result = 12
4383 '``xor``' Instruction
4384 ^^^^^^^^^^^^^^^^^^^^^
4391 <result> = xor <ty> <op1>, <op2> ; yields {ty}:result
4396 The '``xor``' instruction returns the bitwise logical exclusive or of
4397 its two operands. The ``xor`` is used to implement the "one's
4398 complement" operation, which is the "~" operator in C.
4403 The two arguments to the '``xor``' instruction must be
4404 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4405 arguments must have identical types.
4410 The truth table used for the '``xor``' instruction is:
4427 .. code-block:: llvm
4429 <result> = xor i32 4, %var ; yields {i32}:result = 4 ^ %var
4430 <result> = xor i32 15, 40 ; yields {i32}:result = 39
4431 <result> = xor i32 4, 8 ; yields {i32}:result = 12
4432 <result> = xor i32 %V, -1 ; yields {i32}:result = ~%V
4437 LLVM supports several instructions to represent vector operations in a
4438 target-independent manner. These instructions cover the element-access
4439 and vector-specific operations needed to process vectors effectively.
4440 While LLVM does directly support these vector operations, many
4441 sophisticated algorithms will want to use target-specific intrinsics to
4442 take full advantage of a specific target.
4444 .. _i_extractelement:
4446 '``extractelement``' Instruction
4447 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4454 <result> = extractelement <n x <ty>> <val>, i32 <idx> ; yields <ty>
4459 The '``extractelement``' instruction extracts a single scalar element
4460 from a vector at a specified index.
4465 The first operand of an '``extractelement``' instruction is a value of
4466 :ref:`vector <t_vector>` type. The second operand is an index indicating
4467 the position from which to extract the element. The index may be a
4473 The result is a scalar of the same type as the element type of ``val``.
4474 Its value is the value at position ``idx`` of ``val``. If ``idx``
4475 exceeds the length of ``val``, the results are undefined.
4480 .. code-block:: llvm
4482 <result> = extractelement <4 x i32> %vec, i32 0 ; yields i32
4484 .. _i_insertelement:
4486 '``insertelement``' Instruction
4487 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4494 <result> = insertelement <n x <ty>> <val>, <ty> <elt>, i32 <idx> ; yields <n x <ty>>
4499 The '``insertelement``' instruction inserts a scalar element into a
4500 vector at a specified index.
4505 The first operand of an '``insertelement``' instruction is a value of
4506 :ref:`vector <t_vector>` type. The second operand is a scalar value whose
4507 type must equal the element type of the first operand. The third operand
4508 is an index indicating the position at which to insert the value. The
4509 index may be a variable.
4514 The result is a vector of the same type as ``val``. Its element values
4515 are those of ``val`` except at position ``idx``, where it gets the value
4516 ``elt``. If ``idx`` exceeds the length of ``val``, the results are
4522 .. code-block:: llvm
4524 <result> = insertelement <4 x i32> %vec, i32 1, i32 0 ; yields <4 x i32>
4526 .. _i_shufflevector:
4528 '``shufflevector``' Instruction
4529 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4536 <result> = shufflevector <n x <ty>> <v1>, <n x <ty>> <v2>, <m x i32> <mask> ; yields <m x <ty>>
4541 The '``shufflevector``' instruction constructs a permutation of elements
4542 from two input vectors, returning a vector with the same element type as
4543 the input and length that is the same as the shuffle mask.
4548 The first two operands of a '``shufflevector``' instruction are vectors
4549 with the same type. The third argument is a shuffle mask whose element
4550 type is always 'i32'. The result of the instruction is a vector whose
4551 length is the same as the shuffle mask and whose element type is the
4552 same as the element type of the first two operands.
4554 The shuffle mask operand is required to be a constant vector with either
4555 constant integer or undef values.
4560 The elements of the two input vectors are numbered from left to right
4561 across both of the vectors. The shuffle mask operand specifies, for each
4562 element of the result vector, which element of the two input vectors the
4563 result element gets. The element selector may be undef (meaning "don't
4564 care") and the second operand may be undef if performing a shuffle from
4570 .. code-block:: llvm
4572 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
4573 <4 x i32> <i32 0, i32 4, i32 1, i32 5> ; yields <4 x i32>
4574 <result> = shufflevector <4 x i32> %v1, <4 x i32> undef,
4575 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32> - Identity shuffle.
4576 <result> = shufflevector <8 x i32> %v1, <8 x i32> undef,
4577 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32>
4578 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
4579 <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7 > ; yields <8 x i32>
4581 Aggregate Operations
4582 --------------------
4584 LLVM supports several instructions for working with
4585 :ref:`aggregate <t_aggregate>` values.
4589 '``extractvalue``' Instruction
4590 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4597 <result> = extractvalue <aggregate type> <val>, <idx>{, <idx>}*
4602 The '``extractvalue``' instruction extracts the value of a member field
4603 from an :ref:`aggregate <t_aggregate>` value.
4608 The first operand of an '``extractvalue``' instruction is a value of
4609 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The operands are
4610 constant indices to specify which value to extract in a similar manner
4611 as indices in a '``getelementptr``' instruction.
4613 The major differences to ``getelementptr`` indexing are:
4615 - Since the value being indexed is not a pointer, the first index is
4616 omitted and assumed to be zero.
4617 - At least one index must be specified.
4618 - Not only struct indices but also array indices must be in bounds.
4623 The result is the value at the position in the aggregate specified by
4629 .. code-block:: llvm
4631 <result> = extractvalue {i32, float} %agg, 0 ; yields i32
4635 '``insertvalue``' Instruction
4636 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4643 <result> = insertvalue <aggregate type> <val>, <ty> <elt>, <idx>{, <idx>}* ; yields <aggregate type>
4648 The '``insertvalue``' instruction inserts a value into a member field in
4649 an :ref:`aggregate <t_aggregate>` value.
4654 The first operand of an '``insertvalue``' instruction is a value of
4655 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The second operand is
4656 a first-class value to insert. The following operands are constant
4657 indices indicating the position at which to insert the value in a
4658 similar manner as indices in a '``extractvalue``' instruction. The value
4659 to insert must have the same type as the value identified by the
4665 The result is an aggregate of the same type as ``val``. Its value is
4666 that of ``val`` except that the value at the position specified by the
4667 indices is that of ``elt``.
4672 .. code-block:: llvm
4674 %agg1 = insertvalue {i32, float} undef, i32 1, 0 ; yields {i32 1, float undef}
4675 %agg2 = insertvalue {i32, float} %agg1, float %val, 1 ; yields {i32 1, float %val}
4676 %agg3 = insertvalue {i32, {float}} %agg1, float %val, 1, 0 ; yields {i32 1, float %val}
4680 Memory Access and Addressing Operations
4681 ---------------------------------------
4683 A key design point of an SSA-based representation is how it represents
4684 memory. In LLVM, no memory locations are in SSA form, which makes things
4685 very simple. This section describes how to read, write, and allocate
4690 '``alloca``' Instruction
4691 ^^^^^^^^^^^^^^^^^^^^^^^^
4698 <result> = alloca <type>[, inalloca][, <ty> <NumElements>][, align <alignment>] ; yields {type*}:result
4703 The '``alloca``' instruction allocates memory on the stack frame of the
4704 currently executing function, to be automatically released when this
4705 function returns to its caller. The object is always allocated in the
4706 generic address space (address space zero).
4711 The '``alloca``' instruction allocates ``sizeof(<type>)*NumElements``
4712 bytes of memory on the runtime stack, returning a pointer of the
4713 appropriate type to the program. If "NumElements" is specified, it is
4714 the number of elements allocated, otherwise "NumElements" is defaulted
4715 to be one. If a constant alignment is specified, the value result of the
4716 allocation is guaranteed to be aligned to at least that boundary. If not
4717 specified, or if zero, the target can choose to align the allocation on
4718 any convenient boundary compatible with the type.
4720 '``type``' may be any sized type.
4725 Memory is allocated; a pointer is returned. The operation is undefined
4726 if there is insufficient stack space for the allocation. '``alloca``'d
4727 memory is automatically released when the function returns. The
4728 '``alloca``' instruction is commonly used to represent automatic
4729 variables that must have an address available. When the function returns
4730 (either with the ``ret`` or ``resume`` instructions), the memory is
4731 reclaimed. Allocating zero bytes is legal, but the result is undefined.
4732 The order in which memory is allocated (ie., which way the stack grows)
4738 .. code-block:: llvm
4740 %ptr = alloca i32 ; yields {i32*}:ptr
4741 %ptr = alloca i32, i32 4 ; yields {i32*}:ptr
4742 %ptr = alloca i32, i32 4, align 1024 ; yields {i32*}:ptr
4743 %ptr = alloca i32, align 1024 ; yields {i32*}:ptr
4747 '``load``' Instruction
4748 ^^^^^^^^^^^^^^^^^^^^^^
4755 <result> = load [volatile] <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>][, !invariant.load !<index>]
4756 <result> = load atomic [volatile] <ty>* <pointer> [singlethread] <ordering>, align <alignment>
4757 !<index> = !{ i32 1 }
4762 The '``load``' instruction is used to read from memory.
4767 The argument to the ``load`` instruction specifies the memory address
4768 from which to load. The pointer must point to a :ref:`first
4769 class <t_firstclass>` type. If the ``load`` is marked as ``volatile``,
4770 then the optimizer is not allowed to modify the number or order of
4771 execution of this ``load`` with other :ref:`volatile
4772 operations <volatile>`.
4774 If the ``load`` is marked as ``atomic``, it takes an extra
4775 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
4776 ``release`` and ``acq_rel`` orderings are not valid on ``load``
4777 instructions. Atomic loads produce :ref:`defined <memmodel>` results
4778 when they may see multiple atomic stores. The type of the pointee must
4779 be an integer type whose bit width is a power of two greater than or
4780 equal to eight and less than or equal to a target-specific size limit.
4781 ``align`` must be explicitly specified on atomic loads, and the load has
4782 undefined behavior if the alignment is not set to a value which is at
4783 least the size in bytes of the pointee. ``!nontemporal`` does not have
4784 any defined semantics for atomic loads.
4786 The optional constant ``align`` argument specifies the alignment of the
4787 operation (that is, the alignment of the memory address). A value of 0
4788 or an omitted ``align`` argument means that the operation has the ABI
4789 alignment for the target. It is the responsibility of the code emitter
4790 to ensure that the alignment information is correct. Overestimating the
4791 alignment results in undefined behavior. Underestimating the alignment
4792 may produce less efficient code. An alignment of 1 is always safe.
4794 The optional ``!nontemporal`` metadata must reference a single
4795 metadata name ``<index>`` corresponding to a metadata node with one
4796 ``i32`` entry of value 1. The existence of the ``!nontemporal``
4797 metadata on the instruction tells the optimizer and code generator
4798 that this load is not expected to be reused in the cache. The code
4799 generator may select special instructions to save cache bandwidth, such
4800 as the ``MOVNT`` instruction on x86.
4802 The optional ``!invariant.load`` metadata must reference a single
4803 metadata name ``<index>`` corresponding to a metadata node with no
4804 entries. The existence of the ``!invariant.load`` metadata on the
4805 instruction tells the optimizer and code generator that this load
4806 address points to memory which does not change value during program
4807 execution. The optimizer may then move this load around, for example, by
4808 hoisting it out of loops using loop invariant code motion.
4813 The location of memory pointed to is loaded. If the value being loaded
4814 is of scalar type then the number of bytes read does not exceed the
4815 minimum number of bytes needed to hold all bits of the type. For
4816 example, loading an ``i24`` reads at most three bytes. When loading a
4817 value of a type like ``i20`` with a size that is not an integral number
4818 of bytes, the result is undefined if the value was not originally
4819 written using a store of the same type.
4824 .. code-block:: llvm
4826 %ptr = alloca i32 ; yields {i32*}:ptr
4827 store i32 3, i32* %ptr ; yields {void}
4828 %val = load i32* %ptr ; yields {i32}:val = i32 3
4832 '``store``' Instruction
4833 ^^^^^^^^^^^^^^^^^^^^^^^
4840 store [volatile] <ty> <value>, <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>] ; yields {void}
4841 store atomic [volatile] <ty> <value>, <ty>* <pointer> [singlethread] <ordering>, align <alignment> ; yields {void}
4846 The '``store``' instruction is used to write to memory.
4851 There are two arguments to the ``store`` instruction: a value to store
4852 and an address at which to store it. The type of the ``<pointer>``
4853 operand must be a pointer to the :ref:`first class <t_firstclass>` type of
4854 the ``<value>`` operand. If the ``store`` is marked as ``volatile``,
4855 then the optimizer is not allowed to modify the number or order of
4856 execution of this ``store`` with other :ref:`volatile
4857 operations <volatile>`.
4859 If the ``store`` is marked as ``atomic``, it takes an extra
4860 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
4861 ``acquire`` and ``acq_rel`` orderings aren't valid on ``store``
4862 instructions. Atomic loads produce :ref:`defined <memmodel>` results
4863 when they may see multiple atomic stores. The type of the pointee must
4864 be an integer type whose bit width is a power of two greater than or
4865 equal to eight and less than or equal to a target-specific size limit.
4866 ``align`` must be explicitly specified on atomic stores, and the store
4867 has undefined behavior if the alignment is not set to a value which is
4868 at least the size in bytes of the pointee. ``!nontemporal`` does not
4869 have any defined semantics for atomic stores.
4871 The optional constant ``align`` argument specifies the alignment of the
4872 operation (that is, the alignment of the memory address). A value of 0
4873 or an omitted ``align`` argument means that the operation has the ABI
4874 alignment for the target. It is the responsibility of the code emitter
4875 to ensure that the alignment information is correct. Overestimating the
4876 alignment results in undefined behavior. Underestimating the
4877 alignment may produce less efficient code. An alignment of 1 is always
4880 The optional ``!nontemporal`` metadata must reference a single metadata
4881 name ``<index>`` corresponding to a metadata node with one ``i32`` entry of
4882 value 1. The existence of the ``!nontemporal`` metadata on the instruction
4883 tells the optimizer and code generator that this load is not expected to
4884 be reused in the cache. The code generator may select special
4885 instructions to save cache bandwidth, such as the MOVNT instruction on
4891 The contents of memory are updated to contain ``<value>`` at the
4892 location specified by the ``<pointer>`` operand. If ``<value>`` is
4893 of scalar type then the number of bytes written does not exceed the
4894 minimum number of bytes needed to hold all bits of the type. For
4895 example, storing an ``i24`` writes at most three bytes. When writing a
4896 value of a type like ``i20`` with a size that is not an integral number
4897 of bytes, it is unspecified what happens to the extra bits that do not
4898 belong to the type, but they will typically be overwritten.
4903 .. code-block:: llvm
4905 %ptr = alloca i32 ; yields {i32*}:ptr
4906 store i32 3, i32* %ptr ; yields {void}
4907 %val = load i32* %ptr ; yields {i32}:val = i32 3
4911 '``fence``' Instruction
4912 ^^^^^^^^^^^^^^^^^^^^^^^
4919 fence [singlethread] <ordering> ; yields {void}
4924 The '``fence``' instruction is used to introduce happens-before edges
4930 '``fence``' instructions take an :ref:`ordering <ordering>` argument which
4931 defines what *synchronizes-with* edges they add. They can only be given
4932 ``acquire``, ``release``, ``acq_rel``, and ``seq_cst`` orderings.
4937 A fence A which has (at least) ``release`` ordering semantics
4938 *synchronizes with* a fence B with (at least) ``acquire`` ordering
4939 semantics if and only if there exist atomic operations X and Y, both
4940 operating on some atomic object M, such that A is sequenced before X, X
4941 modifies M (either directly or through some side effect of a sequence
4942 headed by X), Y is sequenced before B, and Y observes M. This provides a
4943 *happens-before* dependency between A and B. Rather than an explicit
4944 ``fence``, one (but not both) of the atomic operations X or Y might
4945 provide a ``release`` or ``acquire`` (resp.) ordering constraint and
4946 still *synchronize-with* the explicit ``fence`` and establish the
4947 *happens-before* edge.
4949 A ``fence`` which has ``seq_cst`` ordering, in addition to having both
4950 ``acquire`` and ``release`` semantics specified above, participates in
4951 the global program order of other ``seq_cst`` operations and/or fences.
4953 The optional ":ref:`singlethread <singlethread>`" argument specifies
4954 that the fence only synchronizes with other fences in the same thread.
4955 (This is useful for interacting with signal handlers.)
4960 .. code-block:: llvm
4962 fence acquire ; yields {void}
4963 fence singlethread seq_cst ; yields {void}
4967 '``cmpxchg``' Instruction
4968 ^^^^^^^^^^^^^^^^^^^^^^^^^
4975 cmpxchg [volatile] <ty>* <pointer>, <ty> <cmp>, <ty> <new> [singlethread] <ordering> ; yields {ty}
4980 The '``cmpxchg``' instruction is used to atomically modify memory. It
4981 loads a value in memory and compares it to a given value. If they are
4982 equal, it stores a new value into the memory.
4987 There are three arguments to the '``cmpxchg``' instruction: an address
4988 to operate on, a value to compare to the value currently be at that
4989 address, and a new value to place at that address if the compared values
4990 are equal. The type of '<cmp>' must be an integer type whose bit width
4991 is a power of two greater than or equal to eight and less than or equal
4992 to a target-specific size limit. '<cmp>' and '<new>' must have the same
4993 type, and the type of '<pointer>' must be a pointer to that type. If the
4994 ``cmpxchg`` is marked as ``volatile``, then the optimizer is not allowed
4995 to modify the number or order of execution of this ``cmpxchg`` with
4996 other :ref:`volatile operations <volatile>`.
4998 The :ref:`ordering <ordering>` argument specifies how this ``cmpxchg``
4999 synchronizes with other atomic operations.
5001 The optional "``singlethread``" argument declares that the ``cmpxchg``
5002 is only atomic with respect to code (usually signal handlers) running in
5003 the same thread as the ``cmpxchg``. Otherwise the cmpxchg is atomic with
5004 respect to all other code in the system.
5006 The pointer passed into cmpxchg must have alignment greater than or
5007 equal to the size in memory of the operand.
5012 The contents of memory at the location specified by the '``<pointer>``'
5013 operand is read and compared to '``<cmp>``'; if the read value is the
5014 equal, '``<new>``' is written. The original value at the location is
5017 A successful ``cmpxchg`` is a read-modify-write instruction for the purpose
5018 of identifying release sequences. A failed ``cmpxchg`` is equivalent to an
5019 atomic load with an ordering parameter determined by dropping any
5020 ``release`` part of the ``cmpxchg``'s ordering.
5025 .. code-block:: llvm
5028 %orig = atomic load i32* %ptr unordered ; yields {i32}
5032 %cmp = phi i32 [ %orig, %entry ], [%old, %loop]
5033 %squared = mul i32 %cmp, %cmp
5034 %old = cmpxchg i32* %ptr, i32 %cmp, i32 %squared ; yields {i32}
5035 %success = icmp eq i32 %cmp, %old
5036 br i1 %success, label %done, label %loop
5043 '``atomicrmw``' Instruction
5044 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
5051 atomicrmw [volatile] <operation> <ty>* <pointer>, <ty> <value> [singlethread] <ordering> ; yields {ty}
5056 The '``atomicrmw``' instruction is used to atomically modify memory.
5061 There are three arguments to the '``atomicrmw``' instruction: an
5062 operation to apply, an address whose value to modify, an argument to the
5063 operation. The operation must be one of the following keywords:
5077 The type of '<value>' must be an integer type whose bit width is a power
5078 of two greater than or equal to eight and less than or equal to a
5079 target-specific size limit. The type of the '``<pointer>``' operand must
5080 be a pointer to that type. If the ``atomicrmw`` is marked as
5081 ``volatile``, then the optimizer is not allowed to modify the number or
5082 order of execution of this ``atomicrmw`` with other :ref:`volatile
5083 operations <volatile>`.
5088 The contents of memory at the location specified by the '``<pointer>``'
5089 operand are atomically read, modified, and written back. The original
5090 value at the location is returned. The modification is specified by the
5093 - xchg: ``*ptr = val``
5094 - add: ``*ptr = *ptr + val``
5095 - sub: ``*ptr = *ptr - val``
5096 - and: ``*ptr = *ptr & val``
5097 - nand: ``*ptr = ~(*ptr & val)``
5098 - or: ``*ptr = *ptr | val``
5099 - xor: ``*ptr = *ptr ^ val``
5100 - max: ``*ptr = *ptr > val ? *ptr : val`` (using a signed comparison)
5101 - min: ``*ptr = *ptr < val ? *ptr : val`` (using a signed comparison)
5102 - umax: ``*ptr = *ptr > val ? *ptr : val`` (using an unsigned
5104 - umin: ``*ptr = *ptr < val ? *ptr : val`` (using an unsigned
5110 .. code-block:: llvm
5112 %old = atomicrmw add i32* %ptr, i32 1 acquire ; yields {i32}
5114 .. _i_getelementptr:
5116 '``getelementptr``' Instruction
5117 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5124 <result> = getelementptr <pty>* <ptrval>{, <ty> <idx>}*
5125 <result> = getelementptr inbounds <pty>* <ptrval>{, <ty> <idx>}*
5126 <result> = getelementptr <ptr vector> ptrval, <vector index type> idx
5131 The '``getelementptr``' instruction is used to get the address of a
5132 subelement of an :ref:`aggregate <t_aggregate>` data structure. It performs
5133 address calculation only and does not access memory.
5138 The first argument is always a pointer or a vector of pointers, and
5139 forms the basis of the calculation. The remaining arguments are indices
5140 that indicate which of the elements of the aggregate object are indexed.
5141 The interpretation of each index is dependent on the type being indexed
5142 into. The first index always indexes the pointer value given as the
5143 first argument, the second index indexes a value of the type pointed to
5144 (not necessarily the value directly pointed to, since the first index
5145 can be non-zero), etc. The first type indexed into must be a pointer
5146 value, subsequent types can be arrays, vectors, and structs. Note that
5147 subsequent types being indexed into can never be pointers, since that
5148 would require loading the pointer before continuing calculation.
5150 The type of each index argument depends on the type it is indexing into.
5151 When indexing into a (optionally packed) structure, only ``i32`` integer
5152 **constants** are allowed (when using a vector of indices they must all
5153 be the **same** ``i32`` integer constant). When indexing into an array,
5154 pointer or vector, integers of any width are allowed, and they are not
5155 required to be constant. These integers are treated as signed values
5158 For example, let's consider a C code fragment and how it gets compiled
5174 int *foo(struct ST *s) {
5175 return &s[1].Z.B[5][13];
5178 The LLVM code generated by Clang is:
5180 .. code-block:: llvm
5182 %struct.RT = type { i8, [10 x [20 x i32]], i8 }
5183 %struct.ST = type { i32, double, %struct.RT }
5185 define i32* @foo(%struct.ST* %s) nounwind uwtable readnone optsize ssp {
5187 %arrayidx = getelementptr inbounds %struct.ST* %s, i64 1, i32 2, i32 1, i64 5, i64 13
5194 In the example above, the first index is indexing into the
5195 '``%struct.ST*``' type, which is a pointer, yielding a '``%struct.ST``'
5196 = '``{ i32, double, %struct.RT }``' type, a structure. The second index
5197 indexes into the third element of the structure, yielding a
5198 '``%struct.RT``' = '``{ i8 , [10 x [20 x i32]], i8 }``' type, another
5199 structure. The third index indexes into the second element of the
5200 structure, yielding a '``[10 x [20 x i32]]``' type, an array. The two
5201 dimensions of the array are subscripted into, yielding an '``i32``'
5202 type. The '``getelementptr``' instruction returns a pointer to this
5203 element, thus computing a value of '``i32*``' type.
5205 Note that it is perfectly legal to index partially through a structure,
5206 returning a pointer to an inner element. Because of this, the LLVM code
5207 for the given testcase is equivalent to:
5209 .. code-block:: llvm
5211 define i32* @foo(%struct.ST* %s) {
5212 %t1 = getelementptr %struct.ST* %s, i32 1 ; yields %struct.ST*:%t1
5213 %t2 = getelementptr %struct.ST* %t1, i32 0, i32 2 ; yields %struct.RT*:%t2
5214 %t3 = getelementptr %struct.RT* %t2, i32 0, i32 1 ; yields [10 x [20 x i32]]*:%t3
5215 %t4 = getelementptr [10 x [20 x i32]]* %t3, i32 0, i32 5 ; yields [20 x i32]*:%t4
5216 %t5 = getelementptr [20 x i32]* %t4, i32 0, i32 13 ; yields i32*:%t5
5220 If the ``inbounds`` keyword is present, the result value of the
5221 ``getelementptr`` is a :ref:`poison value <poisonvalues>` if the base
5222 pointer is not an *in bounds* address of an allocated object, or if any
5223 of the addresses that would be formed by successive addition of the
5224 offsets implied by the indices to the base address with infinitely
5225 precise signed arithmetic are not an *in bounds* address of that
5226 allocated object. The *in bounds* addresses for an allocated object are
5227 all the addresses that point into the object, plus the address one byte
5228 past the end. In cases where the base is a vector of pointers the
5229 ``inbounds`` keyword applies to each of the computations element-wise.
5231 If the ``inbounds`` keyword is not present, the offsets are added to the
5232 base address with silently-wrapping two's complement arithmetic. If the
5233 offsets have a different width from the pointer, they are sign-extended
5234 or truncated to the width of the pointer. The result value of the
5235 ``getelementptr`` may be outside the object pointed to by the base
5236 pointer. The result value may not necessarily be used to access memory
5237 though, even if it happens to point into allocated storage. See the
5238 :ref:`Pointer Aliasing Rules <pointeraliasing>` section for more
5241 The getelementptr instruction is often confusing. For some more insight
5242 into how it works, see :doc:`the getelementptr FAQ <GetElementPtr>`.
5247 .. code-block:: llvm
5249 ; yields [12 x i8]*:aptr
5250 %aptr = getelementptr {i32, [12 x i8]}* %saptr, i64 0, i32 1
5252 %vptr = getelementptr {i32, <2 x i8>}* %svptr, i64 0, i32 1, i32 1
5254 %eptr = getelementptr [12 x i8]* %aptr, i64 0, i32 1
5256 %iptr = getelementptr [10 x i32]* @arr, i16 0, i16 0
5258 In cases where the pointer argument is a vector of pointers, each index
5259 must be a vector with the same number of elements. For example:
5261 .. code-block:: llvm
5263 %A = getelementptr <4 x i8*> %ptrs, <4 x i64> %offsets,
5265 Conversion Operations
5266 ---------------------
5268 The instructions in this category are the conversion instructions
5269 (casting) which all take a single operand and a type. They perform
5270 various bit conversions on the operand.
5272 '``trunc .. to``' Instruction
5273 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5280 <result> = trunc <ty> <value> to <ty2> ; yields ty2
5285 The '``trunc``' instruction truncates its operand to the type ``ty2``.
5290 The '``trunc``' instruction takes a value to trunc, and a type to trunc
5291 it to. Both types must be of :ref:`integer <t_integer>` types, or vectors
5292 of the same number of integers. The bit size of the ``value`` must be
5293 larger than the bit size of the destination type, ``ty2``. Equal sized
5294 types are not allowed.
5299 The '``trunc``' instruction truncates the high order bits in ``value``
5300 and converts the remaining bits to ``ty2``. Since the source size must
5301 be larger than the destination size, ``trunc`` cannot be a *no-op cast*.
5302 It will always truncate bits.
5307 .. code-block:: llvm
5309 %X = trunc i32 257 to i8 ; yields i8:1
5310 %Y = trunc i32 123 to i1 ; yields i1:true
5311 %Z = trunc i32 122 to i1 ; yields i1:false
5312 %W = trunc <2 x i16> <i16 8, i16 7> to <2 x i8> ; yields <i8 8, i8 7>
5314 '``zext .. to``' Instruction
5315 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5322 <result> = zext <ty> <value> to <ty2> ; yields ty2
5327 The '``zext``' instruction zero extends its operand to type ``ty2``.
5332 The '``zext``' instruction takes a value to cast, and a type to cast it
5333 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
5334 the same number of integers. The bit size of the ``value`` must be
5335 smaller than the bit size of the destination type, ``ty2``.
5340 The ``zext`` fills the high order bits of the ``value`` with zero bits
5341 until it reaches the size of the destination type, ``ty2``.
5343 When zero extending from i1, the result will always be either 0 or 1.
5348 .. code-block:: llvm
5350 %X = zext i32 257 to i64 ; yields i64:257
5351 %Y = zext i1 true to i32 ; yields i32:1
5352 %Z = zext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
5354 '``sext .. to``' Instruction
5355 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5362 <result> = sext <ty> <value> to <ty2> ; yields ty2
5367 The '``sext``' sign extends ``value`` to the type ``ty2``.
5372 The '``sext``' instruction takes a value to cast, and a type to cast it
5373 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
5374 the same number of integers. The bit size of the ``value`` must be
5375 smaller than the bit size of the destination type, ``ty2``.
5380 The '``sext``' instruction performs a sign extension by copying the sign
5381 bit (highest order bit) of the ``value`` until it reaches the bit size
5382 of the type ``ty2``.
5384 When sign extending from i1, the extension always results in -1 or 0.
5389 .. code-block:: llvm
5391 %X = sext i8 -1 to i16 ; yields i16 :65535
5392 %Y = sext i1 true to i32 ; yields i32:-1
5393 %Z = sext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
5395 '``fptrunc .. to``' Instruction
5396 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5403 <result> = fptrunc <ty> <value> to <ty2> ; yields ty2
5408 The '``fptrunc``' instruction truncates ``value`` to type ``ty2``.
5413 The '``fptrunc``' instruction takes a :ref:`floating point <t_floating>`
5414 value to cast and a :ref:`floating point <t_floating>` type to cast it to.
5415 The size of ``value`` must be larger than the size of ``ty2``. This
5416 implies that ``fptrunc`` cannot be used to make a *no-op cast*.
5421 The '``fptrunc``' instruction truncates a ``value`` from a larger
5422 :ref:`floating point <t_floating>` type to a smaller :ref:`floating
5423 point <t_floating>` type. If the value cannot fit within the
5424 destination type, ``ty2``, then the results are undefined.
5429 .. code-block:: llvm
5431 %X = fptrunc double 123.0 to float ; yields float:123.0
5432 %Y = fptrunc double 1.0E+300 to float ; yields undefined
5434 '``fpext .. to``' Instruction
5435 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5442 <result> = fpext <ty> <value> to <ty2> ; yields ty2
5447 The '``fpext``' extends a floating point ``value`` to a larger floating
5453 The '``fpext``' instruction takes a :ref:`floating point <t_floating>`
5454 ``value`` to cast, and a :ref:`floating point <t_floating>` type to cast it
5455 to. The source type must be smaller than the destination type.
5460 The '``fpext``' instruction extends the ``value`` from a smaller
5461 :ref:`floating point <t_floating>` type to a larger :ref:`floating
5462 point <t_floating>` type. The ``fpext`` cannot be used to make a
5463 *no-op cast* because it always changes bits. Use ``bitcast`` to make a
5464 *no-op cast* for a floating point cast.
5469 .. code-block:: llvm
5471 %X = fpext float 3.125 to double ; yields double:3.125000e+00
5472 %Y = fpext double %X to fp128 ; yields fp128:0xL00000000000000004000900000000000
5474 '``fptoui .. to``' Instruction
5475 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5482 <result> = fptoui <ty> <value> to <ty2> ; yields ty2
5487 The '``fptoui``' converts a floating point ``value`` to its unsigned
5488 integer equivalent of type ``ty2``.
5493 The '``fptoui``' instruction takes a value to cast, which must be a
5494 scalar or vector :ref:`floating point <t_floating>` value, and a type to
5495 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
5496 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
5497 type with the same number of elements as ``ty``
5502 The '``fptoui``' instruction converts its :ref:`floating
5503 point <t_floating>` operand into the nearest (rounding towards zero)
5504 unsigned integer value. If the value cannot fit in ``ty2``, the results
5510 .. code-block:: llvm
5512 %X = fptoui double 123.0 to i32 ; yields i32:123
5513 %Y = fptoui float 1.0E+300 to i1 ; yields undefined:1
5514 %Z = fptoui float 1.04E+17 to i8 ; yields undefined:1
5516 '``fptosi .. to``' Instruction
5517 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5524 <result> = fptosi <ty> <value> to <ty2> ; yields ty2
5529 The '``fptosi``' instruction converts :ref:`floating point <t_floating>`
5530 ``value`` to type ``ty2``.
5535 The '``fptosi``' instruction takes a value to cast, which must be a
5536 scalar or vector :ref:`floating point <t_floating>` value, and a type to
5537 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
5538 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
5539 type with the same number of elements as ``ty``
5544 The '``fptosi``' instruction converts its :ref:`floating
5545 point <t_floating>` operand into the nearest (rounding towards zero)
5546 signed integer value. If the value cannot fit in ``ty2``, the results
5552 .. code-block:: llvm
5554 %X = fptosi double -123.0 to i32 ; yields i32:-123
5555 %Y = fptosi float 1.0E-247 to i1 ; yields undefined:1
5556 %Z = fptosi float 1.04E+17 to i8 ; yields undefined:1
5558 '``uitofp .. to``' Instruction
5559 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5566 <result> = uitofp <ty> <value> to <ty2> ; yields ty2
5571 The '``uitofp``' instruction regards ``value`` as an unsigned integer
5572 and converts that value to the ``ty2`` type.
5577 The '``uitofp``' instruction takes a value to cast, which must be a
5578 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
5579 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
5580 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
5581 type with the same number of elements as ``ty``
5586 The '``uitofp``' instruction interprets its operand as an unsigned
5587 integer quantity and converts it to the corresponding floating point
5588 value. If the value cannot fit in the floating point value, the results
5594 .. code-block:: llvm
5596 %X = uitofp i32 257 to float ; yields float:257.0
5597 %Y = uitofp i8 -1 to double ; yields double:255.0
5599 '``sitofp .. to``' Instruction
5600 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5607 <result> = sitofp <ty> <value> to <ty2> ; yields ty2
5612 The '``sitofp``' instruction regards ``value`` as a signed integer and
5613 converts that value to the ``ty2`` type.
5618 The '``sitofp``' instruction takes a value to cast, which must be a
5619 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
5620 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
5621 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
5622 type with the same number of elements as ``ty``
5627 The '``sitofp``' instruction interprets its operand as a signed integer
5628 quantity and converts it to the corresponding floating point value. If
5629 the value cannot fit in the floating point value, the results are
5635 .. code-block:: llvm
5637 %X = sitofp i32 257 to float ; yields float:257.0
5638 %Y = sitofp i8 -1 to double ; yields double:-1.0
5642 '``ptrtoint .. to``' Instruction
5643 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5650 <result> = ptrtoint <ty> <value> to <ty2> ; yields ty2
5655 The '``ptrtoint``' instruction converts the pointer or a vector of
5656 pointers ``value`` to the integer (or vector of integers) type ``ty2``.
5661 The '``ptrtoint``' instruction takes a ``value`` to cast, which must be
5662 a a value of type :ref:`pointer <t_pointer>` or a vector of pointers, and a
5663 type to cast it to ``ty2``, which must be an :ref:`integer <t_integer>` or
5664 a vector of integers type.
5669 The '``ptrtoint``' instruction converts ``value`` to integer type
5670 ``ty2`` by interpreting the pointer value as an integer and either
5671 truncating or zero extending that value to the size of the integer type.
5672 If ``value`` is smaller than ``ty2`` then a zero extension is done. If
5673 ``value`` is larger than ``ty2`` then a truncation is done. If they are
5674 the same size, then nothing is done (*no-op cast*) other than a type
5680 .. code-block:: llvm
5682 %X = ptrtoint i32* %P to i8 ; yields truncation on 32-bit architecture
5683 %Y = ptrtoint i32* %P to i64 ; yields zero extension on 32-bit architecture
5684 %Z = ptrtoint <4 x i32*> %P to <4 x i64>; yields vector zero extension for a vector of addresses on 32-bit architecture
5688 '``inttoptr .. to``' Instruction
5689 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5696 <result> = inttoptr <ty> <value> to <ty2> ; yields ty2
5701 The '``inttoptr``' instruction converts an integer ``value`` to a
5702 pointer type, ``ty2``.
5707 The '``inttoptr``' instruction takes an :ref:`integer <t_integer>` value to
5708 cast, and a type to cast it to, which must be a :ref:`pointer <t_pointer>`
5714 The '``inttoptr``' instruction converts ``value`` to type ``ty2`` by
5715 applying either a zero extension or a truncation depending on the size
5716 of the integer ``value``. If ``value`` is larger than the size of a
5717 pointer then a truncation is done. If ``value`` is smaller than the size
5718 of a pointer then a zero extension is done. If they are the same size,
5719 nothing is done (*no-op cast*).
5724 .. code-block:: llvm
5726 %X = inttoptr i32 255 to i32* ; yields zero extension on 64-bit architecture
5727 %Y = inttoptr i32 255 to i32* ; yields no-op on 32-bit architecture
5728 %Z = inttoptr i64 0 to i32* ; yields truncation on 32-bit architecture
5729 %Z = inttoptr <4 x i32> %G to <4 x i8*>; yields truncation of vector G to four pointers
5733 '``bitcast .. to``' Instruction
5734 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5741 <result> = bitcast <ty> <value> to <ty2> ; yields ty2
5746 The '``bitcast``' instruction converts ``value`` to type ``ty2`` without
5752 The '``bitcast``' instruction takes a value to cast, which must be a
5753 non-aggregate first class value, and a type to cast it to, which must
5754 also be a non-aggregate :ref:`first class <t_firstclass>` type. The
5755 bit sizes of ``value`` and the destination type, ``ty2``, must be
5756 identical. If the source type is a pointer, the destination type must
5757 also be a pointer of the same size. This instruction supports bitwise
5758 conversion of vectors to integers and to vectors of other types (as
5759 long as they have the same size).
5764 The '``bitcast``' instruction converts ``value`` to type ``ty2``. It
5765 is always a *no-op cast* because no bits change with this
5766 conversion. The conversion is done as if the ``value`` had been stored
5767 to memory and read back as type ``ty2``. Pointer (or vector of
5768 pointers) types may only be converted to other pointer (or vector of
5769 pointers) types with the same address space through this instruction.
5770 To convert pointers to other types, use the :ref:`inttoptr <i_inttoptr>`
5771 or :ref:`ptrtoint <i_ptrtoint>` instructions first.
5776 .. code-block:: llvm
5778 %X = bitcast i8 255 to i8 ; yields i8 :-1
5779 %Y = bitcast i32* %x to sint* ; yields sint*:%x
5780 %Z = bitcast <2 x int> %V to i64; ; yields i64: %V
5781 %Z = bitcast <2 x i32*> %V to <2 x i64*> ; yields <2 x i64*>
5783 .. _i_addrspacecast:
5785 '``addrspacecast .. to``' Instruction
5786 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5793 <result> = addrspacecast <pty> <ptrval> to <pty2> ; yields pty2
5798 The '``addrspacecast``' instruction converts ``ptrval`` from ``pty`` in
5799 address space ``n`` to type ``pty2`` in address space ``m``.
5804 The '``addrspacecast``' instruction takes a pointer or vector of pointer value
5805 to cast and a pointer type to cast it to, which must have a different
5811 The '``addrspacecast``' instruction converts the pointer value
5812 ``ptrval`` to type ``pty2``. It can be a *no-op cast* or a complex
5813 value modification, depending on the target and the address space
5814 pair. Pointer conversions within the same address space must be
5815 performed with the ``bitcast`` instruction. Note that if the address space
5816 conversion is legal then both result and operand refer to the same memory
5822 .. code-block:: llvm
5824 %X = addrspacecast i32* %x to i32 addrspace(1)* ; yields i32 addrspace(1)*:%x
5825 %Y = addrspacecast i32 addrspace(1)* %y to i64 addrspace(2)* ; yields i64 addrspace(2)*:%y
5826 %Z = addrspacecast <4 x i32*> %z to <4 x float addrspace(3)*> ; yields <4 x float addrspace(3)*>:%z
5833 The instructions in this category are the "miscellaneous" instructions,
5834 which defy better classification.
5838 '``icmp``' Instruction
5839 ^^^^^^^^^^^^^^^^^^^^^^
5846 <result> = icmp <cond> <ty> <op1>, <op2> ; yields {i1} or {<N x i1>}:result
5851 The '``icmp``' instruction returns a boolean value or a vector of
5852 boolean values based on comparison of its two integer, integer vector,
5853 pointer, or pointer vector operands.
5858 The '``icmp``' instruction takes three operands. The first operand is
5859 the condition code indicating the kind of comparison to perform. It is
5860 not a value, just a keyword. The possible condition code are:
5863 #. ``ne``: not equal
5864 #. ``ugt``: unsigned greater than
5865 #. ``uge``: unsigned greater or equal
5866 #. ``ult``: unsigned less than
5867 #. ``ule``: unsigned less or equal
5868 #. ``sgt``: signed greater than
5869 #. ``sge``: signed greater or equal
5870 #. ``slt``: signed less than
5871 #. ``sle``: signed less or equal
5873 The remaining two arguments must be :ref:`integer <t_integer>` or
5874 :ref:`pointer <t_pointer>` or integer :ref:`vector <t_vector>` typed. They
5875 must also be identical types.
5880 The '``icmp``' compares ``op1`` and ``op2`` according to the condition
5881 code given as ``cond``. The comparison performed always yields either an
5882 :ref:`i1 <t_integer>` or vector of ``i1`` result, as follows:
5884 #. ``eq``: yields ``true`` if the operands are equal, ``false``
5885 otherwise. No sign interpretation is necessary or performed.
5886 #. ``ne``: yields ``true`` if the operands are unequal, ``false``
5887 otherwise. No sign interpretation is necessary or performed.
5888 #. ``ugt``: interprets the operands as unsigned values and yields
5889 ``true`` if ``op1`` is greater than ``op2``.
5890 #. ``uge``: interprets the operands as unsigned values and yields
5891 ``true`` if ``op1`` is greater than or equal to ``op2``.
5892 #. ``ult``: interprets the operands as unsigned values and yields
5893 ``true`` if ``op1`` is less than ``op2``.
5894 #. ``ule``: interprets the operands as unsigned values and yields
5895 ``true`` if ``op1`` is less than or equal to ``op2``.
5896 #. ``sgt``: interprets the operands as signed values and yields ``true``
5897 if ``op1`` is greater than ``op2``.
5898 #. ``sge``: interprets the operands as signed values and yields ``true``
5899 if ``op1`` is greater than or equal to ``op2``.
5900 #. ``slt``: interprets the operands as signed values and yields ``true``
5901 if ``op1`` is less than ``op2``.
5902 #. ``sle``: interprets the operands as signed values and yields ``true``
5903 if ``op1`` is less than or equal to ``op2``.
5905 If the operands are :ref:`pointer <t_pointer>` typed, the pointer values
5906 are compared as if they were integers.
5908 If the operands are integer vectors, then they are compared element by
5909 element. The result is an ``i1`` vector with the same number of elements
5910 as the values being compared. Otherwise, the result is an ``i1``.
5915 .. code-block:: llvm
5917 <result> = icmp eq i32 4, 5 ; yields: result=false
5918 <result> = icmp ne float* %X, %X ; yields: result=false
5919 <result> = icmp ult i16 4, 5 ; yields: result=true
5920 <result> = icmp sgt i16 4, 5 ; yields: result=false
5921 <result> = icmp ule i16 -4, 5 ; yields: result=false
5922 <result> = icmp sge i16 4, 5 ; yields: result=false
5924 Note that the code generator does not yet support vector types with the
5925 ``icmp`` instruction.
5929 '``fcmp``' Instruction
5930 ^^^^^^^^^^^^^^^^^^^^^^
5937 <result> = fcmp <cond> <ty> <op1>, <op2> ; yields {i1} or {<N x i1>}:result
5942 The '``fcmp``' instruction returns a boolean value or vector of boolean
5943 values based on comparison of its operands.
5945 If the operands are floating point scalars, then the result type is a
5946 boolean (:ref:`i1 <t_integer>`).
5948 If the operands are floating point vectors, then the result type is a
5949 vector of boolean with the same number of elements as the operands being
5955 The '``fcmp``' instruction takes three operands. The first operand is
5956 the condition code indicating the kind of comparison to perform. It is
5957 not a value, just a keyword. The possible condition code are:
5959 #. ``false``: no comparison, always returns false
5960 #. ``oeq``: ordered and equal
5961 #. ``ogt``: ordered and greater than
5962 #. ``oge``: ordered and greater than or equal
5963 #. ``olt``: ordered and less than
5964 #. ``ole``: ordered and less than or equal
5965 #. ``one``: ordered and not equal
5966 #. ``ord``: ordered (no nans)
5967 #. ``ueq``: unordered or equal
5968 #. ``ugt``: unordered or greater than
5969 #. ``uge``: unordered or greater than or equal
5970 #. ``ult``: unordered or less than
5971 #. ``ule``: unordered or less than or equal
5972 #. ``une``: unordered or not equal
5973 #. ``uno``: unordered (either nans)
5974 #. ``true``: no comparison, always returns true
5976 *Ordered* means that neither operand is a QNAN while *unordered* means
5977 that either operand may be a QNAN.
5979 Each of ``val1`` and ``val2`` arguments must be either a :ref:`floating
5980 point <t_floating>` type or a :ref:`vector <t_vector>` of floating point
5981 type. They must have identical types.
5986 The '``fcmp``' instruction compares ``op1`` and ``op2`` according to the
5987 condition code given as ``cond``. If the operands are vectors, then the
5988 vectors are compared element by element. Each comparison performed
5989 always yields an :ref:`i1 <t_integer>` result, as follows:
5991 #. ``false``: always yields ``false``, regardless of operands.
5992 #. ``oeq``: yields ``true`` if both operands are not a QNAN and ``op1``
5993 is equal to ``op2``.
5994 #. ``ogt``: yields ``true`` if both operands are not a QNAN and ``op1``
5995 is greater than ``op2``.
5996 #. ``oge``: yields ``true`` if both operands are not a QNAN and ``op1``
5997 is greater than or equal to ``op2``.
5998 #. ``olt``: yields ``true`` if both operands are not a QNAN and ``op1``
5999 is less than ``op2``.
6000 #. ``ole``: yields ``true`` if both operands are not a QNAN and ``op1``
6001 is less than or equal to ``op2``.
6002 #. ``one``: yields ``true`` if both operands are not a QNAN and ``op1``
6003 is not equal to ``op2``.
6004 #. ``ord``: yields ``true`` if both operands are not a QNAN.
6005 #. ``ueq``: yields ``true`` if either operand is a QNAN or ``op1`` is
6007 #. ``ugt``: yields ``true`` if either operand is a QNAN or ``op1`` is
6008 greater than ``op2``.
6009 #. ``uge``: yields ``true`` if either operand is a QNAN or ``op1`` is
6010 greater than or equal to ``op2``.
6011 #. ``ult``: yields ``true`` if either operand is a QNAN or ``op1`` is
6013 #. ``ule``: yields ``true`` if either operand is a QNAN or ``op1`` is
6014 less than or equal to ``op2``.
6015 #. ``une``: yields ``true`` if either operand is a QNAN or ``op1`` is
6016 not equal to ``op2``.
6017 #. ``uno``: yields ``true`` if either operand is a QNAN.
6018 #. ``true``: always yields ``true``, regardless of operands.
6023 .. code-block:: llvm
6025 <result> = fcmp oeq float 4.0, 5.0 ; yields: result=false
6026 <result> = fcmp one float 4.0, 5.0 ; yields: result=true
6027 <result> = fcmp olt float 4.0, 5.0 ; yields: result=true
6028 <result> = fcmp ueq double 1.0, 2.0 ; yields: result=false
6030 Note that the code generator does not yet support vector types with the
6031 ``fcmp`` instruction.
6035 '``phi``' Instruction
6036 ^^^^^^^^^^^^^^^^^^^^^
6043 <result> = phi <ty> [ <val0>, <label0>], ...
6048 The '``phi``' instruction is used to implement the φ node in the SSA
6049 graph representing the function.
6054 The type of the incoming values is specified with the first type field.
6055 After this, the '``phi``' instruction takes a list of pairs as
6056 arguments, with one pair for each predecessor basic block of the current
6057 block. Only values of :ref:`first class <t_firstclass>` type may be used as
6058 the value arguments to the PHI node. Only labels may be used as the
6061 There must be no non-phi instructions between the start of a basic block
6062 and the PHI instructions: i.e. PHI instructions must be first in a basic
6065 For the purposes of the SSA form, the use of each incoming value is
6066 deemed to occur on the edge from the corresponding predecessor block to
6067 the current block (but after any definition of an '``invoke``'
6068 instruction's return value on the same edge).
6073 At runtime, the '``phi``' instruction logically takes on the value
6074 specified by the pair corresponding to the predecessor basic block that
6075 executed just prior to the current block.
6080 .. code-block:: llvm
6082 Loop: ; Infinite loop that counts from 0 on up...
6083 %indvar = phi i32 [ 0, %LoopHeader ], [ %nextindvar, %Loop ]
6084 %nextindvar = add i32 %indvar, 1
6089 '``select``' Instruction
6090 ^^^^^^^^^^^^^^^^^^^^^^^^
6097 <result> = select selty <cond>, <ty> <val1>, <ty> <val2> ; yields ty
6099 selty is either i1 or {<N x i1>}
6104 The '``select``' instruction is used to choose one value based on a
6105 condition, without branching.
6110 The '``select``' instruction requires an 'i1' value or a vector of 'i1'
6111 values indicating the condition, and two values of the same :ref:`first
6112 class <t_firstclass>` type. If the val1/val2 are vectors and the
6113 condition is a scalar, then entire vectors are selected, not individual
6119 If the condition is an i1 and it evaluates to 1, the instruction returns
6120 the first value argument; otherwise, it returns the second value
6123 If the condition is a vector of i1, then the value arguments must be
6124 vectors of the same size, and the selection is done element by element.
6129 .. code-block:: llvm
6131 %X = select i1 true, i8 17, i8 42 ; yields i8:17
6135 '``call``' Instruction
6136 ^^^^^^^^^^^^^^^^^^^^^^
6143 <result> = [tail] call [cconv] [ret attrs] <ty> [<fnty>*] <fnptrval>(<function args>) [fn attrs]
6148 The '``call``' instruction represents a simple function call.
6153 This instruction requires several arguments:
6155 #. The optional "tail" marker indicates that the callee function does
6156 not access any allocas or varargs in the caller. Note that calls may
6157 be marked "tail" even if they do not occur before a
6158 :ref:`ret <i_ret>` instruction. If the "tail" marker is present, the
6159 function call is eligible for tail call optimization, but `might not
6160 in fact be optimized into a jump <CodeGenerator.html#tailcallopt>`_.
6161 The code generator may optimize calls marked "tail" with either 1)
6162 automatic `sibling call
6163 optimization <CodeGenerator.html#sibcallopt>`_ when the caller and
6164 callee have matching signatures, or 2) forced tail call optimization
6165 when the following extra requirements are met:
6167 - Caller and callee both have the calling convention ``fastcc``.
6168 - The call is in tail position (ret immediately follows call and ret
6169 uses value of call or is void).
6170 - Option ``-tailcallopt`` is enabled, or
6171 ``llvm::GuaranteedTailCallOpt`` is ``true``.
6172 - `Platform specific constraints are
6173 met. <CodeGenerator.html#tailcallopt>`_
6175 #. The optional "cconv" marker indicates which :ref:`calling
6176 convention <callingconv>` the call should use. If none is
6177 specified, the call defaults to using C calling conventions. The
6178 calling convention of the call must match the calling convention of
6179 the target function, or else the behavior is undefined.
6180 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
6181 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
6183 #. '``ty``': the type of the call instruction itself which is also the
6184 type of the return value. Functions that return no value are marked
6186 #. '``fnty``': shall be the signature of the pointer to function value
6187 being invoked. The argument types must match the types implied by
6188 this signature. This type can be omitted if the function is not
6189 varargs and if the function type does not return a pointer to a
6191 #. '``fnptrval``': An LLVM value containing a pointer to a function to
6192 be invoked. In most cases, this is a direct function invocation, but
6193 indirect ``call``'s are just as possible, calling an arbitrary pointer
6195 #. '``function args``': argument list whose types match the function
6196 signature argument types and parameter attributes. All arguments must
6197 be of :ref:`first class <t_firstclass>` type. If the function signature
6198 indicates the function accepts a variable number of arguments, the
6199 extra arguments can be specified.
6200 #. The optional :ref:`function attributes <fnattrs>` list. Only
6201 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
6202 attributes are valid here.
6207 The '``call``' instruction is used to cause control flow to transfer to
6208 a specified function, with its incoming arguments bound to the specified
6209 values. Upon a '``ret``' instruction in the called function, control
6210 flow continues with the instruction after the function call, and the
6211 return value of the function is bound to the result argument.
6216 .. code-block:: llvm
6218 %retval = call i32 @test(i32 %argc)
6219 call i32 (i8*, ...)* @printf(i8* %msg, i32 12, i8 42) ; yields i32
6220 %X = tail call i32 @foo() ; yields i32
6221 %Y = tail call fastcc i32 @foo() ; yields i32
6222 call void %foo(i8 97 signext)
6224 %struct.A = type { i32, i8 }
6225 %r = call %struct.A @foo() ; yields { 32, i8 }
6226 %gr = extractvalue %struct.A %r, 0 ; yields i32
6227 %gr1 = extractvalue %struct.A %r, 1 ; yields i8
6228 %Z = call void @foo() noreturn ; indicates that %foo never returns normally
6229 %ZZ = call zeroext i32 @bar() ; Return value is %zero extended
6231 llvm treats calls to some functions with names and arguments that match
6232 the standard C99 library as being the C99 library functions, and may
6233 perform optimizations or generate code for them under that assumption.
6234 This is something we'd like to change in the future to provide better
6235 support for freestanding environments and non-C-based languages.
6239 '``va_arg``' Instruction
6240 ^^^^^^^^^^^^^^^^^^^^^^^^
6247 <resultval> = va_arg <va_list*> <arglist>, <argty>
6252 The '``va_arg``' instruction is used to access arguments passed through
6253 the "variable argument" area of a function call. It is used to implement
6254 the ``va_arg`` macro in C.
6259 This instruction takes a ``va_list*`` value and the type of the
6260 argument. It returns a value of the specified argument type and
6261 increments the ``va_list`` to point to the next argument. The actual
6262 type of ``va_list`` is target specific.
6267 The '``va_arg``' instruction loads an argument of the specified type
6268 from the specified ``va_list`` and causes the ``va_list`` to point to
6269 the next argument. For more information, see the variable argument
6270 handling :ref:`Intrinsic Functions <int_varargs>`.
6272 It is legal for this instruction to be called in a function which does
6273 not take a variable number of arguments, for example, the ``vfprintf``
6276 ``va_arg`` is an LLVM instruction instead of an :ref:`intrinsic
6277 function <intrinsics>` because it takes a type as an argument.
6282 See the :ref:`variable argument processing <int_varargs>` section.
6284 Note that the code generator does not yet fully support va\_arg on many
6285 targets. Also, it does not currently support va\_arg with aggregate
6286 types on any target.
6290 '``landingpad``' Instruction
6291 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6298 <resultval> = landingpad <resultty> personality <type> <pers_fn> <clause>+
6299 <resultval> = landingpad <resultty> personality <type> <pers_fn> cleanup <clause>*
6301 <clause> := catch <type> <value>
6302 <clause> := filter <array constant type> <array constant>
6307 The '``landingpad``' instruction is used by `LLVM's exception handling
6308 system <ExceptionHandling.html#overview>`_ to specify that a basic block
6309 is a landing pad --- one where the exception lands, and corresponds to the
6310 code found in the ``catch`` portion of a ``try``/``catch`` sequence. It
6311 defines values supplied by the personality function (``pers_fn``) upon
6312 re-entry to the function. The ``resultval`` has the type ``resultty``.
6317 This instruction takes a ``pers_fn`` value. This is the personality
6318 function associated with the unwinding mechanism. The optional
6319 ``cleanup`` flag indicates that the landing pad block is a cleanup.
6321 A ``clause`` begins with the clause type --- ``catch`` or ``filter`` --- and
6322 contains the global variable representing the "type" that may be caught
6323 or filtered respectively. Unlike the ``catch`` clause, the ``filter``
6324 clause takes an array constant as its argument. Use
6325 "``[0 x i8**] undef``" for a filter which cannot throw. The
6326 '``landingpad``' instruction must contain *at least* one ``clause`` or
6327 the ``cleanup`` flag.
6332 The '``landingpad``' instruction defines the values which are set by the
6333 personality function (``pers_fn``) upon re-entry to the function, and
6334 therefore the "result type" of the ``landingpad`` instruction. As with
6335 calling conventions, how the personality function results are
6336 represented in LLVM IR is target specific.
6338 The clauses are applied in order from top to bottom. If two
6339 ``landingpad`` instructions are merged together through inlining, the
6340 clauses from the calling function are appended to the list of clauses.
6341 When the call stack is being unwound due to an exception being thrown,
6342 the exception is compared against each ``clause`` in turn. If it doesn't
6343 match any of the clauses, and the ``cleanup`` flag is not set, then
6344 unwinding continues further up the call stack.
6346 The ``landingpad`` instruction has several restrictions:
6348 - A landing pad block is a basic block which is the unwind destination
6349 of an '``invoke``' instruction.
6350 - A landing pad block must have a '``landingpad``' instruction as its
6351 first non-PHI instruction.
6352 - There can be only one '``landingpad``' instruction within the landing
6354 - A basic block that is not a landing pad block may not include a
6355 '``landingpad``' instruction.
6356 - All '``landingpad``' instructions in a function must have the same
6357 personality function.
6362 .. code-block:: llvm
6364 ;; A landing pad which can catch an integer.
6365 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6367 ;; A landing pad that is a cleanup.
6368 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6370 ;; A landing pad which can catch an integer and can only throw a double.
6371 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6373 filter [1 x i8**] [@_ZTId]
6380 LLVM supports the notion of an "intrinsic function". These functions
6381 have well known names and semantics and are required to follow certain
6382 restrictions. Overall, these intrinsics represent an extension mechanism
6383 for the LLVM language that does not require changing all of the
6384 transformations in LLVM when adding to the language (or the bitcode
6385 reader/writer, the parser, etc...).
6387 Intrinsic function names must all start with an "``llvm.``" prefix. This
6388 prefix is reserved in LLVM for intrinsic names; thus, function names may
6389 not begin with this prefix. Intrinsic functions must always be external
6390 functions: you cannot define the body of intrinsic functions. Intrinsic
6391 functions may only be used in call or invoke instructions: it is illegal
6392 to take the address of an intrinsic function. Additionally, because
6393 intrinsic functions are part of the LLVM language, it is required if any
6394 are added that they be documented here.
6396 Some intrinsic functions can be overloaded, i.e., the intrinsic
6397 represents a family of functions that perform the same operation but on
6398 different data types. Because LLVM can represent over 8 million
6399 different integer types, overloading is used commonly to allow an
6400 intrinsic function to operate on any integer type. One or more of the
6401 argument types or the result type can be overloaded to accept any
6402 integer type. Argument types may also be defined as exactly matching a
6403 previous argument's type or the result type. This allows an intrinsic
6404 function which accepts multiple arguments, but needs all of them to be
6405 of the same type, to only be overloaded with respect to a single
6406 argument or the result.
6408 Overloaded intrinsics will have the names of its overloaded argument
6409 types encoded into its function name, each preceded by a period. Only
6410 those types which are overloaded result in a name suffix. Arguments
6411 whose type is matched against another type do not. For example, the
6412 ``llvm.ctpop`` function can take an integer of any width and returns an
6413 integer of exactly the same integer width. This leads to a family of
6414 functions such as ``i8 @llvm.ctpop.i8(i8 %val)`` and
6415 ``i29 @llvm.ctpop.i29(i29 %val)``. Only one type, the return type, is
6416 overloaded, and only one type suffix is required. Because the argument's
6417 type is matched against the return type, it does not require its own
6420 To learn how to add an intrinsic function, please see the `Extending
6421 LLVM Guide <ExtendingLLVM.html>`_.
6425 Variable Argument Handling Intrinsics
6426 -------------------------------------
6428 Variable argument support is defined in LLVM with the
6429 :ref:`va_arg <i_va_arg>` instruction and these three intrinsic
6430 functions. These functions are related to the similarly named macros
6431 defined in the ``<stdarg.h>`` header file.
6433 All of these functions operate on arguments that use a target-specific
6434 value type "``va_list``". The LLVM assembly language reference manual
6435 does not define what this type is, so all transformations should be
6436 prepared to handle these functions regardless of the type used.
6438 This example shows how the :ref:`va_arg <i_va_arg>` instruction and the
6439 variable argument handling intrinsic functions are used.
6441 .. code-block:: llvm
6443 define i32 @test(i32 %X, ...) {
6444 ; Initialize variable argument processing
6446 %ap2 = bitcast i8** %ap to i8*
6447 call void @llvm.va_start(i8* %ap2)
6449 ; Read a single integer argument
6450 %tmp = va_arg i8** %ap, i32
6452 ; Demonstrate usage of llvm.va_copy and llvm.va_end
6454 %aq2 = bitcast i8** %aq to i8*
6455 call void @llvm.va_copy(i8* %aq2, i8* %ap2)
6456 call void @llvm.va_end(i8* %aq2)
6458 ; Stop processing of arguments.
6459 call void @llvm.va_end(i8* %ap2)
6463 declare void @llvm.va_start(i8*)
6464 declare void @llvm.va_copy(i8*, i8*)
6465 declare void @llvm.va_end(i8*)
6469 '``llvm.va_start``' Intrinsic
6470 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6477 declare void @llvm.va_start(i8* <arglist>)
6482 The '``llvm.va_start``' intrinsic initializes ``*<arglist>`` for
6483 subsequent use by ``va_arg``.
6488 The argument is a pointer to a ``va_list`` element to initialize.
6493 The '``llvm.va_start``' intrinsic works just like the ``va_start`` macro
6494 available in C. In a target-dependent way, it initializes the
6495 ``va_list`` element to which the argument points, so that the next call
6496 to ``va_arg`` will produce the first variable argument passed to the
6497 function. Unlike the C ``va_start`` macro, this intrinsic does not need
6498 to know the last argument of the function as the compiler can figure
6501 '``llvm.va_end``' Intrinsic
6502 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6509 declare void @llvm.va_end(i8* <arglist>)
6514 The '``llvm.va_end``' intrinsic destroys ``*<arglist>``, which has been
6515 initialized previously with ``llvm.va_start`` or ``llvm.va_copy``.
6520 The argument is a pointer to a ``va_list`` to destroy.
6525 The '``llvm.va_end``' intrinsic works just like the ``va_end`` macro
6526 available in C. In a target-dependent way, it destroys the ``va_list``
6527 element to which the argument points. Calls to
6528 :ref:`llvm.va_start <int_va_start>` and
6529 :ref:`llvm.va_copy <int_va_copy>` must be matched exactly with calls to
6534 '``llvm.va_copy``' Intrinsic
6535 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6542 declare void @llvm.va_copy(i8* <destarglist>, i8* <srcarglist>)
6547 The '``llvm.va_copy``' intrinsic copies the current argument position
6548 from the source argument list to the destination argument list.
6553 The first argument is a pointer to a ``va_list`` element to initialize.
6554 The second argument is a pointer to a ``va_list`` element to copy from.
6559 The '``llvm.va_copy``' intrinsic works just like the ``va_copy`` macro
6560 available in C. In a target-dependent way, it copies the source
6561 ``va_list`` element into the destination ``va_list`` element. This
6562 intrinsic is necessary because the `` llvm.va_start`` intrinsic may be
6563 arbitrarily complex and require, for example, memory allocation.
6565 Accurate Garbage Collection Intrinsics
6566 --------------------------------------
6568 LLVM support for `Accurate Garbage Collection <GarbageCollection.html>`_
6569 (GC) requires the implementation and generation of these intrinsics.
6570 These intrinsics allow identification of :ref:`GC roots on the
6571 stack <int_gcroot>`, as well as garbage collector implementations that
6572 require :ref:`read <int_gcread>` and :ref:`write <int_gcwrite>` barriers.
6573 Front-ends for type-safe garbage collected languages should generate
6574 these intrinsics to make use of the LLVM garbage collectors. For more
6575 details, see `Accurate Garbage Collection with
6576 LLVM <GarbageCollection.html>`_.
6578 The garbage collection intrinsics only operate on objects in the generic
6579 address space (address space zero).
6583 '``llvm.gcroot``' Intrinsic
6584 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6591 declare void @llvm.gcroot(i8** %ptrloc, i8* %metadata)
6596 The '``llvm.gcroot``' intrinsic declares the existence of a GC root to
6597 the code generator, and allows some metadata to be associated with it.
6602 The first argument specifies the address of a stack object that contains
6603 the root pointer. The second pointer (which must be either a constant or
6604 a global value address) contains the meta-data to be associated with the
6610 At runtime, a call to this intrinsic stores a null pointer into the
6611 "ptrloc" location. At compile-time, the code generator generates
6612 information to allow the runtime to find the pointer at GC safe points.
6613 The '``llvm.gcroot``' intrinsic may only be used in a function which
6614 :ref:`specifies a GC algorithm <gc>`.
6618 '``llvm.gcread``' Intrinsic
6619 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6626 declare i8* @llvm.gcread(i8* %ObjPtr, i8** %Ptr)
6631 The '``llvm.gcread``' intrinsic identifies reads of references from heap
6632 locations, allowing garbage collector implementations that require read
6638 The second argument is the address to read from, which should be an
6639 address allocated from the garbage collector. The first object is a
6640 pointer to the start of the referenced object, if needed by the language
6641 runtime (otherwise null).
6646 The '``llvm.gcread``' intrinsic has the same semantics as a load
6647 instruction, but may be replaced with substantially more complex code by
6648 the garbage collector runtime, as needed. The '``llvm.gcread``'
6649 intrinsic may only be used in a function which :ref:`specifies a GC
6654 '``llvm.gcwrite``' Intrinsic
6655 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6662 declare void @llvm.gcwrite(i8* %P1, i8* %Obj, i8** %P2)
6667 The '``llvm.gcwrite``' intrinsic identifies writes of references to heap
6668 locations, allowing garbage collector implementations that require write
6669 barriers (such as generational or reference counting collectors).
6674 The first argument is the reference to store, the second is the start of
6675 the object to store it to, and the third is the address of the field of
6676 Obj to store to. If the runtime does not require a pointer to the
6677 object, Obj may be null.
6682 The '``llvm.gcwrite``' intrinsic has the same semantics as a store
6683 instruction, but may be replaced with substantially more complex code by
6684 the garbage collector runtime, as needed. The '``llvm.gcwrite``'
6685 intrinsic may only be used in a function which :ref:`specifies a GC
6688 Code Generator Intrinsics
6689 -------------------------
6691 These intrinsics are provided by LLVM to expose special features that
6692 may only be implemented with code generator support.
6694 '``llvm.returnaddress``' Intrinsic
6695 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6702 declare i8 *@llvm.returnaddress(i32 <level>)
6707 The '``llvm.returnaddress``' intrinsic attempts to compute a
6708 target-specific value indicating the return address of the current
6709 function or one of its callers.
6714 The argument to this intrinsic indicates which function to return the
6715 address for. Zero indicates the calling function, one indicates its
6716 caller, etc. The argument is **required** to be a constant integer
6722 The '``llvm.returnaddress``' intrinsic either returns a pointer
6723 indicating the return address of the specified call frame, or zero if it
6724 cannot be identified. The value returned by this intrinsic is likely to
6725 be incorrect or 0 for arguments other than zero, so it should only be
6726 used for debugging purposes.
6728 Note that calling this intrinsic does not prevent function inlining or
6729 other aggressive transformations, so the value returned may not be that
6730 of the obvious source-language caller.
6732 '``llvm.frameaddress``' Intrinsic
6733 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6740 declare i8* @llvm.frameaddress(i32 <level>)
6745 The '``llvm.frameaddress``' intrinsic attempts to return the
6746 target-specific frame pointer value for the specified stack frame.
6751 The argument to this intrinsic indicates which function to return the
6752 frame pointer for. Zero indicates the calling function, one indicates
6753 its caller, etc. The argument is **required** to be a constant integer
6759 The '``llvm.frameaddress``' intrinsic either returns a pointer
6760 indicating the frame address of the specified call frame, or zero if it
6761 cannot be identified. The value returned by this intrinsic is likely to
6762 be incorrect or 0 for arguments other than zero, so it should only be
6763 used for debugging purposes.
6765 Note that calling this intrinsic does not prevent function inlining or
6766 other aggressive transformations, so the value returned may not be that
6767 of the obvious source-language caller.
6771 '``llvm.stacksave``' Intrinsic
6772 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6779 declare i8* @llvm.stacksave()
6784 The '``llvm.stacksave``' intrinsic is used to remember the current state
6785 of the function stack, for use with
6786 :ref:`llvm.stackrestore <int_stackrestore>`. This is useful for
6787 implementing language features like scoped automatic variable sized
6793 This intrinsic returns a opaque pointer value that can be passed to
6794 :ref:`llvm.stackrestore <int_stackrestore>`. When an
6795 ``llvm.stackrestore`` intrinsic is executed with a value saved from
6796 ``llvm.stacksave``, it effectively restores the state of the stack to
6797 the state it was in when the ``llvm.stacksave`` intrinsic executed. In
6798 practice, this pops any :ref:`alloca <i_alloca>` blocks from the stack that
6799 were allocated after the ``llvm.stacksave`` was executed.
6801 .. _int_stackrestore:
6803 '``llvm.stackrestore``' Intrinsic
6804 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6811 declare void @llvm.stackrestore(i8* %ptr)
6816 The '``llvm.stackrestore``' intrinsic is used to restore the state of
6817 the function stack to the state it was in when the corresponding
6818 :ref:`llvm.stacksave <int_stacksave>` intrinsic executed. This is
6819 useful for implementing language features like scoped automatic variable
6820 sized arrays in C99.
6825 See the description for :ref:`llvm.stacksave <int_stacksave>`.
6827 '``llvm.prefetch``' Intrinsic
6828 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6835 declare void @llvm.prefetch(i8* <address>, i32 <rw>, i32 <locality>, i32 <cache type>)
6840 The '``llvm.prefetch``' intrinsic is a hint to the code generator to
6841 insert a prefetch instruction if supported; otherwise, it is a noop.
6842 Prefetches have no effect on the behavior of the program but can change
6843 its performance characteristics.
6848 ``address`` is the address to be prefetched, ``rw`` is the specifier
6849 determining if the fetch should be for a read (0) or write (1), and
6850 ``locality`` is a temporal locality specifier ranging from (0) - no
6851 locality, to (3) - extremely local keep in cache. The ``cache type``
6852 specifies whether the prefetch is performed on the data (1) or
6853 instruction (0) cache. The ``rw``, ``locality`` and ``cache type``
6854 arguments must be constant integers.
6859 This intrinsic does not modify the behavior of the program. In
6860 particular, prefetches cannot trap and do not produce a value. On
6861 targets that support this intrinsic, the prefetch can provide hints to
6862 the processor cache for better performance.
6864 '``llvm.pcmarker``' Intrinsic
6865 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6872 declare void @llvm.pcmarker(i32 <id>)
6877 The '``llvm.pcmarker``' intrinsic is a method to export a Program
6878 Counter (PC) in a region of code to simulators and other tools. The
6879 method is target specific, but it is expected that the marker will use
6880 exported symbols to transmit the PC of the marker. The marker makes no
6881 guarantees that it will remain with any specific instruction after
6882 optimizations. It is possible that the presence of a marker will inhibit
6883 optimizations. The intended use is to be inserted after optimizations to
6884 allow correlations of simulation runs.
6889 ``id`` is a numerical id identifying the marker.
6894 This intrinsic does not modify the behavior of the program. Backends
6895 that do not support this intrinsic may ignore it.
6897 '``llvm.readcyclecounter``' Intrinsic
6898 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6905 declare i64 @llvm.readcyclecounter()
6910 The '``llvm.readcyclecounter``' intrinsic provides access to the cycle
6911 counter register (or similar low latency, high accuracy clocks) on those
6912 targets that support it. On X86, it should map to RDTSC. On Alpha, it
6913 should map to RPCC. As the backing counters overflow quickly (on the
6914 order of 9 seconds on alpha), this should only be used for small
6920 When directly supported, reading the cycle counter should not modify any
6921 memory. Implementations are allowed to either return a application
6922 specific value or a system wide value. On backends without support, this
6923 is lowered to a constant 0.
6925 Note that runtime support may be conditional on the privilege-level code is
6926 running at and the host platform.
6928 Standard C Library Intrinsics
6929 -----------------------------
6931 LLVM provides intrinsics for a few important standard C library
6932 functions. These intrinsics allow source-language front-ends to pass
6933 information about the alignment of the pointer arguments to the code
6934 generator, providing opportunity for more efficient code generation.
6938 '``llvm.memcpy``' Intrinsic
6939 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6944 This is an overloaded intrinsic. You can use ``llvm.memcpy`` on any
6945 integer bit width and for different address spaces. Not all targets
6946 support all bit widths however.
6950 declare void @llvm.memcpy.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
6951 i32 <len>, i32 <align>, i1 <isvolatile>)
6952 declare void @llvm.memcpy.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
6953 i64 <len>, i32 <align>, i1 <isvolatile>)
6958 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
6959 source location to the destination location.
6961 Note that, unlike the standard libc function, the ``llvm.memcpy.*``
6962 intrinsics do not return a value, takes extra alignment/isvolatile
6963 arguments and the pointers can be in specified address spaces.
6968 The first argument is a pointer to the destination, the second is a
6969 pointer to the source. The third argument is an integer argument
6970 specifying the number of bytes to copy, the fourth argument is the
6971 alignment of the source and destination locations, and the fifth is a
6972 boolean indicating a volatile access.
6974 If the call to this intrinsic has an alignment value that is not 0 or 1,
6975 then the caller guarantees that both the source and destination pointers
6976 are aligned to that boundary.
6978 If the ``isvolatile`` parameter is ``true``, the ``llvm.memcpy`` call is
6979 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
6980 very cleanly specified and it is unwise to depend on it.
6985 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
6986 source location to the destination location, which are not allowed to
6987 overlap. It copies "len" bytes of memory over. If the argument is known
6988 to be aligned to some boundary, this can be specified as the fourth
6989 argument, otherwise it should be set to 0 or 1 (both meaning no alignment).
6991 '``llvm.memmove``' Intrinsic
6992 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6997 This is an overloaded intrinsic. You can use llvm.memmove on any integer
6998 bit width and for different address space. Not all targets support all
7003 declare void @llvm.memmove.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
7004 i32 <len>, i32 <align>, i1 <isvolatile>)
7005 declare void @llvm.memmove.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
7006 i64 <len>, i32 <align>, i1 <isvolatile>)
7011 The '``llvm.memmove.*``' intrinsics move a block of memory from the
7012 source location to the destination location. It is similar to the
7013 '``llvm.memcpy``' intrinsic but allows the two memory locations to
7016 Note that, unlike the standard libc function, the ``llvm.memmove.*``
7017 intrinsics do not return a value, takes extra alignment/isvolatile
7018 arguments and the pointers can be in specified address spaces.
7023 The first argument is a pointer to the destination, the second is a
7024 pointer to the source. The third argument is an integer argument
7025 specifying the number of bytes to copy, the fourth argument is the
7026 alignment of the source and destination locations, and the fifth is a
7027 boolean indicating a volatile access.
7029 If the call to this intrinsic has an alignment value that is not 0 or 1,
7030 then the caller guarantees that the source and destination pointers are
7031 aligned to that boundary.
7033 If the ``isvolatile`` parameter is ``true``, the ``llvm.memmove`` call
7034 is a :ref:`volatile operation <volatile>`. The detailed access behavior is
7035 not very cleanly specified and it is unwise to depend on it.
7040 The '``llvm.memmove.*``' intrinsics copy a block of memory from the
7041 source location to the destination location, which may overlap. It
7042 copies "len" bytes of memory over. If the argument is known to be
7043 aligned to some boundary, this can be specified as the fourth argument,
7044 otherwise it should be set to 0 or 1 (both meaning no alignment).
7046 '``llvm.memset.*``' Intrinsics
7047 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7052 This is an overloaded intrinsic. You can use llvm.memset on any integer
7053 bit width and for different address spaces. However, not all targets
7054 support all bit widths.
7058 declare void @llvm.memset.p0i8.i32(i8* <dest>, i8 <val>,
7059 i32 <len>, i32 <align>, i1 <isvolatile>)
7060 declare void @llvm.memset.p0i8.i64(i8* <dest>, i8 <val>,
7061 i64 <len>, i32 <align>, i1 <isvolatile>)
7066 The '``llvm.memset.*``' intrinsics fill a block of memory with a
7067 particular byte value.
7069 Note that, unlike the standard libc function, the ``llvm.memset``
7070 intrinsic does not return a value and takes extra alignment/volatile
7071 arguments. Also, the destination can be in an arbitrary address space.
7076 The first argument is a pointer to the destination to fill, the second
7077 is the byte value with which to fill it, the third argument is an
7078 integer argument specifying the number of bytes to fill, and the fourth
7079 argument is the known alignment of the destination location.
7081 If the call to this intrinsic has an alignment value that is not 0 or 1,
7082 then the caller guarantees that the destination pointer is aligned to
7085 If the ``isvolatile`` parameter is ``true``, the ``llvm.memset`` call is
7086 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
7087 very cleanly specified and it is unwise to depend on it.
7092 The '``llvm.memset.*``' intrinsics fill "len" bytes of memory starting
7093 at the destination location. If the argument is known to be aligned to
7094 some boundary, this can be specified as the fourth argument, otherwise
7095 it should be set to 0 or 1 (both meaning no alignment).
7097 '``llvm.sqrt.*``' Intrinsic
7098 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7103 This is an overloaded intrinsic. You can use ``llvm.sqrt`` on any
7104 floating point or vector of floating point type. Not all targets support
7109 declare float @llvm.sqrt.f32(float %Val)
7110 declare double @llvm.sqrt.f64(double %Val)
7111 declare x86_fp80 @llvm.sqrt.f80(x86_fp80 %Val)
7112 declare fp128 @llvm.sqrt.f128(fp128 %Val)
7113 declare ppc_fp128 @llvm.sqrt.ppcf128(ppc_fp128 %Val)
7118 The '``llvm.sqrt``' intrinsics return the sqrt of the specified operand,
7119 returning the same value as the libm '``sqrt``' functions would. Unlike
7120 ``sqrt`` in libm, however, ``llvm.sqrt`` has undefined behavior for
7121 negative numbers other than -0.0 (which allows for better optimization,
7122 because there is no need to worry about errno being set).
7123 ``llvm.sqrt(-0.0)`` is defined to return -0.0 like IEEE sqrt.
7128 The argument and return value are floating point numbers of the same
7134 This function returns the sqrt of the specified operand if it is a
7135 nonnegative floating point number.
7137 '``llvm.powi.*``' Intrinsic
7138 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7143 This is an overloaded intrinsic. You can use ``llvm.powi`` on any
7144 floating point or vector of floating point type. Not all targets support
7149 declare float @llvm.powi.f32(float %Val, i32 %power)
7150 declare double @llvm.powi.f64(double %Val, i32 %power)
7151 declare x86_fp80 @llvm.powi.f80(x86_fp80 %Val, i32 %power)
7152 declare fp128 @llvm.powi.f128(fp128 %Val, i32 %power)
7153 declare ppc_fp128 @llvm.powi.ppcf128(ppc_fp128 %Val, i32 %power)
7158 The '``llvm.powi.*``' intrinsics return the first operand raised to the
7159 specified (positive or negative) power. The order of evaluation of
7160 multiplications is not defined. When a vector of floating point type is
7161 used, the second argument remains a scalar integer value.
7166 The second argument is an integer power, and the first is a value to
7167 raise to that power.
7172 This function returns the first value raised to the second power with an
7173 unspecified sequence of rounding operations.
7175 '``llvm.sin.*``' Intrinsic
7176 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7181 This is an overloaded intrinsic. You can use ``llvm.sin`` on any
7182 floating point or vector of floating point type. Not all targets support
7187 declare float @llvm.sin.f32(float %Val)
7188 declare double @llvm.sin.f64(double %Val)
7189 declare x86_fp80 @llvm.sin.f80(x86_fp80 %Val)
7190 declare fp128 @llvm.sin.f128(fp128 %Val)
7191 declare ppc_fp128 @llvm.sin.ppcf128(ppc_fp128 %Val)
7196 The '``llvm.sin.*``' intrinsics return the sine of the operand.
7201 The argument and return value are floating point numbers of the same
7207 This function returns the sine of the specified operand, returning the
7208 same values as the libm ``sin`` functions would, and handles error
7209 conditions in the same way.
7211 '``llvm.cos.*``' Intrinsic
7212 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7217 This is an overloaded intrinsic. You can use ``llvm.cos`` on any
7218 floating point or vector of floating point type. Not all targets support
7223 declare float @llvm.cos.f32(float %Val)
7224 declare double @llvm.cos.f64(double %Val)
7225 declare x86_fp80 @llvm.cos.f80(x86_fp80 %Val)
7226 declare fp128 @llvm.cos.f128(fp128 %Val)
7227 declare ppc_fp128 @llvm.cos.ppcf128(ppc_fp128 %Val)
7232 The '``llvm.cos.*``' intrinsics return the cosine of the operand.
7237 The argument and return value are floating point numbers of the same
7243 This function returns the cosine of the specified operand, returning the
7244 same values as the libm ``cos`` functions would, and handles error
7245 conditions in the same way.
7247 '``llvm.pow.*``' Intrinsic
7248 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7253 This is an overloaded intrinsic. You can use ``llvm.pow`` on any
7254 floating point or vector of floating point type. Not all targets support
7259 declare float @llvm.pow.f32(float %Val, float %Power)
7260 declare double @llvm.pow.f64(double %Val, double %Power)
7261 declare x86_fp80 @llvm.pow.f80(x86_fp80 %Val, x86_fp80 %Power)
7262 declare fp128 @llvm.pow.f128(fp128 %Val, fp128 %Power)
7263 declare ppc_fp128 @llvm.pow.ppcf128(ppc_fp128 %Val, ppc_fp128 Power)
7268 The '``llvm.pow.*``' intrinsics return the first operand raised to the
7269 specified (positive or negative) power.
7274 The second argument is a floating point power, and the first is a value
7275 to raise to that power.
7280 This function returns the first value raised to the second power,
7281 returning the same values as the libm ``pow`` functions would, and
7282 handles error conditions in the same way.
7284 '``llvm.exp.*``' Intrinsic
7285 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7290 This is an overloaded intrinsic. You can use ``llvm.exp`` on any
7291 floating point or vector of floating point type. Not all targets support
7296 declare float @llvm.exp.f32(float %Val)
7297 declare double @llvm.exp.f64(double %Val)
7298 declare x86_fp80 @llvm.exp.f80(x86_fp80 %Val)
7299 declare fp128 @llvm.exp.f128(fp128 %Val)
7300 declare ppc_fp128 @llvm.exp.ppcf128(ppc_fp128 %Val)
7305 The '``llvm.exp.*``' intrinsics perform the exp function.
7310 The argument and return value are floating point numbers of the same
7316 This function returns the same values as the libm ``exp`` functions
7317 would, and handles error conditions in the same way.
7319 '``llvm.exp2.*``' Intrinsic
7320 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7325 This is an overloaded intrinsic. You can use ``llvm.exp2`` on any
7326 floating point or vector of floating point type. Not all targets support
7331 declare float @llvm.exp2.f32(float %Val)
7332 declare double @llvm.exp2.f64(double %Val)
7333 declare x86_fp80 @llvm.exp2.f80(x86_fp80 %Val)
7334 declare fp128 @llvm.exp2.f128(fp128 %Val)
7335 declare ppc_fp128 @llvm.exp2.ppcf128(ppc_fp128 %Val)
7340 The '``llvm.exp2.*``' intrinsics perform the exp2 function.
7345 The argument and return value are floating point numbers of the same
7351 This function returns the same values as the libm ``exp2`` functions
7352 would, and handles error conditions in the same way.
7354 '``llvm.log.*``' Intrinsic
7355 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7360 This is an overloaded intrinsic. You can use ``llvm.log`` on any
7361 floating point or vector of floating point type. Not all targets support
7366 declare float @llvm.log.f32(float %Val)
7367 declare double @llvm.log.f64(double %Val)
7368 declare x86_fp80 @llvm.log.f80(x86_fp80 %Val)
7369 declare fp128 @llvm.log.f128(fp128 %Val)
7370 declare ppc_fp128 @llvm.log.ppcf128(ppc_fp128 %Val)
7375 The '``llvm.log.*``' intrinsics perform the log function.
7380 The argument and return value are floating point numbers of the same
7386 This function returns the same values as the libm ``log`` functions
7387 would, and handles error conditions in the same way.
7389 '``llvm.log10.*``' Intrinsic
7390 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7395 This is an overloaded intrinsic. You can use ``llvm.log10`` on any
7396 floating point or vector of floating point type. Not all targets support
7401 declare float @llvm.log10.f32(float %Val)
7402 declare double @llvm.log10.f64(double %Val)
7403 declare x86_fp80 @llvm.log10.f80(x86_fp80 %Val)
7404 declare fp128 @llvm.log10.f128(fp128 %Val)
7405 declare ppc_fp128 @llvm.log10.ppcf128(ppc_fp128 %Val)
7410 The '``llvm.log10.*``' intrinsics perform the log10 function.
7415 The argument and return value are floating point numbers of the same
7421 This function returns the same values as the libm ``log10`` functions
7422 would, and handles error conditions in the same way.
7424 '``llvm.log2.*``' Intrinsic
7425 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7430 This is an overloaded intrinsic. You can use ``llvm.log2`` on any
7431 floating point or vector of floating point type. Not all targets support
7436 declare float @llvm.log2.f32(float %Val)
7437 declare double @llvm.log2.f64(double %Val)
7438 declare x86_fp80 @llvm.log2.f80(x86_fp80 %Val)
7439 declare fp128 @llvm.log2.f128(fp128 %Val)
7440 declare ppc_fp128 @llvm.log2.ppcf128(ppc_fp128 %Val)
7445 The '``llvm.log2.*``' intrinsics perform the log2 function.
7450 The argument and return value are floating point numbers of the same
7456 This function returns the same values as the libm ``log2`` functions
7457 would, and handles error conditions in the same way.
7459 '``llvm.fma.*``' Intrinsic
7460 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7465 This is an overloaded intrinsic. You can use ``llvm.fma`` on any
7466 floating point or vector of floating point type. Not all targets support
7471 declare float @llvm.fma.f32(float %a, float %b, float %c)
7472 declare double @llvm.fma.f64(double %a, double %b, double %c)
7473 declare x86_fp80 @llvm.fma.f80(x86_fp80 %a, x86_fp80 %b, x86_fp80 %c)
7474 declare fp128 @llvm.fma.f128(fp128 %a, fp128 %b, fp128 %c)
7475 declare ppc_fp128 @llvm.fma.ppcf128(ppc_fp128 %a, ppc_fp128 %b, ppc_fp128 %c)
7480 The '``llvm.fma.*``' intrinsics perform the fused multiply-add
7486 The argument and return value are floating point numbers of the same
7492 This function returns the same values as the libm ``fma`` functions
7493 would, and does not set errno.
7495 '``llvm.fabs.*``' Intrinsic
7496 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7501 This is an overloaded intrinsic. You can use ``llvm.fabs`` on any
7502 floating point or vector of floating point type. Not all targets support
7507 declare float @llvm.fabs.f32(float %Val)
7508 declare double @llvm.fabs.f64(double %Val)
7509 declare x86_fp80 @llvm.fabs.f80(x86_fp80 %Val)
7510 declare fp128 @llvm.fabs.f128(fp128 %Val)
7511 declare ppc_fp128 @llvm.fabs.ppcf128(ppc_fp128 %Val)
7516 The '``llvm.fabs.*``' intrinsics return the absolute value of the
7522 The argument and return value are floating point numbers of the same
7528 This function returns the same values as the libm ``fabs`` functions
7529 would, and handles error conditions in the same way.
7531 '``llvm.copysign.*``' Intrinsic
7532 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7537 This is an overloaded intrinsic. You can use ``llvm.copysign`` on any
7538 floating point or vector of floating point type. Not all targets support
7543 declare float @llvm.copysign.f32(float %Mag, float %Sgn)
7544 declare double @llvm.copysign.f64(double %Mag, double %Sgn)
7545 declare x86_fp80 @llvm.copysign.f80(x86_fp80 %Mag, x86_fp80 %Sgn)
7546 declare fp128 @llvm.copysign.f128(fp128 %Mag, fp128 %Sgn)
7547 declare ppc_fp128 @llvm.copysign.ppcf128(ppc_fp128 %Mag, ppc_fp128 %Sgn)
7552 The '``llvm.copysign.*``' intrinsics return a value with the magnitude of the
7553 first operand and the sign of the second operand.
7558 The arguments and return value are floating point numbers of the same
7564 This function returns the same values as the libm ``copysign``
7565 functions would, and handles error conditions in the same way.
7567 '``llvm.floor.*``' Intrinsic
7568 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7573 This is an overloaded intrinsic. You can use ``llvm.floor`` on any
7574 floating point or vector of floating point type. Not all targets support
7579 declare float @llvm.floor.f32(float %Val)
7580 declare double @llvm.floor.f64(double %Val)
7581 declare x86_fp80 @llvm.floor.f80(x86_fp80 %Val)
7582 declare fp128 @llvm.floor.f128(fp128 %Val)
7583 declare ppc_fp128 @llvm.floor.ppcf128(ppc_fp128 %Val)
7588 The '``llvm.floor.*``' intrinsics return the floor of the operand.
7593 The argument and return value are floating point numbers of the same
7599 This function returns the same values as the libm ``floor`` functions
7600 would, and handles error conditions in the same way.
7602 '``llvm.ceil.*``' Intrinsic
7603 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7608 This is an overloaded intrinsic. You can use ``llvm.ceil`` on any
7609 floating point or vector of floating point type. Not all targets support
7614 declare float @llvm.ceil.f32(float %Val)
7615 declare double @llvm.ceil.f64(double %Val)
7616 declare x86_fp80 @llvm.ceil.f80(x86_fp80 %Val)
7617 declare fp128 @llvm.ceil.f128(fp128 %Val)
7618 declare ppc_fp128 @llvm.ceil.ppcf128(ppc_fp128 %Val)
7623 The '``llvm.ceil.*``' intrinsics return the ceiling of the operand.
7628 The argument and return value are floating point numbers of the same
7634 This function returns the same values as the libm ``ceil`` functions
7635 would, and handles error conditions in the same way.
7637 '``llvm.trunc.*``' Intrinsic
7638 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7643 This is an overloaded intrinsic. You can use ``llvm.trunc`` on any
7644 floating point or vector of floating point type. Not all targets support
7649 declare float @llvm.trunc.f32(float %Val)
7650 declare double @llvm.trunc.f64(double %Val)
7651 declare x86_fp80 @llvm.trunc.f80(x86_fp80 %Val)
7652 declare fp128 @llvm.trunc.f128(fp128 %Val)
7653 declare ppc_fp128 @llvm.trunc.ppcf128(ppc_fp128 %Val)
7658 The '``llvm.trunc.*``' intrinsics returns the operand rounded to the
7659 nearest integer not larger in magnitude than the operand.
7664 The argument and return value are floating point numbers of the same
7670 This function returns the same values as the libm ``trunc`` functions
7671 would, and handles error conditions in the same way.
7673 '``llvm.rint.*``' Intrinsic
7674 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7679 This is an overloaded intrinsic. You can use ``llvm.rint`` on any
7680 floating point or vector of floating point type. Not all targets support
7685 declare float @llvm.rint.f32(float %Val)
7686 declare double @llvm.rint.f64(double %Val)
7687 declare x86_fp80 @llvm.rint.f80(x86_fp80 %Val)
7688 declare fp128 @llvm.rint.f128(fp128 %Val)
7689 declare ppc_fp128 @llvm.rint.ppcf128(ppc_fp128 %Val)
7694 The '``llvm.rint.*``' intrinsics returns the operand rounded to the
7695 nearest integer. It may raise an inexact floating-point exception if the
7696 operand isn't an integer.
7701 The argument and return value are floating point numbers of the same
7707 This function returns the same values as the libm ``rint`` functions
7708 would, and handles error conditions in the same way.
7710 '``llvm.nearbyint.*``' Intrinsic
7711 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7716 This is an overloaded intrinsic. You can use ``llvm.nearbyint`` on any
7717 floating point or vector of floating point type. Not all targets support
7722 declare float @llvm.nearbyint.f32(float %Val)
7723 declare double @llvm.nearbyint.f64(double %Val)
7724 declare x86_fp80 @llvm.nearbyint.f80(x86_fp80 %Val)
7725 declare fp128 @llvm.nearbyint.f128(fp128 %Val)
7726 declare ppc_fp128 @llvm.nearbyint.ppcf128(ppc_fp128 %Val)
7731 The '``llvm.nearbyint.*``' intrinsics returns the operand rounded to the
7737 The argument and return value are floating point numbers of the same
7743 This function returns the same values as the libm ``nearbyint``
7744 functions would, and handles error conditions in the same way.
7746 '``llvm.round.*``' Intrinsic
7747 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7752 This is an overloaded intrinsic. You can use ``llvm.round`` on any
7753 floating point or vector of floating point type. Not all targets support
7758 declare float @llvm.round.f32(float %Val)
7759 declare double @llvm.round.f64(double %Val)
7760 declare x86_fp80 @llvm.round.f80(x86_fp80 %Val)
7761 declare fp128 @llvm.round.f128(fp128 %Val)
7762 declare ppc_fp128 @llvm.round.ppcf128(ppc_fp128 %Val)
7767 The '``llvm.round.*``' intrinsics returns the operand rounded to the
7773 The argument and return value are floating point numbers of the same
7779 This function returns the same values as the libm ``round``
7780 functions would, and handles error conditions in the same way.
7782 Bit Manipulation Intrinsics
7783 ---------------------------
7785 LLVM provides intrinsics for a few important bit manipulation
7786 operations. These allow efficient code generation for some algorithms.
7788 '``llvm.bswap.*``' Intrinsics
7789 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7794 This is an overloaded intrinsic function. You can use bswap on any
7795 integer type that is an even number of bytes (i.e. BitWidth % 16 == 0).
7799 declare i16 @llvm.bswap.i16(i16 <id>)
7800 declare i32 @llvm.bswap.i32(i32 <id>)
7801 declare i64 @llvm.bswap.i64(i64 <id>)
7806 The '``llvm.bswap``' family of intrinsics is used to byte swap integer
7807 values with an even number of bytes (positive multiple of 16 bits).
7808 These are useful for performing operations on data that is not in the
7809 target's native byte order.
7814 The ``llvm.bswap.i16`` intrinsic returns an i16 value that has the high
7815 and low byte of the input i16 swapped. Similarly, the ``llvm.bswap.i32``
7816 intrinsic returns an i32 value that has the four bytes of the input i32
7817 swapped, so that if the input bytes are numbered 0, 1, 2, 3 then the
7818 returned i32 will have its bytes in 3, 2, 1, 0 order. The
7819 ``llvm.bswap.i48``, ``llvm.bswap.i64`` and other intrinsics extend this
7820 concept to additional even-byte lengths (6 bytes, 8 bytes and more,
7823 '``llvm.ctpop.*``' Intrinsic
7824 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7829 This is an overloaded intrinsic. You can use llvm.ctpop on any integer
7830 bit width, or on any vector with integer elements. Not all targets
7831 support all bit widths or vector types, however.
7835 declare i8 @llvm.ctpop.i8(i8 <src>)
7836 declare i16 @llvm.ctpop.i16(i16 <src>)
7837 declare i32 @llvm.ctpop.i32(i32 <src>)
7838 declare i64 @llvm.ctpop.i64(i64 <src>)
7839 declare i256 @llvm.ctpop.i256(i256 <src>)
7840 declare <2 x i32> @llvm.ctpop.v2i32(<2 x i32> <src>)
7845 The '``llvm.ctpop``' family of intrinsics counts the number of bits set
7851 The only argument is the value to be counted. The argument may be of any
7852 integer type, or a vector with integer elements. The return type must
7853 match the argument type.
7858 The '``llvm.ctpop``' intrinsic counts the 1's in a variable, or within
7859 each element of a vector.
7861 '``llvm.ctlz.*``' Intrinsic
7862 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7867 This is an overloaded intrinsic. You can use ``llvm.ctlz`` on any
7868 integer bit width, or any vector whose elements are integers. Not all
7869 targets support all bit widths or vector types, however.
7873 declare i8 @llvm.ctlz.i8 (i8 <src>, i1 <is_zero_undef>)
7874 declare i16 @llvm.ctlz.i16 (i16 <src>, i1 <is_zero_undef>)
7875 declare i32 @llvm.ctlz.i32 (i32 <src>, i1 <is_zero_undef>)
7876 declare i64 @llvm.ctlz.i64 (i64 <src>, i1 <is_zero_undef>)
7877 declare i256 @llvm.ctlz.i256(i256 <src>, i1 <is_zero_undef>)
7878 declase <2 x i32> @llvm.ctlz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
7883 The '``llvm.ctlz``' family of intrinsic functions counts the number of
7884 leading zeros in a variable.
7889 The first argument is the value to be counted. This argument may be of
7890 any integer type, or a vectory with integer element type. The return
7891 type must match the first argument type.
7893 The second argument must be a constant and is a flag to indicate whether
7894 the intrinsic should ensure that a zero as the first argument produces a
7895 defined result. Historically some architectures did not provide a
7896 defined result for zero values as efficiently, and many algorithms are
7897 now predicated on avoiding zero-value inputs.
7902 The '``llvm.ctlz``' intrinsic counts the leading (most significant)
7903 zeros in a variable, or within each element of the vector. If
7904 ``src == 0`` then the result is the size in bits of the type of ``src``
7905 if ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
7906 ``llvm.ctlz(i32 2) = 30``.
7908 '``llvm.cttz.*``' Intrinsic
7909 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7914 This is an overloaded intrinsic. You can use ``llvm.cttz`` on any
7915 integer bit width, or any vector of integer elements. Not all targets
7916 support all bit widths or vector types, however.
7920 declare i8 @llvm.cttz.i8 (i8 <src>, i1 <is_zero_undef>)
7921 declare i16 @llvm.cttz.i16 (i16 <src>, i1 <is_zero_undef>)
7922 declare i32 @llvm.cttz.i32 (i32 <src>, i1 <is_zero_undef>)
7923 declare i64 @llvm.cttz.i64 (i64 <src>, i1 <is_zero_undef>)
7924 declare i256 @llvm.cttz.i256(i256 <src>, i1 <is_zero_undef>)
7925 declase <2 x i32> @llvm.cttz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
7930 The '``llvm.cttz``' family of intrinsic functions counts the number of
7936 The first argument is the value to be counted. This argument may be of
7937 any integer type, or a vectory with integer element type. The return
7938 type must match the first argument type.
7940 The second argument must be a constant and is a flag to indicate whether
7941 the intrinsic should ensure that a zero as the first argument produces a
7942 defined result. Historically some architectures did not provide a
7943 defined result for zero values as efficiently, and many algorithms are
7944 now predicated on avoiding zero-value inputs.
7949 The '``llvm.cttz``' intrinsic counts the trailing (least significant)
7950 zeros in a variable, or within each element of a vector. If ``src == 0``
7951 then the result is the size in bits of the type of ``src`` if
7952 ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
7953 ``llvm.cttz(2) = 1``.
7955 Arithmetic with Overflow Intrinsics
7956 -----------------------------------
7958 LLVM provides intrinsics for some arithmetic with overflow operations.
7960 '``llvm.sadd.with.overflow.*``' Intrinsics
7961 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7966 This is an overloaded intrinsic. You can use ``llvm.sadd.with.overflow``
7967 on any integer bit width.
7971 declare {i16, i1} @llvm.sadd.with.overflow.i16(i16 %a, i16 %b)
7972 declare {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
7973 declare {i64, i1} @llvm.sadd.with.overflow.i64(i64 %a, i64 %b)
7978 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
7979 a signed addition of the two arguments, and indicate whether an overflow
7980 occurred during the signed summation.
7985 The arguments (%a and %b) and the first element of the result structure
7986 may be of integer types of any bit width, but they must have the same
7987 bit width. The second element of the result structure must be of type
7988 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
7994 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
7995 a signed addition of the two variables. They return a structure --- the
7996 first element of which is the signed summation, and the second element
7997 of which is a bit specifying if the signed summation resulted in an
8003 .. code-block:: llvm
8005 %res = call {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
8006 %sum = extractvalue {i32, i1} %res, 0
8007 %obit = extractvalue {i32, i1} %res, 1
8008 br i1 %obit, label %overflow, label %normal
8010 '``llvm.uadd.with.overflow.*``' Intrinsics
8011 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8016 This is an overloaded intrinsic. You can use ``llvm.uadd.with.overflow``
8017 on any integer bit width.
8021 declare {i16, i1} @llvm.uadd.with.overflow.i16(i16 %a, i16 %b)
8022 declare {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
8023 declare {i64, i1} @llvm.uadd.with.overflow.i64(i64 %a, i64 %b)
8028 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
8029 an unsigned addition of the two arguments, and indicate whether a carry
8030 occurred during the unsigned summation.
8035 The arguments (%a and %b) and the first element of the result structure
8036 may be of integer types of any bit width, but they must have the same
8037 bit width. The second element of the result structure must be of type
8038 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
8044 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
8045 an unsigned addition of the two arguments. They return a structure --- the
8046 first element of which is the sum, and the second element of which is a
8047 bit specifying if the unsigned summation resulted in a carry.
8052 .. code-block:: llvm
8054 %res = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
8055 %sum = extractvalue {i32, i1} %res, 0
8056 %obit = extractvalue {i32, i1} %res, 1
8057 br i1 %obit, label %carry, label %normal
8059 '``llvm.ssub.with.overflow.*``' Intrinsics
8060 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8065 This is an overloaded intrinsic. You can use ``llvm.ssub.with.overflow``
8066 on any integer bit width.
8070 declare {i16, i1} @llvm.ssub.with.overflow.i16(i16 %a, i16 %b)
8071 declare {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
8072 declare {i64, i1} @llvm.ssub.with.overflow.i64(i64 %a, i64 %b)
8077 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
8078 a signed subtraction of the two arguments, and indicate whether an
8079 overflow occurred during the signed subtraction.
8084 The arguments (%a and %b) and the first element of the result structure
8085 may be of integer types of any bit width, but they must have the same
8086 bit width. The second element of the result structure must be of type
8087 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
8093 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
8094 a signed subtraction of the two arguments. They return a structure --- the
8095 first element of which is the subtraction, and the second element of
8096 which is a bit specifying if the signed subtraction resulted in an
8102 .. code-block:: llvm
8104 %res = call {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
8105 %sum = extractvalue {i32, i1} %res, 0
8106 %obit = extractvalue {i32, i1} %res, 1
8107 br i1 %obit, label %overflow, label %normal
8109 '``llvm.usub.with.overflow.*``' Intrinsics
8110 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8115 This is an overloaded intrinsic. You can use ``llvm.usub.with.overflow``
8116 on any integer bit width.
8120 declare {i16, i1} @llvm.usub.with.overflow.i16(i16 %a, i16 %b)
8121 declare {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
8122 declare {i64, i1} @llvm.usub.with.overflow.i64(i64 %a, i64 %b)
8127 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
8128 an unsigned subtraction of the two arguments, and indicate whether an
8129 overflow occurred during the unsigned subtraction.
8134 The arguments (%a and %b) and the first element of the result structure
8135 may be of integer types of any bit width, but they must have the same
8136 bit width. The second element of the result structure must be of type
8137 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
8143 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
8144 an unsigned subtraction of the two arguments. They return a structure ---
8145 the first element of which is the subtraction, and the second element of
8146 which is a bit specifying if the unsigned subtraction resulted in an
8152 .. code-block:: llvm
8154 %res = call {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
8155 %sum = extractvalue {i32, i1} %res, 0
8156 %obit = extractvalue {i32, i1} %res, 1
8157 br i1 %obit, label %overflow, label %normal
8159 '``llvm.smul.with.overflow.*``' Intrinsics
8160 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8165 This is an overloaded intrinsic. You can use ``llvm.smul.with.overflow``
8166 on any integer bit width.
8170 declare {i16, i1} @llvm.smul.with.overflow.i16(i16 %a, i16 %b)
8171 declare {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
8172 declare {i64, i1} @llvm.smul.with.overflow.i64(i64 %a, i64 %b)
8177 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
8178 a signed multiplication of the two arguments, and indicate whether an
8179 overflow occurred during the signed multiplication.
8184 The arguments (%a and %b) and the first element of the result structure
8185 may be of integer types of any bit width, but they must have the same
8186 bit width. The second element of the result structure must be of type
8187 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
8193 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
8194 a signed multiplication of the two arguments. They return a structure ---
8195 the first element of which is the multiplication, and the second element
8196 of which is a bit specifying if the signed multiplication resulted in an
8202 .. code-block:: llvm
8204 %res = call {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
8205 %sum = extractvalue {i32, i1} %res, 0
8206 %obit = extractvalue {i32, i1} %res, 1
8207 br i1 %obit, label %overflow, label %normal
8209 '``llvm.umul.with.overflow.*``' Intrinsics
8210 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8215 This is an overloaded intrinsic. You can use ``llvm.umul.with.overflow``
8216 on any integer bit width.
8220 declare {i16, i1} @llvm.umul.with.overflow.i16(i16 %a, i16 %b)
8221 declare {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
8222 declare {i64, i1} @llvm.umul.with.overflow.i64(i64 %a, i64 %b)
8227 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
8228 a unsigned multiplication of the two arguments, and indicate whether an
8229 overflow occurred during the unsigned multiplication.
8234 The arguments (%a and %b) and the first element of the result structure
8235 may be of integer types of any bit width, but they must have the same
8236 bit width. The second element of the result structure must be of type
8237 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
8243 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
8244 an unsigned multiplication of the two arguments. They return a structure ---
8245 the first element of which is the multiplication, and the second
8246 element of which is a bit specifying if the unsigned multiplication
8247 resulted in an overflow.
8252 .. code-block:: llvm
8254 %res = call {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
8255 %sum = extractvalue {i32, i1} %res, 0
8256 %obit = extractvalue {i32, i1} %res, 1
8257 br i1 %obit, label %overflow, label %normal
8259 Specialised Arithmetic Intrinsics
8260 ---------------------------------
8262 '``llvm.fmuladd.*``' Intrinsic
8263 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8270 declare float @llvm.fmuladd.f32(float %a, float %b, float %c)
8271 declare double @llvm.fmuladd.f64(double %a, double %b, double %c)
8276 The '``llvm.fmuladd.*``' intrinsic functions represent multiply-add
8277 expressions that can be fused if the code generator determines that (a) the
8278 target instruction set has support for a fused operation, and (b) that the
8279 fused operation is more efficient than the equivalent, separate pair of mul
8280 and add instructions.
8285 The '``llvm.fmuladd.*``' intrinsics each take three arguments: two
8286 multiplicands, a and b, and an addend c.
8295 %0 = call float @llvm.fmuladd.f32(%a, %b, %c)
8297 is equivalent to the expression a \* b + c, except that rounding will
8298 not be performed between the multiplication and addition steps if the
8299 code generator fuses the operations. Fusion is not guaranteed, even if
8300 the target platform supports it. If a fused multiply-add is required the
8301 corresponding llvm.fma.\* intrinsic function should be used
8302 instead. This never sets errno, just as '``llvm.fma.*``'.
8307 .. code-block:: llvm
8309 %r2 = call float @llvm.fmuladd.f32(float %a, float %b, float %c) ; yields {float}:r2 = (a * b) + c
8311 Half Precision Floating Point Intrinsics
8312 ----------------------------------------
8314 For most target platforms, half precision floating point is a
8315 storage-only format. This means that it is a dense encoding (in memory)
8316 but does not support computation in the format.
8318 This means that code must first load the half-precision floating point
8319 value as an i16, then convert it to float with
8320 :ref:`llvm.convert.from.fp16 <int_convert_from_fp16>`. Computation can
8321 then be performed on the float value (including extending to double
8322 etc). To store the value back to memory, it is first converted to float
8323 if needed, then converted to i16 with
8324 :ref:`llvm.convert.to.fp16 <int_convert_to_fp16>`, then storing as an
8327 .. _int_convert_to_fp16:
8329 '``llvm.convert.to.fp16``' Intrinsic
8330 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8337 declare i16 @llvm.convert.to.fp16(f32 %a)
8342 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion
8343 from single precision floating point format to half precision floating
8349 The intrinsic function contains single argument - the value to be
8355 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion
8356 from single precision floating point format to half precision floating
8357 point format. The return value is an ``i16`` which contains the
8363 .. code-block:: llvm
8365 %res = call i16 @llvm.convert.to.fp16(f32 %a)
8366 store i16 %res, i16* @x, align 2
8368 .. _int_convert_from_fp16:
8370 '``llvm.convert.from.fp16``' Intrinsic
8371 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8378 declare f32 @llvm.convert.from.fp16(i16 %a)
8383 The '``llvm.convert.from.fp16``' intrinsic function performs a
8384 conversion from half precision floating point format to single precision
8385 floating point format.
8390 The intrinsic function contains single argument - the value to be
8396 The '``llvm.convert.from.fp16``' intrinsic function performs a
8397 conversion from half single precision floating point format to single
8398 precision floating point format. The input half-float value is
8399 represented by an ``i16`` value.
8404 .. code-block:: llvm
8406 %a = load i16* @x, align 2
8407 %res = call f32 @llvm.convert.from.fp16(i16 %a)
8412 The LLVM debugger intrinsics (which all start with ``llvm.dbg.``
8413 prefix), are described in the `LLVM Source Level
8414 Debugging <SourceLevelDebugging.html#format_common_intrinsics>`_
8417 Exception Handling Intrinsics
8418 -----------------------------
8420 The LLVM exception handling intrinsics (which all start with
8421 ``llvm.eh.`` prefix), are described in the `LLVM Exception
8422 Handling <ExceptionHandling.html#format_common_intrinsics>`_ document.
8426 Trampoline Intrinsics
8427 ---------------------
8429 These intrinsics make it possible to excise one parameter, marked with
8430 the :ref:`nest <nest>` attribute, from a function. The result is a
8431 callable function pointer lacking the nest parameter - the caller does
8432 not need to provide a value for it. Instead, the value to use is stored
8433 in advance in a "trampoline", a block of memory usually allocated on the
8434 stack, which also contains code to splice the nest value into the
8435 argument list. This is used to implement the GCC nested function address
8438 For example, if the function is ``i32 f(i8* nest %c, i32 %x, i32 %y)``
8439 then the resulting function pointer has signature ``i32 (i32, i32)*``.
8440 It can be created as follows:
8442 .. code-block:: llvm
8444 %tramp = alloca [10 x i8], align 4 ; size and alignment only correct for X86
8445 %tramp1 = getelementptr [10 x i8]* %tramp, i32 0, i32 0
8446 call i8* @llvm.init.trampoline(i8* %tramp1, i8* bitcast (i32 (i8*, i32, i32)* @f to i8*), i8* %nval)
8447 %p = call i8* @llvm.adjust.trampoline(i8* %tramp1)
8448 %fp = bitcast i8* %p to i32 (i32, i32)*
8450 The call ``%val = call i32 %fp(i32 %x, i32 %y)`` is then equivalent to
8451 ``%val = call i32 %f(i8* %nval, i32 %x, i32 %y)``.
8455 '``llvm.init.trampoline``' Intrinsic
8456 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8463 declare void @llvm.init.trampoline(i8* <tramp>, i8* <func>, i8* <nval>)
8468 This fills the memory pointed to by ``tramp`` with executable code,
8469 turning it into a trampoline.
8474 The ``llvm.init.trampoline`` intrinsic takes three arguments, all
8475 pointers. The ``tramp`` argument must point to a sufficiently large and
8476 sufficiently aligned block of memory; this memory is written to by the
8477 intrinsic. Note that the size and the alignment are target-specific -
8478 LLVM currently provides no portable way of determining them, so a
8479 front-end that generates this intrinsic needs to have some
8480 target-specific knowledge. The ``func`` argument must hold a function
8481 bitcast to an ``i8*``.
8486 The block of memory pointed to by ``tramp`` is filled with target
8487 dependent code, turning it into a function. Then ``tramp`` needs to be
8488 passed to :ref:`llvm.adjust.trampoline <int_at>` to get a pointer which can
8489 be :ref:`bitcast (to a new function) and called <int_trampoline>`. The new
8490 function's signature is the same as that of ``func`` with any arguments
8491 marked with the ``nest`` attribute removed. At most one such ``nest``
8492 argument is allowed, and it must be of pointer type. Calling the new
8493 function is equivalent to calling ``func`` with the same argument list,
8494 but with ``nval`` used for the missing ``nest`` argument. If, after
8495 calling ``llvm.init.trampoline``, the memory pointed to by ``tramp`` is
8496 modified, then the effect of any later call to the returned function
8497 pointer is undefined.
8501 '``llvm.adjust.trampoline``' Intrinsic
8502 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8509 declare i8* @llvm.adjust.trampoline(i8* <tramp>)
8514 This performs any required machine-specific adjustment to the address of
8515 a trampoline (passed as ``tramp``).
8520 ``tramp`` must point to a block of memory which already has trampoline
8521 code filled in by a previous call to
8522 :ref:`llvm.init.trampoline <int_it>`.
8527 On some architectures the address of the code to be executed needs to be
8528 different to the address where the trampoline is actually stored. This
8529 intrinsic returns the executable address corresponding to ``tramp``
8530 after performing the required machine specific adjustments. The pointer
8531 returned can then be :ref:`bitcast and executed <int_trampoline>`.
8536 This class of intrinsics exists to information about the lifetime of
8537 memory objects and ranges where variables are immutable.
8541 '``llvm.lifetime.start``' Intrinsic
8542 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8549 declare void @llvm.lifetime.start(i64 <size>, i8* nocapture <ptr>)
8554 The '``llvm.lifetime.start``' intrinsic specifies the start of a memory
8560 The first argument is a constant integer representing the size of the
8561 object, or -1 if it is variable sized. The second argument is a pointer
8567 This intrinsic indicates that before this point in the code, the value
8568 of the memory pointed to by ``ptr`` is dead. This means that it is known
8569 to never be used and has an undefined value. A load from the pointer
8570 that precedes this intrinsic can be replaced with ``'undef'``.
8574 '``llvm.lifetime.end``' Intrinsic
8575 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8582 declare void @llvm.lifetime.end(i64 <size>, i8* nocapture <ptr>)
8587 The '``llvm.lifetime.end``' intrinsic specifies the end of a memory
8593 The first argument is a constant integer representing the size of the
8594 object, or -1 if it is variable sized. The second argument is a pointer
8600 This intrinsic indicates that after this point in the code, the value of
8601 the memory pointed to by ``ptr`` is dead. This means that it is known to
8602 never be used and has an undefined value. Any stores into the memory
8603 object following this intrinsic may be removed as dead.
8605 '``llvm.invariant.start``' Intrinsic
8606 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8613 declare {}* @llvm.invariant.start(i64 <size>, i8* nocapture <ptr>)
8618 The '``llvm.invariant.start``' intrinsic specifies that the contents of
8619 a memory object will not change.
8624 The first argument is a constant integer representing the size of the
8625 object, or -1 if it is variable sized. The second argument is a pointer
8631 This intrinsic indicates that until an ``llvm.invariant.end`` that uses
8632 the return value, the referenced memory location is constant and
8635 '``llvm.invariant.end``' Intrinsic
8636 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8643 declare void @llvm.invariant.end({}* <start>, i64 <size>, i8* nocapture <ptr>)
8648 The '``llvm.invariant.end``' intrinsic specifies that the contents of a
8649 memory object are mutable.
8654 The first argument is the matching ``llvm.invariant.start`` intrinsic.
8655 The second argument is a constant integer representing the size of the
8656 object, or -1 if it is variable sized and the third argument is a
8657 pointer to the object.
8662 This intrinsic indicates that the memory is mutable again.
8667 This class of intrinsics is designed to be generic and has no specific
8670 '``llvm.var.annotation``' Intrinsic
8671 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8678 declare void @llvm.var.annotation(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
8683 The '``llvm.var.annotation``' intrinsic.
8688 The first argument is a pointer to a value, the second is a pointer to a
8689 global string, the third is a pointer to a global string which is the
8690 source file name, and the last argument is the line number.
8695 This intrinsic allows annotation of local variables with arbitrary
8696 strings. This can be useful for special purpose optimizations that want
8697 to look for these annotations. These have no other defined use; they are
8698 ignored by code generation and optimization.
8700 '``llvm.ptr.annotation.*``' Intrinsic
8701 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8706 This is an overloaded intrinsic. You can use '``llvm.ptr.annotation``' on a
8707 pointer to an integer of any width. *NOTE* you must specify an address space for
8708 the pointer. The identifier for the default address space is the integer
8713 declare i8* @llvm.ptr.annotation.p<address space>i8(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
8714 declare i16* @llvm.ptr.annotation.p<address space>i16(i16* <val>, i8* <str>, i8* <str>, i32 <int>)
8715 declare i32* @llvm.ptr.annotation.p<address space>i32(i32* <val>, i8* <str>, i8* <str>, i32 <int>)
8716 declare i64* @llvm.ptr.annotation.p<address space>i64(i64* <val>, i8* <str>, i8* <str>, i32 <int>)
8717 declare i256* @llvm.ptr.annotation.p<address space>i256(i256* <val>, i8* <str>, i8* <str>, i32 <int>)
8722 The '``llvm.ptr.annotation``' intrinsic.
8727 The first argument is a pointer to an integer value of arbitrary bitwidth
8728 (result of some expression), the second is a pointer to a global string, the
8729 third is a pointer to a global string which is the source file name, and the
8730 last argument is the line number. It returns the value of the first argument.
8735 This intrinsic allows annotation of a pointer to an integer with arbitrary
8736 strings. This can be useful for special purpose optimizations that want to look
8737 for these annotations. These have no other defined use; they are ignored by code
8738 generation and optimization.
8740 '``llvm.annotation.*``' Intrinsic
8741 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8746 This is an overloaded intrinsic. You can use '``llvm.annotation``' on
8747 any integer bit width.
8751 declare i8 @llvm.annotation.i8(i8 <val>, i8* <str>, i8* <str>, i32 <int>)
8752 declare i16 @llvm.annotation.i16(i16 <val>, i8* <str>, i8* <str>, i32 <int>)
8753 declare i32 @llvm.annotation.i32(i32 <val>, i8* <str>, i8* <str>, i32 <int>)
8754 declare i64 @llvm.annotation.i64(i64 <val>, i8* <str>, i8* <str>, i32 <int>)
8755 declare i256 @llvm.annotation.i256(i256 <val>, i8* <str>, i8* <str>, i32 <int>)
8760 The '``llvm.annotation``' intrinsic.
8765 The first argument is an integer value (result of some expression), the
8766 second is a pointer to a global string, the third is a pointer to a
8767 global string which is the source file name, and the last argument is
8768 the line number. It returns the value of the first argument.
8773 This intrinsic allows annotations to be put on arbitrary expressions
8774 with arbitrary strings. This can be useful for special purpose
8775 optimizations that want to look for these annotations. These have no
8776 other defined use; they are ignored by code generation and optimization.
8778 '``llvm.trap``' Intrinsic
8779 ^^^^^^^^^^^^^^^^^^^^^^^^^
8786 declare void @llvm.trap() noreturn nounwind
8791 The '``llvm.trap``' intrinsic.
8801 This intrinsic is lowered to the target dependent trap instruction. If
8802 the target does not have a trap instruction, this intrinsic will be
8803 lowered to a call of the ``abort()`` function.
8805 '``llvm.debugtrap``' Intrinsic
8806 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8813 declare void @llvm.debugtrap() nounwind
8818 The '``llvm.debugtrap``' intrinsic.
8828 This intrinsic is lowered to code which is intended to cause an
8829 execution trap with the intention of requesting the attention of a
8832 '``llvm.stackprotector``' Intrinsic
8833 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8840 declare void @llvm.stackprotector(i8* <guard>, i8** <slot>)
8845 The ``llvm.stackprotector`` intrinsic takes the ``guard`` and stores it
8846 onto the stack at ``slot``. The stack slot is adjusted to ensure that it
8847 is placed on the stack before local variables.
8852 The ``llvm.stackprotector`` intrinsic requires two pointer arguments.
8853 The first argument is the value loaded from the stack guard
8854 ``@__stack_chk_guard``. The second variable is an ``alloca`` that has
8855 enough space to hold the value of the guard.
8860 This intrinsic causes the prologue/epilogue inserter to force the position of
8861 the ``AllocaInst`` stack slot to be before local variables on the stack. This is
8862 to ensure that if a local variable on the stack is overwritten, it will destroy
8863 the value of the guard. When the function exits, the guard on the stack is
8864 checked against the original guard by ``llvm.stackprotectorcheck``. If they are
8865 different, then ``llvm.stackprotectorcheck`` causes the program to abort by
8866 calling the ``__stack_chk_fail()`` function.
8868 '``llvm.stackprotectorcheck``' Intrinsic
8869 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8876 declare void @llvm.stackprotectorcheck(i8** <guard>)
8881 The ``llvm.stackprotectorcheck`` intrinsic compares ``guard`` against an already
8882 created stack protector and if they are not equal calls the
8883 ``__stack_chk_fail()`` function.
8888 The ``llvm.stackprotectorcheck`` intrinsic requires one pointer argument, the
8889 the variable ``@__stack_chk_guard``.
8894 This intrinsic is provided to perform the stack protector check by comparing
8895 ``guard`` with the stack slot created by ``llvm.stackprotector`` and if the
8896 values do not match call the ``__stack_chk_fail()`` function.
8898 The reason to provide this as an IR level intrinsic instead of implementing it
8899 via other IR operations is that in order to perform this operation at the IR
8900 level without an intrinsic, one would need to create additional basic blocks to
8901 handle the success/failure cases. This makes it difficult to stop the stack
8902 protector check from disrupting sibling tail calls in Codegen. With this
8903 intrinsic, we are able to generate the stack protector basic blocks late in
8904 codegen after the tail call decision has occurred.
8906 '``llvm.objectsize``' Intrinsic
8907 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8914 declare i32 @llvm.objectsize.i32(i8* <object>, i1 <min>)
8915 declare i64 @llvm.objectsize.i64(i8* <object>, i1 <min>)
8920 The ``llvm.objectsize`` intrinsic is designed to provide information to
8921 the optimizers to determine at compile time whether a) an operation
8922 (like memcpy) will overflow a buffer that corresponds to an object, or
8923 b) that a runtime check for overflow isn't necessary. An object in this
8924 context means an allocation of a specific class, structure, array, or
8930 The ``llvm.objectsize`` intrinsic takes two arguments. The first
8931 argument is a pointer to or into the ``object``. The second argument is
8932 a boolean and determines whether ``llvm.objectsize`` returns 0 (if true)
8933 or -1 (if false) when the object size is unknown. The second argument
8934 only accepts constants.
8939 The ``llvm.objectsize`` intrinsic is lowered to a constant representing
8940 the size of the object concerned. If the size cannot be determined at
8941 compile time, ``llvm.objectsize`` returns ``i32/i64 -1 or 0`` (depending
8942 on the ``min`` argument).
8944 '``llvm.expect``' Intrinsic
8945 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8952 declare i32 @llvm.expect.i32(i32 <val>, i32 <expected_val>)
8953 declare i64 @llvm.expect.i64(i64 <val>, i64 <expected_val>)
8958 The ``llvm.expect`` intrinsic provides information about expected (the
8959 most probable) value of ``val``, which can be used by optimizers.
8964 The ``llvm.expect`` intrinsic takes two arguments. The first argument is
8965 a value. The second argument is an expected value, this needs to be a
8966 constant value, variables are not allowed.
8971 This intrinsic is lowered to the ``val``.
8973 '``llvm.donothing``' Intrinsic
8974 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8981 declare void @llvm.donothing() nounwind readnone
8986 The ``llvm.donothing`` intrinsic doesn't perform any operation. It's the
8987 only intrinsic that can be called with an invoke instruction.
8997 This intrinsic does nothing, and it's removed by optimizers and ignored
9000 Stack Map Intrinsics
9001 --------------------
9003 LLVM provides experimental intrinsics to support runtime patching
9004 mechanisms commonly desired in dynamic language JITs. These intrinsics
9005 are described in :doc:`StackMaps`.